diff --git "a/unet/RyzenAI/.cache/down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_meta.json" "b/unet/RyzenAI/.cache/down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_meta.json" new file mode 100644--- /dev/null +++ "b/unet/RyzenAI/.cache/down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_meta.json" @@ -0,0 +1,58307 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 2, + "state_table_updates": [], + "op_list": [ + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5660" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5661" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5690" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5691" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5720" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5721" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5750" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5751" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5780" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5781" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5810" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5811" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5848" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5849" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5895" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5896" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5925" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5926" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_17" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5955" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5956" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_19" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5986" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_20" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_5987" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6016" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_22" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6017" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_23" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6046" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_24" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6047" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_25" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6077" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_26" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6078" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_27" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6107" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6108" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6137" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_30" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "encoder_hidden_states.out2_6_0" + ], + "const_args": [ + "onnx::MatMul_6138" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_31" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "77", + "768" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "77", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "768", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_0-/conv_in/Conv", + "type": "SDConv", + "in_args": [ + "Transpose_44_out-sample.out0_1_0" + ], + "const_args": [ + "NhwcConv_0_weight_NHWC" + ], + "out_args": [ + "NhwcConv_0_out-/conv_in/Conv_output_0.out0_1_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "4" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "4" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_0", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_0_out-/conv_in/Conv_output_0.out0_1_0" + ], + "const_args": [ + "GroupNorm_0_wts_7_1_0" + ], + "out_args": [ + "GroupNorm_0.out4_0.out7_1_0" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_0", + "type": "SDSilu", + "in_args": [ + "GroupNorm_0.out4_0.out7_1_0" + ], + "const_args": [ + "Sigmoid_0.weights10_0" + ], + "out_args": [ + "GroupNorm_0_sigmoid_out.4_0.out10_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_1-/down_blocks.0/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_0_sigmoid_out.4_0.out10_0" + ], + "const_args": [ + "NhwcConv_1_weight_NHWC" + ], + "out_args": [ + "NhwcConv_1_out-/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_embedding/linear_1/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_proj/Concat_1_output_0.out2_3_0" + ], + "const_args": [ + "time_embedding.linear_1.weight_8_0" + ], + "out_args": [ + "/time_embedding/linear_1/Gemm_output_0.out2_3_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_embedding/act/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_embedding/linear_1/Gemm_output_0.out2_3_0" + ], + "const_args": [ + "/time_embedding/act/Sigmoid.weights10_1" + ], + "out_args": [ + "/time_embedding/act/Sigmoid_output_0.out10_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/time_embedding/linear_2/Gemm", + "type": "SDGemm", + "in_args": [ + "/time_embedding/act/Sigmoid_output_0.out10_1" + ], + "const_args": [ + "time_embedding.linear_2.weight_8_1" + ], + "out_args": [ + "/time_embedding/linear_2/Gemm_output_0.out2_3_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.0/act_1/Sigmoid", + "type": "SDSilu", + "in_args": [ + "/time_embedding/linear_2/Gemm_output_0.out2_3_1" + ], + "const_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid.weights10_2" + ], + "out_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.0.resnets.0.time_emb_proj.weight_8_2" + ], + "out_args": [ + "/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.0.resnets.1.time_emb_proj.weight_8_3" + ], + "out_args": [ + "/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.1.resnets.0.time_emb_proj.weight_8_4" + ], + "out_args": [ + "/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.1.resnets.1.time_emb_proj.weight_8_5" + ], + "out_args": [ + "/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.2.resnets.0.time_emb_proj.weight_8_6" + ], + "out_args": [ + "/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.2.resnets.1.time_emb_proj.weight_8_7" + ], + "out_args": [ + "/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.3/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.3.resnets.0.time_emb_proj.weight_8_8" + ], + "out_args": [ + "/down_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.3/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "down_blocks.3.resnets.1.time_emb_proj.weight_8_9" + ], + "out_args": [ + "/down_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "mid_block.resnets.0.time_emb_proj.weight_8_10" + ], + "out_args": [ + "/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "mid_block.resnets.1.time_emb_proj.weight_8_11" + ], + "out_args": [ + "/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.0.resnets.0.time_emb_proj.weight_8_12" + ], + "out_args": [ + "/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.0.resnets.1.time_emb_proj.weight_8_13" + ], + "out_args": [ + "/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.0.resnets.2.time_emb_proj.weight_8_14" + ], + "out_args": [ + "/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.1.resnets.0.time_emb_proj.weight_8_15" + ], + "out_args": [ + "/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.1.resnets.1.time_emb_proj.weight_8_16" + ], + "out_args": [ + "/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.1.resnets.2.time_emb_proj.weight_8_17" + ], + "out_args": [ + "/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.2.resnets.0.time_emb_proj.weight_8_18" + ], + "out_args": [ + "/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.2.resnets.1.time_emb_proj.weight_8_19" + ], + "out_args": [ + "/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_17" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.2.resnets.2.time_emb_proj.weight_8_20" + ], + "out_args": [ + "/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.0/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.3.resnets.0.time_emb_proj.weight_8_21" + ], + "out_args": [ + "/up_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_19" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.1/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.3.resnets.1.time_emb_proj.weight_8_22" + ], + "out_args": [ + "/up_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_20" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.2/time_emb_proj/Gemm", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2" + ], + "const_args": [ + "up_blocks.3.resnets.2.time_emb_proj.weight_8_23" + ], + "out_args": [ + "/up_blocks.3/resnets.2/Unsqueeze_1_output_0.out2_0_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_1_out-/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1", + "/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/resnets.0/Add.out_12_1_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_1", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.0/resnets.0/Add.out_12_1_0" + ], + "const_args": [ + "GroupNorm_1_wts_7_1_1" + ], + "out_args": [ + "GroupNorm_1.out4_1.out7_1_1" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_1", + "type": "SDSilu", + "in_args": [ + "GroupNorm_1.out4_1.out7_1_1" + ], + "const_args": [ + "Sigmoid_1.weights10_3" + ], + "out_args": [ + "GroupNorm_1_sigmoid_out.4_1.out10_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_2-/down_blocks.0/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_1_sigmoid_out.4_1.out10_3" + ], + "const_args": [ + "NhwcConv_2_weight_NHWC" + ], + "out_args": [ + "NhwcConv_2_out-/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_0_out-/conv_in/Conv_output_0.out0_1_0", + "NhwcConv_2_out-/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/resnets.0/Add_1.out_12_1_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_2", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.0/resnets.0/Add_1.out_12_1_1" + ], + "const_args": [ + "GroupNorm_2_wts_7_1_2" + ], + "out_args": [ + "/down_blocks.0/attentions.0/norm/Add_output_0_NHWC.out7_1_2" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_3-/down_blocks.0/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.0/attentions.0/norm/Add_output_0_NHWC.out7_1_2" + ], + "const_args": [ + "NhwcConv_3_weight_NHWC" + ], + "out_args": [ + "/down_blocks.0/attentions.0/Reshape_output_0.out0_0_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.0/attentions.0/Reshape_output_0.out0_0_0" + ], + "const_args": [ + "down_blocks.0.attentions.0.transformer_blocks.0.norm1.weight", + "down_blocks.0.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_0" + ], + "const_args": [ + "onnx::MatMul_5649" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_32" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_0" + ], + "const_args": [ + "onnx::MatMul_5648" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_33" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_0" + ], + "const_args": [ + "onnx::MatMul_5650" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_34" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_0", + "type": "SDMHA", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_33", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_32", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_34" + ], + "const_args": [ + "MultiHeadAttention_0_mask.8_0" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0" + ], + "const_args": [ + "onnx::MatMul_5658" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_2", + "/down_blocks.0/attentions.0/Reshape_output_0.out0_0_0" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_2" + ], + "const_args": [ + "down_blocks.0.attentions.0.transformer_blocks.0.norm2.weight", + "down_blocks.0.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_1" + ], + "const_args": [ + "onnx::MatMul_5659" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_35" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_16", + "type": "SDMHA", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_35", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1" + ], + "const_args": [ + "MultiHeadAttention_16_mask.8_1" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1" + ], + "const_args": [ + "onnx::MatMul_5669" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_3", + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_2" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_3" + ], + "const_args": [ + "down_blocks.0.attentions.0.transformer_blocks.0.norm3.weight", + "down_blocks.0.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_0_0", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_2" + ], + "const_args": [ + "onnx::MatMul_5670_2_0_0" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_0_1", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_2" + ], + "const_args": [ + "onnx::MatMul_5670_2_0_1" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_4", + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_1" + ], + "const_args": [ + "onnx::MatMul_5671" + ], + "out_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_5", + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_3" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.0/Reshape_1_output_0.out_12_0_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_4-/down_blocks.0/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.0/attentions.0/Reshape_1_output_0.out_12_0_0" + ], + "const_args": [ + "NhwcConv_4_weight_NHWC" + ], + "out_args": [ + "NhwcConv_4_out-/down_blocks.0/attentions.0/proj_out/Conv_output_0.out0_1_3" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_4_out-/down_blocks.0/attentions.0/proj_out/Conv_output_0.out0_1_3", + "/down_blocks.0/resnets.0/Add_1.out_12_1_1" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.0/Add.out_12_1_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_3", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.0/attentions.0/Add.out_12_1_4" + ], + "const_args": [ + "GroupNorm_3_wts_7_1_3" + ], + "out_args": [ + "GroupNorm_3.out4_3.out7_1_3" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_2", + "type": "SDSilu", + "in_args": [ + "GroupNorm_3.out4_3.out7_1_3" + ], + "const_args": [ + "Sigmoid_2.weights10_4" + ], + "out_args": [ + "GroupNorm_3_sigmoid_out.4_3.out10_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_5-/down_blocks.0/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_3_sigmoid_out.4_3.out10_4" + ], + "const_args": [ + "NhwcConv_5_weight_NHWC" + ], + "out_args": [ + "NhwcConv_5_out-/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_4" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_5_out-/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_4", + "/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/resnets.1/Add.out_12_1_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_4", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.0/resnets.1/Add.out_12_1_5" + ], + "const_args": [ + "GroupNorm_4_wts_7_1_4" + ], + "out_args": [ + "GroupNorm_4.out4_4.out7_1_4" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_3", + "type": "SDSilu", + "in_args": [ + "GroupNorm_4.out4_4.out7_1_4" + ], + "const_args": [ + "Sigmoid_3.weights10_5" + ], + "out_args": [ + "GroupNorm_4_sigmoid_out.4_4.out10_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_6-/down_blocks.0/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_4_sigmoid_out.4_4.out10_5" + ], + "const_args": [ + "NhwcConv_6_weight_NHWC" + ], + "out_args": [ + "NhwcConv_6_out-/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_5" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.0/Add.out_12_1_4", + "NhwcConv_6_out-/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_5" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/resnets.1/Add_1.out_12_1_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_5", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.0/resnets.1/Add_1.out_12_1_6" + ], + "const_args": [ + "GroupNorm_5_wts_7_1_5" + ], + "out_args": [ + "/down_blocks.0/attentions.1/norm/Add_output_0_NHWC.out7_1_5" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_7-/down_blocks.0/attentions.1/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.0/attentions.1/norm/Add_output_0_NHWC.out7_1_5" + ], + "const_args": [ + "NhwcConv_7_weight_NHWC" + ], + "out_args": [ + "/down_blocks.0/attentions.1/Reshape_output_0.out0_0_1" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.0/attentions.1/Reshape_output_0.out0_0_1" + ], + "const_args": [ + "down_blocks.0.attentions.1.transformer_blocks.0.norm1.weight", + "down_blocks.0.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_3" + ], + "const_args": [ + "onnx::MatMul_5679" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_36" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_3" + ], + "const_args": [ + "onnx::MatMul_5678" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_37" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_3" + ], + "const_args": [ + "onnx::MatMul_5680" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_38" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_1", + "type": "SDMHA", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_37", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_36", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_38" + ], + "const_args": [ + "MultiHeadAttention_1_mask.8_2" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_2" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_2" + ], + "const_args": [ + "onnx::MatMul_5688" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_6", + "/down_blocks.0/attentions.1/Reshape_output_0.out0_0_1" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_7" + ], + "const_args": [ + "down_blocks.0.attentions.1.transformer_blocks.0.norm2.weight", + "down_blocks.0.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_4" + ], + "const_args": [ + "onnx::MatMul_5689" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_39" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_17", + "type": "SDMHA", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_39", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_2", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_3" + ], + "const_args": [ + "MultiHeadAttention_17_mask.8_3" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_3" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_3" + ], + "const_args": [ + "onnx::MatMul_5699" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_7", + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_7" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_8" + ], + "const_args": [ + "down_blocks.0.attentions.1.transformer_blocks.0.norm3.weight", + "down_blocks.0.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_1_0", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_5" + ], + "const_args": [ + "onnx::MatMul_5700_2_1_0" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_1_1", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_5" + ], + "const_args": [ + "onnx::MatMul_5700_2_1_1" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_8", + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_1" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_2" + ], + "const_args": [ + "onnx::MatMul_5701" + ], + "out_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_9", + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_8" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.1/Reshape_1_output_0.out_12_0_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_8-/down_blocks.0/attentions.1/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.0/attentions.1/Reshape_1_output_0.out_12_0_1" + ], + "const_args": [ + "NhwcConv_8_weight_NHWC" + ], + "out_args": [ + "NhwcConv_8_out-/down_blocks.0/attentions.1/proj_out/Conv_output_0.out0_1_6" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.0/attentions.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_8_out-/down_blocks.0/attentions.1/proj_out/Conv_output_0.out0_1_6", + "/down_blocks.0/resnets.1/Add_1.out_12_1_6" + ], + "const_args": [], + "out_args": [ + "/down_blocks.0/attentions.1/Add.out_12_1_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_9-/down_blocks.0/downsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.0/attentions.1/Add.out_12_1_9" + ], + "const_args": [ + "NhwcConv_9_weight_NHWC" + ], + "out_args": [ + "NhwcConv_9_out-/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_7" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_6", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_9_out-/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_7" + ], + "const_args": [ + "GroupNorm_6_wts_7_1_6" + ], + "out_args": [ + "GroupNorm_6.out4_6.out7_1_6" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_12-/down_blocks.1/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_9_out-/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_7" + ], + "const_args": [ + "NhwcConv_12_weight_NHWC" + ], + "out_args": [ + "NhwcConv_12_out-/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_8" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_4", + "type": "SDSilu", + "in_args": [ + "GroupNorm_6.out4_6.out7_1_6" + ], + "const_args": [ + "Sigmoid_4.weights10_6" + ], + "out_args": [ + "GroupNorm_6_sigmoid_out.4_6.out10_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_10-/down_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_6_sigmoid_out.4_6.out10_6" + ], + "const_args": [ + "NhwcConv_10_weight_NHWC" + ], + "out_args": [ + "NhwcConv_10_out-/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_9" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_10_out-/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_9", + "/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/resnets.0/Add.out_12_1_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_7", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.1/resnets.0/Add.out_12_1_10" + ], + "const_args": [ + "GroupNorm_7_wts_7_1_7" + ], + "out_args": [ + "GroupNorm_7.out4_7.out7_1_7" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_5", + "type": "SDSilu", + "in_args": [ + "GroupNorm_7.out4_7.out7_1_7" + ], + "const_args": [ + "Sigmoid_5.weights10_7" + ], + "out_args": [ + "GroupNorm_7_sigmoid_out.4_7.out10_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_11-/down_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_7_sigmoid_out.4_7.out10_7" + ], + "const_args": [ + "NhwcConv_11_weight_NHWC" + ], + "out_args": [ + "NhwcConv_11_out-/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_10" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_12_out-/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_8", + "NhwcConv_11_out-/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_10" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/resnets.0/Add_1.out_12_1_11" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_8", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.1/resnets.0/Add_1.out_12_1_11" + ], + "const_args": [ + "GroupNorm_8_wts_7_1_8" + ], + "out_args": [ + "/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_8" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_13-/down_blocks.1/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_8" + ], + "const_args": [ + "NhwcConv_13_weight_NHWC" + ], + "out_args": [ + "/down_blocks.1/attentions.0/Reshape_output_0.out0_0_2" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.1/attentions.0/Reshape_output_0.out0_0_2" + ], + "const_args": [ + "down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_6" + ], + "const_args": [ + "onnx::MatMul_5709" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_40" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_6" + ], + "const_args": [ + "onnx::MatMul_5708" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_41" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_6" + ], + "const_args": [ + "onnx::MatMul_5710" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_42" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_2", + "type": "SDMHA", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_41", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_40", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_42" + ], + "const_args": [ + "MultiHeadAttention_2_mask.8_4" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4" + ], + "const_args": [ + "onnx::MatMul_5718" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_10", + "/down_blocks.1/attentions.0/Reshape_output_0.out0_0_2" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_12" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_12" + ], + "const_args": [ + "down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_7" + ], + "const_args": [ + "onnx::MatMul_5719" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_43" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_18", + "type": "SDMHA", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_43", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5" + ], + "const_args": [ + "MultiHeadAttention_18_mask.8_5" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5" + ], + "const_args": [ + "onnx::MatMul_5729" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_11", + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_12" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_13" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_13" + ], + "const_args": [ + "down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_2_0", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_8" + ], + "const_args": [ + "onnx::MatMul_5730_2_2_0" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_2_1", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_8" + ], + "const_args": [ + "onnx::MatMul_5730_2_2_1" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_12", + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_3" + ], + "const_args": [ + "onnx::MatMul_5731" + ], + "out_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_13", + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_13" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_14-/down_blocks.1/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_2" + ], + "const_args": [ + "NhwcConv_14_weight_NHWC" + ], + "out_args": [ + "NhwcConv_14_out-/down_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_11" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_14_out-/down_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_11", + "/down_blocks.1/resnets.0/Add_1.out_12_1_11" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.0/Add.out_12_1_14" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_9", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.1/attentions.0/Add.out_12_1_14" + ], + "const_args": [ + "GroupNorm_9_wts_7_1_9" + ], + "out_args": [ + "GroupNorm_9.out4_9.out7_1_9" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_6", + "type": "SDSilu", + "in_args": [ + "GroupNorm_9.out4_9.out7_1_9" + ], + "const_args": [ + "Sigmoid_6.weights10_8" + ], + "out_args": [ + "GroupNorm_9_sigmoid_out.4_9.out10_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_15-/down_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_9_sigmoid_out.4_9.out10_8" + ], + "const_args": [ + "NhwcConv_15_weight_NHWC" + ], + "out_args": [ + "NhwcConv_15_out-/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_12" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_15_out-/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_12", + "/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/resnets.1/Add.out_12_1_15" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_10", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.1/resnets.1/Add.out_12_1_15" + ], + "const_args": [ + "GroupNorm_10_wts_7_1_10" + ], + "out_args": [ + "GroupNorm_10.out4_10.out7_1_10" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_7", + "type": "SDSilu", + "in_args": [ + "GroupNorm_10.out4_10.out7_1_10" + ], + "const_args": [ + "Sigmoid_7.weights10_9" + ], + "out_args": [ + "GroupNorm_10_sigmoid_out.4_10.out10_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_16-/down_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_10_sigmoid_out.4_10.out10_9" + ], + "const_args": [ + "NhwcConv_16_weight_NHWC" + ], + "out_args": [ + "NhwcConv_16_out-/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_13" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.0/Add.out_12_1_14", + "NhwcConv_16_out-/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_13" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/resnets.1/Add_1.out_12_1_16" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_11", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.1/resnets.1/Add_1.out_12_1_16" + ], + "const_args": [ + "GroupNorm_11_wts_7_1_11" + ], + "out_args": [ + "/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_11" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_17-/down_blocks.1/attentions.1/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_11" + ], + "const_args": [ + "NhwcConv_17_weight_NHWC" + ], + "out_args": [ + "/down_blocks.1/attentions.1/Reshape_output_0.out0_0_3" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.1/attentions.1/Reshape_output_0.out0_0_3" + ], + "const_args": [ + "down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_9" + ], + "const_args": [ + "onnx::MatMul_5739" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_44" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_9" + ], + "const_args": [ + "onnx::MatMul_5738" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_45" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_9" + ], + "const_args": [ + "onnx::MatMul_5740" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_46" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_3", + "type": "SDMHA", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_45", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_44", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_46" + ], + "const_args": [ + "MultiHeadAttention_3_mask.8_6" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_6" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_6" + ], + "const_args": [ + "onnx::MatMul_5748" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14", + "/down_blocks.1/attentions.1/Reshape_output_0.out0_0_3" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_17" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_17" + ], + "const_args": [ + "down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_10" + ], + "const_args": [ + "onnx::MatMul_5749" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_47" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_19", + "type": "SDMHA", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_47", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_6", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_7" + ], + "const_args": [ + "MultiHeadAttention_19_mask.8_7" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_7" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_7" + ], + "const_args": [ + "onnx::MatMul_5759" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15", + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_17" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_18" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_18" + ], + "const_args": [ + "down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_3_0", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_11" + ], + "const_args": [ + "onnx::MatMul_5760_2_3_0" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_3" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_3_1", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_11" + ], + "const_args": [ + "onnx::MatMul_5760_2_3_1" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16", + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_3" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_4" + ], + "const_args": [ + "onnx::MatMul_5761" + ], + "out_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17", + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_18" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_18-/down_blocks.1/attentions.1/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_3" + ], + "const_args": [ + "NhwcConv_18_weight_NHWC" + ], + "out_args": [ + "NhwcConv_18_out-/down_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_14" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.1/attentions.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_18_out-/down_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_14", + "/down_blocks.1/resnets.1/Add_1.out_12_1_16" + ], + "const_args": [], + "out_args": [ + "/down_blocks.1/attentions.1/Add.out_12_1_19" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_19-/down_blocks.1/downsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.1/attentions.1/Add.out_12_1_19" + ], + "const_args": [ + "NhwcConv_19_weight_NHWC" + ], + "out_args": [ + "NhwcConv_19_out-/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_15" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_22-/down_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_19_out-/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_15" + ], + "const_args": [ + "NhwcConv_22_weight_NHWC" + ], + "out_args": [ + "NhwcConv_22_out-/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_16" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_12", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_19_out-/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_15" + ], + "const_args": [ + "GroupNorm_12_wts_7_1_12" + ], + "out_args": [ + "GroupNorm_12.out4_12.out7_1_12" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_8", + "type": "SDSilu", + "in_args": [ + "GroupNorm_12.out4_12.out7_1_12" + ], + "const_args": [ + "Sigmoid_8.weights10_10" + ], + "out_args": [ + "GroupNorm_12_sigmoid_out.4_12.out10_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_20-/down_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_12_sigmoid_out.4_12.out10_10" + ], + "const_args": [ + "NhwcConv_20_weight_NHWC" + ], + "out_args": [ + "NhwcConv_20_out-/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_17" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_20_out-/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_17", + "/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/resnets.0/Add.out_12_1_20" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_13", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.2/resnets.0/Add.out_12_1_20" + ], + "const_args": [ + "GroupNorm_13_wts_7_1_13" + ], + "out_args": [ + "GroupNorm_13.out4_13.out7_1_13" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_9", + "type": "SDSilu", + "in_args": [ + "GroupNorm_13.out4_13.out7_1_13" + ], + "const_args": [ + "Sigmoid_9.weights10_11" + ], + "out_args": [ + "GroupNorm_13_sigmoid_out.4_13.out10_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_21-/down_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_13_sigmoid_out.4_13.out10_11" + ], + "const_args": [ + "NhwcConv_21_weight_NHWC" + ], + "out_args": [ + "NhwcConv_21_out-/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_18" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_22_out-/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_16", + "NhwcConv_21_out-/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_18" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/resnets.0/Add_1.out_12_1_21" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_14", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.2/resnets.0/Add_1.out_12_1_21" + ], + "const_args": [ + "GroupNorm_14_wts_7_1_14" + ], + "out_args": [ + "/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_14" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_23-/down_blocks.2/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_14" + ], + "const_args": [ + "NhwcConv_23_weight_NHWC" + ], + "out_args": [ + "/down_blocks.2/attentions.0/Reshape_output_0.out0_0_4" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.2/attentions.0/Reshape_output_0.out0_0_4" + ], + "const_args": [ + "down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", + "down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_12" + ], + "const_args": [ + "onnx::MatMul_5769" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_48" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_12" + ], + "const_args": [ + "onnx::MatMul_5768" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_49" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_12" + ], + "const_args": [ + "onnx::MatMul_5770" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_50" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_4", + "type": "SDMHA", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_49", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_48", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_50" + ], + "const_args": [ + "MultiHeadAttention_4_mask.8_8" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8" + ], + "const_args": [ + "onnx::MatMul_5778" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_18", + "/down_blocks.2/attentions.0/Reshape_output_0.out0_0_4" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_22" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_22" + ], + "const_args": [ + "down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", + "down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_13" + ], + "const_args": [ + "onnx::MatMul_5779" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_51" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_20", + "type": "SDMHA", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_51", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9" + ], + "const_args": [ + "MultiHeadAttention_20_mask.8_9" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9" + ], + "const_args": [ + "onnx::MatMul_5789" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_19" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_19", + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_22" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_23" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_23" + ], + "const_args": [ + "down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", + "down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_4_0", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_14" + ], + "const_args": [ + "onnx::MatMul_5790_2_4_0" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_4_1", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_14" + ], + "const_args": [ + "onnx::MatMul_5790_2_4_1" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_20" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_20", + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_5" + ], + "const_args": [ + "onnx::MatMul_5791" + ], + "out_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_21", + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_23" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_24-/down_blocks.2/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_4" + ], + "const_args": [ + "NhwcConv_24_weight_NHWC" + ], + "out_args": [ + "NhwcConv_24_out-/down_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_19" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_24_out-/down_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_19", + "/down_blocks.2/resnets.0/Add_1.out_12_1_21" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.0/Add.out_12_1_24" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_15", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.2/attentions.0/Add.out_12_1_24" + ], + "const_args": [ + "GroupNorm_15_wts_7_1_15" + ], + "out_args": [ + "GroupNorm_15.out4_15.out7_1_15" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_10", + "type": "SDSilu", + "in_args": [ + "GroupNorm_15.out4_15.out7_1_15" + ], + "const_args": [ + "Sigmoid_10.weights10_12" + ], + "out_args": [ + "GroupNorm_15_sigmoid_out.4_15.out10_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_25-/down_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_15_sigmoid_out.4_15.out10_12" + ], + "const_args": [ + "NhwcConv_25_weight_NHWC" + ], + "out_args": [ + "NhwcConv_25_out-/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_20" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_25_out-/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_20", + "/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/resnets.1/Add.out_12_1_25" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_16", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.2/resnets.1/Add.out_12_1_25" + ], + "const_args": [ + "GroupNorm_16_wts_7_1_16" + ], + "out_args": [ + "GroupNorm_16.out4_16.out7_1_16" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_11", + "type": "SDSilu", + "in_args": [ + "GroupNorm_16.out4_16.out7_1_16" + ], + "const_args": [ + "Sigmoid_11.weights10_13" + ], + "out_args": [ + "GroupNorm_16_sigmoid_out.4_16.out10_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_26-/down_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_16_sigmoid_out.4_16.out10_13" + ], + "const_args": [ + "NhwcConv_26_weight_NHWC" + ], + "out_args": [ + "NhwcConv_26_out-/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_21" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.0/Add.out_12_1_24", + "NhwcConv_26_out-/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_21" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/resnets.1/Add_1.out_12_1_26" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_17", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.2/resnets.1/Add_1.out_12_1_26" + ], + "const_args": [ + "GroupNorm_17_wts_7_1_17" + ], + "out_args": [ + "/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_17" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_27-/down_blocks.2/attentions.1/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_17" + ], + "const_args": [ + "NhwcConv_27_weight_NHWC" + ], + "out_args": [ + "/down_blocks.2/attentions.1/Reshape_output_0.out0_0_5" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.2/attentions.1/Reshape_output_0.out0_0_5" + ], + "const_args": [ + "down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", + "down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_15" + ], + "const_args": [ + "onnx::MatMul_5799" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_52" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_15" + ], + "const_args": [ + "onnx::MatMul_5798" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_53" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_15" + ], + "const_args": [ + "onnx::MatMul_5800" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_54" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_5", + "type": "SDMHA", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_53", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_52", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_54" + ], + "const_args": [ + "MultiHeadAttention_5_mask.8_10" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_10" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_10" + ], + "const_args": [ + "onnx::MatMul_5808" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_22" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_22", + "/down_blocks.2/attentions.1/Reshape_output_0.out0_0_5" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_27" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_27" + ], + "const_args": [ + "down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", + "down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_16" + ], + "const_args": [ + "onnx::MatMul_5809" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_55" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_21", + "type": "SDMHA", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_55", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_10", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_11" + ], + "const_args": [ + "MultiHeadAttention_21_mask.8_11" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_11" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_11" + ], + "const_args": [ + "onnx::MatMul_5819" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_23" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_23", + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_27" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_28" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_28" + ], + "const_args": [ + "down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", + "down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_17" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_5_0", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_17" + ], + "const_args": [ + "onnx::MatMul_5820_2_5_0" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_5" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_5_1", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_17" + ], + "const_args": [ + "onnx::MatMul_5820_2_5_1" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_24" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_24", + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_5" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_6" + ], + "const_args": [ + "onnx::MatMul_5821" + ], + "out_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_25" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_25", + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_28" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_28-/down_blocks.2/attentions.1/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_5" + ], + "const_args": [ + "NhwcConv_28_weight_NHWC" + ], + "out_args": [ + "NhwcConv_28_out-/down_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_22" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.2/attentions.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_28_out-/down_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_22", + "/down_blocks.2/resnets.1/Add_1.out_12_1_26" + ], + "const_args": [], + "out_args": [ + "/down_blocks.2/attentions.1/Add.out_12_1_29" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_29-/down_blocks.2/downsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/down_blocks.2/attentions.1/Add.out_12_1_29" + ], + "const_args": [ + "NhwcConv_29_weight_NHWC" + ], + "out_args": [ + "NhwcConv_29_out-/down_blocks.2/downsamplers.0/conv/Conv_output_0.out0_1_23" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_18", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_29_out-/down_blocks.2/downsamplers.0/conv/Conv_output_0.out0_1_23" + ], + "const_args": [ + "GroupNorm_18_wts_7_1_18" + ], + "out_args": [ + "GroupNorm_18.out4_18.out7_1_18" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_12", + "type": "SDSilu", + "in_args": [ + "GroupNorm_18.out4_18.out7_1_18" + ], + "const_args": [ + "Sigmoid_12.weights10_14" + ], + "out_args": [ + "GroupNorm_18_sigmoid_out.4_18.out10_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_30-/down_blocks.3/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_18_sigmoid_out.4_18.out10_14" + ], + "const_args": [ + "NhwcConv_30_weight_NHWC" + ], + "out_args": [ + "NhwcConv_30_out-/down_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_24" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.3/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_30_out-/down_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_24", + "/down_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_6" + ], + "const_args": [], + "out_args": [ + "/down_blocks.3/resnets.0/Add.out_12_1_30" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_19", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.3/resnets.0/Add.out_12_1_30" + ], + "const_args": [ + "GroupNorm_19_wts_7_1_19" + ], + "out_args": [ + "GroupNorm_19.out4_19.out7_1_19" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_13", + "type": "SDSilu", + "in_args": [ + "GroupNorm_19.out4_19.out7_1_19" + ], + "const_args": [ + "Sigmoid_13.weights10_15" + ], + "out_args": [ + "GroupNorm_19_sigmoid_out.4_19.out10_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_31-/down_blocks.3/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_19_sigmoid_out.4_19.out10_15" + ], + "const_args": [ + "NhwcConv_31_weight_NHWC" + ], + "out_args": [ + "NhwcConv_31_out-/down_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_25" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.3/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_29_out-/down_blocks.2/downsamplers.0/conv/Conv_output_0.out0_1_23", + "NhwcConv_31_out-/down_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_25" + ], + "const_args": [], + "out_args": [ + "/down_blocks.3/resnets.0/Add_1.out_12_1_31" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_20", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.3/resnets.0/Add_1.out_12_1_31" + ], + "const_args": [ + "GroupNorm_20_wts_7_1_20" + ], + "out_args": [ + "GroupNorm_20.out4_20.out7_1_20" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_14", + "type": "SDSilu", + "in_args": [ + "GroupNorm_20.out4_20.out7_1_20" + ], + "const_args": [ + "Sigmoid_14.weights10_16" + ], + "out_args": [ + "GroupNorm_20_sigmoid_out.4_20.out10_16" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_32-/down_blocks.3/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_20_sigmoid_out.4_20.out10_16" + ], + "const_args": [ + "NhwcConv_32_weight_NHWC" + ], + "out_args": [ + "NhwcConv_32_out-/down_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_26" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.3/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_32_out-/down_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_26", + "/down_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_7" + ], + "const_args": [], + "out_args": [ + "/down_blocks.3/resnets.1/Add.out_12_1_32" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_21", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.3/resnets.1/Add.out_12_1_32" + ], + "const_args": [ + "GroupNorm_21_wts_7_1_21" + ], + "out_args": [ + "GroupNorm_21.out4_21.out7_1_21" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_15", + "type": "SDSilu", + "in_args": [ + "GroupNorm_21.out4_21.out7_1_21" + ], + "const_args": [ + "Sigmoid_15.weights10_17" + ], + "out_args": [ + "GroupNorm_21_sigmoid_out.4_21.out10_17" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_33-/down_blocks.3/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_21_sigmoid_out.4_21.out10_17" + ], + "const_args": [ + "NhwcConv_33_weight_NHWC" + ], + "out_args": [ + "NhwcConv_33_out-/down_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_27" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/down_blocks.3/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.3/resnets.0/Add_1.out_12_1_31", + "NhwcConv_33_out-/down_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_27" + ], + "const_args": [], + "out_args": [ + "/down_blocks.3/resnets.1/Add_1.out_12_1_33" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_22", + "type": "SDGroupNorm", + "in_args": [ + "/down_blocks.3/resnets.1/Add_1.out_12_1_33" + ], + "const_args": [ + "GroupNorm_22_wts_7_1_22" + ], + "out_args": [ + "GroupNorm_22.out4_22.out7_1_22" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_16", + "type": "SDSilu", + "in_args": [ + "GroupNorm_22.out4_22.out7_1_22" + ], + "const_args": [ + "Sigmoid_16.weights10_18" + ], + "out_args": [ + "GroupNorm_22_sigmoid_out.4_22.out10_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_34-/mid_block/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_22_sigmoid_out.4_22.out10_18" + ], + "const_args": [ + "NhwcConv_34_weight_NHWC" + ], + "out_args": [ + "NhwcConv_34_out-/mid_block/resnets.0/conv1/Conv_output_0.out0_1_28" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_34_out-/mid_block/resnets.0/conv1/Conv_output_0.out0_1_28", + "/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_8" + ], + "const_args": [], + "out_args": [ + "/mid_block/resnets.0/Add.out_12_1_34" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_23", + "type": "SDGroupNorm", + "in_args": [ + "/mid_block/resnets.0/Add.out_12_1_34" + ], + "const_args": [ + "GroupNorm_23_wts_7_1_23" + ], + "out_args": [ + "GroupNorm_23.out4_23.out7_1_23" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_17", + "type": "SDSilu", + "in_args": [ + "GroupNorm_23.out4_23.out7_1_23" + ], + "const_args": [ + "Sigmoid_17.weights10_19" + ], + "out_args": [ + "GroupNorm_23_sigmoid_out.4_23.out10_19" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_35-/mid_block/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_23_sigmoid_out.4_23.out10_19" + ], + "const_args": [ + "NhwcConv_35_weight_NHWC" + ], + "out_args": [ + "NhwcConv_35_out-/mid_block/resnets.0/conv2/Conv_output_0.out0_1_29" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/down_blocks.3/resnets.1/Add_1.out_12_1_33", + "NhwcConv_35_out-/mid_block/resnets.0/conv2/Conv_output_0.out0_1_29" + ], + "const_args": [], + "out_args": [ + "/mid_block/resnets.0/Add_1.out_12_1_35" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_24", + "type": "SDGroupNorm", + "in_args": [ + "/mid_block/resnets.0/Add_1.out_12_1_35" + ], + "const_args": [ + "GroupNorm_24_wts_7_1_24" + ], + "out_args": [ + "/mid_block/attentions.0/norm/Add_output_0_NHWC.out7_1_24" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_36-/mid_block/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/mid_block/attentions.0/norm/Add_output_0_NHWC.out7_1_24" + ], + "const_args": [ + "NhwcConv_36_weight_NHWC" + ], + "out_args": [ + "/mid_block/attentions.0/Reshape_output_0.out0_0_6" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/mid_block/attentions.0/Reshape_output_0.out0_0_6" + ], + "const_args": [ + "mid_block.attentions.0.transformer_blocks.0.norm1.weight", + "mid_block.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_18" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_18" + ], + "const_args": [ + "onnx::MatMul_5837" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_56" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_18" + ], + "const_args": [ + "onnx::MatMul_5836" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_57" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_18" + ], + "const_args": [ + "onnx::MatMul_5838" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_58" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_6", + "type": "SDMHA", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_57", + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_56", + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_58" + ], + "const_args": [ + "MultiHeadAttention_6_mask.8_12" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_12" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280", + "64" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_12" + ], + "const_args": [ + "onnx::MatMul_5846" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_26" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_26", + "/mid_block/attentions.0/Reshape_output_0.out0_0_6" + ], + "const_args": [], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_36" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_36" + ], + "const_args": [ + "mid_block.attentions.0.transformer_blocks.0.norm2.weight", + "mid_block.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_19" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_19" + ], + "const_args": [ + "onnx::MatMul_5847" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_59" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_22", + "type": "SDMHA", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_59", + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_12", + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_13" + ], + "const_args": [ + "MultiHeadAttention_22_mask.8_13" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_13" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_13" + ], + "const_args": [ + "onnx::MatMul_5857" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_27" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_27", + "/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_36" + ], + "const_args": [], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_37" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_37" + ], + "const_args": [ + "mid_block.attentions.0.transformer_blocks.0.norm3.weight", + "mid_block.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_20" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_6_0", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_20" + ], + "const_args": [ + "onnx::MatMul_5858_2_6_0" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_6" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_6_1", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_20" + ], + "const_args": [ + "onnx::MatMul_5858_2_6_1" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_28" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_28", + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_6" + ], + "const_args": [], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_7" + ], + "const_args": [ + "onnx::MatMul_5859" + ], + "out_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_29" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_29", + "/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_37" + ], + "const_args": [], + "out_args": [ + "/mid_block/attentions.0/Reshape_1_output_0.out_12_0_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_37-/mid_block/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/mid_block/attentions.0/Reshape_1_output_0.out_12_0_6" + ], + "const_args": [ + "NhwcConv_37_weight_NHWC" + ], + "out_args": [ + "NhwcConv_37_out-/mid_block/attentions.0/proj_out/Conv_output_0.out0_1_30" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_37_out-/mid_block/attentions.0/proj_out/Conv_output_0.out0_1_30", + "/mid_block/resnets.0/Add_1.out_12_1_35" + ], + "const_args": [], + "out_args": [ + "/mid_block/attentions.0/Add.out_12_1_38" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_25", + "type": "SDGroupNorm", + "in_args": [ + "/mid_block/attentions.0/Add.out_12_1_38" + ], + "const_args": [ + "GroupNorm_25_wts_7_1_25" + ], + "out_args": [ + "GroupNorm_25.out4_25.out7_1_25" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_18", + "type": "SDSilu", + "in_args": [ + "GroupNorm_25.out4_25.out7_1_25" + ], + "const_args": [ + "Sigmoid_18.weights10_20" + ], + "out_args": [ + "GroupNorm_25_sigmoid_out.4_25.out10_20" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_38-/mid_block/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_25_sigmoid_out.4_25.out10_20" + ], + "const_args": [ + "NhwcConv_38_weight_NHWC" + ], + "out_args": [ + "NhwcConv_38_out-/mid_block/resnets.1/conv1/Conv_output_0.out0_1_31" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_38_out-/mid_block/resnets.1/conv1/Conv_output_0.out0_1_31", + "/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_9" + ], + "const_args": [], + "out_args": [ + "/mid_block/resnets.1/Add.out_12_1_39" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_26", + "type": "SDGroupNorm", + "in_args": [ + "/mid_block/resnets.1/Add.out_12_1_39" + ], + "const_args": [ + "GroupNorm_26_wts_7_1_26" + ], + "out_args": [ + "GroupNorm_26.out4_26.out7_1_26" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_19", + "type": "SDSilu", + "in_args": [ + "GroupNorm_26.out4_26.out7_1_26" + ], + "const_args": [ + "Sigmoid_19.weights10_21" + ], + "out_args": [ + "GroupNorm_26_sigmoid_out.4_26.out10_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_39-/mid_block/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_26_sigmoid_out.4_26.out10_21" + ], + "const_args": [ + "NhwcConv_39_weight_NHWC" + ], + "out_args": [ + "NhwcConv_39_out-/mid_block/resnets.1/conv2/Conv_output_0.out0_1_32" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/mid_block/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "/mid_block/attentions.0/Add.out_12_1_38", + "NhwcConv_39_out-/mid_block/resnets.1/conv2/Conv_output_0.out0_1_32" + ], + "const_args": [], + "out_args": [ + "/mid_block/resnets.1/Add_1.out_12_1_40" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/Concat", + "type": "SDConcat", + "in_args": [ + "/mid_block/resnets.1/Add_1.out_12_1_40", + "/down_blocks.3/resnets.1/Add_1.out_12_1_33" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/Concat.out13_1" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_27", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.0/Concat.out13_1" + ], + "const_args": [ + "GroupNorm_27_wts_7_1_27" + ], + "out_args": [ + "GroupNorm_27.out4_27.out7_1_27" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_42-/up_blocks.0/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.0/Concat.out13_1" + ], + "const_args": [ + "NhwcConv_42_weight_NHWC" + ], + "out_args": [ + "NhwcConv_42_out-/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_33" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_20", + "type": "SDSilu", + "in_args": [ + "GroupNorm_27.out4_27.out7_1_27" + ], + "const_args": [ + "Sigmoid_20.weights10_22" + ], + "out_args": [ + "GroupNorm_27_sigmoid_out.4_27.out10_22" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_40-/up_blocks.0/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_27_sigmoid_out.4_27.out10_22" + ], + "const_args": [ + "NhwcConv_40_weight_NHWC" + ], + "out_args": [ + "NhwcConv_40_out-/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_34" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_40_out-/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_34", + "/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_10" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/resnets.0/Add.out_12_1_41" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_28", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.0/resnets.0/Add.out_12_1_41" + ], + "const_args": [ + "GroupNorm_28_wts_7_1_28" + ], + "out_args": [ + "GroupNorm_28.out4_28.out7_1_28" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_21", + "type": "SDSilu", + "in_args": [ + "GroupNorm_28.out4_28.out7_1_28" + ], + "const_args": [ + "Sigmoid_21.weights10_23" + ], + "out_args": [ + "GroupNorm_28_sigmoid_out.4_28.out10_23" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_41-/up_blocks.0/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_28_sigmoid_out.4_28.out10_23" + ], + "const_args": [ + "NhwcConv_41_weight_NHWC" + ], + "out_args": [ + "NhwcConv_41_out-/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_35" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_42_out-/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_33", + "NhwcConv_41_out-/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_35" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/resnets.0/Add_1.out_12_1_42" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/Concat_1", + "type": "SDConcat", + "in_args": [ + "/up_blocks.0/resnets.0/Add_1.out_12_1_42", + "/down_blocks.3/resnets.0/Add_1.out_12_1_31" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/Concat_1.out13_2" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_29", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.0/Concat_1.out13_2" + ], + "const_args": [ + "GroupNorm_29_wts_7_1_29" + ], + "out_args": [ + "GroupNorm_29.out4_29.out7_1_29" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_45-/up_blocks.0/resnets.1/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.0/Concat_1.out13_2" + ], + "const_args": [ + "NhwcConv_45_weight_NHWC" + ], + "out_args": [ + "NhwcConv_45_out-/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_36" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_22", + "type": "SDSilu", + "in_args": [ + "GroupNorm_29.out4_29.out7_1_29" + ], + "const_args": [ + "Sigmoid_22.weights10_24" + ], + "out_args": [ + "GroupNorm_29_sigmoid_out.4_29.out10_24" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_43-/up_blocks.0/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_29_sigmoid_out.4_29.out10_24" + ], + "const_args": [ + "NhwcConv_43_weight_NHWC" + ], + "out_args": [ + "NhwcConv_43_out-/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_37" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_43_out-/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_37", + "/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_11" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/resnets.1/Add.out_12_1_43" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_30", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.0/resnets.1/Add.out_12_1_43" + ], + "const_args": [ + "GroupNorm_30_wts_7_1_30" + ], + "out_args": [ + "GroupNorm_30.out4_30.out7_1_30" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_23", + "type": "SDSilu", + "in_args": [ + "GroupNorm_30.out4_30.out7_1_30" + ], + "const_args": [ + "Sigmoid_23.weights10_25" + ], + "out_args": [ + "GroupNorm_30_sigmoid_out.4_30.out10_25" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_44-/up_blocks.0/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_30_sigmoid_out.4_30.out10_25" + ], + "const_args": [ + "NhwcConv_44_weight_NHWC" + ], + "out_args": [ + "NhwcConv_44_out-/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_38" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_45_out-/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_36", + "NhwcConv_44_out-/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_38" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/resnets.1/Add_1.out_12_1_44" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/Concat_2", + "type": "SDConcat", + "in_args": [ + "/up_blocks.0/resnets.1/Add_1.out_12_1_44", + "NhwcConv_29_out-/down_blocks.2/downsamplers.0/conv/Conv_output_0.out0_1_23" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/Concat_2.out13_3" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_31", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.0/Concat_2.out13_3" + ], + "const_args": [ + "GroupNorm_31_wts_7_1_31" + ], + "out_args": [ + "GroupNorm_31.out4_31.out7_1_31" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_48-/up_blocks.0/resnets.2/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.0/Concat_2.out13_3" + ], + "const_args": [ + "NhwcConv_48_weight_NHWC" + ], + "out_args": [ + "NhwcConv_48_out-/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_39" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_24", + "type": "SDSilu", + "in_args": [ + "GroupNorm_31.out4_31.out7_1_31" + ], + "const_args": [ + "Sigmoid_24.weights10_26" + ], + "out_args": [ + "GroupNorm_31_sigmoid_out.4_31.out10_26" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_46-/up_blocks.0/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_31_sigmoid_out.4_31.out10_26" + ], + "const_args": [ + "NhwcConv_46_weight_NHWC" + ], + "out_args": [ + "NhwcConv_46_out-/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_40" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_46_out-/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_40", + "/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_12" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/resnets.2/Add.out_12_1_45" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_32", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.0/resnets.2/Add.out_12_1_45" + ], + "const_args": [ + "GroupNorm_32_wts_7_1_32" + ], + "out_args": [ + "GroupNorm_32.out4_32.out7_1_32" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_25", + "type": "SDSilu", + "in_args": [ + "GroupNorm_32.out4_32.out7_1_32" + ], + "const_args": [ + "Sigmoid_25.weights10_27" + ], + "out_args": [ + "GroupNorm_32_sigmoid_out.4_32.out10_27" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_47-/up_blocks.0/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_32_sigmoid_out.4_32.out10_27" + ], + "const_args": [ + "NhwcConv_47_weight_NHWC" + ], + "out_args": [ + "NhwcConv_47_out-/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_41" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/resnets.2/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_48_out-/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_39", + "NhwcConv_47_out-/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_41" + ], + "const_args": [], + "out_args": [ + "/up_blocks.0/resnets.2/Add_1.out_12_1_46" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.0/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/up_blocks.0/resnets.2/Add_1.out_12_1_46" + ], + "const_args": [ + "/up_blocks.0/upsamplers.0/Resize.weights4_0" + ], + "out_args": [ + "/up_blocks.0/upsamplers.0/Resize.out_4_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "8", + "8", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_49-/up_blocks.0/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.0/upsamplers.0/Resize.out_4_0" + ], + "const_args": [ + "NhwcConv_49_weight_NHWC" + ], + "out_args": [ + "NhwcConv_49_out-/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_42" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/Concat", + "type": "SDConcat", + "in_args": [ + "NhwcConv_49_out-/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_42", + "/down_blocks.2/attentions.1/Add.out_12_1_29" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/Concat.out13_4" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_33", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/Concat.out13_4" + ], + "const_args": [ + "GroupNorm_33_wts_7_1_33" + ], + "out_args": [ + "GroupNorm_33.out4_33.out7_1_33" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_52-/up_blocks.1/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/Concat.out13_4" + ], + "const_args": [ + "NhwcConv_52_weight_NHWC" + ], + "out_args": [ + "NhwcConv_52_out-/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_43" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_26", + "type": "SDSilu", + "in_args": [ + "GroupNorm_33.out4_33.out7_1_33" + ], + "const_args": [ + "Sigmoid_26.weights10_28" + ], + "out_args": [ + "GroupNorm_33_sigmoid_out.4_33.out10_28" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_50-/up_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_33_sigmoid_out.4_33.out10_28" + ], + "const_args": [ + "NhwcConv_50_weight_NHWC" + ], + "out_args": [ + "NhwcConv_50_out-/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_44" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_50_out-/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_44", + "/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_13" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/resnets.0/Add.out_12_1_47" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_34", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/resnets.0/Add.out_12_1_47" + ], + "const_args": [ + "GroupNorm_34_wts_7_1_34" + ], + "out_args": [ + "GroupNorm_34.out4_34.out7_1_34" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_27", + "type": "SDSilu", + "in_args": [ + "GroupNorm_34.out4_34.out7_1_34" + ], + "const_args": [ + "Sigmoid_27.weights10_29" + ], + "out_args": [ + "GroupNorm_34_sigmoid_out.4_34.out10_29" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_51-/up_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_34_sigmoid_out.4_34.out10_29" + ], + "const_args": [ + "NhwcConv_51_weight_NHWC" + ], + "out_args": [ + "NhwcConv_51_out-/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_45" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_52_out-/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_43", + "NhwcConv_51_out-/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_45" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/resnets.0/Add_1.out_12_1_48" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_35", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/resnets.0/Add_1.out_12_1_48" + ], + "const_args": [ + "GroupNorm_35_wts_7_1_35" + ], + "out_args": [ + "/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_35" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_53-/up_blocks.1/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_35" + ], + "const_args": [ + "NhwcConv_53_weight_NHWC" + ], + "out_args": [ + "/up_blocks.1/attentions.0/Reshape_output_0.out0_0_7" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.0/Reshape_output_0.out0_0_7" + ], + "const_args": [ + "up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_21" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_21" + ], + "const_args": [ + "onnx::MatMul_5884" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_60" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_21" + ], + "const_args": [ + "onnx::MatMul_5883" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_61" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_21" + ], + "const_args": [ + "onnx::MatMul_5885" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_62" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_7", + "type": "SDMHA", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_61", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_60", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_62" + ], + "const_args": [ + "MultiHeadAttention_7_mask.8_14" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_14" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_14" + ], + "const_args": [ + "onnx::MatMul_5893" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_30" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_30", + "/up_blocks.1/attentions.0/Reshape_output_0.out0_0_7" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_49" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_49" + ], + "const_args": [ + "up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_22" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_22" + ], + "const_args": [ + "onnx::MatMul_5894" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_63" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_23", + "type": "SDMHA", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_63", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_14", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_15" + ], + "const_args": [ + "MultiHeadAttention_23_mask.8_15" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_15" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_15" + ], + "const_args": [ + "onnx::MatMul_5904" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_31" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_31", + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_49" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_50" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_50" + ], + "const_args": [ + "up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_23" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_7_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_23" + ], + "const_args": [ + "onnx::MatMul_5905_2_7_0" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_7" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_7_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_23" + ], + "const_args": [ + "onnx::MatMul_5905_2_7_1" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_32" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_32", + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_7" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_8" + ], + "const_args": [ + "onnx::MatMul_5906" + ], + "out_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_33" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_33", + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_50" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_54-/up_blocks.1/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_7" + ], + "const_args": [ + "NhwcConv_54_weight_NHWC" + ], + "out_args": [ + "NhwcConv_54_out-/up_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_46" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_54_out-/up_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_46", + "/up_blocks.1/resnets.0/Add_1.out_12_1_48" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.0/Add.out_12_1_51" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/Concat_1", + "type": "SDConcat", + "in_args": [ + "/up_blocks.1/attentions.0/Add.out_12_1_51", + "/down_blocks.2/attentions.0/Add.out_12_1_24" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/Concat_1.out13_5" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_36", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/Concat_1.out13_5" + ], + "const_args": [ + "GroupNorm_36_wts_7_1_36" + ], + "out_args": [ + "GroupNorm_36.out4_36.out7_1_36" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_57-/up_blocks.1/resnets.1/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/Concat_1.out13_5" + ], + "const_args": [ + "NhwcConv_57_weight_NHWC" + ], + "out_args": [ + "NhwcConv_57_out-/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_47" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_28", + "type": "SDSilu", + "in_args": [ + "GroupNorm_36.out4_36.out7_1_36" + ], + "const_args": [ + "Sigmoid_28.weights10_30" + ], + "out_args": [ + "GroupNorm_36_sigmoid_out.4_36.out10_30" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_55-/up_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_36_sigmoid_out.4_36.out10_30" + ], + "const_args": [ + "NhwcConv_55_weight_NHWC" + ], + "out_args": [ + "NhwcConv_55_out-/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_48" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "2560" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "2560" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_55_out-/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_48", + "/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_14" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/resnets.1/Add.out_12_1_52" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_37", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/resnets.1/Add.out_12_1_52" + ], + "const_args": [ + "GroupNorm_37_wts_7_1_37" + ], + "out_args": [ + "GroupNorm_37.out4_37.out7_1_37" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_29", + "type": "SDSilu", + "in_args": [ + "GroupNorm_37.out4_37.out7_1_37" + ], + "const_args": [ + "Sigmoid_29.weights10_31" + ], + "out_args": [ + "GroupNorm_37_sigmoid_out.4_37.out10_31" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_56-/up_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_37_sigmoid_out.4_37.out10_31" + ], + "const_args": [ + "NhwcConv_56_weight_NHWC" + ], + "out_args": [ + "NhwcConv_56_out-/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_49" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_57_out-/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_47", + "NhwcConv_56_out-/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_49" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/resnets.1/Add_1.out_12_1_53" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_38", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/resnets.1/Add_1.out_12_1_53" + ], + "const_args": [ + "GroupNorm_38_wts_7_1_38" + ], + "out_args": [ + "/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_38" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_58-/up_blocks.1/attentions.1/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_38" + ], + "const_args": [ + "NhwcConv_58_weight_NHWC" + ], + "out_args": [ + "/up_blocks.1/attentions.1/Reshape_output_0.out0_0_8" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.1/Reshape_output_0.out0_0_8" + ], + "const_args": [ + "up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_24" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_24" + ], + "const_args": [ + "onnx::MatMul_5914" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_64" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_24" + ], + "const_args": [ + "onnx::MatMul_5913" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_65" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_24" + ], + "const_args": [ + "onnx::MatMul_5915" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_66" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_8", + "type": "SDMHA", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_65", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_64", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_66" + ], + "const_args": [ + "MultiHeadAttention_8_mask.8_16" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_16" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_16" + ], + "const_args": [ + "onnx::MatMul_5923" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_34" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_34", + "/up_blocks.1/attentions.1/Reshape_output_0.out0_0_8" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_54" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_54" + ], + "const_args": [ + "up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_25" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_25" + ], + "const_args": [ + "onnx::MatMul_5924" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_67" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_24", + "type": "SDMHA", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_67", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_16", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_17" + ], + "const_args": [ + "MultiHeadAttention_24_mask.8_17" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_17" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_17" + ], + "const_args": [ + "onnx::MatMul_5934" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_35" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_35", + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_54" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_55" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_55" + ], + "const_args": [ + "up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_26" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_8_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_26" + ], + "const_args": [ + "onnx::MatMul_5935_2_8_0" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_8" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_8_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_26" + ], + "const_args": [ + "onnx::MatMul_5935_2_8_1" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_36" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_36", + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_8" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_9" + ], + "const_args": [ + "onnx::MatMul_5936" + ], + "out_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_37" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_37", + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_55" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_59-/up_blocks.1/attentions.1/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_8" + ], + "const_args": [ + "NhwcConv_59_weight_NHWC" + ], + "out_args": [ + "NhwcConv_59_out-/up_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_50" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_59_out-/up_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_50", + "/up_blocks.1/resnets.1/Add_1.out_12_1_53" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.1/Add.out_12_1_56" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/Concat_2", + "type": "SDConcat", + "in_args": [ + "/up_blocks.1/attentions.1/Add.out_12_1_56", + "NhwcConv_19_out-/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_15" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/Concat_2.out13_6" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_39", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/Concat_2.out13_6" + ], + "const_args": [ + "GroupNorm_39_wts_7_1_39" + ], + "out_args": [ + "GroupNorm_39.out4_39.out7_1_39" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "3840" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_62-/up_blocks.1/resnets.2/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/Concat_2.out13_6" + ], + "const_args": [ + "NhwcConv_62_weight_NHWC" + ], + "out_args": [ + "NhwcConv_62_out-/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_51" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_30", + "type": "SDSilu", + "in_args": [ + "GroupNorm_39.out4_39.out7_1_39" + ], + "const_args": [ + "Sigmoid_30.weights10_32" + ], + "out_args": [ + "GroupNorm_39_sigmoid_out.4_39.out10_32" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_60-/up_blocks.1/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_39_sigmoid_out.4_39.out10_32" + ], + "const_args": [ + "NhwcConv_60_weight_NHWC" + ], + "out_args": [ + "NhwcConv_60_out-/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_52" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_60_out-/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_52", + "/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_15" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/resnets.2/Add.out_12_1_57" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_40", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/resnets.2/Add.out_12_1_57" + ], + "const_args": [ + "GroupNorm_40_wts_7_1_40" + ], + "out_args": [ + "GroupNorm_40.out4_40.out7_1_40" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_31", + "type": "SDSilu", + "in_args": [ + "GroupNorm_40.out4_40.out7_1_40" + ], + "const_args": [ + "Sigmoid_31.weights10_33" + ], + "out_args": [ + "GroupNorm_40_sigmoid_out.4_40.out10_33" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_61-/up_blocks.1/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_40_sigmoid_out.4_40.out10_33" + ], + "const_args": [ + "NhwcConv_61_weight_NHWC" + ], + "out_args": [ + "NhwcConv_61_out-/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_53" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/resnets.2/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_62_out-/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_51", + "NhwcConv_61_out-/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_53" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/resnets.2/Add_1.out_12_1_58" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_41", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.1/resnets.2/Add_1.out_12_1_58" + ], + "const_args": [ + "GroupNorm_41_wts_7_1_41" + ], + "out_args": [ + "/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out7_1_41" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_63-/up_blocks.1/attentions.2/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out7_1_41" + ], + "const_args": [ + "NhwcConv_63_weight_NHWC" + ], + "out_args": [ + "/up_blocks.1/attentions.2/Reshape_output_0.out0_0_9" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.2/Reshape_output_0.out0_0_9" + ], + "const_args": [ + "up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight", + "up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_27" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_27" + ], + "const_args": [ + "onnx::MatMul_5944" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_68" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_27" + ], + "const_args": [ + "onnx::MatMul_5943" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_69" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_27" + ], + "const_args": [ + "onnx::MatMul_5945" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_70" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_9", + "type": "SDMHA", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_69", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_68", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_70" + ], + "const_args": [ + "MultiHeadAttention_9_mask.8_18" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_18" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_18" + ], + "const_args": [ + "onnx::MatMul_5953" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_38" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_38", + "/up_blocks.1/attentions.2/Reshape_output_0.out0_0_9" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_59" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_59" + ], + "const_args": [ + "up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight", + "up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_28" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_28" + ], + "const_args": [ + "onnx::MatMul_5954" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_71" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_25", + "type": "SDMHA", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_71", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_18", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_19" + ], + "const_args": [ + "MultiHeadAttention_25_mask.8_19" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_19" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_19" + ], + "const_args": [ + "onnx::MatMul_5964" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_39" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_39", + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_59" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_60" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_60" + ], + "const_args": [ + "up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight", + "up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_29" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul_2_9_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_29" + ], + "const_args": [ + "onnx::MatMul_5965_2_9_0" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_9" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul_2_9_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_29" + ], + "const_args": [ + "onnx::MatMul_5965_2_9_1" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_40" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "5120" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_40", + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_9" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_10" + ], + "const_args": [ + "onnx::MatMul_5966" + ], + "out_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_41" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "256", + "5120" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "5120", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_41", + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_60" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.2/Reshape_1_output_0.out_12_0_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "256", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_64-/up_blocks.1/attentions.2/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/attentions.2/Reshape_1_output_0.out_12_0_9" + ], + "const_args": [ + "NhwcConv_64_weight_NHWC" + ], + "out_args": [ + "NhwcConv_64_out-/up_blocks.1/attentions.2/proj_out/Conv_output_0.out0_1_54" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/attentions.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_64_out-/up_blocks.1/attentions.2/proj_out/Conv_output_0.out0_1_54", + "/up_blocks.1/resnets.2/Add_1.out_12_1_58" + ], + "const_args": [], + "out_args": [ + "/up_blocks.1/attentions.2/Add.out_12_1_61" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.1/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/up_blocks.1/attentions.2/Add.out_12_1_61" + ], + "const_args": [ + "/up_blocks.1/upsamplers.0/Resize.weights4_1" + ], + "out_args": [ + "/up_blocks.1/upsamplers.0/Resize.out_4_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "16", + "16", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_65-/up_blocks.1/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.1/upsamplers.0/Resize.out_4_1" + ], + "const_args": [ + "NhwcConv_65_weight_NHWC" + ], + "out_args": [ + "NhwcConv_65_out-/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_55" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/Concat", + "type": "SDConcat", + "in_args": [ + "NhwcConv_65_out-/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_55", + "/down_blocks.1/attentions.1/Add.out_12_1_19" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/Concat.out13_7" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_42", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/Concat.out13_7" + ], + "const_args": [ + "GroupNorm_42_wts_7_1_42" + ], + "out_args": [ + "GroupNorm_42.out4_42.out7_1_42" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "3840" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_68-/up_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/Concat.out13_7" + ], + "const_args": [ + "NhwcConv_68_weight_NHWC" + ], + "out_args": [ + "NhwcConv_68_out-/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_56" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_32", + "type": "SDSilu", + "in_args": [ + "GroupNorm_42.out4_42.out7_1_42" + ], + "const_args": [ + "Sigmoid_32.weights10_34" + ], + "out_args": [ + "GroupNorm_42_sigmoid_out.4_42.out10_34" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_66-/up_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_42_sigmoid_out.4_42.out10_34" + ], + "const_args": [ + "NhwcConv_66_weight_NHWC" + ], + "out_args": [ + "NhwcConv_66_out-/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_57" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1920" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "1920" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_66_out-/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_57", + "/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_16" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/resnets.0/Add.out_12_1_62" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_43", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/resnets.0/Add.out_12_1_62" + ], + "const_args": [ + "GroupNorm_43_wts_7_1_43" + ], + "out_args": [ + "GroupNorm_43.out4_43.out7_1_43" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_33", + "type": "SDSilu", + "in_args": [ + "GroupNorm_43.out4_43.out7_1_43" + ], + "const_args": [ + "Sigmoid_33.weights10_35" + ], + "out_args": [ + "GroupNorm_43_sigmoid_out.4_43.out10_35" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_67-/up_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_43_sigmoid_out.4_43.out10_35" + ], + "const_args": [ + "NhwcConv_67_weight_NHWC" + ], + "out_args": [ + "NhwcConv_67_out-/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_58" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_68_out-/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_56", + "NhwcConv_67_out-/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_58" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/resnets.0/Add_1.out_12_1_63" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_44", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/resnets.0/Add_1.out_12_1_63" + ], + "const_args": [ + "GroupNorm_44_wts_7_1_44" + ], + "out_args": [ + "/up_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_44" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_69-/up_blocks.2/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_44" + ], + "const_args": [ + "NhwcConv_69_weight_NHWC" + ], + "out_args": [ + "/up_blocks.2/attentions.0/Reshape_output_0.out0_0_10" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.0/Reshape_output_0.out0_0_10" + ], + "const_args": [ + "up_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", + "up_blocks.2.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_30" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_30" + ], + "const_args": [ + "onnx::MatMul_5975" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_72" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_30" + ], + "const_args": [ + "onnx::MatMul_5974" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_73" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_30" + ], + "const_args": [ + "onnx::MatMul_5976" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_74" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_10", + "type": "SDMHA", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_73", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_72", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_74" + ], + "const_args": [ + "MultiHeadAttention_10_mask.8_20" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_20" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_20" + ], + "const_args": [ + "onnx::MatMul_5984" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_42" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_42", + "/up_blocks.2/attentions.0/Reshape_output_0.out0_0_10" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_64" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_64" + ], + "const_args": [ + "up_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", + "up_blocks.2.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_31" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_31" + ], + "const_args": [ + "onnx::MatMul_5985" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_75" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_26", + "type": "SDMHA", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_75", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_20", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_21" + ], + "const_args": [ + "MultiHeadAttention_26_mask.8_21" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_21" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_21" + ], + "const_args": [ + "onnx::MatMul_5995" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_43" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_43", + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_64" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_65" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_65" + ], + "const_args": [ + "up_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", + "up_blocks.2.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_32" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_10_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_32" + ], + "const_args": [ + "onnx::MatMul_5996_2_10_0" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_10" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_10_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_32" + ], + "const_args": [ + "onnx::MatMul_5996_2_10_1" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_44" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_44", + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_10" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_11" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_11" + ], + "const_args": [ + "onnx::MatMul_5997" + ], + "out_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_45" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_45", + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_65" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_70-/up_blocks.2/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_10" + ], + "const_args": [ + "NhwcConv_70_weight_NHWC" + ], + "out_args": [ + "NhwcConv_70_out-/up_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_59" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_70_out-/up_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_59", + "/up_blocks.2/resnets.0/Add_1.out_12_1_63" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.0/Add.out_12_1_66" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/Concat_1", + "type": "SDConcat", + "in_args": [ + "/up_blocks.2/attentions.0/Add.out_12_1_66", + "/down_blocks.1/attentions.0/Add.out_12_1_14" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/Concat_1.out13_8" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_45", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/Concat_1.out13_8" + ], + "const_args": [ + "GroupNorm_45_wts_7_1_45" + ], + "out_args": [ + "GroupNorm_45.out4_45.out7_1_45" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_73-/up_blocks.2/resnets.1/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/Concat_1.out13_8" + ], + "const_args": [ + "NhwcConv_73_weight_NHWC" + ], + "out_args": [ + "NhwcConv_73_out-/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_60" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_34", + "type": "SDSilu", + "in_args": [ + "GroupNorm_45.out4_45.out7_1_45" + ], + "const_args": [ + "Sigmoid_34.weights10_36" + ], + "out_args": [ + "GroupNorm_45_sigmoid_out.4_45.out10_36" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_71-/up_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_45_sigmoid_out.4_45.out10_36" + ], + "const_args": [ + "NhwcConv_71_weight_NHWC" + ], + "out_args": [ + "NhwcConv_71_out-/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_61" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "1280" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "1280" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_71_out-/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_61", + "/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_17" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/resnets.1/Add.out_12_1_67" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_46", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/resnets.1/Add.out_12_1_67" + ], + "const_args": [ + "GroupNorm_46_wts_7_1_46" + ], + "out_args": [ + "GroupNorm_46.out4_46.out7_1_46" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_35", + "type": "SDSilu", + "in_args": [ + "GroupNorm_46.out4_46.out7_1_46" + ], + "const_args": [ + "Sigmoid_35.weights10_37" + ], + "out_args": [ + "GroupNorm_46_sigmoid_out.4_46.out10_37" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_72-/up_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_46_sigmoid_out.4_46.out10_37" + ], + "const_args": [ + "NhwcConv_72_weight_NHWC" + ], + "out_args": [ + "NhwcConv_72_out-/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_62" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_73_out-/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_60", + "NhwcConv_72_out-/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_62" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/resnets.1/Add_1.out_12_1_68" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_47", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/resnets.1/Add_1.out_12_1_68" + ], + "const_args": [ + "GroupNorm_47_wts_7_1_47" + ], + "out_args": [ + "/up_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_47" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_74-/up_blocks.2/attentions.1/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_47" + ], + "const_args": [ + "NhwcConv_74_weight_NHWC" + ], + "out_args": [ + "/up_blocks.2/attentions.1/Reshape_output_0.out0_0_11" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.1/Reshape_output_0.out0_0_11" + ], + "const_args": [ + "up_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", + "up_blocks.2.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_33" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_33" + ], + "const_args": [ + "onnx::MatMul_6005" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_76" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_33" + ], + "const_args": [ + "onnx::MatMul_6004" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_77" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_33" + ], + "const_args": [ + "onnx::MatMul_6006" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_78" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_11", + "type": "SDMHA", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_77", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_76", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_78" + ], + "const_args": [ + "MultiHeadAttention_11_mask.8_22" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_22" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_22" + ], + "const_args": [ + "onnx::MatMul_6014" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_46" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_46", + "/up_blocks.2/attentions.1/Reshape_output_0.out0_0_11" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_69" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_69" + ], + "const_args": [ + "up_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", + "up_blocks.2.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_34" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_34" + ], + "const_args": [ + "onnx::MatMul_6015" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_79" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_27", + "type": "SDMHA", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_79", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_22", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_23" + ], + "const_args": [ + "MultiHeadAttention_27_mask.8_23" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_23" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_23" + ], + "const_args": [ + "onnx::MatMul_6025" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_47" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_47", + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_69" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_70" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_70" + ], + "const_args": [ + "up_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", + "up_blocks.2.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_35" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_11_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_35" + ], + "const_args": [ + "onnx::MatMul_6026_2_11_0" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_11" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_11_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_35" + ], + "const_args": [ + "onnx::MatMul_6026_2_11_1" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_48" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_48", + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_11" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_12" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_12" + ], + "const_args": [ + "onnx::MatMul_6027" + ], + "out_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_49" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_49", + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_70" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_11" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_75-/up_blocks.2/attentions.1/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_11" + ], + "const_args": [ + "NhwcConv_75_weight_NHWC" + ], + "out_args": [ + "NhwcConv_75_out-/up_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_63" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_75_out-/up_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_63", + "/up_blocks.2/resnets.1/Add_1.out_12_1_68" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.1/Add.out_12_1_71" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/Concat_2", + "type": "SDConcat", + "in_args": [ + "/up_blocks.2/attentions.1/Add.out_12_1_71", + "NhwcConv_9_out-/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_7" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/Concat_2.out13_9" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_48", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/Concat_2.out13_9" + ], + "const_args": [ + "GroupNorm_48_wts_7_1_48" + ], + "out_args": [ + "GroupNorm_48.out4_48.out7_1_48" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_78-/up_blocks.2/resnets.2/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/Concat_2.out13_9" + ], + "const_args": [ + "NhwcConv_78_weight_NHWC" + ], + "out_args": [ + "NhwcConv_78_out-/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_64" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_36", + "type": "SDSilu", + "in_args": [ + "GroupNorm_48.out4_48.out7_1_48" + ], + "const_args": [ + "Sigmoid_36.weights10_38" + ], + "out_args": [ + "GroupNorm_48_sigmoid_out.4_48.out10_38" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_76-/up_blocks.2/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_48_sigmoid_out.4_48.out10_38" + ], + "const_args": [ + "NhwcConv_76_weight_NHWC" + ], + "out_args": [ + "NhwcConv_76_out-/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_65" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_76_out-/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_65", + "/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_18" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/resnets.2/Add.out_12_1_72" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_49", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/resnets.2/Add.out_12_1_72" + ], + "const_args": [ + "GroupNorm_49_wts_7_1_49" + ], + "out_args": [ + "GroupNorm_49.out4_49.out7_1_49" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_37", + "type": "SDSilu", + "in_args": [ + "GroupNorm_49.out4_49.out7_1_49" + ], + "const_args": [ + "Sigmoid_37.weights10_39" + ], + "out_args": [ + "GroupNorm_49_sigmoid_out.4_49.out10_39" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_77-/up_blocks.2/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_49_sigmoid_out.4_49.out10_39" + ], + "const_args": [ + "NhwcConv_77_weight_NHWC" + ], + "out_args": [ + "NhwcConv_77_out-/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_66" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/resnets.2/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_78_out-/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_64", + "NhwcConv_77_out-/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_66" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/resnets.2/Add_1.out_12_1_73" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_50", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.2/resnets.2/Add_1.out_12_1_73" + ], + "const_args": [ + "GroupNorm_50_wts_7_1_50" + ], + "out_args": [ + "/up_blocks.2/attentions.2/norm/Add_output_0_NHWC.out7_1_50" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_79-/up_blocks.2/attentions.2/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/attentions.2/norm/Add_output_0_NHWC.out7_1_50" + ], + "const_args": [ + "NhwcConv_79_weight_NHWC" + ], + "out_args": [ + "/up_blocks.2/attentions.2/Reshape_output_0.out0_0_12" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.2/Reshape_output_0.out0_0_12" + ], + "const_args": [ + "up_blocks.2.attentions.2.transformer_blocks.0.norm1.weight", + "up_blocks.2.attentions.2.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_36" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_36" + ], + "const_args": [ + "onnx::MatMul_6035" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_80" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_36" + ], + "const_args": [ + "onnx::MatMul_6034" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_81" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_36" + ], + "const_args": [ + "onnx::MatMul_6036" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_82" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_12", + "type": "SDMHA", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_81", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_80", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_82" + ], + "const_args": [ + "MultiHeadAttention_12_mask.8_24" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_24" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_24" + ], + "const_args": [ + "onnx::MatMul_6044" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_50" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_50", + "/up_blocks.2/attentions.2/Reshape_output_0.out0_0_12" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_74" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_74" + ], + "const_args": [ + "up_blocks.2.attentions.2.transformer_blocks.0.norm2.weight", + "up_blocks.2.attentions.2.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_37" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_37" + ], + "const_args": [ + "onnx::MatMul_6045" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_83" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_28", + "type": "SDMHA", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_83", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_24", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_25" + ], + "const_args": [ + "MultiHeadAttention_28_mask.8_25" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_25" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_25" + ], + "const_args": [ + "onnx::MatMul_6055" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_51" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_51", + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_74" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_75" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_75" + ], + "const_args": [ + "up_blocks.2.attentions.2.transformer_blocks.0.norm3.weight", + "up_blocks.2.attentions.2.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_38" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul_2_12_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_38" + ], + "const_args": [ + "onnx::MatMul_6056_2_12_0" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_12" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul_2_12_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_38" + ], + "const_args": [ + "onnx::MatMul_6056_2_12_1" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_52" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "2560" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_52", + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_12" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_13" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_13" + ], + "const_args": [ + "onnx::MatMul_6057" + ], + "out_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_53" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "1024", + "2560" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "2560", + "640" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_53", + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_75" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.2/Reshape_1_output_0.out_12_0_12" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "1024", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_80-/up_blocks.2/attentions.2/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/attentions.2/Reshape_1_output_0.out_12_0_12" + ], + "const_args": [ + "NhwcConv_80_weight_NHWC" + ], + "out_args": [ + "NhwcConv_80_out-/up_blocks.2/attentions.2/proj_out/Conv_output_0.out0_1_67" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/attentions.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_80_out-/up_blocks.2/attentions.2/proj_out/Conv_output_0.out0_1_67", + "/up_blocks.2/resnets.2/Add_1.out_12_1_73" + ], + "const_args": [], + "out_args": [ + "/up_blocks.2/attentions.2/Add.out_12_1_76" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.2/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/up_blocks.2/attentions.2/Add.out_12_1_76" + ], + "const_args": [ + "/up_blocks.2/upsamplers.0/Resize.weights4_2" + ], + "out_args": [ + "/up_blocks.2/upsamplers.0/Resize.out_4_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "32", + "32", + "640" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_81-/up_blocks.2/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.2/upsamplers.0/Resize.out_4_2" + ], + "const_args": [ + "NhwcConv_81_weight_NHWC" + ], + "out_args": [ + "NhwcConv_81_out-/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_68" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "640", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/Concat", + "type": "SDConcat", + "in_args": [ + "NhwcConv_81_out-/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_68", + "/down_blocks.0/attentions.1/Add.out_12_1_9" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/Concat.out13_10" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_51", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/Concat.out13_10" + ], + "const_args": [ + "GroupNorm_51_wts_7_1_51" + ], + "out_args": [ + "GroupNorm_51.out4_51.out7_1_51" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1920" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_84-/up_blocks.3/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/Concat.out13_10" + ], + "const_args": [ + "NhwcConv_84_weight_NHWC" + ], + "out_args": [ + "NhwcConv_84_out-/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_69" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_38", + "type": "SDSilu", + "in_args": [ + "GroupNorm_51.out4_51.out7_1_51" + ], + "const_args": [ + "Sigmoid_38.weights10_40" + ], + "out_args": [ + "GroupNorm_51_sigmoid_out.4_51.out10_40" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_82-/up_blocks.3/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_51_sigmoid_out.4_51.out10_40" + ], + "const_args": [ + "NhwcConv_82_weight_NHWC" + ], + "out_args": [ + "NhwcConv_82_out-/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_70" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "960" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "960" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_82_out-/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_70", + "/up_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_19" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/resnets.0/Add.out_12_1_77" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_52", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/resnets.0/Add.out_12_1_77" + ], + "const_args": [ + "GroupNorm_52_wts_7_1_52" + ], + "out_args": [ + "GroupNorm_52.out4_52.out7_1_52" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_39", + "type": "SDSilu", + "in_args": [ + "GroupNorm_52.out4_52.out7_1_52" + ], + "const_args": [ + "Sigmoid_39.weights10_41" + ], + "out_args": [ + "GroupNorm_52_sigmoid_out.4_52.out10_41" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_83-/up_blocks.3/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_52_sigmoid_out.4_52.out10_41" + ], + "const_args": [ + "NhwcConv_83_weight_NHWC" + ], + "out_args": [ + "NhwcConv_83_out-/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_71" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.0/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_84_out-/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_69", + "NhwcConv_83_out-/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_71" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/resnets.0/Add_1.out_12_1_78" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_53", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/resnets.0/Add_1.out_12_1_78" + ], + "const_args": [ + "GroupNorm_53_wts_7_1_53" + ], + "out_args": [ + "/up_blocks.3/attentions.0/norm/Add_output_0_NHWC.out7_1_53" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_85-/up_blocks.3/attentions.0/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/attentions.0/norm/Add_output_0_NHWC.out7_1_53" + ], + "const_args": [ + "NhwcConv_85_weight_NHWC" + ], + "out_args": [ + "/up_blocks.3/attentions.0/Reshape_output_0.out0_0_13" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.0/Reshape_output_0.out0_0_13" + ], + "const_args": [ + "up_blocks.3.attentions.0.transformer_blocks.0.norm1.weight", + "up_blocks.3.attentions.0.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_39" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_39" + ], + "const_args": [ + "onnx::MatMul_6066" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_84" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_39" + ], + "const_args": [ + "onnx::MatMul_6065" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_85" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_39" + ], + "const_args": [ + "onnx::MatMul_6067" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_86" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_13", + "type": "SDMHA", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_85", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_84", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_86" + ], + "const_args": [ + "MultiHeadAttention_13_mask.8_26" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_26" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_26" + ], + "const_args": [ + "onnx::MatMul_6075" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_54" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_54", + "/up_blocks.3/attentions.0/Reshape_output_0.out0_0_13" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_79" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_79" + ], + "const_args": [ + "up_blocks.3.attentions.0.transformer_blocks.0.norm2.weight", + "up_blocks.3.attentions.0.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_40" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_40" + ], + "const_args": [ + "onnx::MatMul_6076" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_87" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_29", + "type": "SDMHA", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_87", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_26", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_27" + ], + "const_args": [ + "MultiHeadAttention_29_mask.8_27" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_27" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_27" + ], + "const_args": [ + "onnx::MatMul_6086" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_55" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_55", + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_79" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_80" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_80" + ], + "const_args": [ + "up_blocks.3.attentions.0.transformer_blocks.0.norm3.weight", + "up_blocks.3.attentions.0.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_41" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_13_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_41" + ], + "const_args": [ + "onnx::MatMul_6087_2_13_0" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_13" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/proj/MatMul_2_13_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_41" + ], + "const_args": [ + "onnx::MatMul_6087_2_13_1" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_56" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_56", + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_13" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_14" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_14" + ], + "const_args": [ + "onnx::MatMul_6088" + ], + "out_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_57" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_57", + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_80" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.0/Reshape_1_output_0.out_12_0_13" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_86-/up_blocks.3/attentions.0/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/attentions.0/Reshape_1_output_0.out_12_0_13" + ], + "const_args": [ + "NhwcConv_86_weight_NHWC" + ], + "out_args": [ + "NhwcConv_86_out-/up_blocks.3/attentions.0/proj_out/Conv_output_0.out0_1_72" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_86_out-/up_blocks.3/attentions.0/proj_out/Conv_output_0.out0_1_72", + "/up_blocks.3/resnets.0/Add_1.out_12_1_78" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.0/Add.out_12_1_81" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/Concat_1", + "type": "SDConcat", + "in_args": [ + "/up_blocks.3/attentions.0/Add.out_12_1_81", + "/down_blocks.0/attentions.0/Add.out_12_1_4" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/Concat_1.out13_11" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_54", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/Concat_1.out13_11" + ], + "const_args": [ + "GroupNorm_54_wts_7_1_54" + ], + "out_args": [ + "GroupNorm_54.out4_54.out7_1_54" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_89-/up_blocks.3/resnets.1/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/Concat_1.out13_11" + ], + "const_args": [ + "NhwcConv_89_weight_NHWC" + ], + "out_args": [ + "NhwcConv_89_out-/up_blocks.3/resnets.1/conv_shortcut/Conv_output_0.out0_1_73" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_40", + "type": "SDSilu", + "in_args": [ + "GroupNorm_54.out4_54.out7_1_54" + ], + "const_args": [ + "Sigmoid_40.weights10_42" + ], + "out_args": [ + "GroupNorm_54_sigmoid_out.4_54.out10_42" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_87-/up_blocks.3/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_54_sigmoid_out.4_54.out10_42" + ], + "const_args": [ + "NhwcConv_87_weight_NHWC" + ], + "out_args": [ + "NhwcConv_87_out-/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_74" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_87_out-/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_74", + "/up_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_20" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/resnets.1/Add.out_12_1_82" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_55", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/resnets.1/Add.out_12_1_82" + ], + "const_args": [ + "GroupNorm_55_wts_7_1_55" + ], + "out_args": [ + "GroupNorm_55.out4_55.out7_1_55" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_41", + "type": "SDSilu", + "in_args": [ + "GroupNorm_55.out4_55.out7_1_55" + ], + "const_args": [ + "Sigmoid_41.weights10_43" + ], + "out_args": [ + "GroupNorm_55_sigmoid_out.4_55.out10_43" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_88-/up_blocks.3/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_55_sigmoid_out.4_55.out10_43" + ], + "const_args": [ + "NhwcConv_88_weight_NHWC" + ], + "out_args": [ + "NhwcConv_88_out-/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_75" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.1/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_89_out-/up_blocks.3/resnets.1/conv_shortcut/Conv_output_0.out0_1_73", + "NhwcConv_88_out-/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_75" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/resnets.1/Add_1.out_12_1_83" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_56", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/resnets.1/Add_1.out_12_1_83" + ], + "const_args": [ + "GroupNorm_56_wts_7_1_56" + ], + "out_args": [ + "/up_blocks.3/attentions.1/norm/Add_output_0_NHWC.out7_1_56" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_90-/up_blocks.3/attentions.1/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/attentions.1/norm/Add_output_0_NHWC.out7_1_56" + ], + "const_args": [ + "NhwcConv_90_weight_NHWC" + ], + "out_args": [ + "/up_blocks.3/attentions.1/Reshape_output_0.out0_0_14" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.1/Reshape_output_0.out0_0_14" + ], + "const_args": [ + "up_blocks.3.attentions.1.transformer_blocks.0.norm1.weight", + "up_blocks.3.attentions.1.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_42" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_42" + ], + "const_args": [ + "onnx::MatMul_6096" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_88" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_42" + ], + "const_args": [ + "onnx::MatMul_6095" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_89" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_42" + ], + "const_args": [ + "onnx::MatMul_6097" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_90" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_14", + "type": "SDMHA", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_89", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_88", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_90" + ], + "const_args": [ + "MultiHeadAttention_14_mask.8_28" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28" + ], + "const_args": [ + "onnx::MatMul_6105" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_58" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_58", + "/up_blocks.3/attentions.1/Reshape_output_0.out0_0_14" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_84" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_84" + ], + "const_args": [ + "up_blocks.3.attentions.1.transformer_blocks.0.norm2.weight", + "up_blocks.3.attentions.1.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_43" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_43" + ], + "const_args": [ + "onnx::MatMul_6106" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_91" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_30", + "type": "SDMHA", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_91", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29" + ], + "const_args": [ + "MultiHeadAttention_30_mask.8_29" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29" + ], + "const_args": [ + "onnx::MatMul_6116" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_59" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_59", + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_84" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_85" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_85" + ], + "const_args": [ + "up_blocks.3.attentions.1.transformer_blocks.0.norm3.weight", + "up_blocks.3.attentions.1.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_44" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_14_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_44" + ], + "const_args": [ + "onnx::MatMul_6117_2_14_0" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/proj/MatMul_2_14_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_44" + ], + "const_args": [ + "onnx::MatMul_6117_2_14_1" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_60" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_60", + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_15" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_15" + ], + "const_args": [ + "onnx::MatMul_6118" + ], + "out_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_61" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_61", + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_85" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.1/Reshape_1_output_0.out_12_0_14" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_91-/up_blocks.3/attentions.1/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/attentions.1/Reshape_1_output_0.out_12_0_14" + ], + "const_args": [ + "NhwcConv_91_weight_NHWC" + ], + "out_args": [ + "NhwcConv_91_out-/up_blocks.3/attentions.1/proj_out/Conv_output_0.out0_1_76" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.1/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_91_out-/up_blocks.3/attentions.1/proj_out/Conv_output_0.out0_1_76", + "/up_blocks.3/resnets.1/Add_1.out_12_1_83" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.1/Add.out_12_1_86" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/Concat_2", + "type": "SDConcat", + "in_args": [ + "/up_blocks.3/attentions.1/Add.out_12_1_86", + "NhwcConv_0_out-/conv_in/Conv_output_0.out0_1_0" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/Concat_2.out13_12" + ], + "attrs": { + "axis": { + "type": "int", + "value": [ + "3" + ] + }, + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_57", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/Concat_2.out13_12" + ], + "const_args": [ + "GroupNorm_57_wts_7_1_57" + ], + "out_args": [ + "GroupNorm_57.out4_57.out7_1_57" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_94-/up_blocks.3/resnets.2/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/Concat_2.out13_12" + ], + "const_args": [ + "NhwcConv_94_weight_NHWC" + ], + "out_args": [ + "NhwcConv_94_out-/up_blocks.3/resnets.2/conv_shortcut/Conv_output_0.out0_1_77" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_42", + "type": "SDSilu", + "in_args": [ + "GroupNorm_57.out4_57.out7_1_57" + ], + "const_args": [ + "Sigmoid_42.weights10_44" + ], + "out_args": [ + "GroupNorm_57_sigmoid_out.4_57.out10_44" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_92-/up_blocks.3/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_57_sigmoid_out.4_57.out10_44" + ], + "const_args": [ + "NhwcConv_92_weight_NHWC" + ], + "out_args": [ + "NhwcConv_92_out-/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_78" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "640" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "640" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_92_out-/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_78", + "/up_blocks.3/resnets.2/Unsqueeze_1_output_0.out2_0_21" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/resnets.2/Add.out_12_1_87" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "1", + "1", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_58", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/resnets.2/Add.out_12_1_87" + ], + "const_args": [ + "GroupNorm_58_wts_7_1_58" + ], + "out_args": [ + "GroupNorm_58.out4_58.out7_1_58" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_43", + "type": "SDSilu", + "in_args": [ + "GroupNorm_58.out4_58.out7_1_58" + ], + "const_args": [ + "Sigmoid_43.weights10_45" + ], + "out_args": [ + "GroupNorm_58_sigmoid_out.4_58.out10_45" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_93-/up_blocks.3/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_58_sigmoid_out.4_58.out10_45" + ], + "const_args": [ + "NhwcConv_93_weight_NHWC" + ], + "out_args": [ + "NhwcConv_93_out-/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_79" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/resnets.2/Add_1", + "type": "SDAdd", + "in_args": [ + "NhwcConv_94_out-/up_blocks.3/resnets.2/conv_shortcut/Conv_output_0.out0_1_77", + "NhwcConv_93_out-/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_79" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/resnets.2/Add_1.out_12_1_88" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_59", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/resnets.2/Add_1.out_12_1_88" + ], + "const_args": [ + "GroupNorm_59_wts_7_1_59" + ], + "out_args": [ + "/up_blocks.3/attentions.2/norm/Add_output_0_NHWC.out7_1_59" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_95-/up_blocks.3/attentions.2/proj_in/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/attentions.2/norm/Add_output_0_NHWC.out7_1_59" + ], + "const_args": [ + "NhwcConv_95_weight_NHWC" + ], + "out_args": [ + "/up_blocks.3/attentions.2/Reshape_output_0.out0_0_15" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.2/Reshape_output_0.out0_0_15" + ], + "const_args": [ + "up_blocks.3.attentions.2.transformer_blocks.0.norm1.weight", + "up_blocks.3.attentions.2.transformer_blocks.0.norm1.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_45" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_45" + ], + "const_args": [ + "onnx::MatMul_6126" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_92" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_45" + ], + "const_args": [ + "onnx::MatMul_6125" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_93" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_45" + ], + "const_args": [ + "onnx::MatMul_6127" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_94" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_15", + "type": "SDMHA", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_93", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_92", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_94" + ], + "const_args": [ + "MultiHeadAttention_15_mask.8_30" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_30" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "4096" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_30" + ], + "const_args": [ + "onnx::MatMul_6135" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_62" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/Add", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_62", + "/up_blocks.3/attentions.2/Reshape_output_0.out0_0_15" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_89" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/norm2/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_89" + ], + "const_args": [ + "up_blocks.3.attentions.2.transformer_blocks.0.norm2.weight", + "up_blocks.3.attentions.2.transformer_blocks.0.norm2.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_46" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_46" + ], + "const_args": [ + "onnx::MatMul_6136" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_95" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "MultiHeadAttention_31", + "type": "SDMHA", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_95", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_30", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_31" + ], + "const_args": [ + "MultiHeadAttention_31_mask.8_31" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_31" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "8" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320", + "77" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_31" + ], + "const_args": [ + "onnx::MatMul_6146" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_63" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/Add_1", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_63", + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_89" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_90" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/norm3/LayerNormalization", + "type": "SDLayerNorm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_90" + ], + "const_args": [ + "up_blocks.3.attentions.2.transformer_blocks.0.norm3.weight", + "up_blocks.3.attentions.2.transformer_blocks.0.norm3.bias" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_47" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "320" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul_2_15_0", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_47" + ], + "const_args": [ + "onnx::MatMul_6147_2_15_0" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_15" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Gelu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/proj/MatMul_2_15_1", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_47" + ], + "const_args": [ + "onnx::MatMul_6147_2_15_1" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_64" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1280" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_4", + "type": "SDMul", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_64", + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_15" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_16" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.2/MatMul", + "type": "SDGemm", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_16" + ], + "const_args": [ + "onnx::MatMul_6148" + ], + "out_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_65" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "4096", + "1280" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "1280", + "320" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/transformer_blocks.0/Add_2", + "type": "SDAdd", + "in_args": [ + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_65", + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_90" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.2/Reshape_1_output_0.out_12_0_15" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "4096", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_96-/up_blocks.3/attentions.2/proj_out/Conv", + "type": "SDConv", + "in_args": [ + "/up_blocks.3/attentions.2/Reshape_1_output_0.out_12_0_15" + ], + "const_args": [ + "NhwcConv_96_weight_NHWC" + ], + "out_args": [ + "NhwcConv_96_out-/up_blocks.3/attentions.2/proj_out/Conv_output_0.out0_1_80" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "320", + "1", + "1", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/up_blocks.3/attentions.2/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_96_out-/up_blocks.3/attentions.2/proj_out/Conv_output_0.out0_1_80", + "/up_blocks.3/resnets.2/Add_1.out_12_1_88" + ], + "const_args": [], + "out_args": [ + "/up_blocks.3/attentions.2/Add.out_12_1_91" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_60", + "type": "SDGroupNorm", + "in_args": [ + "/up_blocks.3/attentions.2/Add.out_12_1_91" + ], + "const_args": [ + "GroupNorm_60_wts_7_1_60" + ], + "out_args": [ + "GroupNorm_60.out4_60.out7_1_60" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999747378752e-06" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "640" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "Sigmoid_44", + "type": "SDSilu", + "in_args": [ + "GroupNorm_60.out4_60.out7_1_60" + ], + "const_args": [ + "Sigmoid_44.weights10_46" + ], + "out_args": [ + "GroupNorm_60_sigmoid_out.4_60.out10_46" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_97-/conv_out/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_60_sigmoid_out.4_60.out10_46" + ], + "const_args": [ + "NhwcConv_97_weight_NHWC" + ], + "out_args": [ + "NhwcConv_97_out-out_sample.out0_1_81" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "320" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "2", + "64", + "64", + "4" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "4", + "3", + "3", + "320" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 7636224, + "xrt_arg_id": 0, + "packed_tensors": [ + "encoder_hidden_states.out2_6_0", + "Transpose_44_out-sample.out0_1_0", + "/time_proj/Concat_1_output_0.out2_3_0" + ] + }, + "out": { + "buffer_size": 65536, + "xrt_arg_id": 1, + "packed_tensors": [ + "NhwcConv_97_out-out_sample.out0_1_81" + ] + }, + "scratch": { + "buffer_size": 2037766400, + "xrt_arg_id": 2, + "packed_tensors": [ + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_2", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_3", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_6", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_7", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_10", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_11", + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_12", + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_13", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_14", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_15", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_16", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_17", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_18", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_19", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_20", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_21", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_22", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_23", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_24", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_25", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_26", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_27", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_30", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_31", + "NhwcConv_0_out-/conv_in/Conv_output_0.out0_1_0", + "GroupNorm_0.out4_0.out7_1_0", + "GroupNorm_0_sigmoid_out.4_0.out10_0", + "NhwcConv_1_out-/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1", + "/time_embedding/linear_1/Gemm_output_0.out2_3_0", + "/time_embedding/act/Sigmoid_output_0.out10_1", + "/time_embedding/linear_2/Gemm_output_0.out2_3_1", + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2", + "/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0", + "/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1", + "/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2", + "/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3", + "/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4", + "/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5", + "/down_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_6", + "/down_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_7", + "/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_8", + "/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_9", + "/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_10", + "/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_11", + "/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_12", + "/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_13", + "/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_14", + "/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_15", + "/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_16", + "/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_17", + "/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_18", + "/up_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_19", + "/up_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_20", + "/up_blocks.3/resnets.2/Unsqueeze_1_output_0.out2_0_21", + "/down_blocks.0/resnets.0/Add.out_12_1_0", + "GroupNorm_1.out4_1.out7_1_1", + "GroupNorm_1_sigmoid_out.4_1.out10_3", + "NhwcConv_2_out-/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2", + "/down_blocks.0/resnets.0/Add_1.out_12_1_1", + "/down_blocks.0/attentions.0/norm/Add_output_0_NHWC.out7_1_2", + "/down_blocks.0/attentions.0/Reshape_output_0.out0_0_0", + "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_0", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_32", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_33", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_34", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_2", + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_2", + "/down_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_1", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_35", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1", + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_3", + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_3", + "/down_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_2", + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0", + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_4", + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_1", + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_5", + "/down_blocks.0/attentions.0/Reshape_1_output_0.out_12_0_0", + "NhwcConv_4_out-/down_blocks.0/attentions.0/proj_out/Conv_output_0.out0_1_3", + "/down_blocks.0/attentions.0/Add.out_12_1_4", + "GroupNorm_3.out4_3.out7_1_3", + "GroupNorm_3_sigmoid_out.4_3.out10_4", + "NhwcConv_5_out-/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_4", + "/down_blocks.0/resnets.1/Add.out_12_1_5", + "GroupNorm_4.out4_4.out7_1_4", + "GroupNorm_4_sigmoid_out.4_4.out10_5", + "NhwcConv_6_out-/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_5", + "/down_blocks.0/resnets.1/Add_1.out_12_1_6", + "/down_blocks.0/attentions.1/norm/Add_output_0_NHWC.out7_1_5", + "/down_blocks.0/attentions.1/Reshape_output_0.out0_0_1", + "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_3", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_36", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_37", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_38", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_2", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_6", + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_7", + "/down_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_4", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_39", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_3", + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_7", + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_8", + "/down_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_5", + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_1", + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_8", + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_2", + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_9", + "/down_blocks.0/attentions.1/Reshape_1_output_0.out_12_0_1", + "NhwcConv_8_out-/down_blocks.0/attentions.1/proj_out/Conv_output_0.out0_1_6", + "/down_blocks.0/attentions.1/Add.out_12_1_9", + "NhwcConv_9_out-/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_7", + "GroupNorm_6.out4_6.out7_1_6", + "NhwcConv_12_out-/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_8", + "GroupNorm_6_sigmoid_out.4_6.out10_6", + "NhwcConv_10_out-/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_9", + "/down_blocks.1/resnets.0/Add.out_12_1_10", + "GroupNorm_7.out4_7.out7_1_7", + "GroupNorm_7_sigmoid_out.4_7.out10_7", + "NhwcConv_11_out-/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_10", + "/down_blocks.1/resnets.0/Add_1.out_12_1_11", + "/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_8", + "/down_blocks.1/attentions.0/Reshape_output_0.out0_0_2", + "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_6", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_40", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_41", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_42", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_10", + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_12", + "/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_7", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_43", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5", + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_11", + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_13", + "/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_8", + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2", + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_12", + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_3", + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_13", + "/down_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_2", + "NhwcConv_14_out-/down_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_11", + "/down_blocks.1/attentions.0/Add.out_12_1_14", + "GroupNorm_9.out4_9.out7_1_9", + "GroupNorm_9_sigmoid_out.4_9.out10_8", + "NhwcConv_15_out-/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_12", + "/down_blocks.1/resnets.1/Add.out_12_1_15", + "GroupNorm_10.out4_10.out7_1_10", + "GroupNorm_10_sigmoid_out.4_10.out10_9", + "NhwcConv_16_out-/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_13", + "/down_blocks.1/resnets.1/Add_1.out_12_1_16", + "/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_11", + "/down_blocks.1/attentions.1/Reshape_output_0.out0_0_3", + "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_9", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_44", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_45", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_46", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_6", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14", + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_17", + "/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_10", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_47", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_7", + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15", + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_18", + "/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_11", + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_3", + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16", + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_4", + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17", + "/down_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_3", + "NhwcConv_18_out-/down_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_14", + "/down_blocks.1/attentions.1/Add.out_12_1_19", + "NhwcConv_19_out-/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_15", + "NhwcConv_22_out-/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_16", + "GroupNorm_12.out4_12.out7_1_12", + "GroupNorm_12_sigmoid_out.4_12.out10_10", + "NhwcConv_20_out-/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_17", + "/down_blocks.2/resnets.0/Add.out_12_1_20", + "GroupNorm_13.out4_13.out7_1_13", + "GroupNorm_13_sigmoid_out.4_13.out10_11", + "NhwcConv_21_out-/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_18", + "/down_blocks.2/resnets.0/Add_1.out_12_1_21", + "/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_14", + "/down_blocks.2/attentions.0/Reshape_output_0.out0_0_4", + "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_12", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_48", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_49", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_50", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_18", + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_22", + "/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_13", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_51", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9", + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_19", + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_23", + "/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_14", + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4", + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_20", + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_5", + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_21", + "/down_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_4", + "NhwcConv_24_out-/down_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_19", + "/down_blocks.2/attentions.0/Add.out_12_1_24", + "GroupNorm_15.out4_15.out7_1_15", + "GroupNorm_15_sigmoid_out.4_15.out10_12", + "NhwcConv_25_out-/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_20", + "/down_blocks.2/resnets.1/Add.out_12_1_25", + "GroupNorm_16.out4_16.out7_1_16", + "GroupNorm_16_sigmoid_out.4_16.out10_13", + "NhwcConv_26_out-/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_21", + "/down_blocks.2/resnets.1/Add_1.out_12_1_26", + "/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_17", + "/down_blocks.2/attentions.1/Reshape_output_0.out0_0_5", + "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_15", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_52", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_53", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_54", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_10", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_22", + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_27", + "/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_16", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_55", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_11", + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_23", + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_28", + "/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_17", + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_5", + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_24", + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_6", + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_25", + "/down_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_5", + "NhwcConv_28_out-/down_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_22", + "/down_blocks.2/attentions.1/Add.out_12_1_29", + "NhwcConv_29_out-/down_blocks.2/downsamplers.0/conv/Conv_output_0.out0_1_23", + "GroupNorm_18.out4_18.out7_1_18", + "GroupNorm_18_sigmoid_out.4_18.out10_14", + "NhwcConv_30_out-/down_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_24", + "/down_blocks.3/resnets.0/Add.out_12_1_30", + "GroupNorm_19.out4_19.out7_1_19", + "GroupNorm_19_sigmoid_out.4_19.out10_15", + "NhwcConv_31_out-/down_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_25", + "/down_blocks.3/resnets.0/Add_1.out_12_1_31", + "GroupNorm_20.out4_20.out7_1_20", + "GroupNorm_20_sigmoid_out.4_20.out10_16", + "NhwcConv_32_out-/down_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_26", + "/down_blocks.3/resnets.1/Add.out_12_1_32", + "GroupNorm_21.out4_21.out7_1_21", + "GroupNorm_21_sigmoid_out.4_21.out10_17", + "NhwcConv_33_out-/down_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_27", + "/down_blocks.3/resnets.1/Add_1.out_12_1_33", + "GroupNorm_22.out4_22.out7_1_22", + "GroupNorm_22_sigmoid_out.4_22.out10_18", + "NhwcConv_34_out-/mid_block/resnets.0/conv1/Conv_output_0.out0_1_28", + "/mid_block/resnets.0/Add.out_12_1_34", + "GroupNorm_23.out4_23.out7_1_23", + "GroupNorm_23_sigmoid_out.4_23.out10_19", + "NhwcConv_35_out-/mid_block/resnets.0/conv2/Conv_output_0.out0_1_29", + "/mid_block/resnets.0/Add_1.out_12_1_35", + "/mid_block/attentions.0/norm/Add_output_0_NHWC.out7_1_24", + "/mid_block/attentions.0/Reshape_output_0.out0_0_6", + "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_18", + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_56", + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_57", + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_58", + "/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_12", + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_26", + "/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_36", + "/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_19", + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_59", + "/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_13", + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_27", + "/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_37", + "/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_20", + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_6", + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_28", + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_7", + "/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_29", + "/mid_block/attentions.0/Reshape_1_output_0.out_12_0_6", + "NhwcConv_37_out-/mid_block/attentions.0/proj_out/Conv_output_0.out0_1_30", + "/mid_block/attentions.0/Add.out_12_1_38", + "GroupNorm_25.out4_25.out7_1_25", + "GroupNorm_25_sigmoid_out.4_25.out10_20", + "NhwcConv_38_out-/mid_block/resnets.1/conv1/Conv_output_0.out0_1_31", + "/mid_block/resnets.1/Add.out_12_1_39", + "GroupNorm_26.out4_26.out7_1_26", + "GroupNorm_26_sigmoid_out.4_26.out10_21", + "NhwcConv_39_out-/mid_block/resnets.1/conv2/Conv_output_0.out0_1_32", + "/mid_block/resnets.1/Add_1.out_12_1_40", + "/up_blocks.0/Concat.out13_1", + "GroupNorm_27.out4_27.out7_1_27", + "NhwcConv_42_out-/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_33", + "GroupNorm_27_sigmoid_out.4_27.out10_22", + "NhwcConv_40_out-/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_34", + "/up_blocks.0/resnets.0/Add.out_12_1_41", + "GroupNorm_28.out4_28.out7_1_28", + "GroupNorm_28_sigmoid_out.4_28.out10_23", + "NhwcConv_41_out-/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_35", + "/up_blocks.0/resnets.0/Add_1.out_12_1_42", + "/up_blocks.0/Concat_1.out13_2", + "GroupNorm_29.out4_29.out7_1_29", + "NhwcConv_45_out-/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_36", + "GroupNorm_29_sigmoid_out.4_29.out10_24", + "NhwcConv_43_out-/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_37", + "/up_blocks.0/resnets.1/Add.out_12_1_43", + "GroupNorm_30.out4_30.out7_1_30", + "GroupNorm_30_sigmoid_out.4_30.out10_25", + "NhwcConv_44_out-/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_38", + "/up_blocks.0/resnets.1/Add_1.out_12_1_44", + "/up_blocks.0/Concat_2.out13_3", + "GroupNorm_31.out4_31.out7_1_31", + "NhwcConv_48_out-/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_39", + "GroupNorm_31_sigmoid_out.4_31.out10_26", + "NhwcConv_46_out-/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_40", + "/up_blocks.0/resnets.2/Add.out_12_1_45", + "GroupNorm_32.out4_32.out7_1_32", + "GroupNorm_32_sigmoid_out.4_32.out10_27", + "NhwcConv_47_out-/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_41", + "/up_blocks.0/resnets.2/Add_1.out_12_1_46", + "/up_blocks.0/upsamplers.0/Resize.out_4_0", + "NhwcConv_49_out-/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_42", + "/up_blocks.1/Concat.out13_4", + "GroupNorm_33.out4_33.out7_1_33", + "NhwcConv_52_out-/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_43", + "GroupNorm_33_sigmoid_out.4_33.out10_28", + "NhwcConv_50_out-/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_44", + "/up_blocks.1/resnets.0/Add.out_12_1_47", + "GroupNorm_34.out4_34.out7_1_34", + "GroupNorm_34_sigmoid_out.4_34.out10_29", + "NhwcConv_51_out-/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_45", + "/up_blocks.1/resnets.0/Add_1.out_12_1_48", + "/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_35", + "/up_blocks.1/attentions.0/Reshape_output_0.out0_0_7", + "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_21", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_60", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_61", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_62", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_14", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_30", + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_49", + "/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_22", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_63", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_15", + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_31", + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_50", + "/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_23", + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_7", + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_32", + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_8", + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_33", + "/up_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_7", + "NhwcConv_54_out-/up_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_46", + "/up_blocks.1/attentions.0/Add.out_12_1_51", + "/up_blocks.1/Concat_1.out13_5", + "GroupNorm_36.out4_36.out7_1_36", + "NhwcConv_57_out-/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_47", + "GroupNorm_36_sigmoid_out.4_36.out10_30", + "NhwcConv_55_out-/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_48", + "/up_blocks.1/resnets.1/Add.out_12_1_52", + "GroupNorm_37.out4_37.out7_1_37", + "GroupNorm_37_sigmoid_out.4_37.out10_31", + "NhwcConv_56_out-/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_49", + "/up_blocks.1/resnets.1/Add_1.out_12_1_53", + "/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_38", + "/up_blocks.1/attentions.1/Reshape_output_0.out0_0_8", + "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_24", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_64", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_65", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_66", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_16", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_34", + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_54", + "/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_25", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_67", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_17", + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_35", + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_55", + "/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_26", + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_8", + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_36", + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_9", + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_37", + "/up_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_8", + "NhwcConv_59_out-/up_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_50", + "/up_blocks.1/attentions.1/Add.out_12_1_56", + "/up_blocks.1/Concat_2.out13_6", + "GroupNorm_39.out4_39.out7_1_39", + "NhwcConv_62_out-/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_51", + "GroupNorm_39_sigmoid_out.4_39.out10_32", + "NhwcConv_60_out-/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_52", + "/up_blocks.1/resnets.2/Add.out_12_1_57", + "GroupNorm_40.out4_40.out7_1_40", + "GroupNorm_40_sigmoid_out.4_40.out10_33", + "NhwcConv_61_out-/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_53", + "/up_blocks.1/resnets.2/Add_1.out_12_1_58", + "/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out7_1_41", + "/up_blocks.1/attentions.2/Reshape_output_0.out0_0_9", + "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_27", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_68", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_69", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_70", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_18", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_38", + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_59", + "/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_28", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_71", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_19", + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_39", + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_60", + "/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_29", + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_9", + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_40", + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_10", + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_41", + "/up_blocks.1/attentions.2/Reshape_1_output_0.out_12_0_9", + "NhwcConv_64_out-/up_blocks.1/attentions.2/proj_out/Conv_output_0.out0_1_54", + "/up_blocks.1/attentions.2/Add.out_12_1_61", + "/up_blocks.1/upsamplers.0/Resize.out_4_1", + "NhwcConv_65_out-/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_55", + "/up_blocks.2/Concat.out13_7", + "GroupNorm_42.out4_42.out7_1_42", + "NhwcConv_68_out-/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_56", + "GroupNorm_42_sigmoid_out.4_42.out10_34", + "NhwcConv_66_out-/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_57", + "/up_blocks.2/resnets.0/Add.out_12_1_62", + "GroupNorm_43.out4_43.out7_1_43", + "GroupNorm_43_sigmoid_out.4_43.out10_35", + "NhwcConv_67_out-/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_58", + "/up_blocks.2/resnets.0/Add_1.out_12_1_63", + "/up_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_44", + "/up_blocks.2/attentions.0/Reshape_output_0.out0_0_10", + "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_30", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_72", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_73", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_74", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_20", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_42", + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_64", + "/up_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_31", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_75", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_21", + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_43", + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_65", + "/up_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_32", + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_10", + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_44", + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_11", + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_45", + "/up_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_10", + "NhwcConv_70_out-/up_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_59", + "/up_blocks.2/attentions.0/Add.out_12_1_66", + "/up_blocks.2/Concat_1.out13_8", + "GroupNorm_45.out4_45.out7_1_45", + "NhwcConv_73_out-/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_60", + "GroupNorm_45_sigmoid_out.4_45.out10_36", + "NhwcConv_71_out-/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_61", + "/up_blocks.2/resnets.1/Add.out_12_1_67", + "GroupNorm_46.out4_46.out7_1_46", + "GroupNorm_46_sigmoid_out.4_46.out10_37", + "NhwcConv_72_out-/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_62", + "/up_blocks.2/resnets.1/Add_1.out_12_1_68", + "/up_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_47", + "/up_blocks.2/attentions.1/Reshape_output_0.out0_0_11", + "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_33", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_76", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_77", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_78", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_22", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_46", + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_69", + "/up_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_34", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_79", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_23", + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_47", + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_70", + "/up_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_35", + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_11", + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_48", + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_12", + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_49", + "/up_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_11", + "NhwcConv_75_out-/up_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_63", + "/up_blocks.2/attentions.1/Add.out_12_1_71", + "/up_blocks.2/Concat_2.out13_9", + "GroupNorm_48.out4_48.out7_1_48", + "NhwcConv_78_out-/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_64", + "GroupNorm_48_sigmoid_out.4_48.out10_38", + "NhwcConv_76_out-/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_65", + "/up_blocks.2/resnets.2/Add.out_12_1_72", + "GroupNorm_49.out4_49.out7_1_49", + "GroupNorm_49_sigmoid_out.4_49.out10_39", + "NhwcConv_77_out-/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_66", + "/up_blocks.2/resnets.2/Add_1.out_12_1_73", + "/up_blocks.2/attentions.2/norm/Add_output_0_NHWC.out7_1_50", + "/up_blocks.2/attentions.2/Reshape_output_0.out0_0_12", + "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_36", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_80", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_81", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_82", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_24", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_50", + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_74", + "/up_blocks.2/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_37", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_83", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_25", + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_51", + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_75", + "/up_blocks.2/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_38", + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_12", + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_52", + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_13", + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_53", + "/up_blocks.2/attentions.2/Reshape_1_output_0.out_12_0_12", + "NhwcConv_80_out-/up_blocks.2/attentions.2/proj_out/Conv_output_0.out0_1_67", + "/up_blocks.2/attentions.2/Add.out_12_1_76", + "/up_blocks.2/upsamplers.0/Resize.out_4_2", + "NhwcConv_81_out-/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_68", + "/up_blocks.3/Concat.out13_10", + "GroupNorm_51.out4_51.out7_1_51", + "NhwcConv_84_out-/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_69", + "GroupNorm_51_sigmoid_out.4_51.out10_40", + "NhwcConv_82_out-/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_70", + "/up_blocks.3/resnets.0/Add.out_12_1_77", + "GroupNorm_52.out4_52.out7_1_52", + "GroupNorm_52_sigmoid_out.4_52.out10_41", + "NhwcConv_83_out-/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_71", + "/up_blocks.3/resnets.0/Add_1.out_12_1_78", + "/up_blocks.3/attentions.0/norm/Add_output_0_NHWC.out7_1_53", + "/up_blocks.3/attentions.0/Reshape_output_0.out0_0_13", + "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_39", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_84", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_85", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_86", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_26", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_54", + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_79", + "/up_blocks.3/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_40", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_87", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_27", + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_55", + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_80", + "/up_blocks.3/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_41", + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_13", + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_56", + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_14", + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_57", + "/up_blocks.3/attentions.0/Reshape_1_output_0.out_12_0_13", + "NhwcConv_86_out-/up_blocks.3/attentions.0/proj_out/Conv_output_0.out0_1_72", + "/up_blocks.3/attentions.0/Add.out_12_1_81", + "/up_blocks.3/Concat_1.out13_11", + "GroupNorm_54.out4_54.out7_1_54", + "NhwcConv_89_out-/up_blocks.3/resnets.1/conv_shortcut/Conv_output_0.out0_1_73", + "GroupNorm_54_sigmoid_out.4_54.out10_42", + "NhwcConv_87_out-/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_74", + "/up_blocks.3/resnets.1/Add.out_12_1_82", + "GroupNorm_55.out4_55.out7_1_55", + "GroupNorm_55_sigmoid_out.4_55.out10_43", + "NhwcConv_88_out-/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_75", + "/up_blocks.3/resnets.1/Add_1.out_12_1_83", + "/up_blocks.3/attentions.1/norm/Add_output_0_NHWC.out7_1_56", + "/up_blocks.3/attentions.1/Reshape_output_0.out0_0_14", + "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_42", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_88", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_89", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_90", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_58", + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_84", + "/up_blocks.3/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_43", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_91", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29", + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_59", + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_85", + "/up_blocks.3/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_44", + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14", + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_60", + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_15", + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_61", + "/up_blocks.3/attentions.1/Reshape_1_output_0.out_12_0_14", + "NhwcConv_91_out-/up_blocks.3/attentions.1/proj_out/Conv_output_0.out0_1_76", + "/up_blocks.3/attentions.1/Add.out_12_1_86", + "/up_blocks.3/Concat_2.out13_12", + "GroupNorm_57.out4_57.out7_1_57", + "NhwcConv_94_out-/up_blocks.3/resnets.2/conv_shortcut/Conv_output_0.out0_1_77", + "GroupNorm_57_sigmoid_out.4_57.out10_44", + "NhwcConv_92_out-/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_78", + "/up_blocks.3/resnets.2/Add.out_12_1_87", + "GroupNorm_58.out4_58.out7_1_58", + "GroupNorm_58_sigmoid_out.4_58.out10_45", + "NhwcConv_93_out-/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_79", + "/up_blocks.3/resnets.2/Add_1.out_12_1_88", + "/up_blocks.3/attentions.2/norm/Add_output_0_NHWC.out7_1_59", + "/up_blocks.3/attentions.2/Reshape_output_0.out0_0_15", + "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_45", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_92", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_93", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_94", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_30", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_62", + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_89", + "/up_blocks.3/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_46", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_95", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_31", + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_63", + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_90", + "/up_blocks.3/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_47", + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_15", + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_64", + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_16", + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_65", + "/up_blocks.3/attentions.2/Reshape_1_output_0.out_12_0_15", + "NhwcConv_96_out-/up_blocks.3/attentions.2/proj_out/Conv_output_0.out0_1_80", + "/up_blocks.3/attentions.2/Add.out_12_1_91", + "GroupNorm_60.out4_60.out7_1_60", + "GroupNorm_60_sigmoid_out.4_60.out10_46" + ] + }, + "const": { + "buffer_size": 1001425344, + "xrt_arg_id": 3, + "packed_tensors": [ + "onnx::MatMul_5660", + "onnx::MatMul_5661", + "onnx::MatMul_5690", + "onnx::MatMul_5691", + "onnx::MatMul_5720", + "onnx::MatMul_5721", + "onnx::MatMul_5750", + "onnx::MatMul_5751", + "onnx::MatMul_5780", + "onnx::MatMul_5781", + "onnx::MatMul_5810", + "onnx::MatMul_5811", + "onnx::MatMul_5848", + "onnx::MatMul_5849", + "onnx::MatMul_5895", + "onnx::MatMul_5896", + "onnx::MatMul_5925", + "onnx::MatMul_5926", + "onnx::MatMul_5955", + "onnx::MatMul_5956", + "onnx::MatMul_5986", + "onnx::MatMul_5987", + "onnx::MatMul_6016", + "onnx::MatMul_6017", + "onnx::MatMul_6046", + "onnx::MatMul_6047", + "onnx::MatMul_6077", + "onnx::MatMul_6078", + "onnx::MatMul_6107", + "onnx::MatMul_6108", + "onnx::MatMul_6137", + "onnx::MatMul_6138", + "NhwcConv_0_weight_NHWC", + "GroupNorm_0_wts_7_1_0", + "Sigmoid_0.weights10_0", + "NhwcConv_1_weight_NHWC", + "time_embedding.linear_1.weight_8_0", + "/time_embedding/act/Sigmoid.weights10_1", + "time_embedding.linear_2.weight_8_1", + "/down_blocks.0/resnets.0/act_1/Sigmoid.weights10_2", + "down_blocks.0.resnets.0.time_emb_proj.weight_8_2", + "down_blocks.0.resnets.1.time_emb_proj.weight_8_3", + "down_blocks.1.resnets.0.time_emb_proj.weight_8_4", + "down_blocks.1.resnets.1.time_emb_proj.weight_8_5", + "down_blocks.2.resnets.0.time_emb_proj.weight_8_6", + "down_blocks.2.resnets.1.time_emb_proj.weight_8_7", + "down_blocks.3.resnets.0.time_emb_proj.weight_8_8", + "down_blocks.3.resnets.1.time_emb_proj.weight_8_9", + "mid_block.resnets.0.time_emb_proj.weight_8_10", + "mid_block.resnets.1.time_emb_proj.weight_8_11", + "up_blocks.0.resnets.0.time_emb_proj.weight_8_12", + "up_blocks.0.resnets.1.time_emb_proj.weight_8_13", + "up_blocks.0.resnets.2.time_emb_proj.weight_8_14", + "up_blocks.1.resnets.0.time_emb_proj.weight_8_15", + "up_blocks.1.resnets.1.time_emb_proj.weight_8_16", + "up_blocks.1.resnets.2.time_emb_proj.weight_8_17", + "up_blocks.2.resnets.0.time_emb_proj.weight_8_18", + "up_blocks.2.resnets.1.time_emb_proj.weight_8_19", + "up_blocks.2.resnets.2.time_emb_proj.weight_8_20", + "up_blocks.3.resnets.0.time_emb_proj.weight_8_21", + "up_blocks.3.resnets.1.time_emb_proj.weight_8_22", + "up_blocks.3.resnets.2.time_emb_proj.weight_8_23", + "GroupNorm_1_wts_7_1_1", + "Sigmoid_1.weights10_3", + "NhwcConv_2_weight_NHWC", + "GroupNorm_2_wts_7_1_2", + "NhwcConv_3_weight_NHWC", + "down_blocks.0.attentions.0.transformer_blocks.0.norm1.weight", + "down_blocks.0.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5649", + "onnx::MatMul_5648", + "onnx::MatMul_5650", + "MultiHeadAttention_0_mask.8_0", + "onnx::MatMul_5658", + "down_blocks.0.attentions.0.transformer_blocks.0.norm2.weight", + "down_blocks.0.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5659", + "MultiHeadAttention_16_mask.8_1", + "onnx::MatMul_5669", + "down_blocks.0.attentions.0.transformer_blocks.0.norm3.weight", + "down_blocks.0.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5670_2_0_0", + "onnx::MatMul_5670_2_0_1", + "onnx::MatMul_5671", + "NhwcConv_4_weight_NHWC", + "GroupNorm_3_wts_7_1_3", + "Sigmoid_2.weights10_4", + "NhwcConv_5_weight_NHWC", + "GroupNorm_4_wts_7_1_4", + "Sigmoid_3.weights10_5", + "NhwcConv_6_weight_NHWC", + "GroupNorm_5_wts_7_1_5", + "NhwcConv_7_weight_NHWC", + "down_blocks.0.attentions.1.transformer_blocks.0.norm1.weight", + "down_blocks.0.attentions.1.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5679", + "onnx::MatMul_5678", + "onnx::MatMul_5680", + "MultiHeadAttention_1_mask.8_2", + "onnx::MatMul_5688", + "down_blocks.0.attentions.1.transformer_blocks.0.norm2.weight", + "down_blocks.0.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5689", + "MultiHeadAttention_17_mask.8_3", + "onnx::MatMul_5699", + "down_blocks.0.attentions.1.transformer_blocks.0.norm3.weight", + "down_blocks.0.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5700_2_1_0", + "onnx::MatMul_5700_2_1_1", + "onnx::MatMul_5701", + "NhwcConv_8_weight_NHWC", + "NhwcConv_9_weight_NHWC", + "GroupNorm_6_wts_7_1_6", + "NhwcConv_12_weight_NHWC", + "Sigmoid_4.weights10_6", + "NhwcConv_10_weight_NHWC", + "GroupNorm_7_wts_7_1_7", + "Sigmoid_5.weights10_7", + "NhwcConv_11_weight_NHWC", + "GroupNorm_8_wts_7_1_8", + "NhwcConv_13_weight_NHWC", + "down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5709", + "onnx::MatMul_5708", + "onnx::MatMul_5710", + "MultiHeadAttention_2_mask.8_4", + "onnx::MatMul_5718", + "down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5719", + "MultiHeadAttention_18_mask.8_5", + "onnx::MatMul_5729", + "down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5730_2_2_0", + "onnx::MatMul_5730_2_2_1", + "onnx::MatMul_5731", + "NhwcConv_14_weight_NHWC", + "GroupNorm_9_wts_7_1_9", + "Sigmoid_6.weights10_8", + "NhwcConv_15_weight_NHWC", + "GroupNorm_10_wts_7_1_10", + "Sigmoid_7.weights10_9", + "NhwcConv_16_weight_NHWC", + "GroupNorm_11_wts_7_1_11", + "NhwcConv_17_weight_NHWC", + "down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5739", + "onnx::MatMul_5738", + "onnx::MatMul_5740", + "MultiHeadAttention_3_mask.8_6", + "onnx::MatMul_5748", + "down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5749", + "MultiHeadAttention_19_mask.8_7", + "onnx::MatMul_5759", + "down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5760_2_3_0", + "onnx::MatMul_5760_2_3_1", + "onnx::MatMul_5761", + "NhwcConv_18_weight_NHWC", + "NhwcConv_19_weight_NHWC", + "NhwcConv_22_weight_NHWC", + "GroupNorm_12_wts_7_1_12", + "Sigmoid_8.weights10_10", + "NhwcConv_20_weight_NHWC", + "GroupNorm_13_wts_7_1_13", + "Sigmoid_9.weights10_11", + "NhwcConv_21_weight_NHWC", + "GroupNorm_14_wts_7_1_14", + "NhwcConv_23_weight_NHWC", + "down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", + "down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5769", + "onnx::MatMul_5768", + "onnx::MatMul_5770", + "MultiHeadAttention_4_mask.8_8", + "onnx::MatMul_5778", + "down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", + "down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5779", + "MultiHeadAttention_20_mask.8_9", + "onnx::MatMul_5789", + "down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", + "down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5790_2_4_0", + "onnx::MatMul_5790_2_4_1", + "onnx::MatMul_5791", + "NhwcConv_24_weight_NHWC", + "GroupNorm_15_wts_7_1_15", + "Sigmoid_10.weights10_12", + "NhwcConv_25_weight_NHWC", + "GroupNorm_16_wts_7_1_16", + "Sigmoid_11.weights10_13", + "NhwcConv_26_weight_NHWC", + "GroupNorm_17_wts_7_1_17", + "NhwcConv_27_weight_NHWC", + "down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", + "down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5799", + "onnx::MatMul_5798", + "onnx::MatMul_5800", + "MultiHeadAttention_5_mask.8_10", + "onnx::MatMul_5808", + "down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", + "down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5809", + "MultiHeadAttention_21_mask.8_11", + "onnx::MatMul_5819", + "down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", + "down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5820_2_5_0", + "onnx::MatMul_5820_2_5_1", + "onnx::MatMul_5821", + "NhwcConv_28_weight_NHWC", + "NhwcConv_29_weight_NHWC", + "GroupNorm_18_wts_7_1_18", + "Sigmoid_12.weights10_14", + "NhwcConv_30_weight_NHWC", + "GroupNorm_19_wts_7_1_19", + "Sigmoid_13.weights10_15", + "NhwcConv_31_weight_NHWC", + "GroupNorm_20_wts_7_1_20", + "Sigmoid_14.weights10_16", + "NhwcConv_32_weight_NHWC", + "GroupNorm_21_wts_7_1_21", + "Sigmoid_15.weights10_17", + "NhwcConv_33_weight_NHWC", + "GroupNorm_22_wts_7_1_22", + "Sigmoid_16.weights10_18", + "NhwcConv_34_weight_NHWC", + "GroupNorm_23_wts_7_1_23", + "Sigmoid_17.weights10_19", + "NhwcConv_35_weight_NHWC", + "GroupNorm_24_wts_7_1_24", + "NhwcConv_36_weight_NHWC", + "mid_block.attentions.0.transformer_blocks.0.norm1.weight", + "mid_block.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5837", + "onnx::MatMul_5836", + "onnx::MatMul_5838", + "MultiHeadAttention_6_mask.8_12", + "onnx::MatMul_5846", + "mid_block.attentions.0.transformer_blocks.0.norm2.weight", + "mid_block.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5847", + "MultiHeadAttention_22_mask.8_13", + "onnx::MatMul_5857", + "mid_block.attentions.0.transformer_blocks.0.norm3.weight", + "mid_block.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5858_2_6_0", + "onnx::MatMul_5858_2_6_1", + "onnx::MatMul_5859", + "NhwcConv_37_weight_NHWC", + "GroupNorm_25_wts_7_1_25", + "Sigmoid_18.weights10_20", + "NhwcConv_38_weight_NHWC", + "GroupNorm_26_wts_7_1_26", + "Sigmoid_19.weights10_21", + "NhwcConv_39_weight_NHWC", + "GroupNorm_27_wts_7_1_27", + "NhwcConv_42_weight_NHWC", + "Sigmoid_20.weights10_22", + "NhwcConv_40_weight_NHWC", + "GroupNorm_28_wts_7_1_28", + "Sigmoid_21.weights10_23", + "NhwcConv_41_weight_NHWC", + "GroupNorm_29_wts_7_1_29", + "NhwcConv_45_weight_NHWC", + "Sigmoid_22.weights10_24", + "NhwcConv_43_weight_NHWC", + "GroupNorm_30_wts_7_1_30", + "Sigmoid_23.weights10_25", + "NhwcConv_44_weight_NHWC", + "GroupNorm_31_wts_7_1_31", + "NhwcConv_48_weight_NHWC", + "Sigmoid_24.weights10_26", + "NhwcConv_46_weight_NHWC", + "GroupNorm_32_wts_7_1_32", + "Sigmoid_25.weights10_27", + "NhwcConv_47_weight_NHWC", + "/up_blocks.0/upsamplers.0/Resize.weights4_0", + "NhwcConv_49_weight_NHWC", + "GroupNorm_33_wts_7_1_33", + "NhwcConv_52_weight_NHWC", + "Sigmoid_26.weights10_28", + "NhwcConv_50_weight_NHWC", + "GroupNorm_34_wts_7_1_34", + "Sigmoid_27.weights10_29", + "NhwcConv_51_weight_NHWC", + "GroupNorm_35_wts_7_1_35", + "NhwcConv_53_weight_NHWC", + "up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight", + "up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5884", + "onnx::MatMul_5883", + "onnx::MatMul_5885", + "MultiHeadAttention_7_mask.8_14", + "onnx::MatMul_5893", + "up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight", + "up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5894", + "MultiHeadAttention_23_mask.8_15", + "onnx::MatMul_5904", + "up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight", + "up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5905_2_7_0", + "onnx::MatMul_5905_2_7_1", + "onnx::MatMul_5906", + "NhwcConv_54_weight_NHWC", + "GroupNorm_36_wts_7_1_36", + "NhwcConv_57_weight_NHWC", + "Sigmoid_28.weights10_30", + "NhwcConv_55_weight_NHWC", + "GroupNorm_37_wts_7_1_37", + "Sigmoid_29.weights10_31", + "NhwcConv_56_weight_NHWC", + "GroupNorm_38_wts_7_1_38", + "NhwcConv_58_weight_NHWC", + "up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight", + "up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5914", + "onnx::MatMul_5913", + "onnx::MatMul_5915", + "MultiHeadAttention_8_mask.8_16", + "onnx::MatMul_5923", + "up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight", + "up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5924", + "MultiHeadAttention_24_mask.8_17", + "onnx::MatMul_5934", + "up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight", + "up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5935_2_8_0", + "onnx::MatMul_5935_2_8_1", + "onnx::MatMul_5936", + "NhwcConv_59_weight_NHWC", + "GroupNorm_39_wts_7_1_39", + "NhwcConv_62_weight_NHWC", + "Sigmoid_30.weights10_32", + "NhwcConv_60_weight_NHWC", + "GroupNorm_40_wts_7_1_40", + "Sigmoid_31.weights10_33", + "NhwcConv_61_weight_NHWC", + "GroupNorm_41_wts_7_1_41", + "NhwcConv_63_weight_NHWC", + "up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight", + "up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5944", + "onnx::MatMul_5943", + "onnx::MatMul_5945", + "MultiHeadAttention_9_mask.8_18", + "onnx::MatMul_5953", + "up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight", + "up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5954", + "MultiHeadAttention_25_mask.8_19", + "onnx::MatMul_5964", + "up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight", + "up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5965_2_9_0", + "onnx::MatMul_5965_2_9_1", + "onnx::MatMul_5966", + "NhwcConv_64_weight_NHWC", + "/up_blocks.1/upsamplers.0/Resize.weights4_1", + "NhwcConv_65_weight_NHWC", + "GroupNorm_42_wts_7_1_42", + "NhwcConv_68_weight_NHWC", + "Sigmoid_32.weights10_34", + "NhwcConv_66_weight_NHWC", + "GroupNorm_43_wts_7_1_43", + "Sigmoid_33.weights10_35", + "NhwcConv_67_weight_NHWC", + "GroupNorm_44_wts_7_1_44", + "NhwcConv_69_weight_NHWC", + "up_blocks.2.attentions.0.transformer_blocks.0.norm1.weight", + "up_blocks.2.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_5975", + "onnx::MatMul_5974", + "onnx::MatMul_5976", + "MultiHeadAttention_10_mask.8_20", + "onnx::MatMul_5984", + "up_blocks.2.attentions.0.transformer_blocks.0.norm2.weight", + "up_blocks.2.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_5985", + "MultiHeadAttention_26_mask.8_21", + "onnx::MatMul_5995", + "up_blocks.2.attentions.0.transformer_blocks.0.norm3.weight", + "up_blocks.2.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_5996_2_10_0", + "onnx::MatMul_5996_2_10_1", + "onnx::MatMul_5997", + "NhwcConv_70_weight_NHWC", + "GroupNorm_45_wts_7_1_45", + "NhwcConv_73_weight_NHWC", + "Sigmoid_34.weights10_36", + "NhwcConv_71_weight_NHWC", + "GroupNorm_46_wts_7_1_46", + "Sigmoid_35.weights10_37", + "NhwcConv_72_weight_NHWC", + "GroupNorm_47_wts_7_1_47", + "NhwcConv_74_weight_NHWC", + "up_blocks.2.attentions.1.transformer_blocks.0.norm1.weight", + "up_blocks.2.attentions.1.transformer_blocks.0.norm1.bias", + "onnx::MatMul_6005", + "onnx::MatMul_6004", + "onnx::MatMul_6006", + "MultiHeadAttention_11_mask.8_22", + "onnx::MatMul_6014", + "up_blocks.2.attentions.1.transformer_blocks.0.norm2.weight", + "up_blocks.2.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_6015", + "MultiHeadAttention_27_mask.8_23", + "onnx::MatMul_6025", + "up_blocks.2.attentions.1.transformer_blocks.0.norm3.weight", + "up_blocks.2.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_6026_2_11_0", + "onnx::MatMul_6026_2_11_1", + "onnx::MatMul_6027", + "NhwcConv_75_weight_NHWC", + "GroupNorm_48_wts_7_1_48", + "NhwcConv_78_weight_NHWC", + "Sigmoid_36.weights10_38", + "NhwcConv_76_weight_NHWC", + "GroupNorm_49_wts_7_1_49", + "Sigmoid_37.weights10_39", + "NhwcConv_77_weight_NHWC", + "GroupNorm_50_wts_7_1_50", + "NhwcConv_79_weight_NHWC", + "up_blocks.2.attentions.2.transformer_blocks.0.norm1.weight", + "up_blocks.2.attentions.2.transformer_blocks.0.norm1.bias", + "onnx::MatMul_6035", + "onnx::MatMul_6034", + "onnx::MatMul_6036", + "MultiHeadAttention_12_mask.8_24", + "onnx::MatMul_6044", + "up_blocks.2.attentions.2.transformer_blocks.0.norm2.weight", + "up_blocks.2.attentions.2.transformer_blocks.0.norm2.bias", + "onnx::MatMul_6045", + "MultiHeadAttention_28_mask.8_25", + "onnx::MatMul_6055", + "up_blocks.2.attentions.2.transformer_blocks.0.norm3.weight", + "up_blocks.2.attentions.2.transformer_blocks.0.norm3.bias", + "onnx::MatMul_6056_2_12_0", + "onnx::MatMul_6056_2_12_1", + "onnx::MatMul_6057", + "NhwcConv_80_weight_NHWC", + "/up_blocks.2/upsamplers.0/Resize.weights4_2", + "NhwcConv_81_weight_NHWC", + "GroupNorm_51_wts_7_1_51", + "NhwcConv_84_weight_NHWC", + "Sigmoid_38.weights10_40", + "NhwcConv_82_weight_NHWC", + "GroupNorm_52_wts_7_1_52", + "Sigmoid_39.weights10_41", + "NhwcConv_83_weight_NHWC", + "GroupNorm_53_wts_7_1_53", + "NhwcConv_85_weight_NHWC", + "up_blocks.3.attentions.0.transformer_blocks.0.norm1.weight", + "up_blocks.3.attentions.0.transformer_blocks.0.norm1.bias", + "onnx::MatMul_6066", + "onnx::MatMul_6065", + "onnx::MatMul_6067", + "MultiHeadAttention_13_mask.8_26", + "onnx::MatMul_6075", + "up_blocks.3.attentions.0.transformer_blocks.0.norm2.weight", + "up_blocks.3.attentions.0.transformer_blocks.0.norm2.bias", + "onnx::MatMul_6076", + "MultiHeadAttention_29_mask.8_27", + "onnx::MatMul_6086", + "up_blocks.3.attentions.0.transformer_blocks.0.norm3.weight", + "up_blocks.3.attentions.0.transformer_blocks.0.norm3.bias", + "onnx::MatMul_6087_2_13_0", + "onnx::MatMul_6087_2_13_1", + "onnx::MatMul_6088", + "NhwcConv_86_weight_NHWC", + "GroupNorm_54_wts_7_1_54", + "NhwcConv_89_weight_NHWC", + "Sigmoid_40.weights10_42", + "NhwcConv_87_weight_NHWC", + "GroupNorm_55_wts_7_1_55", + "Sigmoid_41.weights10_43", + "NhwcConv_88_weight_NHWC", + "GroupNorm_56_wts_7_1_56", + "NhwcConv_90_weight_NHWC", + "up_blocks.3.attentions.1.transformer_blocks.0.norm1.weight", + "up_blocks.3.attentions.1.transformer_blocks.0.norm1.bias", + "onnx::MatMul_6096", + "onnx::MatMul_6095", + "onnx::MatMul_6097", + "MultiHeadAttention_14_mask.8_28", + "onnx::MatMul_6105", + "up_blocks.3.attentions.1.transformer_blocks.0.norm2.weight", + "up_blocks.3.attentions.1.transformer_blocks.0.norm2.bias", + "onnx::MatMul_6106", + "MultiHeadAttention_30_mask.8_29", + "onnx::MatMul_6116", + "up_blocks.3.attentions.1.transformer_blocks.0.norm3.weight", + "up_blocks.3.attentions.1.transformer_blocks.0.norm3.bias", + "onnx::MatMul_6117_2_14_0", + "onnx::MatMul_6117_2_14_1", + "onnx::MatMul_6118", + "NhwcConv_91_weight_NHWC", + "GroupNorm_57_wts_7_1_57", + "NhwcConv_94_weight_NHWC", + "Sigmoid_42.weights10_44", + "NhwcConv_92_weight_NHWC", + "GroupNorm_58_wts_7_1_58", + "Sigmoid_43.weights10_45", + "NhwcConv_93_weight_NHWC", + "GroupNorm_59_wts_7_1_59", + "NhwcConv_95_weight_NHWC", + "up_blocks.3.attentions.2.transformer_blocks.0.norm1.weight", + "up_blocks.3.attentions.2.transformer_blocks.0.norm1.bias", + "onnx::MatMul_6126", + "onnx::MatMul_6125", + "onnx::MatMul_6127", + "MultiHeadAttention_15_mask.8_30", + "onnx::MatMul_6135", + "up_blocks.3.attentions.2.transformer_blocks.0.norm2.weight", + "up_blocks.3.attentions.2.transformer_blocks.0.norm2.bias", + "onnx::MatMul_6136", + "MultiHeadAttention_31_mask.8_31", + "onnx::MatMul_6146", + "up_blocks.3.attentions.2.transformer_blocks.0.norm3.weight", + "up_blocks.3.attentions.2.transformer_blocks.0.norm3.bias", + "onnx::MatMul_6147_2_15_0", + "onnx::MatMul_6147_2_15_1", + "onnx::MatMul_6148", + "NhwcConv_96_weight_NHWC", + "GroupNorm_60_wts_7_1_60", + "Sigmoid_44.weights10_46", + "NhwcConv_97_weight_NHWC" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "encoder_hidden_states.out2_6_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 768 + ], + "size_in_bytes": 236544, + "offset": 7332864 + }, + "Transpose_44_out-sample.out0_1_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 4 + ], + "size_in_bytes": 65536, + "offset": 7569408 + }, + "/time_proj/Concat_1_output_0.out2_3_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 2, + 320 + ], + "size_in_bytes": 1280, + "offset": 7634944 + }, + "NhwcConv_97_out-out_sample.out0_1_81": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 4 + ], + "size_in_bytes": 65536, + "offset": 0 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 0 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 98560 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 197120 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 295680 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 394240 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 591360 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 788480 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 985600 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 1182720 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 1576960 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 1971200 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 2365440 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 2759680 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 3153920 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 3548160 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 3942400 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 4336640 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 4730880 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 5125120 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 1280 + ], + "size_in_bytes": 394240, + "offset": 5519360 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 5913600 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 6110720 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 6307840 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 6504960 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 6702080 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 640 + ], + "size_in_bytes": 197120, + "offset": 6899200 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 7096320 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 7194880 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 7293440 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 7392000 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_k/MatMul_output_0.out2_6_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 7490560 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_v/MatMul_output_0.out2_6_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 77, + 320 + ], + "size_in_bytes": 98560, + "offset": 7589120 + }, + "NhwcConv_0_out-/conv_in/Conv_output_0.out0_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 7687680 + }, + "GroupNorm_0.out4_0.out7_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 12930560 + }, + "GroupNorm_0_sigmoid_out.4_0.out10_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 18173440 + }, + "NhwcConv_1_out-/down_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 23416320 + }, + "/time_embedding/linear_1/Gemm_output_0.out2_3_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28659200 + }, + "/time_embedding/act/Sigmoid_output_0.out10_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28664320 + }, + "/time_embedding/linear_2/Gemm_output_0.out2_3_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28669440 + }, + "/down_blocks.0/resnets.0/act_1/Sigmoid_output_0.out10_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28674560 + }, + "/down_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "offset": 28679680 + }, + "/down_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "offset": 28680960 + }, + "/down_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "offset": 28682240 + }, + "/down_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "offset": 28684800 + }, + "/down_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28687360 + }, + "/down_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28692480 + }, + "/down_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28697600 + }, + "/down_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28702720 + }, + "/mid_block/resnets.0/Unsqueeze_1_output_0.out2_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28707840 + }, + "/mid_block/resnets.1/Unsqueeze_1_output_0.out2_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28712960 + }, + "/up_blocks.0/resnets.0/Unsqueeze_1_output_0.out2_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28718080 + }, + "/up_blocks.0/resnets.1/Unsqueeze_1_output_0.out2_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28723200 + }, + "/up_blocks.0/resnets.2/Unsqueeze_1_output_0.out2_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28728320 + }, + "/up_blocks.1/resnets.0/Unsqueeze_1_output_0.out2_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28733440 + }, + "/up_blocks.1/resnets.1/Unsqueeze_1_output_0.out2_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28738560 + }, + "/up_blocks.1/resnets.2/Unsqueeze_1_output_0.out2_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 1280 + ], + "size_in_bytes": 5120, + "offset": 28743680 + }, + "/up_blocks.2/resnets.0/Unsqueeze_1_output_0.out2_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "offset": 28748800 + }, + "/up_blocks.2/resnets.1/Unsqueeze_1_output_0.out2_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "offset": 28751360 + }, + "/up_blocks.2/resnets.2/Unsqueeze_1_output_0.out2_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 640 + ], + "size_in_bytes": 2560, + "offset": 28753920 + }, + "/up_blocks.3/resnets.0/Unsqueeze_1_output_0.out2_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "offset": 28756480 + }, + "/up_blocks.3/resnets.1/Unsqueeze_1_output_0.out2_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "offset": 28757760 + }, + "/up_blocks.3/resnets.2/Unsqueeze_1_output_0.out2_0_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1, + 1, + 320 + ], + "size_in_bytes": 1280, + "offset": 28759040 + }, + "/down_blocks.0/resnets.0/Add.out_12_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 28760320 + }, + "GroupNorm_1.out4_1.out7_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 34003200 + }, + "GroupNorm_1_sigmoid_out.4_1.out10_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 39246080 + }, + "NhwcConv_2_out-/down_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 44488960 + }, + "/down_blocks.0/resnets.0/Add_1.out_12_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 49731840 + }, + "/down_blocks.0/attentions.0/norm/Add_output_0_NHWC.out7_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 54974720 + }, + "/down_blocks.0/attentions.0/Reshape_output_0.out0_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 60217600 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 65460480 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 70703360 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 75946240 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 81189120 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 86432000 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 91674880 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 96917760 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 102160640 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 107403520 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 112646400 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 117889280 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 123132160 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 128375040 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 133617920 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 154589440 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 175560960 + }, + "/down_blocks.0/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 196532480 + }, + "/down_blocks.0/attentions.0/Reshape_1_output_0.out_12_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 201775360 + }, + "NhwcConv_4_out-/down_blocks.0/attentions.0/proj_out/Conv_output_0.out0_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 207018240 + }, + "/down_blocks.0/attentions.0/Add.out_12_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 212261120 + }, + "GroupNorm_3.out4_3.out7_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 217504000 + }, + "GroupNorm_3_sigmoid_out.4_3.out10_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 222746880 + }, + "NhwcConv_5_out-/down_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 227989760 + }, + "/down_blocks.0/resnets.1/Add.out_12_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 233232640 + }, + "GroupNorm_4.out4_4.out7_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 238475520 + }, + "GroupNorm_4_sigmoid_out.4_4.out10_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 243718400 + }, + "NhwcConv_6_out-/down_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 248961280 + }, + "/down_blocks.0/resnets.1/Add_1.out_12_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 254204160 + }, + "/down_blocks.0/attentions.1/norm/Add_output_0_NHWC.out7_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 259447040 + }, + "/down_blocks.0/attentions.1/Reshape_output_0.out0_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 264689920 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 269932800 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 275175680 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 280418560 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 285661440 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 290904320 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 296147200 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 301390080 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 306632960 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 311875840 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 317118720 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 322361600 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 327604480 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 332847360 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 338090240 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 359061760 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 380033280 + }, + "/down_blocks.0/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 401004800 + }, + "/down_blocks.0/attentions.1/Reshape_1_output_0.out_12_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 406247680 + }, + "NhwcConv_8_out-/down_blocks.0/attentions.1/proj_out/Conv_output_0.out0_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 411490560 + }, + "/down_blocks.0/attentions.1/Add.out_12_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 416733440 + }, + "NhwcConv_9_out-/down_blocks.0/downsamplers.0/conv/Conv_output_0.out0_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 320 + ], + "size_in_bytes": 1310720, + "offset": 421976320 + }, + "GroupNorm_6.out4_6.out7_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 320 + ], + "size_in_bytes": 1310720, + "offset": 423287040 + }, + "NhwcConv_12_out-/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 424597760 + }, + "GroupNorm_6_sigmoid_out.4_6.out10_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 320 + ], + "size_in_bytes": 1310720, + "offset": 427219200 + }, + "NhwcConv_10_out-/down_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 428529920 + }, + "/down_blocks.1/resnets.0/Add.out_12_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 431151360 + }, + "GroupNorm_7.out4_7.out7_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 433772800 + }, + "GroupNorm_7_sigmoid_out.4_7.out10_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 436394240 + }, + "NhwcConv_11_out-/down_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 439015680 + }, + "/down_blocks.1/resnets.0/Add_1.out_12_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 441637120 + }, + "/down_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 444258560 + }, + "/down_blocks.1/attentions.0/Reshape_output_0.out0_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 446880000 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 449501440 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 452122880 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 454744320 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 457365760 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 459987200 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 462608640 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 465230080 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 467851520 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 470472960 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 473094400 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 475715840 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 478337280 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 480958720 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 483580160 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 494065920 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 504551680 + }, + "/down_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 515037440 + }, + "/down_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 517658880 + }, + "NhwcConv_14_out-/down_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 520280320 + }, + "/down_blocks.1/attentions.0/Add.out_12_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 522901760 + }, + "GroupNorm_9.out4_9.out7_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 525523200 + }, + "GroupNorm_9_sigmoid_out.4_9.out10_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 528144640 + }, + "NhwcConv_15_out-/down_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 530766080 + }, + "/down_blocks.1/resnets.1/Add.out_12_1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 533387520 + }, + "GroupNorm_10.out4_10.out7_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 536008960 + }, + "GroupNorm_10_sigmoid_out.4_10.out10_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 538630400 + }, + "NhwcConv_16_out-/down_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 541251840 + }, + "/down_blocks.1/resnets.1/Add_1.out_12_1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 543873280 + }, + "/down_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 546494720 + }, + "/down_blocks.1/attentions.1/Reshape_output_0.out0_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 549116160 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 551737600 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 554359040 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 556980480 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 559601920 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 562223360 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 564844800 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 567466240 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 570087680 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 572709120 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 575330560 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 577952000 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 580573440 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 583194880 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 585816320 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 596302080 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 606787840 + }, + "/down_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 617273600 + }, + "/down_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 619895040 + }, + "NhwcConv_18_out-/down_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 622516480 + }, + "/down_blocks.1/attentions.1/Add.out_12_1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 625137920 + }, + "NhwcConv_19_out-/down_blocks.1/downsamplers.0/conv/Conv_output_0.out0_1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 640 + ], + "size_in_bytes": 655360, + "offset": 627759360 + }, + "NhwcConv_22_out-/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 628414720 + }, + "GroupNorm_12.out4_12.out7_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 640 + ], + "size_in_bytes": 655360, + "offset": 629725440 + }, + "GroupNorm_12_sigmoid_out.4_12.out10_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 640 + ], + "size_in_bytes": 655360, + "offset": 630380800 + }, + "NhwcConv_20_out-/down_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 631036160 + }, + "/down_blocks.2/resnets.0/Add.out_12_1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 632346880 + }, + "GroupNorm_13.out4_13.out7_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 633657600 + }, + "GroupNorm_13_sigmoid_out.4_13.out10_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 634968320 + }, + "NhwcConv_21_out-/down_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 636279040 + }, + "/down_blocks.2/resnets.0/Add_1.out_12_1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 637589760 + }, + "/down_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 638900480 + }, + "/down_blocks.2/attentions.0/Reshape_output_0.out0_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 640211200 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 641521920 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 642832640 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 644143360 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 645454080 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 646764800 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 648075520 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 649386240 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 650696960 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 652007680 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 653318400 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 654629120 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 655939840 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 657250560 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 658561280 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 663804160 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 669047040 + }, + "/down_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 674289920 + }, + "/down_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 675600640 + }, + "NhwcConv_24_out-/down_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 676911360 + }, + "/down_blocks.2/attentions.0/Add.out_12_1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 678222080 + }, + "GroupNorm_15.out4_15.out7_1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 679532800 + }, + "GroupNorm_15_sigmoid_out.4_15.out10_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 680843520 + }, + "NhwcConv_25_out-/down_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 682154240 + }, + "/down_blocks.2/resnets.1/Add.out_12_1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 683464960 + }, + "GroupNorm_16.out4_16.out7_1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 684775680 + }, + "GroupNorm_16_sigmoid_out.4_16.out10_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 686086400 + }, + "NhwcConv_26_out-/down_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 687397120 + }, + "/down_blocks.2/resnets.1/Add_1.out_12_1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 688707840 + }, + "/down_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 690018560 + }, + "/down_blocks.2/attentions.1/Reshape_output_0.out0_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 691329280 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 692640000 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 693950720 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 695261440 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 696572160 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 697882880 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 699193600 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 700504320 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 701815040 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 703125760 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 704436480 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 705747200 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 707057920 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 708368640 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 709679360 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 714922240 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 720165120 + }, + "/down_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 725408000 + }, + "/down_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 726718720 + }, + "NhwcConv_28_out-/down_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 728029440 + }, + "/down_blocks.2/attentions.1/Add.out_12_1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 729340160 + }, + "NhwcConv_29_out-/down_blocks.2/downsamplers.0/conv/Conv_output_0.out0_1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 730650880 + }, + "GroupNorm_18.out4_18.out7_1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 730978560 + }, + "GroupNorm_18_sigmoid_out.4_18.out10_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 731306240 + }, + "NhwcConv_30_out-/down_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 731633920 + }, + "/down_blocks.3/resnets.0/Add.out_12_1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 731961600 + }, + "GroupNorm_19.out4_19.out7_1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 732289280 + }, + "GroupNorm_19_sigmoid_out.4_19.out10_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 732616960 + }, + "NhwcConv_31_out-/down_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 732944640 + }, + "/down_blocks.3/resnets.0/Add_1.out_12_1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 733272320 + }, + "GroupNorm_20.out4_20.out7_1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 733600000 + }, + "GroupNorm_20_sigmoid_out.4_20.out10_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 733927680 + }, + "NhwcConv_32_out-/down_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 734255360 + }, + "/down_blocks.3/resnets.1/Add.out_12_1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 734583040 + }, + "GroupNorm_21.out4_21.out7_1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 734910720 + }, + "GroupNorm_21_sigmoid_out.4_21.out10_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 735238400 + }, + "NhwcConv_33_out-/down_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 735566080 + }, + "/down_blocks.3/resnets.1/Add_1.out_12_1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 735893760 + }, + "GroupNorm_22.out4_22.out7_1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 736221440 + }, + "GroupNorm_22_sigmoid_out.4_22.out10_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 736549120 + }, + "NhwcConv_34_out-/mid_block/resnets.0/conv1/Conv_output_0.out0_1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 736876800 + }, + "/mid_block/resnets.0/Add.out_12_1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 737204480 + }, + "GroupNorm_23.out4_23.out7_1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 737532160 + }, + "GroupNorm_23_sigmoid_out.4_23.out10_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 737859840 + }, + "NhwcConv_35_out-/mid_block/resnets.0/conv2/Conv_output_0.out0_1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 738187520 + }, + "/mid_block/resnets.0/Add_1.out_12_1_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 738515200 + }, + "/mid_block/attentions.0/norm/Add_output_0_NHWC.out7_1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 738842880 + }, + "/mid_block/attentions.0/Reshape_output_0.out0_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 739170560 + }, + "/mid_block/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 739498240 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 739825920 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 740153600 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 740481280 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 740808960 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 741136640 + }, + "/mid_block/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 741464320 + }, + "/mid_block/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 741792000 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 742119680 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 742447360 + }, + "/mid_block/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 742775040 + }, + "/mid_block/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 743102720 + }, + "/mid_block/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 743430400 + }, + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 5120 + ], + "size_in_bytes": 1310720, + "offset": 743758080 + }, + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 5120 + ], + "size_in_bytes": 1310720, + "offset": 745068800 + }, + "/mid_block/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 5120 + ], + "size_in_bytes": 1310720, + "offset": 746379520 + }, + "/mid_block/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 1280 + ], + "size_in_bytes": 327680, + "offset": 747690240 + }, + "/mid_block/attentions.0/Reshape_1_output_0.out_12_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 748017920 + }, + "NhwcConv_37_out-/mid_block/attentions.0/proj_out/Conv_output_0.out0_1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 748345600 + }, + "/mid_block/attentions.0/Add.out_12_1_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 748673280 + }, + "GroupNorm_25.out4_25.out7_1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 749000960 + }, + "GroupNorm_25_sigmoid_out.4_25.out10_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 749328640 + }, + "NhwcConv_38_out-/mid_block/resnets.1/conv1/Conv_output_0.out0_1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 749656320 + }, + "/mid_block/resnets.1/Add.out_12_1_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 749984000 + }, + "GroupNorm_26.out4_26.out7_1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 750311680 + }, + "GroupNorm_26_sigmoid_out.4_26.out10_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 750639360 + }, + "NhwcConv_39_out-/mid_block/resnets.1/conv2/Conv_output_0.out0_1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 750967040 + }, + "/mid_block/resnets.1/Add_1.out_12_1_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 751294720 + }, + "/up_blocks.0/Concat.out13_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 751622400 + }, + "GroupNorm_27.out4_27.out7_1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 752277760 + }, + "NhwcConv_42_out-/up_blocks.0/resnets.0/conv_shortcut/Conv_output_0.out0_1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 752933120 + }, + "GroupNorm_27_sigmoid_out.4_27.out10_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 753260800 + }, + "NhwcConv_40_out-/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 753916160 + }, + "/up_blocks.0/resnets.0/Add.out_12_1_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 754243840 + }, + "GroupNorm_28.out4_28.out7_1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 754571520 + }, + "GroupNorm_28_sigmoid_out.4_28.out10_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 754899200 + }, + "NhwcConv_41_out-/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 755226880 + }, + "/up_blocks.0/resnets.0/Add_1.out_12_1_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 755554560 + }, + "/up_blocks.0/Concat_1.out13_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 755882240 + }, + "GroupNorm_29.out4_29.out7_1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 756537600 + }, + "NhwcConv_45_out-/up_blocks.0/resnets.1/conv_shortcut/Conv_output_0.out0_1_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 757192960 + }, + "GroupNorm_29_sigmoid_out.4_29.out10_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 757520640 + }, + "NhwcConv_43_out-/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 758176000 + }, + "/up_blocks.0/resnets.1/Add.out_12_1_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 758503680 + }, + "GroupNorm_30.out4_30.out7_1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 758831360 + }, + "GroupNorm_30_sigmoid_out.4_30.out10_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 759159040 + }, + "NhwcConv_44_out-/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 759486720 + }, + "/up_blocks.0/resnets.1/Add_1.out_12_1_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 759814400 + }, + "/up_blocks.0/Concat_2.out13_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 760142080 + }, + "GroupNorm_31.out4_31.out7_1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 760797440 + }, + "NhwcConv_48_out-/up_blocks.0/resnets.2/conv_shortcut/Conv_output_0.out0_1_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 761452800 + }, + "GroupNorm_31_sigmoid_out.4_31.out10_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 2560 + ], + "size_in_bytes": 655360, + "offset": 761780480 + }, + "NhwcConv_46_out-/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 762435840 + }, + "/up_blocks.0/resnets.2/Add.out_12_1_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 762763520 + }, + "GroupNorm_32.out4_32.out7_1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 763091200 + }, + "GroupNorm_32_sigmoid_out.4_32.out10_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 763418880 + }, + "NhwcConv_47_out-/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 763746560 + }, + "/up_blocks.0/resnets.2/Add_1.out_12_1_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 8, + 8, + 1280 + ], + "size_in_bytes": 327680, + "offset": 764074240 + }, + "/up_blocks.0/upsamplers.0/Resize.out_4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 764401920 + }, + "NhwcConv_49_out-/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 765712640 + }, + "/up_blocks.1/Concat.out13_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 2560 + ], + "size_in_bytes": 2621440, + "offset": 767023360 + }, + "GroupNorm_33.out4_33.out7_1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 2560 + ], + "size_in_bytes": 2621440, + "offset": 769644800 + }, + "NhwcConv_52_out-/up_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out0_1_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 772266240 + }, + "GroupNorm_33_sigmoid_out.4_33.out10_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 2560 + ], + "size_in_bytes": 2621440, + "offset": 773576960 + }, + "NhwcConv_50_out-/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 776198400 + }, + "/up_blocks.1/resnets.0/Add.out_12_1_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 777509120 + }, + "GroupNorm_34.out4_34.out7_1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 778819840 + }, + "GroupNorm_34_sigmoid_out.4_34.out10_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 780130560 + }, + "NhwcConv_51_out-/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 781441280 + }, + "/up_blocks.1/resnets.0/Add_1.out_12_1_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 782752000 + }, + "/up_blocks.1/attentions.0/norm/Add_output_0_NHWC.out7_1_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 784062720 + }, + "/up_blocks.1/attentions.0/Reshape_output_0.out0_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 785373440 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 786684160 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 787994880 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 789305600 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 790616320 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 791927040 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 793237760 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 794548480 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 795859200 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 797169920 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 798480640 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 799791360 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 801102080 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 802412800 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 803723520 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 808966400 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 814209280 + }, + "/up_blocks.1/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 819452160 + }, + "/up_blocks.1/attentions.0/Reshape_1_output_0.out_12_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 820762880 + }, + "NhwcConv_54_out-/up_blocks.1/attentions.0/proj_out/Conv_output_0.out0_1_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 822073600 + }, + "/up_blocks.1/attentions.0/Add.out_12_1_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 823384320 + }, + "/up_blocks.1/Concat_1.out13_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 2560 + ], + "size_in_bytes": 2621440, + "offset": 824695040 + }, + "GroupNorm_36.out4_36.out7_1_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 2560 + ], + "size_in_bytes": 2621440, + "offset": 827316480 + }, + "NhwcConv_57_out-/up_blocks.1/resnets.1/conv_shortcut/Conv_output_0.out0_1_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 829937920 + }, + "GroupNorm_36_sigmoid_out.4_36.out10_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 2560 + ], + "size_in_bytes": 2621440, + "offset": 831248640 + }, + "NhwcConv_55_out-/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 833870080 + }, + "/up_blocks.1/resnets.1/Add.out_12_1_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 835180800 + }, + "GroupNorm_37.out4_37.out7_1_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 836491520 + }, + "GroupNorm_37_sigmoid_out.4_37.out10_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 837802240 + }, + "NhwcConv_56_out-/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 839112960 + }, + "/up_blocks.1/resnets.1/Add_1.out_12_1_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 840423680 + }, + "/up_blocks.1/attentions.1/norm/Add_output_0_NHWC.out7_1_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 841734400 + }, + "/up_blocks.1/attentions.1/Reshape_output_0.out0_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 843045120 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 844355840 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 845666560 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 846977280 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 848288000 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 849598720 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 850909440 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 852220160 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 853530880 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 854841600 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 856152320 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 857463040 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 858773760 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 860084480 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 861395200 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 866638080 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 871880960 + }, + "/up_blocks.1/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 877123840 + }, + "/up_blocks.1/attentions.1/Reshape_1_output_0.out_12_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 878434560 + }, + "NhwcConv_59_out-/up_blocks.1/attentions.1/proj_out/Conv_output_0.out0_1_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 879745280 + }, + "/up_blocks.1/attentions.1/Add.out_12_1_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 881056000 + }, + "/up_blocks.1/Concat_2.out13_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1920 + ], + "size_in_bytes": 1966080, + "offset": 882366720 + }, + "GroupNorm_39.out4_39.out7_1_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1920 + ], + "size_in_bytes": 1966080, + "offset": 884332800 + }, + "NhwcConv_62_out-/up_blocks.1/resnets.2/conv_shortcut/Conv_output_0.out0_1_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 886298880 + }, + "GroupNorm_39_sigmoid_out.4_39.out10_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1920 + ], + "size_in_bytes": 1966080, + "offset": 887609600 + }, + "NhwcConv_60_out-/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 889575680 + }, + "/up_blocks.1/resnets.2/Add.out_12_1_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 890886400 + }, + "GroupNorm_40.out4_40.out7_1_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 892197120 + }, + "GroupNorm_40_sigmoid_out.4_40.out10_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 893507840 + }, + "NhwcConv_61_out-/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 894818560 + }, + "/up_blocks.1/resnets.2/Add_1.out_12_1_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 896129280 + }, + "/up_blocks.1/attentions.2/norm/Add_output_0_NHWC.out7_1_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 897440000 + }, + "/up_blocks.1/attentions.2/Reshape_output_0.out0_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 898750720 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 900061440 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 901372160 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 902682880 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 903993600 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 905304320 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 906615040 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 907925760 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 909236480 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 910547200 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 911857920 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 913168640 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 914479360 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 915790080 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 917100800 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 922343680 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 5120 + ], + "size_in_bytes": 5242880, + "offset": 927586560 + }, + "/up_blocks.1/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 256, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 932829440 + }, + "/up_blocks.1/attentions.2/Reshape_1_output_0.out_12_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 934140160 + }, + "NhwcConv_64_out-/up_blocks.1/attentions.2/proj_out/Conv_output_0.out0_1_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 935450880 + }, + "/up_blocks.1/attentions.2/Add.out_12_1_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 16, + 16, + 1280 + ], + "size_in_bytes": 1310720, + "offset": 936761600 + }, + "/up_blocks.1/upsamplers.0/Resize.out_4_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 5242880, + "offset": 938072320 + }, + "NhwcConv_65_out-/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 5242880, + "offset": 943315200 + }, + "/up_blocks.2/Concat.out13_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1920 + ], + "size_in_bytes": 7864320, + "offset": 948558080 + }, + "GroupNorm_42.out4_42.out7_1_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1920 + ], + "size_in_bytes": 7864320, + "offset": 956422400 + }, + "NhwcConv_68_out-/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 964286720 + }, + "GroupNorm_42_sigmoid_out.4_42.out10_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1920 + ], + "size_in_bytes": 7864320, + "offset": 966908160 + }, + "NhwcConv_66_out-/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 974772480 + }, + "/up_blocks.2/resnets.0/Add.out_12_1_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 977393920 + }, + "GroupNorm_43.out4_43.out7_1_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 980015360 + }, + "GroupNorm_43_sigmoid_out.4_43.out10_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 982636800 + }, + "NhwcConv_67_out-/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 985258240 + }, + "/up_blocks.2/resnets.0/Add_1.out_12_1_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 987879680 + }, + "/up_blocks.2/attentions.0/norm/Add_output_0_NHWC.out7_1_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 990501120 + }, + "/up_blocks.2/attentions.0/Reshape_output_0.out0_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 993122560 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 995744000 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 998365440 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1000986880 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1003608320 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1006229760 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1008851200 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1011472640 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1014094080 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1016715520 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1019336960 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1021958400 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1024579840 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1027201280 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1029822720 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1040308480 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1050794240 + }, + "/up_blocks.2/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1061280000 + }, + "/up_blocks.2/attentions.0/Reshape_1_output_0.out_12_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1063901440 + }, + "NhwcConv_70_out-/up_blocks.2/attentions.0/proj_out/Conv_output_0.out0_1_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1066522880 + }, + "/up_blocks.2/attentions.0/Add.out_12_1_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1069144320 + }, + "/up_blocks.2/Concat_1.out13_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 5242880, + "offset": 1071765760 + }, + "GroupNorm_45.out4_45.out7_1_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 5242880, + "offset": 1077008640 + }, + "NhwcConv_73_out-/up_blocks.2/resnets.1/conv_shortcut/Conv_output_0.out0_1_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1082251520 + }, + "GroupNorm_45_sigmoid_out.4_45.out10_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 1280 + ], + "size_in_bytes": 5242880, + "offset": 1084872960 + }, + "NhwcConv_71_out-/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1090115840 + }, + "/up_blocks.2/resnets.1/Add.out_12_1_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1092737280 + }, + "GroupNorm_46.out4_46.out7_1_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1095358720 + }, + "GroupNorm_46_sigmoid_out.4_46.out10_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1097980160 + }, + "NhwcConv_72_out-/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1100601600 + }, + "/up_blocks.2/resnets.1/Add_1.out_12_1_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1103223040 + }, + "/up_blocks.2/attentions.1/norm/Add_output_0_NHWC.out7_1_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1105844480 + }, + "/up_blocks.2/attentions.1/Reshape_output_0.out0_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1108465920 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1111087360 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1113708800 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1116330240 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1118951680 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1121573120 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1124194560 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1126816000 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1129437440 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1132058880 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1134680320 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1137301760 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1139923200 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_35": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1142544640 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1145166080 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1155651840 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1166137600 + }, + "/up_blocks.2/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1176623360 + }, + "/up_blocks.2/attentions.1/Reshape_1_output_0.out_12_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1179244800 + }, + "NhwcConv_75_out-/up_blocks.2/attentions.1/proj_out/Conv_output_0.out0_1_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1181866240 + }, + "/up_blocks.2/attentions.1/Add.out_12_1_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1184487680 + }, + "/up_blocks.2/Concat_2.out13_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 960 + ], + "size_in_bytes": 3932160, + "offset": 1187109120 + }, + "GroupNorm_48.out4_48.out7_1_48": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 960 + ], + "size_in_bytes": 3932160, + "offset": 1191041280 + }, + "NhwcConv_78_out-/up_blocks.2/resnets.2/conv_shortcut/Conv_output_0.out0_1_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1194973440 + }, + "GroupNorm_48_sigmoid_out.4_48.out10_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 960 + ], + "size_in_bytes": 3932160, + "offset": 1197594880 + }, + "NhwcConv_76_out-/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1201527040 + }, + "/up_blocks.2/resnets.2/Add.out_12_1_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1204148480 + }, + "GroupNorm_49.out4_49.out7_1_49": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1206769920 + }, + "GroupNorm_49_sigmoid_out.4_49.out10_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1209391360 + }, + "NhwcConv_77_out-/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_66": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1212012800 + }, + "/up_blocks.2/resnets.2/Add_1.out_12_1_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1214634240 + }, + "/up_blocks.2/attentions.2/norm/Add_output_0_NHWC.out7_1_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1217255680 + }, + "/up_blocks.2/attentions.2/Reshape_output_0.out0_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1219877120 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_36": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1222498560 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1225120000 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1227741440 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1230362880 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1232984320 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_50": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1235605760 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1238227200 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_37": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1240848640 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1243470080 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1246091520 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1248712960 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1251334400 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_38": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1253955840 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1256577280 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1267063040 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 2560 + ], + "size_in_bytes": 10485760, + "offset": 1277548800 + }, + "/up_blocks.2/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 1024, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1288034560 + }, + "/up_blocks.2/attentions.2/Reshape_1_output_0.out_12_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1290656000 + }, + "NhwcConv_80_out-/up_blocks.2/attentions.2/proj_out/Conv_output_0.out0_1_67": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1293277440 + }, + "/up_blocks.2/attentions.2/Add.out_12_1_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 32, + 32, + 640 + ], + "size_in_bytes": 2621440, + "offset": 1295898880 + }, + "/up_blocks.2/upsamplers.0/Resize.out_4_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1298520320 + }, + "NhwcConv_81_out-/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_68": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1309006080 + }, + "/up_blocks.3/Concat.out13_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 960 + ], + "size_in_bytes": 15728640, + "offset": 1319491840 + }, + "GroupNorm_51.out4_51.out7_1_51": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 960 + ], + "size_in_bytes": 15728640, + "offset": 1335220480 + }, + "NhwcConv_84_out-/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_69": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1350949120 + }, + "GroupNorm_51_sigmoid_out.4_51.out10_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 960 + ], + "size_in_bytes": 15728640, + "offset": 1356192000 + }, + "NhwcConv_82_out-/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_70": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1371920640 + }, + "/up_blocks.3/resnets.0/Add.out_12_1_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1377163520 + }, + "GroupNorm_52.out4_52.out7_1_52": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1382406400 + }, + "GroupNorm_52_sigmoid_out.4_52.out10_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1387649280 + }, + "NhwcConv_83_out-/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_71": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1392892160 + }, + "/up_blocks.3/resnets.0/Add_1.out_12_1_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1398135040 + }, + "/up_blocks.3/attentions.0/norm/Add_output_0_NHWC.out7_1_53": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1403377920 + }, + "/up_blocks.3/attentions.0/Reshape_output_0.out0_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1408620800 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_39": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1413863680 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1419106560 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1424349440 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1429592320 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/Reshape_3_output_0.out8_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1434835200 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1440078080 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_output_0.out_12_1_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1445320960 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_40": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1450563840 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1455806720 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/Reshape_3_output_0.out8_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1461049600 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1466292480 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/Add_1_output_0.out_12_1_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1471535360 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_41": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1476778240 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1482021120 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1502992640 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1523964160 + }, + "/up_blocks.3/attentions.0/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1544935680 + }, + "/up_blocks.3/attentions.0/Reshape_1_output_0.out_12_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1550178560 + }, + "NhwcConv_86_out-/up_blocks.3/attentions.0/proj_out/Conv_output_0.out0_1_72": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1555421440 + }, + "/up_blocks.3/attentions.0/Add.out_12_1_81": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1560664320 + }, + "/up_blocks.3/Concat_1.out13_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1565907200 + }, + "GroupNorm_54.out4_54.out7_1_54": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1576392960 + }, + "NhwcConv_89_out-/up_blocks.3/resnets.1/conv_shortcut/Conv_output_0.out0_1_73": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1586878720 + }, + "GroupNorm_54_sigmoid_out.4_54.out10_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1592121600 + }, + "NhwcConv_87_out-/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_74": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1602607360 + }, + "/up_blocks.3/resnets.1/Add.out_12_1_82": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1607850240 + }, + "GroupNorm_55.out4_55.out7_1_55": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1613093120 + }, + "GroupNorm_55_sigmoid_out.4_55.out10_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1618336000 + }, + "NhwcConv_88_out-/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_75": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1623578880 + }, + "/up_blocks.3/resnets.1/Add_1.out_12_1_83": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1628821760 + }, + "/up_blocks.3/attentions.1/norm/Add_output_0_NHWC.out7_1_56": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1634064640 + }, + "/up_blocks.3/attentions.1/Reshape_output_0.out0_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1639307520 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_42": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1644550400 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1649793280 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1655036160 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1660279040 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/Reshape_3_output_0.out8_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1665521920 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1670764800 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_output_0.out_12_1_84": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1676007680 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_43": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1681250560 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1686493440 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/Reshape_3_output_0.out8_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1691736320 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1696979200 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/Add_1_output_0.out_12_1_85": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1702222080 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1707464960 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1712707840 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1733679360 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1754650880 + }, + "/up_blocks.3/attentions.1/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_61": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1775622400 + }, + "/up_blocks.3/attentions.1/Reshape_1_output_0.out_12_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1780865280 + }, + "NhwcConv_91_out-/up_blocks.3/attentions.1/proj_out/Conv_output_0.out0_1_76": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1786108160 + }, + "/up_blocks.3/attentions.1/Add.out_12_1_86": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1791351040 + }, + "/up_blocks.3/Concat_2.out13_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1796593920 + }, + "GroupNorm_57.out4_57.out7_1_57": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1807079680 + }, + "NhwcConv_94_out-/up_blocks.3/resnets.2/conv_shortcut/Conv_output_0.out0_1_77": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1817565440 + }, + "GroupNorm_57_sigmoid_out.4_57.out10_44": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 640 + ], + "size_in_bytes": 10485760, + "offset": 1822808320 + }, + "NhwcConv_92_out-/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_78": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1833294080 + }, + "/up_blocks.3/resnets.2/Add.out_12_1_87": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1838536960 + }, + "GroupNorm_58.out4_58.out7_1_58": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1843779840 + }, + "GroupNorm_58_sigmoid_out.4_58.out10_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1849022720 + }, + "NhwcConv_93_out-/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_79": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1854265600 + }, + "/up_blocks.3/resnets.2/Add_1.out_12_1_88": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1859508480 + }, + "/up_blocks.3/attentions.2/norm/Add_output_0_NHWC.out7_1_59": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1864751360 + }, + "/up_blocks.3/attentions.2/Reshape_output_0.out0_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1869994240 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/norm1/LayerNormalization_output_0.out9_45": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1875237120 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_k/MatMul_output_0.out2_6_92": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1880480000 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_q/MatMul_output_0.out2_6_93": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1885722880 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_v/MatMul_output_0.out2_6_94": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1890965760 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/Reshape_3_output_0.out8_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1896208640 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn1/to_out.0/Add_output_0.out2_3_62": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1901451520 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_output_0.out_12_1_89": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1906694400 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/norm2/LayerNormalization_output_0.out9_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1911937280 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_q/MatMul_output_0.out2_6_95": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1917180160 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/Reshape_3_output_0.out8_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1922423040 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/attn2/to_out.0/Add_output_0.out2_3_63": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1927665920 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/Add_1_output_0.out_12_1_90": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1932908800 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/norm3/LayerNormalization_output_0.out9_47": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 1938151680 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_3_output_0.out2_2_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1943394560 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Slice_output_0.out2_3_64": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1964366080 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.0/Mul_4_output_0.out14_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 1280 + ], + "size_in_bytes": 20971520, + "offset": 1985337600 + }, + "/up_blocks.3/attentions.2/transformer_blocks.0/ff/net.2/Add_output_0.out2_3_65": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 4096, + 320 + ], + "size_in_bytes": 5242880, + "offset": 2006309120 + }, + "/up_blocks.3/attentions.2/Reshape_1_output_0.out_12_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 2011552000 + }, + "NhwcConv_96_out-/up_blocks.3/attentions.2/proj_out/Conv_output_0.out0_1_80": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 2016794880 + }, + "/up_blocks.3/attentions.2/Add.out_12_1_91": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 2022037760 + }, + "GroupNorm_60.out4_60.out7_1_60": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 2027280640 + }, + "GroupNorm_60_sigmoid_out.4_60.out10_46": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 2, + 64, + 64, + 320 + ], + "size_in_bytes": 5242880, + "offset": 2032523520 + }, + "onnx::MatMul_5660": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 0, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_0.const", + "file_size": 281600 + }, + "onnx::MatMul_5661": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 281600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_1.const", + "file_size": 281600 + }, + "onnx::MatMul_5690": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 563200, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_2.const", + "file_size": 281600 + }, + "onnx::MatMul_5691": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 844800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_3.const", + "file_size": 281600 + }, + "onnx::MatMul_5720": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 1126400, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_4.const", + "file_size": 563200 + }, + "onnx::MatMul_5721": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 1689600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_5.const", + "file_size": 563200 + }, + "onnx::MatMul_5750": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 2252800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_6.const", + "file_size": 563200 + }, + "onnx::MatMul_5751": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 2816000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_7.const", + "file_size": 563200 + }, + "onnx::MatMul_5780": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 3379200, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_8.const", + "file_size": 1121280 + }, + "onnx::MatMul_5781": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 4500480, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_9.const", + "file_size": 1121280 + }, + "onnx::MatMul_5810": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 5621760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_10.const", + "file_size": 1121280 + }, + "onnx::MatMul_5811": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 6743040, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_11.const", + "file_size": 1121280 + }, + "onnx::MatMul_5848": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 7864320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_12.const", + "file_size": 1121280 + }, + "onnx::MatMul_5849": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 8985600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_13.const", + "file_size": 1121280 + }, + "onnx::MatMul_5895": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 10106880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_14.const", + "file_size": 1121280 + }, + "onnx::MatMul_5896": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 11228160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_15.const", + "file_size": 1121280 + }, + "onnx::MatMul_5925": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 12349440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_16.const", + "file_size": 1121280 + }, + "onnx::MatMul_5926": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 13470720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_17.const", + "file_size": 1121280 + }, + "onnx::MatMul_5955": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 14592000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_18.const", + "file_size": 1121280 + }, + "onnx::MatMul_5956": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1121280 + ], + "size_in_bytes": 1121280, + "offset": 15713280, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_19.const", + "file_size": 1121280 + }, + "onnx::MatMul_5986": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 16834560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_20.const", + "file_size": 563200 + }, + "onnx::MatMul_5987": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 17397760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_21.const", + "file_size": 563200 + }, + "onnx::MatMul_6016": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 17960960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_22.const", + "file_size": 563200 + }, + "onnx::MatMul_6017": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 18524160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_23.const", + "file_size": 563200 + }, + "onnx::MatMul_6046": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 19087360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_24.const", + "file_size": 563200 + }, + "onnx::MatMul_6047": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 563200 + ], + "size_in_bytes": 563200, + "offset": 19650560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_25.const", + "file_size": 563200 + }, + "onnx::MatMul_6077": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 20213760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_26.const", + "file_size": 281600 + }, + "onnx::MatMul_6078": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 20495360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_27.const", + "file_size": 281600 + }, + "onnx::MatMul_6107": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 20776960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_28.const", + "file_size": 281600 + }, + "onnx::MatMul_6108": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 21058560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_29.const", + "file_size": 281600 + }, + "onnx::MatMul_6137": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 21340160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_30.const", + "file_size": 281600 + }, + "onnx::MatMul_6138": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 281600 + ], + "size_in_bytes": 281600, + "offset": 21621760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_31.const", + "file_size": 281600 + }, + "NhwcConv_0_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 31040 + ], + "size_in_bytes": 31040, + "offset": 21903360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_32.const", + "file_size": 31040 + }, + "GroupNorm_0_wts_7_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 21934400, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_33.const", + "file_size": 1280 + }, + "Sigmoid_0.weights10_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 21935680, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_34.const", + "file_size": 256 + }, + "NhwcConv_1_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 21935936, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_35.const", + "file_size": 1088000 + }, + "time_embedding.linear_1.weight_8_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 23023936, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_36.const", + "file_size": 471040 + }, + "/time_embedding/act/Sigmoid.weights10_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 23494976, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_37.const", + "file_size": 256 + }, + "time_embedding.linear_2.weight_8_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 23495232, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_38.const", + "file_size": 1868800 + }, + "/down_blocks.0/resnets.0/act_1/Sigmoid.weights10_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 25364032, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_39.const", + "file_size": 256 + }, + "down_blocks.0.resnets.0.time_emb_proj.weight_8_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "offset": 25364288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_40.const", + "file_size": 467200 + }, + "down_blocks.0.resnets.1.time_emb_proj.weight_8_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "offset": 25831488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_41.const", + "file_size": 467200 + }, + "down_blocks.1.resnets.0.time_emb_proj.weight_8_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "offset": 26298688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_42.const", + "file_size": 934400 + }, + "down_blocks.1.resnets.1.time_emb_proj.weight_8_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "offset": 27233088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_43.const", + "file_size": 934400 + }, + "down_blocks.2.resnets.0.time_emb_proj.weight_8_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 28167488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_44.const", + "file_size": 1868800 + }, + "down_blocks.2.resnets.1.time_emb_proj.weight_8_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 30036288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_45.const", + "file_size": 1868800 + }, + "down_blocks.3.resnets.0.time_emb_proj.weight_8_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 31905088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_46.const", + "file_size": 1868800 + }, + "down_blocks.3.resnets.1.time_emb_proj.weight_8_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 33773888, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_47.const", + "file_size": 1868800 + }, + "mid_block.resnets.0.time_emb_proj.weight_8_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 35642688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_48.const", + "file_size": 1868800 + }, + "mid_block.resnets.1.time_emb_proj.weight_8_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 37511488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_49.const", + "file_size": 1868800 + }, + "up_blocks.0.resnets.0.time_emb_proj.weight_8_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 39380288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_50.const", + "file_size": 1868800 + }, + "up_blocks.0.resnets.1.time_emb_proj.weight_8_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 41249088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_51.const", + "file_size": 1868800 + }, + "up_blocks.0.resnets.2.time_emb_proj.weight_8_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 43117888, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_52.const", + "file_size": 1868800 + }, + "up_blocks.1.resnets.0.time_emb_proj.weight_8_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 44986688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_53.const", + "file_size": 1868800 + }, + "up_blocks.1.resnets.1.time_emb_proj.weight_8_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 46855488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_54.const", + "file_size": 1868800 + }, + "up_blocks.1.resnets.2.time_emb_proj.weight_8_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 48724288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_55.const", + "file_size": 1868800 + }, + "up_blocks.2.resnets.0.time_emb_proj.weight_8_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "offset": 50593088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_56.const", + "file_size": 934400 + }, + "up_blocks.2.resnets.1.time_emb_proj.weight_8_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "offset": 51527488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_57.const", + "file_size": 934400 + }, + "up_blocks.2.resnets.2.time_emb_proj.weight_8_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 934400 + ], + "size_in_bytes": 934400, + "offset": 52461888, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_58.const", + "file_size": 934400 + }, + "up_blocks.3.resnets.0.time_emb_proj.weight_8_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "offset": 53396288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_59.const", + "file_size": 467200 + }, + "up_blocks.3.resnets.1.time_emb_proj.weight_8_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "offset": 53863488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_60.const", + "file_size": 467200 + }, + "up_blocks.3.resnets.2.time_emb_proj.weight_8_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 467200 + ], + "size_in_bytes": 467200, + "offset": 54330688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_61.const", + "file_size": 467200 + }, + "GroupNorm_1_wts_7_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 54797888, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_62.const", + "file_size": 1280 + }, + "Sigmoid_1.weights10_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 54799168, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_63.const", + "file_size": 256 + }, + "NhwcConv_2_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 54799424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_64.const", + "file_size": 1088000 + }, + "GroupNorm_2_wts_7_1_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 55887424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_65.const", + "file_size": 1280 + }, + "NhwcConv_3_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 55888704, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_66.const", + "file_size": 135680 + }, + "down_blocks.0.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 56024384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_67.const", + "file_size": 640 + }, + "down_blocks.0.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 56025024, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_68.const", + "file_size": 640 + }, + "onnx::MatMul_5649": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 56025664, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_69.const", + "file_size": 117760 + }, + "onnx::MatMul_5648": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 56143424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_70.const", + "file_size": 117760 + }, + "onnx::MatMul_5650": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 56261184, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_71.const", + "file_size": 117760 + }, + "MultiHeadAttention_0_mask.8_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 4096 + ], + "size_in_bytes": 8192, + "offset": 56378944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_72.const", + "file_size": 8192 + }, + "onnx::MatMul_5658": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 56387136, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_73.const", + "file_size": 117760 + }, + "down_blocks.0.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 56504896, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_74.const", + "file_size": 640 + }, + "down_blocks.0.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 56505536, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_75.const", + "file_size": 640 + }, + "onnx::MatMul_5659": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 56506176, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_76.const", + "file_size": 117760 + }, + "MultiHeadAttention_16_mask.8_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 56623936, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_77.const", + "file_size": 512 + }, + "onnx::MatMul_5669": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 56624448, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_78.const", + "file_size": 117760 + }, + "down_blocks.0.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 56742208, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_79.const", + "file_size": 640 + }, + "down_blocks.0.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 56742848, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_80.const", + "file_size": 640 + }, + "onnx::MatMul_5670_2_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 56743488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_81.const", + "file_size": 471040 + }, + "onnx::MatMul_5670_2_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 473600 + ], + "size_in_bytes": 473600, + "offset": 57214528, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_82.const", + "file_size": 473600 + }, + "onnx::MatMul_5671": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 57688128, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_83.const", + "file_size": 471040 + }, + "NhwcConv_4_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 58159168, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_84.const", + "file_size": 135680 + }, + "GroupNorm_3_wts_7_1_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 58294848, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_85.const", + "file_size": 1280 + }, + "Sigmoid_2.weights10_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 58296128, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_86.const", + "file_size": 256 + }, + "NhwcConv_5_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 58296384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_87.const", + "file_size": 1088000 + }, + "GroupNorm_4_wts_7_1_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 59384384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_88.const", + "file_size": 1280 + }, + "Sigmoid_3.weights10_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 59385664, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_89.const", + "file_size": 256 + }, + "NhwcConv_6_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 59385920, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_90.const", + "file_size": 1088000 + }, + "GroupNorm_5_wts_7_1_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 60473920, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_91.const", + "file_size": 1280 + }, + "NhwcConv_7_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 60475200, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_92.const", + "file_size": 135680 + }, + "down_blocks.0.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 60610880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_93.const", + "file_size": 640 + }, + "down_blocks.0.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 60611520, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_94.const", + "file_size": 640 + }, + "onnx::MatMul_5679": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 60612160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_95.const", + "file_size": 117760 + }, + "onnx::MatMul_5678": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 60729920, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_96.const", + "file_size": 117760 + }, + "onnx::MatMul_5680": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 60847680, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_97.const", + "file_size": 117760 + }, + "MultiHeadAttention_1_mask.8_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 4096 + ], + "size_in_bytes": 8192, + "offset": 60965440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_98.const", + "file_size": 8192 + }, + "onnx::MatMul_5688": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 60973632, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_99.const", + "file_size": 117760 + }, + "down_blocks.0.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 61091392, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_100.const", + "file_size": 640 + }, + "down_blocks.0.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 61092032, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_101.const", + "file_size": 640 + }, + "onnx::MatMul_5689": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 61092672, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_102.const", + "file_size": 117760 + }, + "MultiHeadAttention_17_mask.8_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 61210432, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_103.const", + "file_size": 512 + }, + "onnx::MatMul_5699": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 61210944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_104.const", + "file_size": 117760 + }, + "down_blocks.0.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 61328704, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_105.const", + "file_size": 640 + }, + "down_blocks.0.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 61329344, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_106.const", + "file_size": 640 + }, + "onnx::MatMul_5700_2_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 61329984, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_107.const", + "file_size": 471040 + }, + "onnx::MatMul_5700_2_1_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 473600 + ], + "size_in_bytes": 473600, + "offset": 61801024, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_108.const", + "file_size": 473600 + }, + "onnx::MatMul_5701": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 62274624, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_109.const", + "file_size": 471040 + }, + "NhwcConv_8_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 62745664, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_110.const", + "file_size": 135680 + }, + "NhwcConv_9_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1077760 + ], + "size_in_bytes": 1077760, + "offset": 62881344, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_111.const", + "file_size": 1077760 + }, + "GroupNorm_6_wts_7_1_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 63959104, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_112.const", + "file_size": 1280 + }, + "NhwcConv_12_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 271360 + ], + "size_in_bytes": 271360, + "offset": 63960384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_113.const", + "file_size": 271360 + }, + "Sigmoid_4.weights10_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 64231744, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_114.const", + "file_size": 256 + }, + "NhwcConv_10_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2176000 + ], + "size_in_bytes": 2176000, + "offset": 64232000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_115.const", + "file_size": 2176000 + }, + "GroupNorm_7_wts_7_1_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 66408000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_116.const", + "file_size": 2560 + }, + "Sigmoid_5.weights10_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 66410560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_117.const", + "file_size": 256 + }, + "NhwcConv_11_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 66410816, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_118.const", + "file_size": 4352000 + }, + "GroupNorm_8_wts_7_1_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 70762816, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_119.const", + "file_size": 2560 + }, + "NhwcConv_13_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 70765376, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_120.const", + "file_size": 512000 + }, + "down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 71277376, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_121.const", + "file_size": 1280 + }, + "down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 71278656, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_122.const", + "file_size": 1280 + }, + "onnx::MatMul_5709": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 71279936, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_123.const", + "file_size": 471040 + }, + "onnx::MatMul_5708": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 71750976, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_124.const", + "file_size": 471040 + }, + "onnx::MatMul_5710": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 72222016, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_125.const", + "file_size": 471040 + }, + "MultiHeadAttention_2_mask.8_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "offset": 72693056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_126.const", + "file_size": 2048 + }, + "onnx::MatMul_5718": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 72695104, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_127.const", + "file_size": 471040 + }, + "down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 73166144, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_128.const", + "file_size": 1280 + }, + "down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 73167424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_129.const", + "file_size": 1280 + }, + "onnx::MatMul_5719": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 73168704, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_130.const", + "file_size": 471040 + }, + "MultiHeadAttention_18_mask.8_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 73639744, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_131.const", + "file_size": 512 + }, + "onnx::MatMul_5729": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 73640256, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_132.const", + "file_size": 471040 + }, + "down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 74111296, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_133.const", + "file_size": 1280 + }, + "down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 74112576, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_134.const", + "file_size": 1280 + }, + "onnx::MatMul_5730_2_2_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 74113856, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_135.const", + "file_size": 1894400 + }, + "onnx::MatMul_5730_2_2_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 76008256, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_136.const", + "file_size": 1894400 + }, + "onnx::MatMul_5731": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 77902656, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_137.const", + "file_size": 1868800 + }, + "NhwcConv_14_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 79771456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_138.const", + "file_size": 512000 + }, + "GroupNorm_9_wts_7_1_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 80283456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_139.const", + "file_size": 2560 + }, + "Sigmoid_6.weights10_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 80286016, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_140.const", + "file_size": 256 + }, + "NhwcConv_15_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 80286272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_141.const", + "file_size": 4352000 + }, + "GroupNorm_10_wts_7_1_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 84638272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_142.const", + "file_size": 2560 + }, + "Sigmoid_7.weights10_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 84640832, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_143.const", + "file_size": 256 + }, + "NhwcConv_16_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 84641088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_144.const", + "file_size": 4352000 + }, + "GroupNorm_11_wts_7_1_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 88993088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_145.const", + "file_size": 2560 + }, + "NhwcConv_17_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 88995648, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_146.const", + "file_size": 512000 + }, + "down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 89507648, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_147.const", + "file_size": 1280 + }, + "down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 89508928, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_148.const", + "file_size": 1280 + }, + "onnx::MatMul_5739": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 89510208, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_149.const", + "file_size": 471040 + }, + "onnx::MatMul_5738": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 89981248, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_150.const", + "file_size": 471040 + }, + "onnx::MatMul_5740": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 90452288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_151.const", + "file_size": 471040 + }, + "MultiHeadAttention_3_mask.8_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "offset": 90923328, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_152.const", + "file_size": 2048 + }, + "onnx::MatMul_5748": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 90925376, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_153.const", + "file_size": 471040 + }, + "down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 91396416, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_154.const", + "file_size": 1280 + }, + "down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 91397696, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_155.const", + "file_size": 1280 + }, + "onnx::MatMul_5749": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 91398976, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_156.const", + "file_size": 471040 + }, + "MultiHeadAttention_19_mask.8_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 91870016, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_157.const", + "file_size": 512 + }, + "onnx::MatMul_5759": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 91870528, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_158.const", + "file_size": 471040 + }, + "down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 92341568, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_159.const", + "file_size": 1280 + }, + "down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 92342848, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_160.const", + "file_size": 1280 + }, + "onnx::MatMul_5760_2_3_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 92344128, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_161.const", + "file_size": 1894400 + }, + "onnx::MatMul_5760_2_3_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 94238528, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_162.const", + "file_size": 1894400 + }, + "onnx::MatMul_5761": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 96132928, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_163.const", + "file_size": 1868800 + }, + "NhwcConv_18_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 98001728, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_164.const", + "file_size": 512000 + }, + "NhwcConv_19_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4311040 + ], + "size_in_bytes": 4311040, + "offset": 98513728, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_165.const", + "file_size": 4311040 + }, + "NhwcConv_22_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1024000 + ], + "size_in_bytes": 1024000, + "offset": 102824768, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_166.const", + "file_size": 1024000 + }, + "GroupNorm_12_wts_7_1_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 103848768, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_167.const", + "file_size": 2560 + }, + "Sigmoid_8.weights10_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 103851328, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_168.const", + "file_size": 256 + }, + "NhwcConv_20_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 8622080 + ], + "size_in_bytes": 8622080, + "offset": 103851584, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_169.const", + "file_size": 8622080 + }, + "GroupNorm_13_wts_7_1_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 112473664, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_170.const", + "file_size": 5120 + }, + "Sigmoid_9.weights10_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 112478784, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_171.const", + "file_size": 256 + }, + "NhwcConv_21_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 112479040, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_172.const", + "file_size": 17244160 + }, + "GroupNorm_14_wts_7_1_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 129723200, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_173.const", + "file_size": 5120 + }, + "NhwcConv_23_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 129728320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_174.const", + "file_size": 2048000 + }, + "down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 131776320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_175.const", + "file_size": 2560 + }, + "down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 131778880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_176.const", + "file_size": 2560 + }, + "onnx::MatMul_5769": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 131781440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_177.const", + "file_size": 1868800 + }, + "onnx::MatMul_5768": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 133650240, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_178.const", + "file_size": 1868800 + }, + "onnx::MatMul_5770": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 135519040, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_179.const", + "file_size": 1868800 + }, + "MultiHeadAttention_4_mask.8_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 137387840, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_180.const", + "file_size": 512 + }, + "onnx::MatMul_5778": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 137388352, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_181.const", + "file_size": 1868800 + }, + "down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 139257152, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_182.const", + "file_size": 2560 + }, + "down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 139259712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_183.const", + "file_size": 2560 + }, + "onnx::MatMul_5779": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 139262272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_184.const", + "file_size": 1868800 + }, + "MultiHeadAttention_20_mask.8_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 141131072, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_185.const", + "file_size": 512 + }, + "onnx::MatMul_5789": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 141131584, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_186.const", + "file_size": 1868800 + }, + "down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 143000384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_187.const", + "file_size": 2560 + }, + "down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 143002944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_188.const", + "file_size": 2560 + }, + "onnx::MatMul_5790_2_4_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 143005504, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_189.const", + "file_size": 7536640 + }, + "onnx::MatMul_5790_2_4_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 150542144, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_190.const", + "file_size": 7577600 + }, + "onnx::MatMul_5791": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 158119744, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_191.const", + "file_size": 7577600 + }, + "NhwcConv_24_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 165697344, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_192.const", + "file_size": 2048000 + }, + "GroupNorm_15_wts_7_1_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 167745344, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_193.const", + "file_size": 5120 + }, + "Sigmoid_10.weights10_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 167750464, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_194.const", + "file_size": 256 + }, + "NhwcConv_25_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 167750720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_195.const", + "file_size": 17244160 + }, + "GroupNorm_16_wts_7_1_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 184994880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_196.const", + "file_size": 5120 + }, + "Sigmoid_11.weights10_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 185000000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_197.const", + "file_size": 256 + }, + "NhwcConv_26_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 185000256, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_198.const", + "file_size": 17244160 + }, + "GroupNorm_17_wts_7_1_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 202244416, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_199.const", + "file_size": 5120 + }, + "NhwcConv_27_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 202249536, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_200.const", + "file_size": 2048000 + }, + "down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 204297536, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_201.const", + "file_size": 2560 + }, + "down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 204300096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_202.const", + "file_size": 2560 + }, + "onnx::MatMul_5799": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 204302656, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_203.const", + "file_size": 1868800 + }, + "onnx::MatMul_5798": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 206171456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_204.const", + "file_size": 1868800 + }, + "onnx::MatMul_5800": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 208040256, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_205.const", + "file_size": 1868800 + }, + "MultiHeadAttention_5_mask.8_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 209909056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_206.const", + "file_size": 512 + }, + "onnx::MatMul_5808": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 209909568, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_207.const", + "file_size": 1868800 + }, + "down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 211778368, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_208.const", + "file_size": 2560 + }, + "down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 211780928, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_209.const", + "file_size": 2560 + }, + "onnx::MatMul_5809": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 211783488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_210.const", + "file_size": 1868800 + }, + "MultiHeadAttention_21_mask.8_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 213652288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_211.const", + "file_size": 512 + }, + "onnx::MatMul_5819": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 213652800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_212.const", + "file_size": 1868800 + }, + "down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 215521600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_213.const", + "file_size": 2560 + }, + "down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 215524160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_214.const", + "file_size": 2560 + }, + "onnx::MatMul_5820_2_5_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 215526720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_215.const", + "file_size": 7536640 + }, + "onnx::MatMul_5820_2_5_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 223063360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_216.const", + "file_size": 7577600 + }, + "onnx::MatMul_5821": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 230640960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_217.const", + "file_size": 7577600 + }, + "NhwcConv_28_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 238218560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_218.const", + "file_size": 2048000 + }, + "NhwcConv_29_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 240266560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_219.const", + "file_size": 17244160 + }, + "GroupNorm_18_wts_7_1_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 257510720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_220.const", + "file_size": 5120 + }, + "Sigmoid_12.weights10_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 257515840, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_221.const", + "file_size": 256 + }, + "NhwcConv_30_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 257516096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_222.const", + "file_size": 16998400 + }, + "GroupNorm_19_wts_7_1_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 274514496, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_223.const", + "file_size": 5120 + }, + "Sigmoid_13.weights10_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 274519616, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_224.const", + "file_size": 256 + }, + "NhwcConv_31_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 274519872, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_225.const", + "file_size": 16998400 + }, + "GroupNorm_20_wts_7_1_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 291518272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_226.const", + "file_size": 5120 + }, + "Sigmoid_14.weights10_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 291523392, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_227.const", + "file_size": 256 + }, + "NhwcConv_32_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 291523648, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_228.const", + "file_size": 16998400 + }, + "GroupNorm_21_wts_7_1_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 308522048, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_229.const", + "file_size": 5120 + }, + "Sigmoid_15.weights10_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 308527168, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_230.const", + "file_size": 256 + }, + "NhwcConv_33_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 308527424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_231.const", + "file_size": 16998400 + }, + "GroupNorm_22_wts_7_1_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 325525824, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_232.const", + "file_size": 5120 + }, + "Sigmoid_16.weights10_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 325530944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_233.const", + "file_size": 256 + }, + "NhwcConv_34_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 325531200, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_234.const", + "file_size": 16998400 + }, + "GroupNorm_23_wts_7_1_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 342529600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_235.const", + "file_size": 5120 + }, + "Sigmoid_17.weights10_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 342534720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_236.const", + "file_size": 256 + }, + "NhwcConv_35_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 342534976, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_237.const", + "file_size": 16998400 + }, + "GroupNorm_24_wts_7_1_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 359533376, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_238.const", + "file_size": 5120 + }, + "NhwcConv_36_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 359538496, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_239.const", + "file_size": 2048000 + }, + "mid_block.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 361586496, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_240.const", + "file_size": 2560 + }, + "mid_block.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 361589056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_241.const", + "file_size": 2560 + }, + "onnx::MatMul_5837": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "offset": 361591616, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_242.const", + "file_size": 1884160 + }, + "onnx::MatMul_5836": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "offset": 363475776, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_243.const", + "file_size": 1884160 + }, + "onnx::MatMul_5838": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "offset": 365359936, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_244.const", + "file_size": 1884160 + }, + "MultiHeadAttention_6_mask.8_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 367244096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_245.const", + "file_size": 512 + }, + "onnx::MatMul_5846": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "offset": 367244608, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_246.const", + "file_size": 1884160 + }, + "mid_block.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 369128768, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_247.const", + "file_size": 2560 + }, + "mid_block.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 369131328, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_248.const", + "file_size": 2560 + }, + "onnx::MatMul_5847": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "offset": 369133888, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_249.const", + "file_size": 1884160 + }, + "MultiHeadAttention_22_mask.8_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 371018048, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_250.const", + "file_size": 512 + }, + "onnx::MatMul_5857": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1884160 + ], + "size_in_bytes": 1884160, + "offset": 371018560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_251.const", + "file_size": 1884160 + }, + "mid_block.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 372902720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_252.const", + "file_size": 2560 + }, + "mid_block.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 372905280, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_253.const", + "file_size": 2560 + }, + "onnx::MatMul_5858_2_6_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7475200 + ], + "size_in_bytes": 7475200, + "offset": 372907840, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_254.const", + "file_size": 7475200 + }, + "onnx::MatMul_5858_2_6_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 380383040, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_255.const", + "file_size": 7536640 + }, + "onnx::MatMul_5859": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 387919680, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_256.const", + "file_size": 7536640 + }, + "NhwcConv_37_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 395456320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_257.const", + "file_size": 2048000 + }, + "GroupNorm_25_wts_7_1_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 397504320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_258.const", + "file_size": 5120 + }, + "Sigmoid_18.weights10_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 397509440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_259.const", + "file_size": 256 + }, + "NhwcConv_38_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 397509696, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_260.const", + "file_size": 16998400 + }, + "GroupNorm_26_wts_7_1_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 414508096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_261.const", + "file_size": 5120 + }, + "Sigmoid_19.weights10_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 414513216, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_262.const", + "file_size": 256 + }, + "NhwcConv_39_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 414513472, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_263.const", + "file_size": 16998400 + }, + "GroupNorm_27_wts_7_1_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "offset": 431511872, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_264.const", + "file_size": 10240 + }, + "NhwcConv_42_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4096000 + ], + "size_in_bytes": 4096000, + "offset": 431522112, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_265.const", + "file_size": 4096000 + }, + "Sigmoid_20.weights10_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 435618112, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_266.const", + "file_size": 256 + }, + "NhwcConv_40_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 34488320 + ], + "size_in_bytes": 34488320, + "offset": 435618368, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_267.const", + "file_size": 34488320 + }, + "GroupNorm_28_wts_7_1_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 470106688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_268.const", + "file_size": 5120 + }, + "Sigmoid_21.weights10_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 470111808, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_269.const", + "file_size": 256 + }, + "NhwcConv_41_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 470112064, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_270.const", + "file_size": 16998400 + }, + "GroupNorm_29_wts_7_1_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "offset": 487110464, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_271.const", + "file_size": 10240 + }, + "NhwcConv_45_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4096000 + ], + "size_in_bytes": 4096000, + "offset": 487120704, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_272.const", + "file_size": 4096000 + }, + "Sigmoid_22.weights10_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 491216704, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_273.const", + "file_size": 256 + }, + "NhwcConv_43_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 34488320 + ], + "size_in_bytes": 34488320, + "offset": 491216960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_274.const", + "file_size": 34488320 + }, + "GroupNorm_30_wts_7_1_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 525705280, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_275.const", + "file_size": 5120 + }, + "Sigmoid_23.weights10_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 525710400, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_276.const", + "file_size": 256 + }, + "NhwcConv_44_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 525710656, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_277.const", + "file_size": 16998400 + }, + "GroupNorm_31_wts_7_1_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "offset": 542709056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_278.const", + "file_size": 10240 + }, + "NhwcConv_48_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4096000 + ], + "size_in_bytes": 4096000, + "offset": 542719296, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_279.const", + "file_size": 4096000 + }, + "Sigmoid_24.weights10_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 546815296, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_280.const", + "file_size": 256 + }, + "NhwcConv_46_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 34488320 + ], + "size_in_bytes": 34488320, + "offset": 546815552, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_281.const", + "file_size": 34488320 + }, + "GroupNorm_32_wts_7_1_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 581303872, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_282.const", + "file_size": 5120 + }, + "Sigmoid_25.weights10_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 581308992, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_283.const", + "file_size": 256 + }, + "NhwcConv_47_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 16998400 + ], + "size_in_bytes": 16998400, + "offset": 581309248, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_284.const", + "file_size": 16998400 + }, + "/up_blocks.0/upsamplers.0/Resize.weights4_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 598307648, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_285.const", + "file_size": 256 + }, + "NhwcConv_49_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 598307904, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_286.const", + "file_size": 17244160 + }, + "GroupNorm_33_wts_7_1_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "offset": 615552064, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_287.const", + "file_size": 10240 + }, + "NhwcConv_52_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4096000 + ], + "size_in_bytes": 4096000, + "offset": 615562304, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_288.const", + "file_size": 4096000 + }, + "Sigmoid_26.weights10_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 619658304, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_289.const", + "file_size": 256 + }, + "NhwcConv_50_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 34488320 + ], + "size_in_bytes": 34488320, + "offset": 619658560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_290.const", + "file_size": 34488320 + }, + "GroupNorm_34_wts_7_1_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 654146880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_291.const", + "file_size": 5120 + }, + "Sigmoid_27.weights10_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 654152000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_292.const", + "file_size": 256 + }, + "NhwcConv_51_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 654152256, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_293.const", + "file_size": 17244160 + }, + "GroupNorm_35_wts_7_1_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 671396416, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_294.const", + "file_size": 5120 + }, + "NhwcConv_53_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 671401536, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_295.const", + "file_size": 2048000 + }, + "up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 673449536, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_296.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 673452096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_297.const", + "file_size": 2560 + }, + "onnx::MatMul_5884": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 673454656, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_298.const", + "file_size": 1868800 + }, + "onnx::MatMul_5883": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 675323456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_299.const", + "file_size": 1868800 + }, + "onnx::MatMul_5885": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 677192256, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_300.const", + "file_size": 1868800 + }, + "MultiHeadAttention_7_mask.8_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 679061056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_301.const", + "file_size": 512 + }, + "onnx::MatMul_5893": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 679061568, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_302.const", + "file_size": 1868800 + }, + "up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 680930368, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_303.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 680932928, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_304.const", + "file_size": 2560 + }, + "onnx::MatMul_5894": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 680935488, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_305.const", + "file_size": 1868800 + }, + "MultiHeadAttention_23_mask.8_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 682804288, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_306.const", + "file_size": 512 + }, + "onnx::MatMul_5904": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 682804800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_307.const", + "file_size": 1868800 + }, + "up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 684673600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_308.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 684676160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_309.const", + "file_size": 2560 + }, + "onnx::MatMul_5905_2_7_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 684678720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_310.const", + "file_size": 7536640 + }, + "onnx::MatMul_5905_2_7_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 692215360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_311.const", + "file_size": 7577600 + }, + "onnx::MatMul_5906": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 699792960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_312.const", + "file_size": 7577600 + }, + "NhwcConv_54_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 707370560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_313.const", + "file_size": 2048000 + }, + "GroupNorm_36_wts_7_1_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 5120 + ], + "size_in_bytes": 10240, + "offset": 709418560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_314.const", + "file_size": 10240 + }, + "NhwcConv_57_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4096000 + ], + "size_in_bytes": 4096000, + "offset": 709428800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_315.const", + "file_size": 4096000 + }, + "Sigmoid_28.weights10_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 713524800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_316.const", + "file_size": 256 + }, + "NhwcConv_55_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 34488320 + ], + "size_in_bytes": 34488320, + "offset": 713525056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_317.const", + "file_size": 34488320 + }, + "GroupNorm_37_wts_7_1_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 748013376, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_318.const", + "file_size": 5120 + }, + "Sigmoid_29.weights10_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 748018496, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_319.const", + "file_size": 256 + }, + "NhwcConv_56_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 748018752, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_320.const", + "file_size": 17244160 + }, + "GroupNorm_38_wts_7_1_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 765262912, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_321.const", + "file_size": 5120 + }, + "NhwcConv_58_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 765268032, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_322.const", + "file_size": 2048000 + }, + "up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 767316032, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_323.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 767318592, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_324.const", + "file_size": 2560 + }, + "onnx::MatMul_5914": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 767321152, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_325.const", + "file_size": 1868800 + }, + "onnx::MatMul_5913": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 769189952, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_326.const", + "file_size": 1868800 + }, + "onnx::MatMul_5915": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 771058752, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_327.const", + "file_size": 1868800 + }, + "MultiHeadAttention_8_mask.8_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 772927552, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_328.const", + "file_size": 512 + }, + "onnx::MatMul_5923": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 772928064, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_329.const", + "file_size": 1868800 + }, + "up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 774796864, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_330.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 774799424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_331.const", + "file_size": 2560 + }, + "onnx::MatMul_5924": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 774801984, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_332.const", + "file_size": 1868800 + }, + "MultiHeadAttention_24_mask.8_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 776670784, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_333.const", + "file_size": 512 + }, + "onnx::MatMul_5934": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 776671296, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_334.const", + "file_size": 1868800 + }, + "up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 778540096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_335.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 778542656, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_336.const", + "file_size": 2560 + }, + "onnx::MatMul_5935_2_8_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 778545216, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_337.const", + "file_size": 7536640 + }, + "onnx::MatMul_5935_2_8_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 786081856, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_338.const", + "file_size": 7577600 + }, + "onnx::MatMul_5936": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 793659456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_339.const", + "file_size": 7577600 + }, + "NhwcConv_59_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 801237056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_340.const", + "file_size": 2048000 + }, + "GroupNorm_39_wts_7_1_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 3840 + ], + "size_in_bytes": 7680, + "offset": 803285056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_341.const", + "file_size": 7680 + }, + "NhwcConv_62_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3072000 + ], + "size_in_bytes": 3072000, + "offset": 803292736, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_342.const", + "file_size": 3072000 + }, + "Sigmoid_30.weights10_32": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 806364736, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_343.const", + "file_size": 256 + }, + "NhwcConv_60_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 25702400 + ], + "size_in_bytes": 25702400, + "offset": 806364992, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_344.const", + "file_size": 25702400 + }, + "GroupNorm_40_wts_7_1_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 832067392, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_345.const", + "file_size": 5120 + }, + "Sigmoid_31.weights10_33": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 832072512, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_346.const", + "file_size": 256 + }, + "NhwcConv_61_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17244160 + ], + "size_in_bytes": 17244160, + "offset": 832072768, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_347.const", + "file_size": 17244160 + }, + "GroupNorm_41_wts_7_1_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 849316928, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_348.const", + "file_size": 5120 + }, + "NhwcConv_63_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 849322048, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_349.const", + "file_size": 2048000 + }, + "up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 851370048, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_350.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 851372608, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_351.const", + "file_size": 2560 + }, + "onnx::MatMul_5944": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 851375168, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_352.const", + "file_size": 1868800 + }, + "onnx::MatMul_5943": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 853243968, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_353.const", + "file_size": 1868800 + }, + "onnx::MatMul_5945": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 855112768, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_354.const", + "file_size": 1868800 + }, + "MultiHeadAttention_9_mask.8_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 856981568, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_355.const", + "file_size": 512 + }, + "onnx::MatMul_5953": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 856982080, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_356.const", + "file_size": 1868800 + }, + "up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 858850880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_357.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 858853440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_358.const", + "file_size": 2560 + }, + "onnx::MatMul_5954": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 858856000, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_359.const", + "file_size": 1868800 + }, + "MultiHeadAttention_25_mask.8_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 860724800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_360.const", + "file_size": 512 + }, + "onnx::MatMul_5964": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 860725312, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_361.const", + "file_size": 1868800 + }, + "up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 862594112, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_362.const", + "file_size": 2560 + }, + "up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 862596672, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_363.const", + "file_size": 2560 + }, + "onnx::MatMul_5965_2_9_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7536640 + ], + "size_in_bytes": 7536640, + "offset": 862599232, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_364.const", + "file_size": 7536640 + }, + "onnx::MatMul_5965_2_9_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 870135872, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_365.const", + "file_size": 7577600 + }, + "onnx::MatMul_5966": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 7577600 + ], + "size_in_bytes": 7577600, + "offset": 877713472, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_366.const", + "file_size": 7577600 + }, + "NhwcConv_64_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2048000 + ], + "size_in_bytes": 2048000, + "offset": 885291072, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_367.const", + "file_size": 2048000 + }, + "/up_blocks.1/upsamplers.0/Resize.weights4_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 887339072, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_368.const", + "file_size": 256 + }, + "NhwcConv_65_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 17408000 + ], + "size_in_bytes": 17408000, + "offset": 887339328, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_369.const", + "file_size": 17408000 + }, + "GroupNorm_42_wts_7_1_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 3840 + ], + "size_in_bytes": 7680, + "offset": 904747328, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_370.const", + "file_size": 7680 + }, + "NhwcConv_68_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1546240 + ], + "size_in_bytes": 1546240, + "offset": 904755008, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_371.const", + "file_size": 1546240 + }, + "Sigmoid_32.weights10_34": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 906301248, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_372.const", + "file_size": 256 + }, + "NhwcConv_66_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 13056000 + ], + "size_in_bytes": 13056000, + "offset": 906301504, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_373.const", + "file_size": 13056000 + }, + "GroupNorm_43_wts_7_1_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 919357504, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_374.const", + "file_size": 2560 + }, + "Sigmoid_33.weights10_35": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 919360064, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_375.const", + "file_size": 256 + }, + "NhwcConv_67_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 919360320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_376.const", + "file_size": 4352000 + }, + "GroupNorm_44_wts_7_1_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 923712320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_377.const", + "file_size": 2560 + }, + "NhwcConv_69_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 923714880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_378.const", + "file_size": 512000 + }, + "up_blocks.2.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 924226880, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_379.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 924228160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_380.const", + "file_size": 1280 + }, + "onnx::MatMul_5975": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 924229440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_381.const", + "file_size": 471040 + }, + "onnx::MatMul_5974": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 924700480, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_382.const", + "file_size": 471040 + }, + "onnx::MatMul_5976": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 925171520, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_383.const", + "file_size": 471040 + }, + "MultiHeadAttention_10_mask.8_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "offset": 925642560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_384.const", + "file_size": 2048 + }, + "onnx::MatMul_5984": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 925644608, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_385.const", + "file_size": 471040 + }, + "up_blocks.2.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 926115648, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_386.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 926116928, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_387.const", + "file_size": 1280 + }, + "onnx::MatMul_5985": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 926118208, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_388.const", + "file_size": 471040 + }, + "MultiHeadAttention_26_mask.8_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 926589248, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_389.const", + "file_size": 512 + }, + "onnx::MatMul_5995": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 926589760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_390.const", + "file_size": 471040 + }, + "up_blocks.2.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 927060800, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_391.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 927062080, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_392.const", + "file_size": 1280 + }, + "onnx::MatMul_5996_2_10_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 927063360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_393.const", + "file_size": 1894400 + }, + "onnx::MatMul_5996_2_10_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 928957760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_394.const", + "file_size": 1894400 + }, + "onnx::MatMul_5997": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 930852160, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_395.const", + "file_size": 1868800 + }, + "NhwcConv_70_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 932720960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_396.const", + "file_size": 512000 + }, + "GroupNorm_45_wts_7_1_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 2560 + ], + "size_in_bytes": 5120, + "offset": 933232960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_397.const", + "file_size": 5120 + }, + "NhwcConv_73_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1024000 + ], + "size_in_bytes": 1024000, + "offset": 933238080, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_398.const", + "file_size": 1024000 + }, + "Sigmoid_34.weights10_36": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 934262080, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_399.const", + "file_size": 256 + }, + "NhwcConv_71_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 8704000 + ], + "size_in_bytes": 8704000, + "offset": 934262336, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_400.const", + "file_size": 8704000 + }, + "GroupNorm_46_wts_7_1_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 942966336, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_401.const", + "file_size": 2560 + }, + "Sigmoid_35.weights10_37": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 942968896, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_402.const", + "file_size": 256 + }, + "NhwcConv_72_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 942969152, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_403.const", + "file_size": 4352000 + }, + "GroupNorm_47_wts_7_1_47": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 947321152, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_404.const", + "file_size": 2560 + }, + "NhwcConv_74_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 947323712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_405.const", + "file_size": 512000 + }, + "up_blocks.2.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 947835712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_406.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 947836992, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_407.const", + "file_size": 1280 + }, + "onnx::MatMul_6005": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 947838272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_408.const", + "file_size": 471040 + }, + "onnx::MatMul_6004": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 948309312, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_409.const", + "file_size": 471040 + }, + "onnx::MatMul_6006": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 948780352, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_410.const", + "file_size": 471040 + }, + "MultiHeadAttention_11_mask.8_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "offset": 949251392, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_411.const", + "file_size": 2048 + }, + "onnx::MatMul_6014": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 949253440, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_412.const", + "file_size": 471040 + }, + "up_blocks.2.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 949724480, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_413.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 949725760, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_414.const", + "file_size": 1280 + }, + "onnx::MatMul_6015": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 949727040, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_415.const", + "file_size": 471040 + }, + "MultiHeadAttention_27_mask.8_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 950198080, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_416.const", + "file_size": 512 + }, + "onnx::MatMul_6025": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 950198592, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_417.const", + "file_size": 471040 + }, + "up_blocks.2.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 950669632, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_418.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 950670912, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_419.const", + "file_size": 1280 + }, + "onnx::MatMul_6026_2_11_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 950672192, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_420.const", + "file_size": 1894400 + }, + "onnx::MatMul_6026_2_11_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 952566592, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_421.const", + "file_size": 1894400 + }, + "onnx::MatMul_6027": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 954460992, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_422.const", + "file_size": 1868800 + }, + "NhwcConv_75_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 956329792, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_423.const", + "file_size": 512000 + }, + "GroupNorm_48_wts_7_1_48": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1920 + ], + "size_in_bytes": 3840, + "offset": 956841792, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_424.const", + "file_size": 3840 + }, + "NhwcConv_78_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 773120 + ], + "size_in_bytes": 773120, + "offset": 956845632, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_425.const", + "file_size": 773120 + }, + "Sigmoid_36.weights10_38": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 957618752, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_426.const", + "file_size": 256 + }, + "NhwcConv_76_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 6528000 + ], + "size_in_bytes": 6528000, + "offset": 957619008, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_427.const", + "file_size": 6528000 + }, + "GroupNorm_49_wts_7_1_49": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 964147008, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_428.const", + "file_size": 2560 + }, + "Sigmoid_37.weights10_39": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 964149568, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_429.const", + "file_size": 256 + }, + "NhwcConv_77_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 964149824, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_430.const", + "file_size": 4352000 + }, + "GroupNorm_50_wts_7_1_50": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 968501824, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_431.const", + "file_size": 2560 + }, + "NhwcConv_79_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 968504384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_432.const", + "file_size": 512000 + }, + "up_blocks.2.attentions.2.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 969016384, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_433.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.2.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 969017664, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_434.const", + "file_size": 1280 + }, + "onnx::MatMul_6035": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 969018944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_435.const", + "file_size": 471040 + }, + "onnx::MatMul_6034": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 969489984, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_436.const", + "file_size": 471040 + }, + "onnx::MatMul_6036": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 969961024, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_437.const", + "file_size": 471040 + }, + "MultiHeadAttention_12_mask.8_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "offset": 970432064, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_438.const", + "file_size": 2048 + }, + "onnx::MatMul_6044": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 970434112, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_439.const", + "file_size": 471040 + }, + "up_blocks.2.attentions.2.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 970905152, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_440.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.2.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 970906432, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_441.const", + "file_size": 1280 + }, + "onnx::MatMul_6045": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 970907712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_442.const", + "file_size": 471040 + }, + "MultiHeadAttention_28_mask.8_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 971378752, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_443.const", + "file_size": 512 + }, + "onnx::MatMul_6055": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 971379264, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_444.const", + "file_size": 471040 + }, + "up_blocks.2.attentions.2.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 971850304, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_445.const", + "file_size": 1280 + }, + "up_blocks.2.attentions.2.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 971851584, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_446.const", + "file_size": 1280 + }, + "onnx::MatMul_6056_2_12_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 971852864, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_447.const", + "file_size": 1894400 + }, + "onnx::MatMul_6056_2_12_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1894400 + ], + "size_in_bytes": 1894400, + "offset": 973747264, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_448.const", + "file_size": 1894400 + }, + "onnx::MatMul_6057": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1868800 + ], + "size_in_bytes": 1868800, + "offset": 975641664, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_449.const", + "file_size": 1868800 + }, + "NhwcConv_80_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 512000 + ], + "size_in_bytes": 512000, + "offset": 977510464, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_450.const", + "file_size": 512000 + }, + "/up_blocks.2/upsamplers.0/Resize.weights4_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 978022464, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_451.const", + "file_size": 256 + }, + "NhwcConv_81_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 4352000 + ], + "size_in_bytes": 4352000, + "offset": 978022720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_452.const", + "file_size": 4352000 + }, + "GroupNorm_51_wts_7_1_51": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1920 + ], + "size_in_bytes": 3840, + "offset": 982374720, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_453.const", + "file_size": 3840 + }, + "NhwcConv_84_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 386560 + ], + "size_in_bytes": 386560, + "offset": 982378560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_454.const", + "file_size": 386560 + }, + "Sigmoid_38.weights10_40": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 982765120, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_455.const", + "file_size": 256 + }, + "NhwcConv_82_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 3212800 + ], + "size_in_bytes": 3212800, + "offset": 982765376, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_456.const", + "file_size": 3212800 + }, + "GroupNorm_52_wts_7_1_52": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 985978176, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_457.const", + "file_size": 1280 + }, + "Sigmoid_39.weights10_41": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 985979456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_458.const", + "file_size": 256 + }, + "NhwcConv_83_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 985979712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_459.const", + "file_size": 1088000 + }, + "GroupNorm_53_wts_7_1_53": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 987067712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_460.const", + "file_size": 1280 + }, + "NhwcConv_85_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 987068992, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_461.const", + "file_size": 135680 + }, + "up_blocks.3.attentions.0.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 987204672, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_462.const", + "file_size": 640 + }, + "up_blocks.3.attentions.0.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 987205312, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_463.const", + "file_size": 640 + }, + "onnx::MatMul_6066": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 987205952, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_464.const", + "file_size": 117760 + }, + "onnx::MatMul_6065": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 987323712, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_465.const", + "file_size": 117760 + }, + "onnx::MatMul_6067": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 987441472, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_466.const", + "file_size": 117760 + }, + "MultiHeadAttention_13_mask.8_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 4096 + ], + "size_in_bytes": 8192, + "offset": 987559232, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_467.const", + "file_size": 8192 + }, + "onnx::MatMul_6075": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 987567424, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_468.const", + "file_size": 117760 + }, + "up_blocks.3.attentions.0.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 987685184, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_469.const", + "file_size": 640 + }, + "up_blocks.3.attentions.0.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 987685824, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_470.const", + "file_size": 640 + }, + "onnx::MatMul_6076": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 987686464, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_471.const", + "file_size": 117760 + }, + "MultiHeadAttention_29_mask.8_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 987804224, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_472.const", + "file_size": 512 + }, + "onnx::MatMul_6086": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 987804736, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_473.const", + "file_size": 117760 + }, + "up_blocks.3.attentions.0.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 987922496, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_474.const", + "file_size": 640 + }, + "up_blocks.3.attentions.0.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 987923136, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_475.const", + "file_size": 640 + }, + "onnx::MatMul_6087_2_13_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 987923776, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_476.const", + "file_size": 471040 + }, + "onnx::MatMul_6087_2_13_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 473600 + ], + "size_in_bytes": 473600, + "offset": 988394816, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_477.const", + "file_size": 473600 + }, + "onnx::MatMul_6088": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 988868416, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_478.const", + "file_size": 471040 + }, + "NhwcConv_86_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 989339456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_479.const", + "file_size": 135680 + }, + "GroupNorm_54_wts_7_1_54": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 989475136, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_480.const", + "file_size": 2560 + }, + "NhwcConv_89_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 271360 + ], + "size_in_bytes": 271360, + "offset": 989477696, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_481.const", + "file_size": 271360 + }, + "Sigmoid_40.weights10_42": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 989749056, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_482.const", + "file_size": 256 + }, + "NhwcConv_87_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2176000 + ], + "size_in_bytes": 2176000, + "offset": 989749312, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_483.const", + "file_size": 2176000 + }, + "GroupNorm_55_wts_7_1_55": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 991925312, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_484.const", + "file_size": 1280 + }, + "Sigmoid_41.weights10_43": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 991926592, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_485.const", + "file_size": 256 + }, + "NhwcConv_88_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 991926848, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_486.const", + "file_size": 1088000 + }, + "GroupNorm_56_wts_7_1_56": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 993014848, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_487.const", + "file_size": 1280 + }, + "NhwcConv_90_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 993016128, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_488.const", + "file_size": 135680 + }, + "up_blocks.3.attentions.1.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 993151808, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_489.const", + "file_size": 640 + }, + "up_blocks.3.attentions.1.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 993152448, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_490.const", + "file_size": 640 + }, + "onnx::MatMul_6096": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 993153088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_491.const", + "file_size": 117760 + }, + "onnx::MatMul_6095": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 993270848, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_492.const", + "file_size": 117760 + }, + "onnx::MatMul_6097": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 993388608, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_493.const", + "file_size": 117760 + }, + "MultiHeadAttention_14_mask.8_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 4096 + ], + "size_in_bytes": 8192, + "offset": 993506368, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_494.const", + "file_size": 8192 + }, + "onnx::MatMul_6105": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 993514560, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_495.const", + "file_size": 117760 + }, + "up_blocks.3.attentions.1.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 993632320, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_496.const", + "file_size": 640 + }, + "up_blocks.3.attentions.1.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 993632960, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_497.const", + "file_size": 640 + }, + "onnx::MatMul_6106": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 993633600, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_498.const", + "file_size": 117760 + }, + "MultiHeadAttention_30_mask.8_29": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 993751360, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_499.const", + "file_size": 512 + }, + "onnx::MatMul_6116": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 993751872, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_500.const", + "file_size": 117760 + }, + "up_blocks.3.attentions.1.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 993869632, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_501.const", + "file_size": 640 + }, + "up_blocks.3.attentions.1.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 993870272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_502.const", + "file_size": 640 + }, + "onnx::MatMul_6117_2_14_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 993870912, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_503.const", + "file_size": 471040 + }, + "onnx::MatMul_6117_2_14_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 473600 + ], + "size_in_bytes": 473600, + "offset": 994341952, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_504.const", + "file_size": 473600 + }, + "onnx::MatMul_6118": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 994815552, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_505.const", + "file_size": 471040 + }, + "NhwcConv_91_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 995286592, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_506.const", + "file_size": 135680 + }, + "GroupNorm_57_wts_7_1_57": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1280 + ], + "size_in_bytes": 2560, + "offset": 995422272, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_507.const", + "file_size": 2560 + }, + "NhwcConv_94_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 271360 + ], + "size_in_bytes": 271360, + "offset": 995424832, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_508.const", + "file_size": 271360 + }, + "Sigmoid_42.weights10_44": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 995696192, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_509.const", + "file_size": 256 + }, + "NhwcConv_92_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2176000 + ], + "size_in_bytes": 2176000, + "offset": 995696448, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_510.const", + "file_size": 2176000 + }, + "GroupNorm_58_wts_7_1_58": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 997872448, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_511.const", + "file_size": 1280 + }, + "Sigmoid_43.weights10_45": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 997873728, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_512.const", + "file_size": 256 + }, + "NhwcConv_93_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1088000 + ], + "size_in_bytes": 1088000, + "offset": 997873984, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_513.const", + "file_size": 1088000 + }, + "GroupNorm_59_wts_7_1_59": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 998961984, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_514.const", + "file_size": 1280 + }, + "NhwcConv_95_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 998963264, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_515.const", + "file_size": 135680 + }, + "up_blocks.3.attentions.2.transformer_blocks.0.norm1.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 999098944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_516.const", + "file_size": 640 + }, + "up_blocks.3.attentions.2.transformer_blocks.0.norm1.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 999099584, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_517.const", + "file_size": 640 + }, + "onnx::MatMul_6126": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 999100224, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_518.const", + "file_size": 117760 + }, + "onnx::MatMul_6125": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 999217984, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_519.const", + "file_size": 117760 + }, + "onnx::MatMul_6127": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 999335744, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_520.const", + "file_size": 117760 + }, + "MultiHeadAttention_15_mask.8_30": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 4096 + ], + "size_in_bytes": 8192, + "offset": 999453504, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_521.const", + "file_size": 8192 + }, + "onnx::MatMul_6135": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 999461696, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_522.const", + "file_size": 117760 + }, + "up_blocks.3.attentions.2.transformer_blocks.0.norm2.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 999579456, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_523.const", + "file_size": 640 + }, + "up_blocks.3.attentions.2.transformer_blocks.0.norm2.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 999580096, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_524.const", + "file_size": 640 + }, + "onnx::MatMul_6136": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 999580736, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_525.const", + "file_size": 117760 + }, + "MultiHeadAttention_31_mask.8_31": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "offset": 999698496, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_526.const", + "file_size": 512 + }, + "onnx::MatMul_6146": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 117760 + ], + "size_in_bytes": 117760, + "offset": 999699008, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_527.const", + "file_size": 117760 + }, + "up_blocks.3.attentions.2.transformer_blocks.0.norm3.weight": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 999816768, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_528.const", + "file_size": 640 + }, + "up_blocks.3.attentions.2.transformer_blocks.0.norm3.bias": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 320 + ], + "size_in_bytes": 640, + "offset": 999817408, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_529.const", + "file_size": 640 + }, + "onnx::MatMul_6147_2_15_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 999818048, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_530.const", + "file_size": 471040 + }, + "onnx::MatMul_6147_2_15_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 473600 + ], + "size_in_bytes": 473600, + "offset": 1000289088, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_531.const", + "file_size": 473600 + }, + "onnx::MatMul_6148": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 471040 + ], + "size_in_bytes": 471040, + "offset": 1000762688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_532.const", + "file_size": 471040 + }, + "NhwcConv_96_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 135680 + ], + "size_in_bytes": 135680, + "offset": 1001233728, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_533.const", + "file_size": 135680 + }, + "GroupNorm_60_wts_7_1_60": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 640 + ], + "size_in_bytes": 1280, + "offset": 1001369408, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_534.const", + "file_size": 1280 + }, + "Sigmoid_44.weights10_46": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "offset": 1001370688, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_535.const", + "file_size": 256 + }, + "NhwcConv_97_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 54400 + ], + "size_in_bytes": 54400, + "offset": 1001370944, + "file_name": ".cache\\down_blocks.0attentions.0transformer_blocks.0attn2to_kMatMul_536.const", + "file_size": 54400 + } + }, + "aux_info": {} +} \ No newline at end of file