program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3402.3.2"}, {"coremlc-version", "3402.4.1"}})]
{
    func main<ios18>(tensor<fp16, [1, 1, ?, ?]> causal_mask, tensor<int32, [1, ?]> input_ids, state<tensor<fp16, [16, 1, 8, 2048, 64]>> key_cache, state<tensor<fp16, [16, 1, 8, 2048, 64]>> value_cache) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"causal_mask", [1, 1, 1, 1]}, {"input_ids", [1, 1]}}), ("RangeDims", {{"causal_mask", [[1, 1], [1, 1], [1, 2048], [1, 2048]]}, {"input_ids", [[1, 1], [1, 2048]]}})))] {
            tensor<int32, [4]> var_7_shape_cast_fp16 = shape(x = causal_mask)[name = string("op_7_shape_cast_fp16")];
            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
            string var_7_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_7_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(3)];
            tensor<int16, [4]> var_7_shape_cast_fp16_to_int16 = cast(dtype = var_7_shape_cast_fp16_to_int16_dtype_0, x = var_7_shape_cast_fp16)[name = string("cast_136")];
            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_7_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_10_shape = shape(x = input_ids)[name = string("op_10_shape")];
            int32 gather_1_axis_0 = const()[name = string("gather_1_axis_0"), val = int32(0)];
            int32 gather_1_batch_dims_0 = const()[name = string("gather_1_batch_dims_0"), val = int32(0)];
            bool gather_1_validate_indices_0 = const()[name = string("gather_1_validate_indices_0"), val = bool(false)];
            string var_10_shape_to_uint16_dtype_0 = const()[name = string("op_10_shape_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_1_to_uint16 = const()[name = string("select_1_to_uint16"), val = uint16(1)];
            tensor<uint16, [2]> var_10_shape_to_uint16 = cast(dtype = var_10_shape_to_uint16_dtype_0, x = var_10_shape)[name = string("cast_135")];
            uint16 gather_1_cast_uint16 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = select_1_to_uint16, validate_indices = gather_1_validate_indices_0, x = var_10_shape_to_uint16)[name = string("gather_1_cast_uint16")];
            string gather_1_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_1_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_133")];
            int32 gather_1_cast_uint16_to_int32 = cast(dtype = gather_1_cast_uint16_to_int32_dtype_0, x = gather_1_cast_uint16)[name = string("cast_134")];
            int32 past_seen_tokens = sub(x = gather_0_cast_uint16_to_int32, y = gather_1_cast_uint16_to_int32)[name = string("past_seen_tokens")];
            int32 var_48 = const()[name = string("op_48"), val = int32(-1)];
            int32 var_59 = const()[name = string("op_59"), val = int32(4)];
            int32 inputs_embeds_axis_0 = const()[name = string("inputs_embeds_axis_0"), val = int32(0)];
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [128256, 2048]> model_model_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [128256, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [128256, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131334272))))[name = string("model_model_embed_tokens_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids, validate_indices = inputs_embeds_validate_indices_0, x = model_model_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            tensor<int32, [3]> var_113_shape_cast_fp16 = shape(x = inputs_embeds_cast_fp16)[name = string("op_113_shape_cast_fp16")];
            int32 gather_2_axis_0 = const()[name = string("gather_2_axis_0"), val = int32(0)];
            int32 gather_2_batch_dims_0 = const()[name = string("gather_2_batch_dims_0"), val = int32(0)];
            bool gather_2_validate_indices_0 = const()[name = string("gather_2_validate_indices_0"), val = bool(false)];
            string var_113_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_113_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_2_to_uint16 = const()[name = string("select_2_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_113_shape_cast_fp16_to_uint16 = cast(dtype = var_113_shape_cast_fp16_to_uint16_dtype_0, x = var_113_shape_cast_fp16)[name = string("cast_132")];
            uint16 gather_2_cast_uint16 = gather(axis = gather_2_axis_0, batch_dims = gather_2_batch_dims_0, indices = select_2_to_uint16, validate_indices = gather_2_validate_indices_0, x = var_113_shape_cast_fp16_to_uint16)[name = string("gather_2_cast_uint16")];
            string gather_2_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_2_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_2_cast_uint16_to_int32 = cast(dtype = gather_2_cast_uint16_to_int32_dtype_0, x = gather_2_cast_uint16)[name = string("cast_131")];
            int32 var_115 = add(x = past_seen_tokens, y = gather_2_cast_uint16_to_int32)[name = string("op_115")];
            int32 const_0 = const()[name = string("const_0"), val = int32(1)];
            tensor<int32, [?]> cache_position = range_1d(end = var_115, start = past_seen_tokens, step = const_0)[name = string("cache_position")];
            tensor<int32, [1]> position_ids_axes_0 = const()[name = string("position_ids_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1, ?]> position_ids = expand_dims(axes = position_ids_axes_0, x = cache_position)[name = string("position_ids")];
            tensor<int32, [1]> var_128_axes_0 = const()[name = string("op_128_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, ?]> var_128 = expand_dims(axes = var_128_axes_0, x = position_ids)[name = string("op_128")];
            bool var_133_transpose_x_0 = const()[name = string("op_133_transpose_x_0"), val = bool(false)];
            bool var_133_transpose_y_0 = const()[name = string("op_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_2_to_fp16 = const()[name = string("const_2_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147751104)))];
            string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, ?]> var_128_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = var_128)[name = string("cast_130")];
            tensor<fp16, [1, 32, ?]> var_133_cast_fp16 = matmul(transpose_x = var_133_transpose_x_0, transpose_y = var_133_transpose_y_0, x = const_2_to_fp16, y = var_128_to_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, ?, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_133_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, ?, 64]> emb_cast_fp16 = concat(axis = var_48, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, ?, 64]> cos_1_cast_fp16 = cos(x = emb_cast_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [1, ?, 64]> sin_1_cast_fp16 = sin(x = emb_cast_fp16)[name = string("sin_1_cast_fp16")];
            fp16 var_54_promoted_to_fp16 = const()[name = string("op_54_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_154_cast_fp16 = pow(x = inputs_embeds_cast_fp16, y = var_54_promoted_to_fp16)[name = string("op_154_cast_fp16")];
            tensor<int32, [1]> variance_1_axes_0 = const()[name = string("variance_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_1_keep_dims_0 = const()[name = string("variance_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_1_cast_fp16 = reduce_mean(axes = variance_1_axes_0, keep_dims = variance_1_keep_dims_0, x = var_154_cast_fp16)[name = string("variance_1_cast_fp16")];
            fp16 var_157_to_fp16 = const()[name = string("op_157_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_158_cast_fp16 = add(x = variance_1_cast_fp16, y = var_157_to_fp16)[name = string("op_158_cast_fp16")];
            fp32 var_159_epsilon_0 = const()[name = string("op_159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_159_cast_fp16 = rsqrt(epsilon = var_159_epsilon_0, x = var_158_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_3_cast_fp16 = mul(x = inputs_embeds_cast_fp16, y = var_159_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147751232)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_7_cast_fp16 = mul(x = model_model_layers_0_input_layernorm_weight_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<int32, [3]> var_170_shape_cast_fp16 = shape(x = hidden_states_7_cast_fp16)[name = string("op_170_shape_cast_fp16")];
            int32 gather_4 = const()[name = string("gather_4"), val = int32(1)];
            int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)];
            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
            string var_170_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_170_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_170_shape_cast_fp16_to_uint16 = cast(dtype = var_170_shape_cast_fp16_to_uint16_dtype_0, x = var_170_shape_cast_fp16)[name = string("cast_129")];
            uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_170_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")];
            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147755392))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149852608))))[name = string("model_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150114816)))];
            tensor<fp16, [1, ?, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_7_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150118976))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150643328))))[name = string("model_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150708928)))];
            tensor<fp16, [1, ?, 512]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_7_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150710016))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151234368))))[name = string("model_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_7_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_179_cast_fp16 = reshape(shape = concat_0x, x = linear_0_cast_fp16)[name = string("op_179_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_182_cast_fp16 = reshape(shape = concat_1x, x = linear_1_cast_fp16)[name = string("op_182_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_185_cast_fp16 = reshape(shape = concat_2x, x = linear_2_cast_fp16)[name = string("op_185_cast_fp16")];
            tensor<int32, [4]> v_state_1_perm_0 = const()[name = string("v_state_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_7_axes_0 = const()[name = string("cos_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, ?, 64]> cos_7_cast_fp16 = expand_dims(axes = cos_7_axes_0, x = cos_1_cast_fp16)[name = string("cos_7_cast_fp16")];
            tensor<int32, [1]> sin_7_axes_0 = const()[name = string("sin_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, ?, 64]> sin_7_cast_fp16 = expand_dims(axes = sin_7_axes_0, x = sin_1_cast_fp16)[name = string("sin_7_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_179_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 32, ?, 64]> var_189_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_7_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_200_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_200_cast_fp16")];
            bool var_202_interleave_0 = const()[name = string("op_202_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_202_cast_fp16 = concat(axis = var_48, interleave = var_202_interleave_0, values = (var_200_cast_fp16, x1_1_cast_fp16))[name = string("op_202_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_203_cast_fp16 = mul(x = var_202_cast_fp16, y = sin_7_cast_fp16)[name = string("op_203_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_3_cast_fp16 = add(x = var_189_cast_fp16, y = var_203_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_182_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, ?, 64]> var_205_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_7_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_216_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_216_cast_fp16")];
            bool var_218_interleave_0 = const()[name = string("op_218_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_218_cast_fp16 = concat(axis = var_48, interleave = var_218_interleave_0, values = (var_216_cast_fp16, x1_3_cast_fp16))[name = string("op_218_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_219_cast_fp16 = mul(x = var_218_cast_fp16, y = sin_7_cast_fp16)[name = string("op_219_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_1_cast_fp16 = add(x = var_205_cast_fp16, y = var_219_cast_fp16)[name = string("k_state_1_cast_fp16")];
            tensor<int32, [1]> var_221_shape = shape(x = cache_position)[name = string("op_221_shape")];
            int32 gather_10_axis_0 = const()[name = string("gather_10_axis_0"), val = int32(0)];
            int32 gather_10_batch_dims_0 = const()[name = string("gather_10_batch_dims_0"), val = int32(0)];
            bool gather_10_validate_indices_0 = const()[name = string("gather_10_validate_indices_0"), val = bool(false)];
            string var_221_shape_to_uint16_dtype_0 = const()[name = string("op_221_shape_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_10_to_uint16 = const()[name = string("select_10_to_uint16"), val = uint16(0)];
            tensor<uint16, [1]> var_221_shape_to_uint16 = cast(dtype = var_221_shape_to_uint16_dtype_0, x = var_221_shape)[name = string("cast_128")];
            uint16 gather_10_cast_uint16 = gather(axis = gather_10_axis_0, batch_dims = gather_10_batch_dims_0, indices = select_10_to_uint16, validate_indices = gather_10_validate_indices_0, x = var_221_shape_to_uint16)[name = string("gather_10_cast_uint16")];
            string gather_10_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_10_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_10_cast_uint16_to_int32 = cast(dtype = gather_10_cast_uint16_to_int32_dtype_0, x = gather_10_cast_uint16)[name = string("cast_127")];
            int32 end_1 = add(x = past_seen_tokens, y = gather_10_cast_uint16_to_int32)[name = string("end_1")];
            tensor<fp16, [16, 1, 8, 2048, 64]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_2_axes_0 = const()[name = string("expand_dims_2_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_2 = expand_dims(axes = expand_dims_2_axes_0, x = past_seen_tokens)[name = string("expand_dims_2")];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([8])];
            tensor<int32, [1]> expand_dims_5_axes_0 = const()[name = string("expand_dims_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_5 = expand_dims(axes = expand_dims_5_axes_0, x = end_1)[name = string("expand_dims_5")];
            tensor<int32, [1]> concat_5_values0_0 = const()[name = string("concat_5_values0_0"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (concat_5_values0_0, expand_dims_0, expand_dims_1, expand_dims_2, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> concat_6_values0_0 = const()[name = string("concat_6_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values4_0 = const()[name = string("concat_6_values4_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (concat_6_values0_0, concat_6_values1_0, expand_dims_4, expand_dims_5, concat_6_values4_0))[name = string("concat_6")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = k_state_1_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_32_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_32 = read_state(input = key_cache)[name = string("coreml_update_state_32")];
            tensor<fp16, [16, 1, 8, 2048, 64]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_1_cast_fp16 = transpose(perm = v_state_1_perm_0, x = var_185_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = v_state_1_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_33_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_33 = read_state(input = value_cache)[name = string("coreml_update_state_33")];
            tensor<int32, [5]> var_242_begin_0 = const()[name = string("op_242_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_242_end_0 = const()[name = string("op_242_end_0"), val = tensor<int32, [5]>([1, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_242_end_mask_0 = const()[name = string("op_242_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_242_squeeze_mask_0 = const()[name = string("op_242_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_242_cast_fp16 = slice_by_index(begin = var_242_begin_0, end = var_242_end_0, end_mask = var_242_end_mask_0, squeeze_mask = var_242_squeeze_mask_0, x = coreml_update_state_32)[name = string("op_242_cast_fp16")];
            int32 concat_11_values0_0 = const()[name = string("concat_11_values0_0"), val = int32(1)];
            int32 concat_11_values1_0 = const()[name = string("concat_11_values1_0"), val = int32(8)];
            int32 concat_11_values3_0 = const()[name = string("concat_11_values3_0"), val = int32(64)];
            int32 concat_11_axis_0 = const()[name = string("concat_11_axis_0"), val = int32(0)];
            bool concat_11_interleave_0 = const()[name = string("concat_11_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (concat_11_values0_0, concat_11_values1_0, end_1, concat_11_values3_0))[name = string("concat_11")];
            tensor<int32, [4]> var_245_begin_0 = const()[name = string("op_245_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_245_end_mask_0 = const()[name = string("op_245_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_245_cast_fp16 = slice_by_index(begin = var_245_begin_0, end = concat_11, end_mask = var_245_end_mask_0, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<int32, [5]> var_247_begin_0 = const()[name = string("op_247_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_247_end_0 = const()[name = string("op_247_end_0"), val = tensor<int32, [5]>([1, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_247_end_mask_0 = const()[name = string("op_247_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_247_squeeze_mask_0 = const()[name = string("op_247_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_247_cast_fp16 = slice_by_index(begin = var_247_begin_0, end = var_247_end_0, end_mask = var_247_end_mask_0, squeeze_mask = var_247_squeeze_mask_0, x = coreml_update_state_33)[name = string("op_247_cast_fp16")];
            tensor<int32, [4]> var_250_begin_0 = const()[name = string("op_250_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_250_end_mask_0 = const()[name = string("op_250_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_250_cast_fp16 = slice_by_index(begin = var_250_begin_0, end = concat_11, end_mask = var_250_end_mask_0, x = var_247_cast_fp16)[name = string("op_250_cast_fp16")];
            tensor<int32, [4]> var_252_shape_cast_fp16 = shape(x = var_245_cast_fp16)[name = string("op_252_shape_cast_fp16")];
            int32 gather_13 = const()[name = string("gather_13"), val = int32(1)];
            int32 gather_14 = const()[name = string("gather_14"), val = int32(8)];
            int32 gather_15_axis_0 = const()[name = string("gather_15_axis_0"), val = int32(0)];
            int32 gather_15_batch_dims_0 = const()[name = string("gather_15_batch_dims_0"), val = int32(0)];
            bool gather_15_validate_indices_0 = const()[name = string("gather_15_validate_indices_0"), val = bool(false)];
            string var_252_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_252_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_15_to_uint16 = const()[name = string("select_15_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_252_shape_cast_fp16_to_uint16 = cast(dtype = var_252_shape_cast_fp16_to_uint16_dtype_0, x = var_252_shape_cast_fp16)[name = string("cast_126")];
            uint16 gather_15_cast_uint16 = gather(axis = gather_15_axis_0, batch_dims = gather_15_batch_dims_0, indices = select_15_to_uint16, validate_indices = gather_15_validate_indices_0, x = var_252_shape_cast_fp16_to_uint16)[name = string("gather_15_cast_uint16")];
            string gather_15_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_15_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_16 = const()[name = string("gather_16"), val = int32(64)];
            tensor<int32, [1]> var_259_axes_0 = const()[name = string("op_259_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_259_cast_fp16 = expand_dims(axes = var_259_axes_0, x = var_245_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<int32, [5]> shape_17_cast_fp16 = shape(x = var_259_cast_fp16)[name = string("shape_17_cast_fp16")];
            int32 concat_13_axis_0 = const()[name = string("concat_13_axis_0"), val = int32(0)];
            bool concat_13_interleave_0 = const()[name = string("concat_13_interleave_0"), val = bool(false)];
            int32 gather_15_cast_uint16_to_int32 = cast(dtype = gather_15_cast_uint16_to_int32_dtype_0, x = gather_15_cast_uint16)[name = string("cast_125")];
            tensor<int32, [5]> concat_13 = concat(axis = concat_13_axis_0, interleave = concat_13_interleave_0, values = (gather_13, gather_14, var_59, gather_15_cast_uint16_to_int32, gather_16))[name = string("concat_13")];
            tensor<int32, [5]> real_div_0 = real_div(x = concat_13, y = shape_17_cast_fp16)[name = string("real_div_0")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_11_cast_fp16 = tile(reps = real_div_0, x = var_259_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_3_cast_fp16 = reshape(shape = concat_14x, x = hidden_states_11_cast_fp16)[name = string("key_states_3_cast_fp16")];
            tensor<int32, [4]> var_269_shape_cast_fp16 = shape(x = var_250_cast_fp16)[name = string("op_269_shape_cast_fp16")];
            int32 gather_17 = const()[name = string("gather_17"), val = int32(1)];
            int32 gather_18 = const()[name = string("gather_18"), val = int32(8)];
            int32 gather_19_axis_0 = const()[name = string("gather_19_axis_0"), val = int32(0)];
            int32 gather_19_batch_dims_0 = const()[name = string("gather_19_batch_dims_0"), val = int32(0)];
            bool gather_19_validate_indices_0 = const()[name = string("gather_19_validate_indices_0"), val = bool(false)];
            string var_269_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_269_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_19_to_uint16 = const()[name = string("select_19_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_269_shape_cast_fp16_to_uint16 = cast(dtype = var_269_shape_cast_fp16_to_uint16_dtype_0, x = var_269_shape_cast_fp16)[name = string("cast_124")];
            uint16 gather_19_cast_uint16 = gather(axis = gather_19_axis_0, batch_dims = gather_19_batch_dims_0, indices = select_19_to_uint16, validate_indices = gather_19_validate_indices_0, x = var_269_shape_cast_fp16_to_uint16)[name = string("gather_19_cast_uint16")];
            string gather_19_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_19_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_20 = const()[name = string("gather_20"), val = int32(64)];
            tensor<int32, [1]> var_276_axes_0 = const()[name = string("op_276_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_276_cast_fp16 = expand_dims(axes = var_276_axes_0, x = var_250_cast_fp16)[name = string("op_276_cast_fp16")];
            tensor<int32, [5]> shape_22_cast_fp16 = shape(x = var_276_cast_fp16)[name = string("shape_22_cast_fp16")];
            int32 concat_15_axis_0 = const()[name = string("concat_15_axis_0"), val = int32(0)];
            bool concat_15_interleave_0 = const()[name = string("concat_15_interleave_0"), val = bool(false)];
            int32 gather_19_cast_uint16_to_int32 = cast(dtype = gather_19_cast_uint16_to_int32_dtype_0, x = gather_19_cast_uint16)[name = string("cast_123")];
            tensor<int32, [5]> concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (gather_17, gather_18, var_59, gather_19_cast_uint16_to_int32, gather_20))[name = string("concat_15")];
            tensor<int32, [5]> real_div_1 = real_div(x = concat_15, y = shape_22_cast_fp16)[name = string("real_div_1")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_15_cast_fp16 = tile(reps = real_div_1, x = var_276_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<int32, [4]> concat_16x = const()[name = string("concat_16x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_3_cast_fp16 = reshape(shape = concat_16x, x = hidden_states_15_cast_fp16)[name = string("value_states_3_cast_fp16")];
            tensor<int32, [4]> var_286_shape_cast_fp16 = shape(x = key_states_3_cast_fp16)[name = string("op_286_shape_cast_fp16")];
            int32 gather_21_axis_0 = const()[name = string("gather_21_axis_0"), val = int32(0)];
            int32 gather_21_batch_dims_0 = const()[name = string("gather_21_batch_dims_0"), val = int32(0)];
            bool gather_21_validate_indices_0 = const()[name = string("gather_21_validate_indices_0"), val = bool(false)];
            string var_286_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_286_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_21_to_uint16 = const()[name = string("select_21_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_286_shape_cast_fp16_to_uint16 = cast(dtype = var_286_shape_cast_fp16_to_uint16_dtype_0, x = var_286_shape_cast_fp16)[name = string("cast_122")];
            uint16 gather_21_cast_uint16 = gather(axis = gather_21_axis_0, batch_dims = gather_21_batch_dims_0, indices = select_21_to_uint16, validate_indices = gather_21_validate_indices_0, x = var_286_shape_cast_fp16_to_uint16)[name = string("gather_21_cast_uint16")];
            string gather_21_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_21_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_17_values0_0 = const()[name = string("concat_17_values0_0"), val = int32(1)];
            int32 concat_17_values1_0 = const()[name = string("concat_17_values1_0"), val = int32(1)];
            int32 concat_17_values2_0 = const()[name = string("concat_17_values2_0"), val = int32(0)];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            int32 gather_21_cast_uint16_to_int32 = cast(dtype = gather_21_cast_uint16_to_int32_dtype_0, x = gather_21_cast_uint16)[name = string("cast_121")];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (concat_17_values0_0, concat_17_values1_0, concat_17_values2_0, gather_21_cast_uint16_to_int32))[name = string("concat_17")];
            tensor<int32, [4]> causal_mask_3_begin_0 = const()[name = string("causal_mask_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_3_end_mask_0 = const()[name = string("causal_mask_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_3_cast_fp16 = slice_by_index(begin = causal_mask_3_begin_0, end = concat_17, end_mask = causal_mask_3_end_mask_0, x = causal_mask)[name = string("causal_mask_3_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_1_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_3_cast_fp16, key = key_states_3_cast_fp16, query = query_states_3_cast_fp16, value = value_states_3_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_292_perm_0 = const()[name = string("op_292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_18_axis_0 = const()[name = string("concat_18_axis_0"), val = int32(0)];
            bool concat_18_interleave_0 = const()[name = string("concat_18_interleave_0"), val = bool(false)];
            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_120")];
            tensor<int32, [3]> concat_18 = concat(axis = concat_18_axis_0, interleave = concat_18_interleave_0, values = (gather_4, gather_5_cast_uint16_to_int32, var_48))[name = string("concat_18")];
            tensor<fp16, [1, ?, 32, 64]> var_292_cast_fp16 = transpose(perm = var_292_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, ?, ?]> input_1_cast_fp16 = reshape(shape = concat_18, x = var_292_cast_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151299968))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153397184))))[name = string("model_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_19_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 var_54_promoted_1_to_fp16 = const()[name = string("op_54_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_301_cast_fp16 = pow(x = hidden_states_19_cast_fp16, y = var_54_promoted_1_to_fp16)[name = string("op_301_cast_fp16")];
            tensor<int32, [1]> variance_3_axes_0 = const()[name = string("variance_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_3_keep_dims_0 = const()[name = string("variance_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_3_cast_fp16 = reduce_mean(axes = variance_3_axes_0, keep_dims = variance_3_keep_dims_0, x = var_301_cast_fp16)[name = string("variance_3_cast_fp16")];
            fp16 var_304_to_fp16 = const()[name = string("op_304_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_305_cast_fp16 = add(x = variance_3_cast_fp16, y = var_304_to_fp16)[name = string("op_305_cast_fp16")];
            fp32 var_306_epsilon_0 = const()[name = string("op_306_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_306_cast_fp16 = rsqrt(epsilon = var_306_epsilon_0, x = var_305_cast_fp16)[name = string("op_306_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_23_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = var_306_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153659392)))];
            tensor<fp16, [1, ?, 2048]> input_3_cast_fp16 = mul(x = model_model_layers_0_post_attention_layernorm_weight_to_fp16, y = hidden_states_23_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153663552))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162052224))))[name = string("model_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [8192]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163100864)))];
            tensor<fp16, [1, ?, 8192]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_318_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_318_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163117312))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171505984))))[name = string("model_model_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_7_cast_fp16 = mul(x = var_318_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172554624))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180943296))))[name = string("model_model_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_29_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = linear_6_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 var_54_promoted_2_to_fp16 = const()[name = string("op_54_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_331_cast_fp16 = pow(x = hidden_states_29_cast_fp16, y = var_54_promoted_2_to_fp16)[name = string("op_331_cast_fp16")];
            tensor<int32, [1]> variance_5_axes_0 = const()[name = string("variance_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_5_keep_dims_0 = const()[name = string("variance_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_5_cast_fp16 = reduce_mean(axes = variance_5_axes_0, keep_dims = variance_5_keep_dims_0, x = var_331_cast_fp16)[name = string("variance_5_cast_fp16")];
            fp16 var_334_to_fp16 = const()[name = string("op_334_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_335_cast_fp16 = add(x = variance_5_cast_fp16, y = var_334_to_fp16)[name = string("op_335_cast_fp16")];
            fp32 var_336_epsilon_0 = const()[name = string("op_336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_336_cast_fp16 = rsqrt(epsilon = var_336_epsilon_0, x = var_335_cast_fp16)[name = string("op_336_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_33_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = var_336_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181991936)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_37_cast_fp16 = mul(x = model_model_layers_1_input_layernorm_weight_to_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [3]> var_347_shape_cast_fp16 = shape(x = hidden_states_37_cast_fp16)[name = string("op_347_shape_cast_fp16")];
            int32 gather_22 = const()[name = string("gather_22"), val = int32(1)];
            int32 gather_23_axis_0 = const()[name = string("gather_23_axis_0"), val = int32(0)];
            int32 gather_23_batch_dims_0 = const()[name = string("gather_23_batch_dims_0"), val = int32(0)];
            bool gather_23_validate_indices_0 = const()[name = string("gather_23_validate_indices_0"), val = bool(false)];
            string var_347_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_347_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_23_to_uint16 = const()[name = string("select_23_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_347_shape_cast_fp16_to_uint16 = cast(dtype = var_347_shape_cast_fp16_to_uint16_dtype_0, x = var_347_shape_cast_fp16)[name = string("cast_119")];
            uint16 gather_23_cast_uint16 = gather(axis = gather_23_axis_0, batch_dims = gather_23_batch_dims_0, indices = select_23_to_uint16, validate_indices = gather_23_validate_indices_0, x = var_347_shape_cast_fp16_to_uint16)[name = string("gather_23_cast_uint16")];
            string gather_23_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_23_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181996096))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184093312))))[name = string("model_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_37_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184355520))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184879872))))[name = string("model_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_37_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184945472))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185469824))))[name = string("model_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_37_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> concat_19x = const()[name = string("concat_19x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_356_cast_fp16 = reshape(shape = concat_19x, x = linear_7_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_20x = const()[name = string("concat_20x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_359_cast_fp16 = reshape(shape = concat_20x, x = linear_8_cast_fp16)[name = string("op_359_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_21x = const()[name = string("concat_21x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_362_cast_fp16 = reshape(shape = concat_21x, x = linear_9_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<int32, [4]> v_state_3_perm_0 = const()[name = string("v_state_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_356_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 32, ?, 64]> var_366_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_7_cast_fp16)[name = string("op_366_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_377_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_377_cast_fp16")];
            bool var_379_interleave_0 = const()[name = string("op_379_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_379_cast_fp16 = concat(axis = var_48, interleave = var_379_interleave_0, values = (var_377_cast_fp16, x1_5_cast_fp16))[name = string("op_379_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_380_cast_fp16 = mul(x = var_379_cast_fp16, y = sin_7_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_7_cast_fp16 = add(x = var_366_cast_fp16, y = var_380_cast_fp16)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_359_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, ?, 64]> var_382_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_7_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_393_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_393_cast_fp16")];
            bool var_395_interleave_0 = const()[name = string("op_395_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_395_cast_fp16 = concat(axis = var_48, interleave = var_395_interleave_0, values = (var_393_cast_fp16, x1_7_cast_fp16))[name = string("op_395_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_396_cast_fp16 = mul(x = var_395_cast_fp16, y = sin_7_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_3_cast_fp16 = add(x = var_382_cast_fp16, y = var_396_cast_fp16)[name = string("k_state_3_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_24_values0_0 = const()[name = string("concat_24_values0_0"), val = tensor<int32, [1]>([1])];
            int32 concat_24_axis_0 = const()[name = string("concat_24_axis_0"), val = int32(0)];
            bool concat_24_interleave_0 = const()[name = string("concat_24_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_24 = concat(axis = concat_24_axis_0, interleave = concat_24_interleave_0, values = (concat_24_values0_0, expand_dims_12, expand_dims_13, expand_dims_2, expand_dims_15))[name = string("concat_24")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_24, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = k_state_3_cast_fp16, x = coreml_update_state_32)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_34_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_34 = read_state(input = key_cache)[name = string("coreml_update_state_34")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_3_cast_fp16 = transpose(perm = v_state_3_perm_0, x = var_362_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_24, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = v_state_3_cast_fp16, x = coreml_update_state_33)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_35_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_35 = read_state(input = value_cache)[name = string("coreml_update_state_35")];
            tensor<int32, [5]> var_419_begin_0 = const()[name = string("op_419_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_419_end_0 = const()[name = string("op_419_end_0"), val = tensor<int32, [5]>([2, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_419_end_mask_0 = const()[name = string("op_419_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_419_squeeze_mask_0 = const()[name = string("op_419_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_419_cast_fp16 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, squeeze_mask = var_419_squeeze_mask_0, x = coreml_update_state_34)[name = string("op_419_cast_fp16")];
            tensor<int32, [4]> var_422_begin_0 = const()[name = string("op_422_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_422_end_mask_0 = const()[name = string("op_422_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_422_cast_fp16 = slice_by_index(begin = var_422_begin_0, end = concat_11, end_mask = var_422_end_mask_0, x = var_419_cast_fp16)[name = string("op_422_cast_fp16")];
            tensor<int32, [5]> var_424_begin_0 = const()[name = string("op_424_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_424_end_0 = const()[name = string("op_424_end_0"), val = tensor<int32, [5]>([2, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_424_end_mask_0 = const()[name = string("op_424_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_424_squeeze_mask_0 = const()[name = string("op_424_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_424_cast_fp16 = slice_by_index(begin = var_424_begin_0, end = var_424_end_0, end_mask = var_424_end_mask_0, squeeze_mask = var_424_squeeze_mask_0, x = coreml_update_state_35)[name = string("op_424_cast_fp16")];
            tensor<int32, [4]> var_427_begin_0 = const()[name = string("op_427_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_427_end_mask_0 = const()[name = string("op_427_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_427_cast_fp16 = slice_by_index(begin = var_427_begin_0, end = concat_11, end_mask = var_427_end_mask_0, x = var_424_cast_fp16)[name = string("op_427_cast_fp16")];
            tensor<int32, [4]> var_429_shape_cast_fp16 = shape(x = var_422_cast_fp16)[name = string("op_429_shape_cast_fp16")];
            int32 gather_31 = const()[name = string("gather_31"), val = int32(1)];
            int32 gather_32 = const()[name = string("gather_32"), val = int32(8)];
            int32 gather_33_axis_0 = const()[name = string("gather_33_axis_0"), val = int32(0)];
            int32 gather_33_batch_dims_0 = const()[name = string("gather_33_batch_dims_0"), val = int32(0)];
            bool gather_33_validate_indices_0 = const()[name = string("gather_33_validate_indices_0"), val = bool(false)];
            string var_429_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_429_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_33_to_uint16 = const()[name = string("select_33_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_429_shape_cast_fp16_to_uint16 = cast(dtype = var_429_shape_cast_fp16_to_uint16_dtype_0, x = var_429_shape_cast_fp16)[name = string("cast_118")];
            uint16 gather_33_cast_uint16 = gather(axis = gather_33_axis_0, batch_dims = gather_33_batch_dims_0, indices = select_33_to_uint16, validate_indices = gather_33_validate_indices_0, x = var_429_shape_cast_fp16_to_uint16)[name = string("gather_33_cast_uint16")];
            string gather_33_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_33_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_34 = const()[name = string("gather_34"), val = int32(64)];
            tensor<int32, [1]> var_436_axes_0 = const()[name = string("op_436_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_436_cast_fp16 = expand_dims(axes = var_436_axes_0, x = var_422_cast_fp16)[name = string("op_436_cast_fp16")];
            tensor<int32, [5]> shape_37_cast_fp16 = shape(x = var_436_cast_fp16)[name = string("shape_37_cast_fp16")];
            int32 concat_32_axis_0 = const()[name = string("concat_32_axis_0"), val = int32(0)];
            bool concat_32_interleave_0 = const()[name = string("concat_32_interleave_0"), val = bool(false)];
            int32 gather_33_cast_uint16_to_int32 = cast(dtype = gather_33_cast_uint16_to_int32_dtype_0, x = gather_33_cast_uint16)[name = string("cast_117")];
            tensor<int32, [5]> concat_32 = concat(axis = concat_32_axis_0, interleave = concat_32_interleave_0, values = (gather_31, gather_32, var_59, gather_33_cast_uint16_to_int32, gather_34))[name = string("concat_32")];
            tensor<int32, [5]> real_div_2 = real_div(x = concat_32, y = shape_37_cast_fp16)[name = string("real_div_2")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_41_cast_fp16 = tile(reps = real_div_2, x = var_436_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
            tensor<int32, [4]> concat_33x = const()[name = string("concat_33x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_7_cast_fp16 = reshape(shape = concat_33x, x = hidden_states_41_cast_fp16)[name = string("key_states_7_cast_fp16")];
            tensor<int32, [4]> var_446_shape_cast_fp16 = shape(x = var_427_cast_fp16)[name = string("op_446_shape_cast_fp16")];
            int32 gather_35 = const()[name = string("gather_35"), val = int32(1)];
            int32 gather_36 = const()[name = string("gather_36"), val = int32(8)];
            int32 gather_37_axis_0 = const()[name = string("gather_37_axis_0"), val = int32(0)];
            int32 gather_37_batch_dims_0 = const()[name = string("gather_37_batch_dims_0"), val = int32(0)];
            bool gather_37_validate_indices_0 = const()[name = string("gather_37_validate_indices_0"), val = bool(false)];
            string var_446_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_446_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_37_to_uint16 = const()[name = string("select_37_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_446_shape_cast_fp16_to_uint16 = cast(dtype = var_446_shape_cast_fp16_to_uint16_dtype_0, x = var_446_shape_cast_fp16)[name = string("cast_116")];
            uint16 gather_37_cast_uint16 = gather(axis = gather_37_axis_0, batch_dims = gather_37_batch_dims_0, indices = select_37_to_uint16, validate_indices = gather_37_validate_indices_0, x = var_446_shape_cast_fp16_to_uint16)[name = string("gather_37_cast_uint16")];
            string gather_37_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_37_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_38 = const()[name = string("gather_38"), val = int32(64)];
            tensor<int32, [1]> var_453_axes_0 = const()[name = string("op_453_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_453_cast_fp16 = expand_dims(axes = var_453_axes_0, x = var_427_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<int32, [5]> shape_42_cast_fp16 = shape(x = var_453_cast_fp16)[name = string("shape_42_cast_fp16")];
            int32 concat_34_axis_0 = const()[name = string("concat_34_axis_0"), val = int32(0)];
            bool concat_34_interleave_0 = const()[name = string("concat_34_interleave_0"), val = bool(false)];
            int32 gather_37_cast_uint16_to_int32 = cast(dtype = gather_37_cast_uint16_to_int32_dtype_0, x = gather_37_cast_uint16)[name = string("cast_115")];
            tensor<int32, [5]> concat_34 = concat(axis = concat_34_axis_0, interleave = concat_34_interleave_0, values = (gather_35, gather_36, var_59, gather_37_cast_uint16_to_int32, gather_38))[name = string("concat_34")];
            tensor<int32, [5]> real_div_3 = real_div(x = concat_34, y = shape_42_cast_fp16)[name = string("real_div_3")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_45_cast_fp16 = tile(reps = real_div_3, x = var_453_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_7_cast_fp16 = reshape(shape = concat_35x, x = hidden_states_45_cast_fp16)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> var_463_shape_cast_fp16 = shape(x = key_states_7_cast_fp16)[name = string("op_463_shape_cast_fp16")];
            int32 gather_39_axis_0 = const()[name = string("gather_39_axis_0"), val = int32(0)];
            int32 gather_39_batch_dims_0 = const()[name = string("gather_39_batch_dims_0"), val = int32(0)];
            bool gather_39_validate_indices_0 = const()[name = string("gather_39_validate_indices_0"), val = bool(false)];
            string var_463_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_463_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_39_to_uint16 = const()[name = string("select_39_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_463_shape_cast_fp16_to_uint16 = cast(dtype = var_463_shape_cast_fp16_to_uint16_dtype_0, x = var_463_shape_cast_fp16)[name = string("cast_114")];
            uint16 gather_39_cast_uint16 = gather(axis = gather_39_axis_0, batch_dims = gather_39_batch_dims_0, indices = select_39_to_uint16, validate_indices = gather_39_validate_indices_0, x = var_463_shape_cast_fp16_to_uint16)[name = string("gather_39_cast_uint16")];
            string gather_39_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_39_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_36_values0_0 = const()[name = string("concat_36_values0_0"), val = int32(1)];
            int32 concat_36_values1_0 = const()[name = string("concat_36_values1_0"), val = int32(1)];
            int32 concat_36_values2_0 = const()[name = string("concat_36_values2_0"), val = int32(0)];
            int32 concat_36_axis_0 = const()[name = string("concat_36_axis_0"), val = int32(0)];
            bool concat_36_interleave_0 = const()[name = string("concat_36_interleave_0"), val = bool(false)];
            int32 gather_39_cast_uint16_to_int32 = cast(dtype = gather_39_cast_uint16_to_int32_dtype_0, x = gather_39_cast_uint16)[name = string("cast_113")];
            tensor<int32, [4]> concat_36 = concat(axis = concat_36_axis_0, interleave = concat_36_interleave_0, values = (concat_36_values0_0, concat_36_values1_0, concat_36_values2_0, gather_39_cast_uint16_to_int32))[name = string("concat_36")];
            tensor<int32, [4]> causal_mask_5_begin_0 = const()[name = string("causal_mask_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_5_end_mask_0 = const()[name = string("causal_mask_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_5_cast_fp16 = slice_by_index(begin = causal_mask_5_begin_0, end = concat_36, end_mask = causal_mask_5_end_mask_0, x = causal_mask)[name = string("causal_mask_5_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_5_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_5_cast_fp16, key = key_states_7_cast_fp16, query = query_states_7_cast_fp16, value = value_states_7_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_469_perm_0 = const()[name = string("op_469_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_37_axis_0 = const()[name = string("concat_37_axis_0"), val = int32(0)];
            bool concat_37_interleave_0 = const()[name = string("concat_37_interleave_0"), val = bool(false)];
            int32 gather_23_cast_uint16_to_int32 = cast(dtype = gather_23_cast_uint16_to_int32_dtype_0, x = gather_23_cast_uint16)[name = string("cast_112")];
            tensor<int32, [3]> concat_37 = concat(axis = concat_37_axis_0, interleave = concat_37_interleave_0, values = (gather_22, gather_23_cast_uint16_to_int32, var_48))[name = string("concat_37")];
            tensor<fp16, [1, ?, 32, 64]> var_469_cast_fp16 = transpose(perm = var_469_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, ?, ?]> input_9_cast_fp16 = reshape(shape = concat_37, x = var_469_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185535424))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187632640))))[name = string("model_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_49_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = linear_10_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 var_54_promoted_3_to_fp16 = const()[name = string("op_54_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_478_cast_fp16 = pow(x = hidden_states_49_cast_fp16, y = var_54_promoted_3_to_fp16)[name = string("op_478_cast_fp16")];
            tensor<int32, [1]> variance_7_axes_0 = const()[name = string("variance_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_7_keep_dims_0 = const()[name = string("variance_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_7_cast_fp16 = reduce_mean(axes = variance_7_axes_0, keep_dims = variance_7_keep_dims_0, x = var_478_cast_fp16)[name = string("variance_7_cast_fp16")];
            fp16 var_481_to_fp16 = const()[name = string("op_481_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_482_cast_fp16 = add(x = variance_7_cast_fp16, y = var_481_to_fp16)[name = string("op_482_cast_fp16")];
            fp32 var_483_epsilon_0 = const()[name = string("op_483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_483_cast_fp16 = rsqrt(epsilon = var_483_epsilon_0, x = var_482_cast_fp16)[name = string("op_483_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_53_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = var_483_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187894848)))];
            tensor<fp16, [1, ?, 2048]> input_11_cast_fp16 = mul(x = model_model_layers_1_post_attention_layernorm_weight_to_fp16, y = hidden_states_53_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187899008))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196287680))))[name = string("model_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_495_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_495_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197336320))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205724992))))[name = string("model_model_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_15_cast_fp16 = mul(x = var_495_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206773632))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215162304))))[name = string("model_model_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_13_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_59_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = linear_13_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 var_54_promoted_4_to_fp16 = const()[name = string("op_54_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_508_cast_fp16 = pow(x = hidden_states_59_cast_fp16, y = var_54_promoted_4_to_fp16)[name = string("op_508_cast_fp16")];
            tensor<int32, [1]> variance_9_axes_0 = const()[name = string("variance_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_9_keep_dims_0 = const()[name = string("variance_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_9_cast_fp16 = reduce_mean(axes = variance_9_axes_0, keep_dims = variance_9_keep_dims_0, x = var_508_cast_fp16)[name = string("variance_9_cast_fp16")];
            fp16 var_511_to_fp16 = const()[name = string("op_511_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_512_cast_fp16 = add(x = variance_9_cast_fp16, y = var_511_to_fp16)[name = string("op_512_cast_fp16")];
            fp32 var_513_epsilon_0 = const()[name = string("op_513_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_513_cast_fp16 = rsqrt(epsilon = var_513_epsilon_0, x = var_512_cast_fp16)[name = string("op_513_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_63_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = var_513_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216210944)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_67_cast_fp16 = mul(x = model_model_layers_2_input_layernorm_weight_to_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<int32, [3]> var_524_shape_cast_fp16 = shape(x = hidden_states_67_cast_fp16)[name = string("op_524_shape_cast_fp16")];
            int32 gather_40 = const()[name = string("gather_40"), val = int32(1)];
            int32 gather_41_axis_0 = const()[name = string("gather_41_axis_0"), val = int32(0)];
            int32 gather_41_batch_dims_0 = const()[name = string("gather_41_batch_dims_0"), val = int32(0)];
            bool gather_41_validate_indices_0 = const()[name = string("gather_41_validate_indices_0"), val = bool(false)];
            string var_524_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_524_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_41_to_uint16 = const()[name = string("select_41_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_524_shape_cast_fp16_to_uint16 = cast(dtype = var_524_shape_cast_fp16_to_uint16_dtype_0, x = var_524_shape_cast_fp16)[name = string("cast_111")];
            uint16 gather_41_cast_uint16 = gather(axis = gather_41_axis_0, batch_dims = gather_41_batch_dims_0, indices = select_41_to_uint16, validate_indices = gather_41_validate_indices_0, x = var_524_shape_cast_fp16_to_uint16)[name = string("gather_41_cast_uint16")];
            string gather_41_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_41_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216215104))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218312320))))[name = string("model_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_67_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218574528))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219098880))))[name = string("model_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_67_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219164480))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219688832))))[name = string("model_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_67_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_533_cast_fp16 = reshape(shape = concat_38x, x = linear_14_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_39x = const()[name = string("concat_39x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_536_cast_fp16 = reshape(shape = concat_39x, x = linear_15_cast_fp16)[name = string("op_536_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_40x = const()[name = string("concat_40x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_539_cast_fp16 = reshape(shape = concat_40x, x = linear_16_cast_fp16)[name = string("op_539_cast_fp16")];
            tensor<int32, [4]> v_state_5_perm_0 = const()[name = string("v_state_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_533_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 32, ?, 64]> var_543_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_7_cast_fp16)[name = string("op_543_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_554_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_554_cast_fp16")];
            bool var_556_interleave_0 = const()[name = string("op_556_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_556_cast_fp16 = concat(axis = var_48, interleave = var_556_interleave_0, values = (var_554_cast_fp16, x1_9_cast_fp16))[name = string("op_556_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_557_cast_fp16 = mul(x = var_556_cast_fp16, y = sin_7_cast_fp16)[name = string("op_557_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_11_cast_fp16 = add(x = var_543_cast_fp16, y = var_557_cast_fp16)[name = string("query_states_11_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_536_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 8, ?, 64]> var_559_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_7_cast_fp16)[name = string("op_559_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_570_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_570_cast_fp16")];
            bool var_572_interleave_0 = const()[name = string("op_572_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_572_cast_fp16 = concat(axis = var_48, interleave = var_572_interleave_0, values = (var_570_cast_fp16, x1_11_cast_fp16))[name = string("op_572_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_573_cast_fp16 = mul(x = var_572_cast_fp16, y = sin_7_cast_fp16)[name = string("op_573_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_5_cast_fp16 = add(x = var_559_cast_fp16, y = var_573_cast_fp16)[name = string("k_state_5_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_43_values0_0 = const()[name = string("concat_43_values0_0"), val = tensor<int32, [1]>([2])];
            int32 concat_43_axis_0 = const()[name = string("concat_43_axis_0"), val = int32(0)];
            bool concat_43_interleave_0 = const()[name = string("concat_43_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_43 = concat(axis = concat_43_axis_0, interleave = concat_43_interleave_0, values = (concat_43_values0_0, expand_dims_24, expand_dims_25, expand_dims_2, expand_dims_27))[name = string("concat_43")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_43, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = k_state_5_cast_fp16, x = coreml_update_state_34)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_36_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_36 = read_state(input = key_cache)[name = string("coreml_update_state_36")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_5_cast_fp16 = transpose(perm = v_state_5_perm_0, x = var_539_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_43, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = v_state_5_cast_fp16, x = coreml_update_state_35)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_37_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_37 = read_state(input = value_cache)[name = string("coreml_update_state_37")];
            tensor<int32, [5]> var_596_begin_0 = const()[name = string("op_596_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> var_596_end_0 = const()[name = string("op_596_end_0"), val = tensor<int32, [5]>([3, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_596_end_mask_0 = const()[name = string("op_596_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_596_squeeze_mask_0 = const()[name = string("op_596_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_596_cast_fp16 = slice_by_index(begin = var_596_begin_0, end = var_596_end_0, end_mask = var_596_end_mask_0, squeeze_mask = var_596_squeeze_mask_0, x = coreml_update_state_36)[name = string("op_596_cast_fp16")];
            tensor<int32, [4]> var_599_begin_0 = const()[name = string("op_599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_599_end_mask_0 = const()[name = string("op_599_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_599_cast_fp16 = slice_by_index(begin = var_599_begin_0, end = concat_11, end_mask = var_599_end_mask_0, x = var_596_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<int32, [5]> var_601_begin_0 = const()[name = string("op_601_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> var_601_end_0 = const()[name = string("op_601_end_0"), val = tensor<int32, [5]>([3, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_601_end_mask_0 = const()[name = string("op_601_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_601_squeeze_mask_0 = const()[name = string("op_601_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_601_cast_fp16 = slice_by_index(begin = var_601_begin_0, end = var_601_end_0, end_mask = var_601_end_mask_0, squeeze_mask = var_601_squeeze_mask_0, x = coreml_update_state_37)[name = string("op_601_cast_fp16")];
            tensor<int32, [4]> var_604_begin_0 = const()[name = string("op_604_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_604_end_mask_0 = const()[name = string("op_604_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_604_cast_fp16 = slice_by_index(begin = var_604_begin_0, end = concat_11, end_mask = var_604_end_mask_0, x = var_601_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [4]> var_606_shape_cast_fp16 = shape(x = var_599_cast_fp16)[name = string("op_606_shape_cast_fp16")];
            int32 gather_49 = const()[name = string("gather_49"), val = int32(1)];
            int32 gather_50 = const()[name = string("gather_50"), val = int32(8)];
            int32 gather_51_axis_0 = const()[name = string("gather_51_axis_0"), val = int32(0)];
            int32 gather_51_batch_dims_0 = const()[name = string("gather_51_batch_dims_0"), val = int32(0)];
            bool gather_51_validate_indices_0 = const()[name = string("gather_51_validate_indices_0"), val = bool(false)];
            string var_606_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_606_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_51_to_uint16 = const()[name = string("select_51_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_606_shape_cast_fp16_to_uint16 = cast(dtype = var_606_shape_cast_fp16_to_uint16_dtype_0, x = var_606_shape_cast_fp16)[name = string("cast_110")];
            uint16 gather_51_cast_uint16 = gather(axis = gather_51_axis_0, batch_dims = gather_51_batch_dims_0, indices = select_51_to_uint16, validate_indices = gather_51_validate_indices_0, x = var_606_shape_cast_fp16_to_uint16)[name = string("gather_51_cast_uint16")];
            string gather_51_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_51_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_52 = const()[name = string("gather_52"), val = int32(64)];
            tensor<int32, [1]> var_613_axes_0 = const()[name = string("op_613_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_613_cast_fp16 = expand_dims(axes = var_613_axes_0, x = var_599_cast_fp16)[name = string("op_613_cast_fp16")];
            tensor<int32, [5]> shape_57_cast_fp16 = shape(x = var_613_cast_fp16)[name = string("shape_57_cast_fp16")];
            int32 concat_51_axis_0 = const()[name = string("concat_51_axis_0"), val = int32(0)];
            bool concat_51_interleave_0 = const()[name = string("concat_51_interleave_0"), val = bool(false)];
            int32 gather_51_cast_uint16_to_int32 = cast(dtype = gather_51_cast_uint16_to_int32_dtype_0, x = gather_51_cast_uint16)[name = string("cast_109")];
            tensor<int32, [5]> concat_51 = concat(axis = concat_51_axis_0, interleave = concat_51_interleave_0, values = (gather_49, gather_50, var_59, gather_51_cast_uint16_to_int32, gather_52))[name = string("concat_51")];
            tensor<int32, [5]> real_div_4 = real_div(x = concat_51, y = shape_57_cast_fp16)[name = string("real_div_4")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_71_cast_fp16 = tile(reps = real_div_4, x = var_613_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<int32, [4]> concat_52x = const()[name = string("concat_52x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_11_cast_fp16 = reshape(shape = concat_52x, x = hidden_states_71_cast_fp16)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [4]> var_623_shape_cast_fp16 = shape(x = var_604_cast_fp16)[name = string("op_623_shape_cast_fp16")];
            int32 gather_53 = const()[name = string("gather_53"), val = int32(1)];
            int32 gather_54 = const()[name = string("gather_54"), val = int32(8)];
            int32 gather_55_axis_0 = const()[name = string("gather_55_axis_0"), val = int32(0)];
            int32 gather_55_batch_dims_0 = const()[name = string("gather_55_batch_dims_0"), val = int32(0)];
            bool gather_55_validate_indices_0 = const()[name = string("gather_55_validate_indices_0"), val = bool(false)];
            string var_623_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_623_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_55_to_uint16 = const()[name = string("select_55_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_623_shape_cast_fp16_to_uint16 = cast(dtype = var_623_shape_cast_fp16_to_uint16_dtype_0, x = var_623_shape_cast_fp16)[name = string("cast_108")];
            uint16 gather_55_cast_uint16 = gather(axis = gather_55_axis_0, batch_dims = gather_55_batch_dims_0, indices = select_55_to_uint16, validate_indices = gather_55_validate_indices_0, x = var_623_shape_cast_fp16_to_uint16)[name = string("gather_55_cast_uint16")];
            string gather_55_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_55_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_56 = const()[name = string("gather_56"), val = int32(64)];
            tensor<int32, [1]> var_630_axes_0 = const()[name = string("op_630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_630_cast_fp16 = expand_dims(axes = var_630_axes_0, x = var_604_cast_fp16)[name = string("op_630_cast_fp16")];
            tensor<int32, [5]> shape_62_cast_fp16 = shape(x = var_630_cast_fp16)[name = string("shape_62_cast_fp16")];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            int32 gather_55_cast_uint16_to_int32 = cast(dtype = gather_55_cast_uint16_to_int32_dtype_0, x = gather_55_cast_uint16)[name = string("cast_107")];
            tensor<int32, [5]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (gather_53, gather_54, var_59, gather_55_cast_uint16_to_int32, gather_56))[name = string("concat_53")];
            tensor<int32, [5]> real_div_5 = real_div(x = concat_53, y = shape_62_cast_fp16)[name = string("real_div_5")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_75_cast_fp16 = tile(reps = real_div_5, x = var_630_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> concat_54x = const()[name = string("concat_54x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_11_cast_fp16 = reshape(shape = concat_54x, x = hidden_states_75_cast_fp16)[name = string("value_states_11_cast_fp16")];
            tensor<int32, [4]> var_640_shape_cast_fp16 = shape(x = key_states_11_cast_fp16)[name = string("op_640_shape_cast_fp16")];
            int32 gather_57_axis_0 = const()[name = string("gather_57_axis_0"), val = int32(0)];
            int32 gather_57_batch_dims_0 = const()[name = string("gather_57_batch_dims_0"), val = int32(0)];
            bool gather_57_validate_indices_0 = const()[name = string("gather_57_validate_indices_0"), val = bool(false)];
            string var_640_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_640_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_57_to_uint16 = const()[name = string("select_57_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_640_shape_cast_fp16_to_uint16 = cast(dtype = var_640_shape_cast_fp16_to_uint16_dtype_0, x = var_640_shape_cast_fp16)[name = string("cast_106")];
            uint16 gather_57_cast_uint16 = gather(axis = gather_57_axis_0, batch_dims = gather_57_batch_dims_0, indices = select_57_to_uint16, validate_indices = gather_57_validate_indices_0, x = var_640_shape_cast_fp16_to_uint16)[name = string("gather_57_cast_uint16")];
            string gather_57_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_57_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_55_values0_0 = const()[name = string("concat_55_values0_0"), val = int32(1)];
            int32 concat_55_values1_0 = const()[name = string("concat_55_values1_0"), val = int32(1)];
            int32 concat_55_values2_0 = const()[name = string("concat_55_values2_0"), val = int32(0)];
            int32 concat_55_axis_0 = const()[name = string("concat_55_axis_0"), val = int32(0)];
            bool concat_55_interleave_0 = const()[name = string("concat_55_interleave_0"), val = bool(false)];
            int32 gather_57_cast_uint16_to_int32 = cast(dtype = gather_57_cast_uint16_to_int32_dtype_0, x = gather_57_cast_uint16)[name = string("cast_105")];
            tensor<int32, [4]> concat_55 = concat(axis = concat_55_axis_0, interleave = concat_55_interleave_0, values = (concat_55_values0_0, concat_55_values1_0, concat_55_values2_0, gather_57_cast_uint16_to_int32))[name = string("concat_55")];
            tensor<int32, [4]> causal_mask_7_begin_0 = const()[name = string("causal_mask_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_7_end_mask_0 = const()[name = string("causal_mask_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_7_cast_fp16 = slice_by_index(begin = causal_mask_7_begin_0, end = concat_55, end_mask = causal_mask_7_end_mask_0, x = causal_mask)[name = string("causal_mask_7_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_9_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_7_cast_fp16, key = key_states_11_cast_fp16, query = query_states_11_cast_fp16, value = value_states_11_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_646_perm_0 = const()[name = string("op_646_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_56_axis_0 = const()[name = string("concat_56_axis_0"), val = int32(0)];
            bool concat_56_interleave_0 = const()[name = string("concat_56_interleave_0"), val = bool(false)];
            int32 gather_41_cast_uint16_to_int32 = cast(dtype = gather_41_cast_uint16_to_int32_dtype_0, x = gather_41_cast_uint16)[name = string("cast_104")];
            tensor<int32, [3]> concat_56 = concat(axis = concat_56_axis_0, interleave = concat_56_interleave_0, values = (gather_40, gather_41_cast_uint16_to_int32, var_48))[name = string("concat_56")];
            tensor<fp16, [1, ?, 32, 64]> var_646_cast_fp16 = transpose(perm = var_646_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, ?, ?]> input_17_cast_fp16 = reshape(shape = concat_56, x = var_646_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219754432))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221851648))))[name = string("model_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_79_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = linear_17_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 var_54_promoted_5_to_fp16 = const()[name = string("op_54_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_655_cast_fp16 = pow(x = hidden_states_79_cast_fp16, y = var_54_promoted_5_to_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [1]> variance_11_axes_0 = const()[name = string("variance_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_11_keep_dims_0 = const()[name = string("variance_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_11_cast_fp16 = reduce_mean(axes = variance_11_axes_0, keep_dims = variance_11_keep_dims_0, x = var_655_cast_fp16)[name = string("variance_11_cast_fp16")];
            fp16 var_658_to_fp16 = const()[name = string("op_658_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_659_cast_fp16 = add(x = variance_11_cast_fp16, y = var_658_to_fp16)[name = string("op_659_cast_fp16")];
            fp32 var_660_epsilon_0 = const()[name = string("op_660_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_660_cast_fp16 = rsqrt(epsilon = var_660_epsilon_0, x = var_659_cast_fp16)[name = string("op_660_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_83_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = var_660_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222113856)))];
            tensor<fp16, [1, ?, 2048]> input_19_cast_fp16 = mul(x = model_model_layers_2_post_attention_layernorm_weight_to_fp16, y = hidden_states_83_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222118016))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230506688))))[name = string("model_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_672_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231555328))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239944000))))[name = string("model_model_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_23_cast_fp16 = mul(x = var_672_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240992640))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249381312))))[name = string("model_model_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_89_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = linear_20_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 var_54_promoted_6_to_fp16 = const()[name = string("op_54_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_685_cast_fp16 = pow(x = hidden_states_89_cast_fp16, y = var_54_promoted_6_to_fp16)[name = string("op_685_cast_fp16")];
            tensor<int32, [1]> variance_13_axes_0 = const()[name = string("variance_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_13_keep_dims_0 = const()[name = string("variance_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_13_cast_fp16 = reduce_mean(axes = variance_13_axes_0, keep_dims = variance_13_keep_dims_0, x = var_685_cast_fp16)[name = string("variance_13_cast_fp16")];
            fp16 var_688_to_fp16 = const()[name = string("op_688_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_689_cast_fp16 = add(x = variance_13_cast_fp16, y = var_688_to_fp16)[name = string("op_689_cast_fp16")];
            fp32 var_690_epsilon_0 = const()[name = string("op_690_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_690_cast_fp16 = rsqrt(epsilon = var_690_epsilon_0, x = var_689_cast_fp16)[name = string("op_690_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_93_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = var_690_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250429952)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_97_cast_fp16 = mul(x = model_model_layers_3_input_layernorm_weight_to_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<int32, [3]> var_701_shape_cast_fp16 = shape(x = hidden_states_97_cast_fp16)[name = string("op_701_shape_cast_fp16")];
            int32 gather_58 = const()[name = string("gather_58"), val = int32(1)];
            int32 gather_59_axis_0 = const()[name = string("gather_59_axis_0"), val = int32(0)];
            int32 gather_59_batch_dims_0 = const()[name = string("gather_59_batch_dims_0"), val = int32(0)];
            bool gather_59_validate_indices_0 = const()[name = string("gather_59_validate_indices_0"), val = bool(false)];
            string var_701_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_701_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_59_to_uint16 = const()[name = string("select_59_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_701_shape_cast_fp16_to_uint16 = cast(dtype = var_701_shape_cast_fp16_to_uint16_dtype_0, x = var_701_shape_cast_fp16)[name = string("cast_103")];
            uint16 gather_59_cast_uint16 = gather(axis = gather_59_axis_0, batch_dims = gather_59_batch_dims_0, indices = select_59_to_uint16, validate_indices = gather_59_validate_indices_0, x = var_701_shape_cast_fp16_to_uint16)[name = string("gather_59_cast_uint16")];
            string gather_59_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_59_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250434112))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252531328))))[name = string("model_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_97_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252793536))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253317888))))[name = string("model_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_97_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253383488))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253907840))))[name = string("model_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_97_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> concat_57x = const()[name = string("concat_57x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_710_cast_fp16 = reshape(shape = concat_57x, x = linear_21_cast_fp16)[name = string("op_710_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_58x = const()[name = string("concat_58x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_713_cast_fp16 = reshape(shape = concat_58x, x = linear_22_cast_fp16)[name = string("op_713_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_716_cast_fp16 = reshape(shape = concat_59x, x = linear_23_cast_fp16)[name = string("op_716_cast_fp16")];
            tensor<int32, [4]> v_state_7_perm_0 = const()[name = string("v_state_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_710_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 32, ?, 64]> var_720_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_7_cast_fp16)[name = string("op_720_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_731_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_731_cast_fp16")];
            bool var_733_interleave_0 = const()[name = string("op_733_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_733_cast_fp16 = concat(axis = var_48, interleave = var_733_interleave_0, values = (var_731_cast_fp16, x1_13_cast_fp16))[name = string("op_733_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_734_cast_fp16 = mul(x = var_733_cast_fp16, y = sin_7_cast_fp16)[name = string("op_734_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_15_cast_fp16 = add(x = var_720_cast_fp16, y = var_734_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_713_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 8, ?, 64]> var_736_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_7_cast_fp16)[name = string("op_736_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_747_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_747_cast_fp16")];
            bool var_749_interleave_0 = const()[name = string("op_749_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_749_cast_fp16 = concat(axis = var_48, interleave = var_749_interleave_0, values = (var_747_cast_fp16, x1_15_cast_fp16))[name = string("op_749_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_750_cast_fp16 = mul(x = var_749_cast_fp16, y = sin_7_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_7_cast_fp16 = add(x = var_736_cast_fp16, y = var_750_cast_fp16)[name = string("k_state_7_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_62_values0_0 = const()[name = string("concat_62_values0_0"), val = tensor<int32, [1]>([3])];
            int32 concat_62_axis_0 = const()[name = string("concat_62_axis_0"), val = int32(0)];
            bool concat_62_interleave_0 = const()[name = string("concat_62_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_62 = concat(axis = concat_62_axis_0, interleave = concat_62_interleave_0, values = (concat_62_values0_0, expand_dims_36, expand_dims_37, expand_dims_2, expand_dims_39))[name = string("concat_62")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_62, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = k_state_7_cast_fp16, x = coreml_update_state_36)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_38_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_38 = read_state(input = key_cache)[name = string("coreml_update_state_38")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_7_cast_fp16 = transpose(perm = v_state_7_perm_0, x = var_716_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_62, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = v_state_7_cast_fp16, x = coreml_update_state_37)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_39_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_39 = read_state(input = value_cache)[name = string("coreml_update_state_39")];
            tensor<int32, [5]> var_773_begin_0 = const()[name = string("op_773_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
            tensor<int32, [5]> var_773_end_0 = const()[name = string("op_773_end_0"), val = tensor<int32, [5]>([4, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_773_end_mask_0 = const()[name = string("op_773_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_773_squeeze_mask_0 = const()[name = string("op_773_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_773_cast_fp16 = slice_by_index(begin = var_773_begin_0, end = var_773_end_0, end_mask = var_773_end_mask_0, squeeze_mask = var_773_squeeze_mask_0, x = coreml_update_state_38)[name = string("op_773_cast_fp16")];
            tensor<int32, [4]> var_776_begin_0 = const()[name = string("op_776_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_776_end_mask_0 = const()[name = string("op_776_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_776_cast_fp16 = slice_by_index(begin = var_776_begin_0, end = concat_11, end_mask = var_776_end_mask_0, x = var_773_cast_fp16)[name = string("op_776_cast_fp16")];
            tensor<int32, [5]> var_778_begin_0 = const()[name = string("op_778_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
            tensor<int32, [5]> var_778_end_0 = const()[name = string("op_778_end_0"), val = tensor<int32, [5]>([4, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_778_end_mask_0 = const()[name = string("op_778_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_778_squeeze_mask_0 = const()[name = string("op_778_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_778_cast_fp16 = slice_by_index(begin = var_778_begin_0, end = var_778_end_0, end_mask = var_778_end_mask_0, squeeze_mask = var_778_squeeze_mask_0, x = coreml_update_state_39)[name = string("op_778_cast_fp16")];
            tensor<int32, [4]> var_781_begin_0 = const()[name = string("op_781_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_781_end_mask_0 = const()[name = string("op_781_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_781_cast_fp16 = slice_by_index(begin = var_781_begin_0, end = concat_11, end_mask = var_781_end_mask_0, x = var_778_cast_fp16)[name = string("op_781_cast_fp16")];
            tensor<int32, [4]> var_783_shape_cast_fp16 = shape(x = var_776_cast_fp16)[name = string("op_783_shape_cast_fp16")];
            int32 gather_67 = const()[name = string("gather_67"), val = int32(1)];
            int32 gather_68 = const()[name = string("gather_68"), val = int32(8)];
            int32 gather_69_axis_0 = const()[name = string("gather_69_axis_0"), val = int32(0)];
            int32 gather_69_batch_dims_0 = const()[name = string("gather_69_batch_dims_0"), val = int32(0)];
            bool gather_69_validate_indices_0 = const()[name = string("gather_69_validate_indices_0"), val = bool(false)];
            string var_783_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_783_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_69_to_uint16 = const()[name = string("select_69_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_783_shape_cast_fp16_to_uint16 = cast(dtype = var_783_shape_cast_fp16_to_uint16_dtype_0, x = var_783_shape_cast_fp16)[name = string("cast_102")];
            uint16 gather_69_cast_uint16 = gather(axis = gather_69_axis_0, batch_dims = gather_69_batch_dims_0, indices = select_69_to_uint16, validate_indices = gather_69_validate_indices_0, x = var_783_shape_cast_fp16_to_uint16)[name = string("gather_69_cast_uint16")];
            string gather_69_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_69_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_70 = const()[name = string("gather_70"), val = int32(64)];
            tensor<int32, [1]> var_790_axes_0 = const()[name = string("op_790_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_790_cast_fp16 = expand_dims(axes = var_790_axes_0, x = var_776_cast_fp16)[name = string("op_790_cast_fp16")];
            tensor<int32, [5]> shape_77_cast_fp16 = shape(x = var_790_cast_fp16)[name = string("shape_77_cast_fp16")];
            int32 concat_70_axis_0 = const()[name = string("concat_70_axis_0"), val = int32(0)];
            bool concat_70_interleave_0 = const()[name = string("concat_70_interleave_0"), val = bool(false)];
            int32 gather_69_cast_uint16_to_int32 = cast(dtype = gather_69_cast_uint16_to_int32_dtype_0, x = gather_69_cast_uint16)[name = string("cast_101")];
            tensor<int32, [5]> concat_70 = concat(axis = concat_70_axis_0, interleave = concat_70_interleave_0, values = (gather_67, gather_68, var_59, gather_69_cast_uint16_to_int32, gather_70))[name = string("concat_70")];
            tensor<int32, [5]> real_div_6 = real_div(x = concat_70, y = shape_77_cast_fp16)[name = string("real_div_6")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_101_cast_fp16 = tile(reps = real_div_6, x = var_790_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_15_cast_fp16 = reshape(shape = concat_71x, x = hidden_states_101_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [4]> var_800_shape_cast_fp16 = shape(x = var_781_cast_fp16)[name = string("op_800_shape_cast_fp16")];
            int32 gather_71 = const()[name = string("gather_71"), val = int32(1)];
            int32 gather_72 = const()[name = string("gather_72"), val = int32(8)];
            int32 gather_73_axis_0 = const()[name = string("gather_73_axis_0"), val = int32(0)];
            int32 gather_73_batch_dims_0 = const()[name = string("gather_73_batch_dims_0"), val = int32(0)];
            bool gather_73_validate_indices_0 = const()[name = string("gather_73_validate_indices_0"), val = bool(false)];
            string var_800_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_800_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_73_to_uint16 = const()[name = string("select_73_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_800_shape_cast_fp16_to_uint16 = cast(dtype = var_800_shape_cast_fp16_to_uint16_dtype_0, x = var_800_shape_cast_fp16)[name = string("cast_100")];
            uint16 gather_73_cast_uint16 = gather(axis = gather_73_axis_0, batch_dims = gather_73_batch_dims_0, indices = select_73_to_uint16, validate_indices = gather_73_validate_indices_0, x = var_800_shape_cast_fp16_to_uint16)[name = string("gather_73_cast_uint16")];
            string gather_73_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_73_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_74 = const()[name = string("gather_74"), val = int32(64)];
            tensor<int32, [1]> var_807_axes_0 = const()[name = string("op_807_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_807_cast_fp16 = expand_dims(axes = var_807_axes_0, x = var_781_cast_fp16)[name = string("op_807_cast_fp16")];
            tensor<int32, [5]> shape_82_cast_fp16 = shape(x = var_807_cast_fp16)[name = string("shape_82_cast_fp16")];
            int32 concat_72_axis_0 = const()[name = string("concat_72_axis_0"), val = int32(0)];
            bool concat_72_interleave_0 = const()[name = string("concat_72_interleave_0"), val = bool(false)];
            int32 gather_73_cast_uint16_to_int32 = cast(dtype = gather_73_cast_uint16_to_int32_dtype_0, x = gather_73_cast_uint16)[name = string("cast_99")];
            tensor<int32, [5]> concat_72 = concat(axis = concat_72_axis_0, interleave = concat_72_interleave_0, values = (gather_71, gather_72, var_59, gather_73_cast_uint16_to_int32, gather_74))[name = string("concat_72")];
            tensor<int32, [5]> real_div_7 = real_div(x = concat_72, y = shape_82_cast_fp16)[name = string("real_div_7")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_105_cast_fp16 = tile(reps = real_div_7, x = var_807_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_15_cast_fp16 = reshape(shape = concat_73x, x = hidden_states_105_cast_fp16)[name = string("value_states_15_cast_fp16")];
            tensor<int32, [4]> var_817_shape_cast_fp16 = shape(x = key_states_15_cast_fp16)[name = string("op_817_shape_cast_fp16")];
            int32 gather_75_axis_0 = const()[name = string("gather_75_axis_0"), val = int32(0)];
            int32 gather_75_batch_dims_0 = const()[name = string("gather_75_batch_dims_0"), val = int32(0)];
            bool gather_75_validate_indices_0 = const()[name = string("gather_75_validate_indices_0"), val = bool(false)];
            string var_817_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_817_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_75_to_uint16 = const()[name = string("select_75_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_817_shape_cast_fp16_to_uint16 = cast(dtype = var_817_shape_cast_fp16_to_uint16_dtype_0, x = var_817_shape_cast_fp16)[name = string("cast_98")];
            uint16 gather_75_cast_uint16 = gather(axis = gather_75_axis_0, batch_dims = gather_75_batch_dims_0, indices = select_75_to_uint16, validate_indices = gather_75_validate_indices_0, x = var_817_shape_cast_fp16_to_uint16)[name = string("gather_75_cast_uint16")];
            string gather_75_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_75_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_74_values0_0 = const()[name = string("concat_74_values0_0"), val = int32(1)];
            int32 concat_74_values1_0 = const()[name = string("concat_74_values1_0"), val = int32(1)];
            int32 concat_74_values2_0 = const()[name = string("concat_74_values2_0"), val = int32(0)];
            int32 concat_74_axis_0 = const()[name = string("concat_74_axis_0"), val = int32(0)];
            bool concat_74_interleave_0 = const()[name = string("concat_74_interleave_0"), val = bool(false)];
            int32 gather_75_cast_uint16_to_int32 = cast(dtype = gather_75_cast_uint16_to_int32_dtype_0, x = gather_75_cast_uint16)[name = string("cast_97")];
            tensor<int32, [4]> concat_74 = concat(axis = concat_74_axis_0, interleave = concat_74_interleave_0, values = (concat_74_values0_0, concat_74_values1_0, concat_74_values2_0, gather_75_cast_uint16_to_int32))[name = string("concat_74")];
            tensor<int32, [4]> causal_mask_9_begin_0 = const()[name = string("causal_mask_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_9_end_mask_0 = const()[name = string("causal_mask_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_9_cast_fp16 = slice_by_index(begin = causal_mask_9_begin_0, end = concat_74, end_mask = causal_mask_9_end_mask_0, x = causal_mask)[name = string("causal_mask_9_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_13_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_9_cast_fp16, key = key_states_15_cast_fp16, query = query_states_15_cast_fp16, value = value_states_15_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_823_perm_0 = const()[name = string("op_823_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_75_axis_0 = const()[name = string("concat_75_axis_0"), val = int32(0)];
            bool concat_75_interleave_0 = const()[name = string("concat_75_interleave_0"), val = bool(false)];
            int32 gather_59_cast_uint16_to_int32 = cast(dtype = gather_59_cast_uint16_to_int32_dtype_0, x = gather_59_cast_uint16)[name = string("cast_96")];
            tensor<int32, [3]> concat_75 = concat(axis = concat_75_axis_0, interleave = concat_75_interleave_0, values = (gather_58, gather_59_cast_uint16_to_int32, var_48))[name = string("concat_75")];
            tensor<fp16, [1, ?, 32, 64]> var_823_cast_fp16 = transpose(perm = var_823_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, ?, ?]> input_25_cast_fp16 = reshape(shape = concat_75, x = var_823_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253973440))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256070656))))[name = string("model_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_109_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = linear_24_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 var_54_promoted_7_to_fp16 = const()[name = string("op_54_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_832_cast_fp16 = pow(x = hidden_states_109_cast_fp16, y = var_54_promoted_7_to_fp16)[name = string("op_832_cast_fp16")];
            tensor<int32, [1]> variance_15_axes_0 = const()[name = string("variance_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_15_keep_dims_0 = const()[name = string("variance_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_15_cast_fp16 = reduce_mean(axes = variance_15_axes_0, keep_dims = variance_15_keep_dims_0, x = var_832_cast_fp16)[name = string("variance_15_cast_fp16")];
            fp16 var_835_to_fp16 = const()[name = string("op_835_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_836_cast_fp16 = add(x = variance_15_cast_fp16, y = var_835_to_fp16)[name = string("op_836_cast_fp16")];
            fp32 var_837_epsilon_0 = const()[name = string("op_837_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_837_cast_fp16 = rsqrt(epsilon = var_837_epsilon_0, x = var_836_cast_fp16)[name = string("op_837_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_113_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = var_837_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256332864)))];
            tensor<fp16, [1, ?, 2048]> input_27_cast_fp16 = mul(x = model_model_layers_3_post_attention_layernorm_weight_to_fp16, y = hidden_states_113_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256337024))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264725696))))[name = string("model_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_849_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265774336))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274163008))))[name = string("model_model_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_31_cast_fp16 = mul(x = var_849_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275211648))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283600320))))[name = string("model_model_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_27_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_119_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = linear_27_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 var_54_promoted_8_to_fp16 = const()[name = string("op_54_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_862_cast_fp16 = pow(x = hidden_states_119_cast_fp16, y = var_54_promoted_8_to_fp16)[name = string("op_862_cast_fp16")];
            tensor<int32, [1]> variance_17_axes_0 = const()[name = string("variance_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_17_keep_dims_0 = const()[name = string("variance_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_17_cast_fp16 = reduce_mean(axes = variance_17_axes_0, keep_dims = variance_17_keep_dims_0, x = var_862_cast_fp16)[name = string("variance_17_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_866_cast_fp16 = add(x = variance_17_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            fp32 var_867_epsilon_0 = const()[name = string("op_867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_867_cast_fp16 = rsqrt(epsilon = var_867_epsilon_0, x = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_123_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = var_867_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284648960)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_127_cast_fp16 = mul(x = model_model_layers_4_input_layernorm_weight_to_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<int32, [3]> var_878_shape_cast_fp16 = shape(x = hidden_states_127_cast_fp16)[name = string("op_878_shape_cast_fp16")];
            int32 gather_76 = const()[name = string("gather_76"), val = int32(1)];
            int32 gather_77_axis_0 = const()[name = string("gather_77_axis_0"), val = int32(0)];
            int32 gather_77_batch_dims_0 = const()[name = string("gather_77_batch_dims_0"), val = int32(0)];
            bool gather_77_validate_indices_0 = const()[name = string("gather_77_validate_indices_0"), val = bool(false)];
            string var_878_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_878_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_77_to_uint16 = const()[name = string("select_77_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_878_shape_cast_fp16_to_uint16 = cast(dtype = var_878_shape_cast_fp16_to_uint16_dtype_0, x = var_878_shape_cast_fp16)[name = string("cast_95")];
            uint16 gather_77_cast_uint16 = gather(axis = gather_77_axis_0, batch_dims = gather_77_batch_dims_0, indices = select_77_to_uint16, validate_indices = gather_77_validate_indices_0, x = var_878_shape_cast_fp16_to_uint16)[name = string("gather_77_cast_uint16")];
            string gather_77_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_77_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284653120))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286750336))))[name = string("model_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287012544))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287536896))))[name = string("model_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287602496))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288126848))))[name = string("model_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> concat_76x = const()[name = string("concat_76x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_887_cast_fp16 = reshape(shape = concat_76x, x = linear_28_cast_fp16)[name = string("op_887_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_77x = const()[name = string("concat_77x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_890_cast_fp16 = reshape(shape = concat_77x, x = linear_29_cast_fp16)[name = string("op_890_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_78x = const()[name = string("concat_78x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_893_cast_fp16 = reshape(shape = concat_78x, x = linear_30_cast_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [4]> v_state_9_perm_0 = const()[name = string("v_state_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_887_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 32, ?, 64]> var_897_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_7_cast_fp16)[name = string("op_897_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_908_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_908_cast_fp16")];
            bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_910_cast_fp16 = concat(axis = var_48, interleave = var_910_interleave_0, values = (var_908_cast_fp16, x1_17_cast_fp16))[name = string("op_910_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_911_cast_fp16 = mul(x = var_910_cast_fp16, y = sin_7_cast_fp16)[name = string("op_911_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_19_cast_fp16 = add(x = var_897_cast_fp16, y = var_911_cast_fp16)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_890_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, ?, 64]> var_913_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_7_cast_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_924_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_924_cast_fp16")];
            bool var_926_interleave_0 = const()[name = string("op_926_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_926_cast_fp16 = concat(axis = var_48, interleave = var_926_interleave_0, values = (var_924_cast_fp16, x1_19_cast_fp16))[name = string("op_926_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_927_cast_fp16 = mul(x = var_926_cast_fp16, y = sin_7_cast_fp16)[name = string("op_927_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_9_cast_fp16 = add(x = var_913_cast_fp16, y = var_927_cast_fp16)[name = string("k_state_9_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_81_values0_0 = const()[name = string("concat_81_values0_0"), val = tensor<int32, [1]>([4])];
            int32 concat_81_axis_0 = const()[name = string("concat_81_axis_0"), val = int32(0)];
            bool concat_81_interleave_0 = const()[name = string("concat_81_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_81 = concat(axis = concat_81_axis_0, interleave = concat_81_interleave_0, values = (concat_81_values0_0, expand_dims_48, expand_dims_49, expand_dims_2, expand_dims_51))[name = string("concat_81")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_81, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = k_state_9_cast_fp16, x = coreml_update_state_38)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_40_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_40 = read_state(input = key_cache)[name = string("coreml_update_state_40")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_9_cast_fp16 = transpose(perm = v_state_9_perm_0, x = var_893_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_81, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = v_state_9_cast_fp16, x = coreml_update_state_39)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_41_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_41 = read_state(input = value_cache)[name = string("coreml_update_state_41")];
            tensor<int32, [5]> var_950_begin_0 = const()[name = string("op_950_begin_0"), val = tensor<int32, [5]>([4, 0, 0, 0, 0])];
            tensor<int32, [5]> var_950_end_0 = const()[name = string("op_950_end_0"), val = tensor<int32, [5]>([5, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_950_end_mask_0 = const()[name = string("op_950_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_950_squeeze_mask_0 = const()[name = string("op_950_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_950_cast_fp16 = slice_by_index(begin = var_950_begin_0, end = var_950_end_0, end_mask = var_950_end_mask_0, squeeze_mask = var_950_squeeze_mask_0, x = coreml_update_state_40)[name = string("op_950_cast_fp16")];
            tensor<int32, [4]> var_953_begin_0 = const()[name = string("op_953_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_953_end_mask_0 = const()[name = string("op_953_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_953_cast_fp16 = slice_by_index(begin = var_953_begin_0, end = concat_11, end_mask = var_953_end_mask_0, x = var_950_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<int32, [5]> var_955_begin_0 = const()[name = string("op_955_begin_0"), val = tensor<int32, [5]>([4, 0, 0, 0, 0])];
            tensor<int32, [5]> var_955_end_0 = const()[name = string("op_955_end_0"), val = tensor<int32, [5]>([5, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_955_end_mask_0 = const()[name = string("op_955_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_955_squeeze_mask_0 = const()[name = string("op_955_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_955_cast_fp16 = slice_by_index(begin = var_955_begin_0, end = var_955_end_0, end_mask = var_955_end_mask_0, squeeze_mask = var_955_squeeze_mask_0, x = coreml_update_state_41)[name = string("op_955_cast_fp16")];
            tensor<int32, [4]> var_958_begin_0 = const()[name = string("op_958_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_958_end_mask_0 = const()[name = string("op_958_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_958_cast_fp16 = slice_by_index(begin = var_958_begin_0, end = concat_11, end_mask = var_958_end_mask_0, x = var_955_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<int32, [4]> var_960_shape_cast_fp16 = shape(x = var_953_cast_fp16)[name = string("op_960_shape_cast_fp16")];
            int32 gather_85 = const()[name = string("gather_85"), val = int32(1)];
            int32 gather_86 = const()[name = string("gather_86"), val = int32(8)];
            int32 gather_87_axis_0 = const()[name = string("gather_87_axis_0"), val = int32(0)];
            int32 gather_87_batch_dims_0 = const()[name = string("gather_87_batch_dims_0"), val = int32(0)];
            bool gather_87_validate_indices_0 = const()[name = string("gather_87_validate_indices_0"), val = bool(false)];
            string var_960_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_960_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_87_to_uint16 = const()[name = string("select_87_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_960_shape_cast_fp16_to_uint16 = cast(dtype = var_960_shape_cast_fp16_to_uint16_dtype_0, x = var_960_shape_cast_fp16)[name = string("cast_94")];
            uint16 gather_87_cast_uint16 = gather(axis = gather_87_axis_0, batch_dims = gather_87_batch_dims_0, indices = select_87_to_uint16, validate_indices = gather_87_validate_indices_0, x = var_960_shape_cast_fp16_to_uint16)[name = string("gather_87_cast_uint16")];
            string gather_87_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_87_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_88 = const()[name = string("gather_88"), val = int32(64)];
            tensor<int32, [1]> var_967_axes_0 = const()[name = string("op_967_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_967_cast_fp16 = expand_dims(axes = var_967_axes_0, x = var_953_cast_fp16)[name = string("op_967_cast_fp16")];
            tensor<int32, [5]> shape_97_cast_fp16 = shape(x = var_967_cast_fp16)[name = string("shape_97_cast_fp16")];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            int32 gather_87_cast_uint16_to_int32 = cast(dtype = gather_87_cast_uint16_to_int32_dtype_0, x = gather_87_cast_uint16)[name = string("cast_93")];
            tensor<int32, [5]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (gather_85, gather_86, var_59, gather_87_cast_uint16_to_int32, gather_88))[name = string("concat_89")];
            tensor<int32, [5]> real_div_8 = real_div(x = concat_89, y = shape_97_cast_fp16)[name = string("real_div_8")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_131_cast_fp16 = tile(reps = real_div_8, x = var_967_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> concat_90x = const()[name = string("concat_90x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_19_cast_fp16 = reshape(shape = concat_90x, x = hidden_states_131_cast_fp16)[name = string("key_states_19_cast_fp16")];
            tensor<int32, [4]> var_977_shape_cast_fp16 = shape(x = var_958_cast_fp16)[name = string("op_977_shape_cast_fp16")];
            int32 gather_89 = const()[name = string("gather_89"), val = int32(1)];
            int32 gather_90 = const()[name = string("gather_90"), val = int32(8)];
            int32 gather_91_axis_0 = const()[name = string("gather_91_axis_0"), val = int32(0)];
            int32 gather_91_batch_dims_0 = const()[name = string("gather_91_batch_dims_0"), val = int32(0)];
            bool gather_91_validate_indices_0 = const()[name = string("gather_91_validate_indices_0"), val = bool(false)];
            string var_977_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_977_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_91_to_uint16 = const()[name = string("select_91_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_977_shape_cast_fp16_to_uint16 = cast(dtype = var_977_shape_cast_fp16_to_uint16_dtype_0, x = var_977_shape_cast_fp16)[name = string("cast_92")];
            uint16 gather_91_cast_uint16 = gather(axis = gather_91_axis_0, batch_dims = gather_91_batch_dims_0, indices = select_91_to_uint16, validate_indices = gather_91_validate_indices_0, x = var_977_shape_cast_fp16_to_uint16)[name = string("gather_91_cast_uint16")];
            string gather_91_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_91_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_92 = const()[name = string("gather_92"), val = int32(64)];
            tensor<int32, [1]> var_984_axes_0 = const()[name = string("op_984_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_984_cast_fp16 = expand_dims(axes = var_984_axes_0, x = var_958_cast_fp16)[name = string("op_984_cast_fp16")];
            tensor<int32, [5]> shape_102_cast_fp16 = shape(x = var_984_cast_fp16)[name = string("shape_102_cast_fp16")];
            int32 concat_91_axis_0 = const()[name = string("concat_91_axis_0"), val = int32(0)];
            bool concat_91_interleave_0 = const()[name = string("concat_91_interleave_0"), val = bool(false)];
            int32 gather_91_cast_uint16_to_int32 = cast(dtype = gather_91_cast_uint16_to_int32_dtype_0, x = gather_91_cast_uint16)[name = string("cast_91")];
            tensor<int32, [5]> concat_91 = concat(axis = concat_91_axis_0, interleave = concat_91_interleave_0, values = (gather_89, gather_90, var_59, gather_91_cast_uint16_to_int32, gather_92))[name = string("concat_91")];
            tensor<int32, [5]> real_div_9 = real_div(x = concat_91, y = shape_102_cast_fp16)[name = string("real_div_9")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_135_cast_fp16 = tile(reps = real_div_9, x = var_984_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> concat_92x = const()[name = string("concat_92x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_19_cast_fp16 = reshape(shape = concat_92x, x = hidden_states_135_cast_fp16)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> var_994_shape_cast_fp16 = shape(x = key_states_19_cast_fp16)[name = string("op_994_shape_cast_fp16")];
            int32 gather_93_axis_0 = const()[name = string("gather_93_axis_0"), val = int32(0)];
            int32 gather_93_batch_dims_0 = const()[name = string("gather_93_batch_dims_0"), val = int32(0)];
            bool gather_93_validate_indices_0 = const()[name = string("gather_93_validate_indices_0"), val = bool(false)];
            string var_994_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_994_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_93_to_uint16 = const()[name = string("select_93_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_994_shape_cast_fp16_to_uint16 = cast(dtype = var_994_shape_cast_fp16_to_uint16_dtype_0, x = var_994_shape_cast_fp16)[name = string("cast_90")];
            uint16 gather_93_cast_uint16 = gather(axis = gather_93_axis_0, batch_dims = gather_93_batch_dims_0, indices = select_93_to_uint16, validate_indices = gather_93_validate_indices_0, x = var_994_shape_cast_fp16_to_uint16)[name = string("gather_93_cast_uint16")];
            string gather_93_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_93_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_93_values0_0 = const()[name = string("concat_93_values0_0"), val = int32(1)];
            int32 concat_93_values1_0 = const()[name = string("concat_93_values1_0"), val = int32(1)];
            int32 concat_93_values2_0 = const()[name = string("concat_93_values2_0"), val = int32(0)];
            int32 concat_93_axis_0 = const()[name = string("concat_93_axis_0"), val = int32(0)];
            bool concat_93_interleave_0 = const()[name = string("concat_93_interleave_0"), val = bool(false)];
            int32 gather_93_cast_uint16_to_int32 = cast(dtype = gather_93_cast_uint16_to_int32_dtype_0, x = gather_93_cast_uint16)[name = string("cast_89")];
            tensor<int32, [4]> concat_93 = concat(axis = concat_93_axis_0, interleave = concat_93_interleave_0, values = (concat_93_values0_0, concat_93_values1_0, concat_93_values2_0, gather_93_cast_uint16_to_int32))[name = string("concat_93")];
            tensor<int32, [4]> causal_mask_11_begin_0 = const()[name = string("causal_mask_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_11_end_mask_0 = const()[name = string("causal_mask_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_11_cast_fp16 = slice_by_index(begin = causal_mask_11_begin_0, end = concat_93, end_mask = causal_mask_11_end_mask_0, x = causal_mask)[name = string("causal_mask_11_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_17_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_11_cast_fp16, key = key_states_19_cast_fp16, query = query_states_19_cast_fp16, value = value_states_19_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_1000_perm_0 = const()[name = string("op_1000_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_94_axis_0 = const()[name = string("concat_94_axis_0"), val = int32(0)];
            bool concat_94_interleave_0 = const()[name = string("concat_94_interleave_0"), val = bool(false)];
            int32 gather_77_cast_uint16_to_int32 = cast(dtype = gather_77_cast_uint16_to_int32_dtype_0, x = gather_77_cast_uint16)[name = string("cast_88")];
            tensor<int32, [3]> concat_94 = concat(axis = concat_94_axis_0, interleave = concat_94_interleave_0, values = (gather_76, gather_77_cast_uint16_to_int32, var_48))[name = string("concat_94")];
            tensor<fp16, [1, ?, 32, 64]> var_1000_cast_fp16 = transpose(perm = var_1000_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, ?, ?]> input_33_cast_fp16 = reshape(shape = concat_94, x = var_1000_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288192448))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290289664))))[name = string("model_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_139_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = linear_31_cast_fp16)[name = string("hidden_states_139_cast_fp16")];
            fp16 var_54_promoted_9_to_fp16 = const()[name = string("op_54_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1009_cast_fp16 = pow(x = hidden_states_139_cast_fp16, y = var_54_promoted_9_to_fp16)[name = string("op_1009_cast_fp16")];
            tensor<int32, [1]> variance_19_axes_0 = const()[name = string("variance_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_19_keep_dims_0 = const()[name = string("variance_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_19_cast_fp16 = reduce_mean(axes = variance_19_axes_0, keep_dims = variance_19_keep_dims_0, x = var_1009_cast_fp16)[name = string("variance_19_cast_fp16")];
            fp16 var_1012_to_fp16 = const()[name = string("op_1012_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1013_cast_fp16 = add(x = variance_19_cast_fp16, y = var_1012_to_fp16)[name = string("op_1013_cast_fp16")];
            fp32 var_1014_epsilon_0 = const()[name = string("op_1014_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1014_cast_fp16 = rsqrt(epsilon = var_1014_epsilon_0, x = var_1013_cast_fp16)[name = string("op_1014_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_143_cast_fp16 = mul(x = hidden_states_139_cast_fp16, y = var_1014_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290551872)))];
            tensor<fp16, [1, ?, 2048]> input_35_cast_fp16 = mul(x = model_model_layers_4_post_attention_layernorm_weight_to_fp16, y = hidden_states_143_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290556032))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298944704))))[name = string("model_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_1026_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_1026_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299993344))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308382016))))[name = string("model_model_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_39_cast_fp16 = mul(x = var_1026_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309430656))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317819328))))[name = string("model_model_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_149_cast_fp16 = add(x = hidden_states_139_cast_fp16, y = linear_34_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            fp16 var_54_promoted_10_to_fp16 = const()[name = string("op_54_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1039_cast_fp16 = pow(x = hidden_states_149_cast_fp16, y = var_54_promoted_10_to_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> variance_21_axes_0 = const()[name = string("variance_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_21_keep_dims_0 = const()[name = string("variance_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_21_cast_fp16 = reduce_mean(axes = variance_21_axes_0, keep_dims = variance_21_keep_dims_0, x = var_1039_cast_fp16)[name = string("variance_21_cast_fp16")];
            fp16 var_1042_to_fp16 = const()[name = string("op_1042_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1043_cast_fp16 = add(x = variance_21_cast_fp16, y = var_1042_to_fp16)[name = string("op_1043_cast_fp16")];
            fp32 var_1044_epsilon_0 = const()[name = string("op_1044_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1044_cast_fp16 = rsqrt(epsilon = var_1044_epsilon_0, x = var_1043_cast_fp16)[name = string("op_1044_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_153_cast_fp16 = mul(x = hidden_states_149_cast_fp16, y = var_1044_cast_fp16)[name = string("hidden_states_153_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318867968)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_157_cast_fp16 = mul(x = model_model_layers_5_input_layernorm_weight_to_fp16, y = hidden_states_153_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [3]> var_1055_shape_cast_fp16 = shape(x = hidden_states_157_cast_fp16)[name = string("op_1055_shape_cast_fp16")];
            int32 gather_94 = const()[name = string("gather_94"), val = int32(1)];
            int32 gather_95_axis_0 = const()[name = string("gather_95_axis_0"), val = int32(0)];
            int32 gather_95_batch_dims_0 = const()[name = string("gather_95_batch_dims_0"), val = int32(0)];
            bool gather_95_validate_indices_0 = const()[name = string("gather_95_validate_indices_0"), val = bool(false)];
            string var_1055_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1055_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_95_to_uint16 = const()[name = string("select_95_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1055_shape_cast_fp16_to_uint16 = cast(dtype = var_1055_shape_cast_fp16_to_uint16_dtype_0, x = var_1055_shape_cast_fp16)[name = string("cast_87")];
            uint16 gather_95_cast_uint16 = gather(axis = gather_95_axis_0, batch_dims = gather_95_batch_dims_0, indices = select_95_to_uint16, validate_indices = gather_95_validate_indices_0, x = var_1055_shape_cast_fp16_to_uint16)[name = string("gather_95_cast_uint16")];
            string gather_95_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_95_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318872128))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320969344))))[name = string("model_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_157_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321231552))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321755904))))[name = string("model_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_157_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321821504))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322345856))))[name = string("model_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_157_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_1064_cast_fp16 = reshape(shape = concat_95x, x = linear_35_cast_fp16)[name = string("op_1064_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1067_cast_fp16 = reshape(shape = concat_96x, x = linear_36_cast_fp16)[name = string("op_1067_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1070_cast_fp16 = reshape(shape = concat_97x, x = linear_37_cast_fp16)[name = string("op_1070_cast_fp16")];
            tensor<int32, [4]> v_state_11_perm_0 = const()[name = string("v_state_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_1064_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 32, ?, 64]> var_1074_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1074_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_1085_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_1085_cast_fp16")];
            bool var_1087_interleave_0 = const()[name = string("op_1087_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_1087_cast_fp16 = concat(axis = var_48, interleave = var_1087_interleave_0, values = (var_1085_cast_fp16, x1_21_cast_fp16))[name = string("op_1087_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_1088_cast_fp16 = mul(x = var_1087_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1088_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_23_cast_fp16 = add(x = var_1074_cast_fp16, y = var_1088_cast_fp16)[name = string("query_states_23_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_1067_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, ?, 64]> var_1090_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1090_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_1101_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_1101_cast_fp16")];
            bool var_1103_interleave_0 = const()[name = string("op_1103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_1103_cast_fp16 = concat(axis = var_48, interleave = var_1103_interleave_0, values = (var_1101_cast_fp16, x1_23_cast_fp16))[name = string("op_1103_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1104_cast_fp16 = mul(x = var_1103_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_11_cast_fp16 = add(x = var_1090_cast_fp16, y = var_1104_cast_fp16)[name = string("k_state_11_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_100_values0_0 = const()[name = string("concat_100_values0_0"), val = tensor<int32, [1]>([5])];
            int32 concat_100_axis_0 = const()[name = string("concat_100_axis_0"), val = int32(0)];
            bool concat_100_interleave_0 = const()[name = string("concat_100_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_100 = concat(axis = concat_100_axis_0, interleave = concat_100_interleave_0, values = (concat_100_values0_0, expand_dims_60, expand_dims_61, expand_dims_2, expand_dims_63))[name = string("concat_100")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_100, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = k_state_11_cast_fp16, x = coreml_update_state_40)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_42_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_42 = read_state(input = key_cache)[name = string("coreml_update_state_42")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_11_cast_fp16 = transpose(perm = v_state_11_perm_0, x = var_1070_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_100, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = v_state_11_cast_fp16, x = coreml_update_state_41)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_43_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_43 = read_state(input = value_cache)[name = string("coreml_update_state_43")];
            tensor<int32, [5]> var_1127_begin_0 = const()[name = string("op_1127_begin_0"), val = tensor<int32, [5]>([5, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1127_end_0 = const()[name = string("op_1127_end_0"), val = tensor<int32, [5]>([6, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1127_end_mask_0 = const()[name = string("op_1127_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1127_squeeze_mask_0 = const()[name = string("op_1127_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1127_cast_fp16 = slice_by_index(begin = var_1127_begin_0, end = var_1127_end_0, end_mask = var_1127_end_mask_0, squeeze_mask = var_1127_squeeze_mask_0, x = coreml_update_state_42)[name = string("op_1127_cast_fp16")];
            tensor<int32, [4]> var_1130_begin_0 = const()[name = string("op_1130_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1130_end_mask_0 = const()[name = string("op_1130_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1130_cast_fp16 = slice_by_index(begin = var_1130_begin_0, end = concat_11, end_mask = var_1130_end_mask_0, x = var_1127_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [5]> var_1132_begin_0 = const()[name = string("op_1132_begin_0"), val = tensor<int32, [5]>([5, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1132_end_0 = const()[name = string("op_1132_end_0"), val = tensor<int32, [5]>([6, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1132_end_mask_0 = const()[name = string("op_1132_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1132_squeeze_mask_0 = const()[name = string("op_1132_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1132_cast_fp16 = slice_by_index(begin = var_1132_begin_0, end = var_1132_end_0, end_mask = var_1132_end_mask_0, squeeze_mask = var_1132_squeeze_mask_0, x = coreml_update_state_43)[name = string("op_1132_cast_fp16")];
            tensor<int32, [4]> var_1135_begin_0 = const()[name = string("op_1135_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1135_end_mask_0 = const()[name = string("op_1135_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1135_cast_fp16 = slice_by_index(begin = var_1135_begin_0, end = concat_11, end_mask = var_1135_end_mask_0, x = var_1132_cast_fp16)[name = string("op_1135_cast_fp16")];
            tensor<int32, [4]> var_1137_shape_cast_fp16 = shape(x = var_1130_cast_fp16)[name = string("op_1137_shape_cast_fp16")];
            int32 gather_103 = const()[name = string("gather_103"), val = int32(1)];
            int32 gather_104 = const()[name = string("gather_104"), val = int32(8)];
            int32 gather_105_axis_0 = const()[name = string("gather_105_axis_0"), val = int32(0)];
            int32 gather_105_batch_dims_0 = const()[name = string("gather_105_batch_dims_0"), val = int32(0)];
            bool gather_105_validate_indices_0 = const()[name = string("gather_105_validate_indices_0"), val = bool(false)];
            string var_1137_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1137_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_105_to_uint16 = const()[name = string("select_105_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1137_shape_cast_fp16_to_uint16 = cast(dtype = var_1137_shape_cast_fp16_to_uint16_dtype_0, x = var_1137_shape_cast_fp16)[name = string("cast_86")];
            uint16 gather_105_cast_uint16 = gather(axis = gather_105_axis_0, batch_dims = gather_105_batch_dims_0, indices = select_105_to_uint16, validate_indices = gather_105_validate_indices_0, x = var_1137_shape_cast_fp16_to_uint16)[name = string("gather_105_cast_uint16")];
            string gather_105_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_105_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_106 = const()[name = string("gather_106"), val = int32(64)];
            tensor<int32, [1]> var_1144_axes_0 = const()[name = string("op_1144_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1144_cast_fp16 = expand_dims(axes = var_1144_axes_0, x = var_1130_cast_fp16)[name = string("op_1144_cast_fp16")];
            tensor<int32, [5]> shape_117_cast_fp16 = shape(x = var_1144_cast_fp16)[name = string("shape_117_cast_fp16")];
            int32 concat_108_axis_0 = const()[name = string("concat_108_axis_0"), val = int32(0)];
            bool concat_108_interleave_0 = const()[name = string("concat_108_interleave_0"), val = bool(false)];
            int32 gather_105_cast_uint16_to_int32 = cast(dtype = gather_105_cast_uint16_to_int32_dtype_0, x = gather_105_cast_uint16)[name = string("cast_85")];
            tensor<int32, [5]> concat_108 = concat(axis = concat_108_axis_0, interleave = concat_108_interleave_0, values = (gather_103, gather_104, var_59, gather_105_cast_uint16_to_int32, gather_106))[name = string("concat_108")];
            tensor<int32, [5]> real_div_10 = real_div(x = concat_108, y = shape_117_cast_fp16)[name = string("real_div_10")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_161_cast_fp16 = tile(reps = real_div_10, x = var_1144_cast_fp16)[name = string("hidden_states_161_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_23_cast_fp16 = reshape(shape = concat_109x, x = hidden_states_161_cast_fp16)[name = string("key_states_23_cast_fp16")];
            tensor<int32, [4]> var_1154_shape_cast_fp16 = shape(x = var_1135_cast_fp16)[name = string("op_1154_shape_cast_fp16")];
            int32 gather_107 = const()[name = string("gather_107"), val = int32(1)];
            int32 gather_108 = const()[name = string("gather_108"), val = int32(8)];
            int32 gather_109_axis_0 = const()[name = string("gather_109_axis_0"), val = int32(0)];
            int32 gather_109_batch_dims_0 = const()[name = string("gather_109_batch_dims_0"), val = int32(0)];
            bool gather_109_validate_indices_0 = const()[name = string("gather_109_validate_indices_0"), val = bool(false)];
            string var_1154_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1154_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_109_to_uint16 = const()[name = string("select_109_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1154_shape_cast_fp16_to_uint16 = cast(dtype = var_1154_shape_cast_fp16_to_uint16_dtype_0, x = var_1154_shape_cast_fp16)[name = string("cast_84")];
            uint16 gather_109_cast_uint16 = gather(axis = gather_109_axis_0, batch_dims = gather_109_batch_dims_0, indices = select_109_to_uint16, validate_indices = gather_109_validate_indices_0, x = var_1154_shape_cast_fp16_to_uint16)[name = string("gather_109_cast_uint16")];
            string gather_109_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_109_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_110 = const()[name = string("gather_110"), val = int32(64)];
            tensor<int32, [1]> var_1161_axes_0 = const()[name = string("op_1161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1161_cast_fp16 = expand_dims(axes = var_1161_axes_0, x = var_1135_cast_fp16)[name = string("op_1161_cast_fp16")];
            tensor<int32, [5]> shape_122_cast_fp16 = shape(x = var_1161_cast_fp16)[name = string("shape_122_cast_fp16")];
            int32 concat_110_axis_0 = const()[name = string("concat_110_axis_0"), val = int32(0)];
            bool concat_110_interleave_0 = const()[name = string("concat_110_interleave_0"), val = bool(false)];
            int32 gather_109_cast_uint16_to_int32 = cast(dtype = gather_109_cast_uint16_to_int32_dtype_0, x = gather_109_cast_uint16)[name = string("cast_83")];
            tensor<int32, [5]> concat_110 = concat(axis = concat_110_axis_0, interleave = concat_110_interleave_0, values = (gather_107, gather_108, var_59, gather_109_cast_uint16_to_int32, gather_110))[name = string("concat_110")];
            tensor<int32, [5]> real_div_11 = real_div(x = concat_110, y = shape_122_cast_fp16)[name = string("real_div_11")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_165_cast_fp16 = tile(reps = real_div_11, x = var_1161_cast_fp16)[name = string("hidden_states_165_cast_fp16")];
            tensor<int32, [4]> concat_111x = const()[name = string("concat_111x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_23_cast_fp16 = reshape(shape = concat_111x, x = hidden_states_165_cast_fp16)[name = string("value_states_23_cast_fp16")];
            tensor<int32, [4]> var_1171_shape_cast_fp16 = shape(x = key_states_23_cast_fp16)[name = string("op_1171_shape_cast_fp16")];
            int32 gather_111_axis_0 = const()[name = string("gather_111_axis_0"), val = int32(0)];
            int32 gather_111_batch_dims_0 = const()[name = string("gather_111_batch_dims_0"), val = int32(0)];
            bool gather_111_validate_indices_0 = const()[name = string("gather_111_validate_indices_0"), val = bool(false)];
            string var_1171_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1171_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_111_to_uint16 = const()[name = string("select_111_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1171_shape_cast_fp16_to_uint16 = cast(dtype = var_1171_shape_cast_fp16_to_uint16_dtype_0, x = var_1171_shape_cast_fp16)[name = string("cast_82")];
            uint16 gather_111_cast_uint16 = gather(axis = gather_111_axis_0, batch_dims = gather_111_batch_dims_0, indices = select_111_to_uint16, validate_indices = gather_111_validate_indices_0, x = var_1171_shape_cast_fp16_to_uint16)[name = string("gather_111_cast_uint16")];
            string gather_111_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_111_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_112_values0_0 = const()[name = string("concat_112_values0_0"), val = int32(1)];
            int32 concat_112_values1_0 = const()[name = string("concat_112_values1_0"), val = int32(1)];
            int32 concat_112_values2_0 = const()[name = string("concat_112_values2_0"), val = int32(0)];
            int32 concat_112_axis_0 = const()[name = string("concat_112_axis_0"), val = int32(0)];
            bool concat_112_interleave_0 = const()[name = string("concat_112_interleave_0"), val = bool(false)];
            int32 gather_111_cast_uint16_to_int32 = cast(dtype = gather_111_cast_uint16_to_int32_dtype_0, x = gather_111_cast_uint16)[name = string("cast_81")];
            tensor<int32, [4]> concat_112 = concat(axis = concat_112_axis_0, interleave = concat_112_interleave_0, values = (concat_112_values0_0, concat_112_values1_0, concat_112_values2_0, gather_111_cast_uint16_to_int32))[name = string("concat_112")];
            tensor<int32, [4]> causal_mask_13_begin_0 = const()[name = string("causal_mask_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_13_end_mask_0 = const()[name = string("causal_mask_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_13_cast_fp16 = slice_by_index(begin = causal_mask_13_begin_0, end = concat_112, end_mask = causal_mask_13_end_mask_0, x = causal_mask)[name = string("causal_mask_13_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_21_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_13_cast_fp16, key = key_states_23_cast_fp16, query = query_states_23_cast_fp16, value = value_states_23_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_1177_perm_0 = const()[name = string("op_1177_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            int32 gather_95_cast_uint16_to_int32 = cast(dtype = gather_95_cast_uint16_to_int32_dtype_0, x = gather_95_cast_uint16)[name = string("cast_80")];
            tensor<int32, [3]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (gather_94, gather_95_cast_uint16_to_int32, var_48))[name = string("concat_113")];
            tensor<fp16, [1, ?, 32, 64]> var_1177_cast_fp16 = transpose(perm = var_1177_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, ?, ?]> input_41_cast_fp16 = reshape(shape = concat_113, x = var_1177_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322411456))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324508672))))[name = string("model_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_169_cast_fp16 = add(x = hidden_states_149_cast_fp16, y = linear_38_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            fp16 var_54_promoted_11_to_fp16 = const()[name = string("op_54_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1186_cast_fp16 = pow(x = hidden_states_169_cast_fp16, y = var_54_promoted_11_to_fp16)[name = string("op_1186_cast_fp16")];
            tensor<int32, [1]> variance_23_axes_0 = const()[name = string("variance_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_23_keep_dims_0 = const()[name = string("variance_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_23_cast_fp16 = reduce_mean(axes = variance_23_axes_0, keep_dims = variance_23_keep_dims_0, x = var_1186_cast_fp16)[name = string("variance_23_cast_fp16")];
            fp16 var_1189_to_fp16 = const()[name = string("op_1189_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1190_cast_fp16 = add(x = variance_23_cast_fp16, y = var_1189_to_fp16)[name = string("op_1190_cast_fp16")];
            fp32 var_1191_epsilon_0 = const()[name = string("op_1191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1191_cast_fp16 = rsqrt(epsilon = var_1191_epsilon_0, x = var_1190_cast_fp16)[name = string("op_1191_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_173_cast_fp16 = mul(x = hidden_states_169_cast_fp16, y = var_1191_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324770880)))];
            tensor<fp16, [1, ?, 2048]> input_43_cast_fp16 = mul(x = model_model_layers_5_post_attention_layernorm_weight_to_fp16, y = hidden_states_173_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324775040))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333163712))))[name = string("model_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_1203_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334212352))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342601024))))[name = string("model_model_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_47_cast_fp16 = mul(x = var_1203_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343649664))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352038336))))[name = string("model_model_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_41_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_179_cast_fp16 = add(x = hidden_states_169_cast_fp16, y = linear_41_cast_fp16)[name = string("hidden_states_179_cast_fp16")];
            fp16 var_54_promoted_12_to_fp16 = const()[name = string("op_54_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1216_cast_fp16 = pow(x = hidden_states_179_cast_fp16, y = var_54_promoted_12_to_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [1]> variance_25_axes_0 = const()[name = string("variance_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_25_keep_dims_0 = const()[name = string("variance_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_25_cast_fp16 = reduce_mean(axes = variance_25_axes_0, keep_dims = variance_25_keep_dims_0, x = var_1216_cast_fp16)[name = string("variance_25_cast_fp16")];
            fp16 var_1219_to_fp16 = const()[name = string("op_1219_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1220_cast_fp16 = add(x = variance_25_cast_fp16, y = var_1219_to_fp16)[name = string("op_1220_cast_fp16")];
            fp32 var_1221_epsilon_0 = const()[name = string("op_1221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1221_cast_fp16 = rsqrt(epsilon = var_1221_epsilon_0, x = var_1220_cast_fp16)[name = string("op_1221_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_183_cast_fp16 = mul(x = hidden_states_179_cast_fp16, y = var_1221_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353086976)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_187_cast_fp16 = mul(x = model_model_layers_6_input_layernorm_weight_to_fp16, y = hidden_states_183_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [3]> var_1232_shape_cast_fp16 = shape(x = hidden_states_187_cast_fp16)[name = string("op_1232_shape_cast_fp16")];
            int32 gather_112 = const()[name = string("gather_112"), val = int32(1)];
            int32 gather_113_axis_0 = const()[name = string("gather_113_axis_0"), val = int32(0)];
            int32 gather_113_batch_dims_0 = const()[name = string("gather_113_batch_dims_0"), val = int32(0)];
            bool gather_113_validate_indices_0 = const()[name = string("gather_113_validate_indices_0"), val = bool(false)];
            string var_1232_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1232_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_113_to_uint16 = const()[name = string("select_113_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1232_shape_cast_fp16_to_uint16 = cast(dtype = var_1232_shape_cast_fp16_to_uint16_dtype_0, x = var_1232_shape_cast_fp16)[name = string("cast_79")];
            uint16 gather_113_cast_uint16 = gather(axis = gather_113_axis_0, batch_dims = gather_113_batch_dims_0, indices = select_113_to_uint16, validate_indices = gather_113_validate_indices_0, x = var_1232_shape_cast_fp16_to_uint16)[name = string("gather_113_cast_uint16")];
            string gather_113_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_113_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353091136))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355188352))))[name = string("model_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_187_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355450560))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355974912))))[name = string("model_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_187_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356040512))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356564864))))[name = string("model_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_187_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> concat_114x = const()[name = string("concat_114x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_1241_cast_fp16 = reshape(shape = concat_114x, x = linear_42_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_115x = const()[name = string("concat_115x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1244_cast_fp16 = reshape(shape = concat_115x, x = linear_43_cast_fp16)[name = string("op_1244_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_116x = const()[name = string("concat_116x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1247_cast_fp16 = reshape(shape = concat_116x, x = linear_44_cast_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [4]> v_state_13_perm_0 = const()[name = string("v_state_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1241_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 32, ?, 64]> var_1251_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1251_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_1262_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_1262_cast_fp16")];
            bool var_1264_interleave_0 = const()[name = string("op_1264_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_1264_cast_fp16 = concat(axis = var_48, interleave = var_1264_interleave_0, values = (var_1262_cast_fp16, x1_25_cast_fp16))[name = string("op_1264_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_1265_cast_fp16 = mul(x = var_1264_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1265_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_27_cast_fp16 = add(x = var_1251_cast_fp16, y = var_1265_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1244_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, ?, 64]> var_1267_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1267_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_1278_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_1278_cast_fp16")];
            bool var_1280_interleave_0 = const()[name = string("op_1280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_1280_cast_fp16 = concat(axis = var_48, interleave = var_1280_interleave_0, values = (var_1278_cast_fp16, x1_27_cast_fp16))[name = string("op_1280_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1281_cast_fp16 = mul(x = var_1280_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_13_cast_fp16 = add(x = var_1267_cast_fp16, y = var_1281_cast_fp16)[name = string("k_state_13_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_119_values0_0 = const()[name = string("concat_119_values0_0"), val = tensor<int32, [1]>([6])];
            int32 concat_119_axis_0 = const()[name = string("concat_119_axis_0"), val = int32(0)];
            bool concat_119_interleave_0 = const()[name = string("concat_119_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_119 = concat(axis = concat_119_axis_0, interleave = concat_119_interleave_0, values = (concat_119_values0_0, expand_dims_72, expand_dims_73, expand_dims_2, expand_dims_75))[name = string("concat_119")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_119, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = k_state_13_cast_fp16, x = coreml_update_state_42)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_44_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_44 = read_state(input = key_cache)[name = string("coreml_update_state_44")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_13_cast_fp16 = transpose(perm = v_state_13_perm_0, x = var_1247_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_119, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = v_state_13_cast_fp16, x = coreml_update_state_43)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_45_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_45 = read_state(input = value_cache)[name = string("coreml_update_state_45")];
            tensor<int32, [5]> var_1304_begin_0 = const()[name = string("op_1304_begin_0"), val = tensor<int32, [5]>([6, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1304_end_0 = const()[name = string("op_1304_end_0"), val = tensor<int32, [5]>([7, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1304_end_mask_0 = const()[name = string("op_1304_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1304_squeeze_mask_0 = const()[name = string("op_1304_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1304_cast_fp16 = slice_by_index(begin = var_1304_begin_0, end = var_1304_end_0, end_mask = var_1304_end_mask_0, squeeze_mask = var_1304_squeeze_mask_0, x = coreml_update_state_44)[name = string("op_1304_cast_fp16")];
            tensor<int32, [4]> var_1307_begin_0 = const()[name = string("op_1307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1307_end_mask_0 = const()[name = string("op_1307_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1307_cast_fp16 = slice_by_index(begin = var_1307_begin_0, end = concat_11, end_mask = var_1307_end_mask_0, x = var_1304_cast_fp16)[name = string("op_1307_cast_fp16")];
            tensor<int32, [5]> var_1309_begin_0 = const()[name = string("op_1309_begin_0"), val = tensor<int32, [5]>([6, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1309_end_0 = const()[name = string("op_1309_end_0"), val = tensor<int32, [5]>([7, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1309_end_mask_0 = const()[name = string("op_1309_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1309_squeeze_mask_0 = const()[name = string("op_1309_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1309_cast_fp16 = slice_by_index(begin = var_1309_begin_0, end = var_1309_end_0, end_mask = var_1309_end_mask_0, squeeze_mask = var_1309_squeeze_mask_0, x = coreml_update_state_45)[name = string("op_1309_cast_fp16")];
            tensor<int32, [4]> var_1312_begin_0 = const()[name = string("op_1312_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1312_end_mask_0 = const()[name = string("op_1312_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = concat_11, end_mask = var_1312_end_mask_0, x = var_1309_cast_fp16)[name = string("op_1312_cast_fp16")];
            tensor<int32, [4]> var_1314_shape_cast_fp16 = shape(x = var_1307_cast_fp16)[name = string("op_1314_shape_cast_fp16")];
            int32 gather_121 = const()[name = string("gather_121"), val = int32(1)];
            int32 gather_122 = const()[name = string("gather_122"), val = int32(8)];
            int32 gather_123_axis_0 = const()[name = string("gather_123_axis_0"), val = int32(0)];
            int32 gather_123_batch_dims_0 = const()[name = string("gather_123_batch_dims_0"), val = int32(0)];
            bool gather_123_validate_indices_0 = const()[name = string("gather_123_validate_indices_0"), val = bool(false)];
            string var_1314_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1314_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_123_to_uint16 = const()[name = string("select_123_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1314_shape_cast_fp16_to_uint16 = cast(dtype = var_1314_shape_cast_fp16_to_uint16_dtype_0, x = var_1314_shape_cast_fp16)[name = string("cast_78")];
            uint16 gather_123_cast_uint16 = gather(axis = gather_123_axis_0, batch_dims = gather_123_batch_dims_0, indices = select_123_to_uint16, validate_indices = gather_123_validate_indices_0, x = var_1314_shape_cast_fp16_to_uint16)[name = string("gather_123_cast_uint16")];
            string gather_123_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_123_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_124 = const()[name = string("gather_124"), val = int32(64)];
            tensor<int32, [1]> var_1321_axes_0 = const()[name = string("op_1321_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1321_cast_fp16 = expand_dims(axes = var_1321_axes_0, x = var_1307_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<int32, [5]> shape_137_cast_fp16 = shape(x = var_1321_cast_fp16)[name = string("shape_137_cast_fp16")];
            int32 concat_127_axis_0 = const()[name = string("concat_127_axis_0"), val = int32(0)];
            bool concat_127_interleave_0 = const()[name = string("concat_127_interleave_0"), val = bool(false)];
            int32 gather_123_cast_uint16_to_int32 = cast(dtype = gather_123_cast_uint16_to_int32_dtype_0, x = gather_123_cast_uint16)[name = string("cast_77")];
            tensor<int32, [5]> concat_127 = concat(axis = concat_127_axis_0, interleave = concat_127_interleave_0, values = (gather_121, gather_122, var_59, gather_123_cast_uint16_to_int32, gather_124))[name = string("concat_127")];
            tensor<int32, [5]> real_div_12 = real_div(x = concat_127, y = shape_137_cast_fp16)[name = string("real_div_12")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_191_cast_fp16 = tile(reps = real_div_12, x = var_1321_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> concat_128x = const()[name = string("concat_128x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_27_cast_fp16 = reshape(shape = concat_128x, x = hidden_states_191_cast_fp16)[name = string("key_states_27_cast_fp16")];
            tensor<int32, [4]> var_1331_shape_cast_fp16 = shape(x = var_1312_cast_fp16)[name = string("op_1331_shape_cast_fp16")];
            int32 gather_125 = const()[name = string("gather_125"), val = int32(1)];
            int32 gather_126 = const()[name = string("gather_126"), val = int32(8)];
            int32 gather_127_axis_0 = const()[name = string("gather_127_axis_0"), val = int32(0)];
            int32 gather_127_batch_dims_0 = const()[name = string("gather_127_batch_dims_0"), val = int32(0)];
            bool gather_127_validate_indices_0 = const()[name = string("gather_127_validate_indices_0"), val = bool(false)];
            string var_1331_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1331_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_127_to_uint16 = const()[name = string("select_127_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1331_shape_cast_fp16_to_uint16 = cast(dtype = var_1331_shape_cast_fp16_to_uint16_dtype_0, x = var_1331_shape_cast_fp16)[name = string("cast_76")];
            uint16 gather_127_cast_uint16 = gather(axis = gather_127_axis_0, batch_dims = gather_127_batch_dims_0, indices = select_127_to_uint16, validate_indices = gather_127_validate_indices_0, x = var_1331_shape_cast_fp16_to_uint16)[name = string("gather_127_cast_uint16")];
            string gather_127_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_127_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_128 = const()[name = string("gather_128"), val = int32(64)];
            tensor<int32, [1]> var_1338_axes_0 = const()[name = string("op_1338_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1338_cast_fp16 = expand_dims(axes = var_1338_axes_0, x = var_1312_cast_fp16)[name = string("op_1338_cast_fp16")];
            tensor<int32, [5]> shape_142_cast_fp16 = shape(x = var_1338_cast_fp16)[name = string("shape_142_cast_fp16")];
            int32 concat_129_axis_0 = const()[name = string("concat_129_axis_0"), val = int32(0)];
            bool concat_129_interleave_0 = const()[name = string("concat_129_interleave_0"), val = bool(false)];
            int32 gather_127_cast_uint16_to_int32 = cast(dtype = gather_127_cast_uint16_to_int32_dtype_0, x = gather_127_cast_uint16)[name = string("cast_75")];
            tensor<int32, [5]> concat_129 = concat(axis = concat_129_axis_0, interleave = concat_129_interleave_0, values = (gather_125, gather_126, var_59, gather_127_cast_uint16_to_int32, gather_128))[name = string("concat_129")];
            tensor<int32, [5]> real_div_13 = real_div(x = concat_129, y = shape_142_cast_fp16)[name = string("real_div_13")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_195_cast_fp16 = tile(reps = real_div_13, x = var_1338_cast_fp16)[name = string("hidden_states_195_cast_fp16")];
            tensor<int32, [4]> concat_130x = const()[name = string("concat_130x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_27_cast_fp16 = reshape(shape = concat_130x, x = hidden_states_195_cast_fp16)[name = string("value_states_27_cast_fp16")];
            tensor<int32, [4]> var_1348_shape_cast_fp16 = shape(x = key_states_27_cast_fp16)[name = string("op_1348_shape_cast_fp16")];
            int32 gather_129_axis_0 = const()[name = string("gather_129_axis_0"), val = int32(0)];
            int32 gather_129_batch_dims_0 = const()[name = string("gather_129_batch_dims_0"), val = int32(0)];
            bool gather_129_validate_indices_0 = const()[name = string("gather_129_validate_indices_0"), val = bool(false)];
            string var_1348_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1348_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_129_to_uint16 = const()[name = string("select_129_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1348_shape_cast_fp16_to_uint16 = cast(dtype = var_1348_shape_cast_fp16_to_uint16_dtype_0, x = var_1348_shape_cast_fp16)[name = string("cast_74")];
            uint16 gather_129_cast_uint16 = gather(axis = gather_129_axis_0, batch_dims = gather_129_batch_dims_0, indices = select_129_to_uint16, validate_indices = gather_129_validate_indices_0, x = var_1348_shape_cast_fp16_to_uint16)[name = string("gather_129_cast_uint16")];
            string gather_129_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_129_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_131_values0_0 = const()[name = string("concat_131_values0_0"), val = int32(1)];
            int32 concat_131_values1_0 = const()[name = string("concat_131_values1_0"), val = int32(1)];
            int32 concat_131_values2_0 = const()[name = string("concat_131_values2_0"), val = int32(0)];
            int32 concat_131_axis_0 = const()[name = string("concat_131_axis_0"), val = int32(0)];
            bool concat_131_interleave_0 = const()[name = string("concat_131_interleave_0"), val = bool(false)];
            int32 gather_129_cast_uint16_to_int32 = cast(dtype = gather_129_cast_uint16_to_int32_dtype_0, x = gather_129_cast_uint16)[name = string("cast_73")];
            tensor<int32, [4]> concat_131 = concat(axis = concat_131_axis_0, interleave = concat_131_interleave_0, values = (concat_131_values0_0, concat_131_values1_0, concat_131_values2_0, gather_129_cast_uint16_to_int32))[name = string("concat_131")];
            tensor<int32, [4]> causal_mask_15_begin_0 = const()[name = string("causal_mask_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_15_end_mask_0 = const()[name = string("causal_mask_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_15_cast_fp16 = slice_by_index(begin = causal_mask_15_begin_0, end = concat_131, end_mask = causal_mask_15_end_mask_0, x = causal_mask)[name = string("causal_mask_15_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_25_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_15_cast_fp16, key = key_states_27_cast_fp16, query = query_states_27_cast_fp16, value = value_states_27_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1354_perm_0 = const()[name = string("op_1354_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_132_axis_0 = const()[name = string("concat_132_axis_0"), val = int32(0)];
            bool concat_132_interleave_0 = const()[name = string("concat_132_interleave_0"), val = bool(false)];
            int32 gather_113_cast_uint16_to_int32 = cast(dtype = gather_113_cast_uint16_to_int32_dtype_0, x = gather_113_cast_uint16)[name = string("cast_72")];
            tensor<int32, [3]> concat_132 = concat(axis = concat_132_axis_0, interleave = concat_132_interleave_0, values = (gather_112, gather_113_cast_uint16_to_int32, var_48))[name = string("concat_132")];
            tensor<fp16, [1, ?, 32, 64]> var_1354_cast_fp16 = transpose(perm = var_1354_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, ?, ?]> input_49_cast_fp16 = reshape(shape = concat_132, x = var_1354_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356630464))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358727680))))[name = string("model_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_199_cast_fp16 = add(x = hidden_states_179_cast_fp16, y = linear_45_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            fp16 var_54_promoted_13_to_fp16 = const()[name = string("op_54_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1363_cast_fp16 = pow(x = hidden_states_199_cast_fp16, y = var_54_promoted_13_to_fp16)[name = string("op_1363_cast_fp16")];
            tensor<int32, [1]> variance_27_axes_0 = const()[name = string("variance_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_27_keep_dims_0 = const()[name = string("variance_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_27_cast_fp16 = reduce_mean(axes = variance_27_axes_0, keep_dims = variance_27_keep_dims_0, x = var_1363_cast_fp16)[name = string("variance_27_cast_fp16")];
            fp16 var_1366_to_fp16 = const()[name = string("op_1366_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1367_cast_fp16 = add(x = variance_27_cast_fp16, y = var_1366_to_fp16)[name = string("op_1367_cast_fp16")];
            fp32 var_1368_epsilon_0 = const()[name = string("op_1368_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1368_cast_fp16 = rsqrt(epsilon = var_1368_epsilon_0, x = var_1367_cast_fp16)[name = string("op_1368_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_203_cast_fp16 = mul(x = hidden_states_199_cast_fp16, y = var_1368_cast_fp16)[name = string("hidden_states_203_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358989888)))];
            tensor<fp16, [1, ?, 2048]> input_51_cast_fp16 = mul(x = model_model_layers_6_post_attention_layernorm_weight_to_fp16, y = hidden_states_203_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358994048))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367382720))))[name = string("model_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_1380_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1380_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368431360))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376820032))))[name = string("model_model_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_55_cast_fp16 = mul(x = var_1380_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377868672))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386257344))))[name = string("model_model_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_209_cast_fp16 = add(x = hidden_states_199_cast_fp16, y = linear_48_cast_fp16)[name = string("hidden_states_209_cast_fp16")];
            fp16 var_54_promoted_14_to_fp16 = const()[name = string("op_54_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1393_cast_fp16 = pow(x = hidden_states_209_cast_fp16, y = var_54_promoted_14_to_fp16)[name = string("op_1393_cast_fp16")];
            tensor<int32, [1]> variance_29_axes_0 = const()[name = string("variance_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_29_keep_dims_0 = const()[name = string("variance_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_29_cast_fp16 = reduce_mean(axes = variance_29_axes_0, keep_dims = variance_29_keep_dims_0, x = var_1393_cast_fp16)[name = string("variance_29_cast_fp16")];
            fp16 var_1396_to_fp16 = const()[name = string("op_1396_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1397_cast_fp16 = add(x = variance_29_cast_fp16, y = var_1396_to_fp16)[name = string("op_1397_cast_fp16")];
            fp32 var_1398_epsilon_0 = const()[name = string("op_1398_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1398_cast_fp16 = rsqrt(epsilon = var_1398_epsilon_0, x = var_1397_cast_fp16)[name = string("op_1398_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_213_cast_fp16 = mul(x = hidden_states_209_cast_fp16, y = var_1398_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387305984)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_217_cast_fp16 = mul(x = model_model_layers_7_input_layernorm_weight_to_fp16, y = hidden_states_213_cast_fp16)[name = string("hidden_states_217_cast_fp16")];
            tensor<int32, [3]> var_1409_shape_cast_fp16 = shape(x = hidden_states_217_cast_fp16)[name = string("op_1409_shape_cast_fp16")];
            int32 gather_130 = const()[name = string("gather_130"), val = int32(1)];
            int32 gather_131_axis_0 = const()[name = string("gather_131_axis_0"), val = int32(0)];
            int32 gather_131_batch_dims_0 = const()[name = string("gather_131_batch_dims_0"), val = int32(0)];
            bool gather_131_validate_indices_0 = const()[name = string("gather_131_validate_indices_0"), val = bool(false)];
            string var_1409_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1409_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_131_to_uint16 = const()[name = string("select_131_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1409_shape_cast_fp16_to_uint16 = cast(dtype = var_1409_shape_cast_fp16_to_uint16_dtype_0, x = var_1409_shape_cast_fp16)[name = string("cast_71")];
            uint16 gather_131_cast_uint16 = gather(axis = gather_131_axis_0, batch_dims = gather_131_batch_dims_0, indices = select_131_to_uint16, validate_indices = gather_131_validate_indices_0, x = var_1409_shape_cast_fp16_to_uint16)[name = string("gather_131_cast_uint16")];
            string gather_131_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_131_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387310144))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389407360))))[name = string("model_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_217_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389669568))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390193920))))[name = string("model_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_217_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390259520))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390783872))))[name = string("model_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_217_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_1418_cast_fp16 = reshape(shape = concat_133x, x = linear_49_cast_fp16)[name = string("op_1418_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1421_cast_fp16 = reshape(shape = concat_134x, x = linear_50_cast_fp16)[name = string("op_1421_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_135x = const()[name = string("concat_135x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1424_cast_fp16 = reshape(shape = concat_135x, x = linear_51_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<int32, [4]> v_state_15_perm_0 = const()[name = string("v_state_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1418_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 32, ?, 64]> var_1428_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1428_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_1439_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_1439_cast_fp16")];
            bool var_1441_interleave_0 = const()[name = string("op_1441_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_1441_cast_fp16 = concat(axis = var_48, interleave = var_1441_interleave_0, values = (var_1439_cast_fp16, x1_29_cast_fp16))[name = string("op_1441_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_1442_cast_fp16 = mul(x = var_1441_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1442_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_31_cast_fp16 = add(x = var_1428_cast_fp16, y = var_1442_cast_fp16)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1421_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 8, ?, 64]> var_1444_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1444_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_1455_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_1455_cast_fp16")];
            bool var_1457_interleave_0 = const()[name = string("op_1457_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_1457_cast_fp16 = concat(axis = var_48, interleave = var_1457_interleave_0, values = (var_1455_cast_fp16, x1_31_cast_fp16))[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1458_cast_fp16 = mul(x = var_1457_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_15_cast_fp16 = add(x = var_1444_cast_fp16, y = var_1458_cast_fp16)[name = string("k_state_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values0_0 = const()[name = string("concat_138_values0_0"), val = tensor<int32, [1]>([7])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (concat_138_values0_0, expand_dims_84, expand_dims_85, expand_dims_2, expand_dims_87))[name = string("concat_138")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_138, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = k_state_15_cast_fp16, x = coreml_update_state_44)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_46_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_46 = read_state(input = key_cache)[name = string("coreml_update_state_46")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_15_cast_fp16 = transpose(perm = v_state_15_perm_0, x = var_1424_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_138, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = v_state_15_cast_fp16, x = coreml_update_state_45)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_47_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_47 = read_state(input = value_cache)[name = string("coreml_update_state_47")];
            tensor<int32, [5]> var_1481_begin_0 = const()[name = string("op_1481_begin_0"), val = tensor<int32, [5]>([7, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1481_end_0 = const()[name = string("op_1481_end_0"), val = tensor<int32, [5]>([8, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1481_end_mask_0 = const()[name = string("op_1481_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1481_squeeze_mask_0 = const()[name = string("op_1481_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1481_cast_fp16 = slice_by_index(begin = var_1481_begin_0, end = var_1481_end_0, end_mask = var_1481_end_mask_0, squeeze_mask = var_1481_squeeze_mask_0, x = coreml_update_state_46)[name = string("op_1481_cast_fp16")];
            tensor<int32, [4]> var_1484_begin_0 = const()[name = string("op_1484_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1484_end_mask_0 = const()[name = string("op_1484_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1484_cast_fp16 = slice_by_index(begin = var_1484_begin_0, end = concat_11, end_mask = var_1484_end_mask_0, x = var_1481_cast_fp16)[name = string("op_1484_cast_fp16")];
            tensor<int32, [5]> var_1486_begin_0 = const()[name = string("op_1486_begin_0"), val = tensor<int32, [5]>([7, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1486_end_0 = const()[name = string("op_1486_end_0"), val = tensor<int32, [5]>([8, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1486_end_mask_0 = const()[name = string("op_1486_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1486_squeeze_mask_0 = const()[name = string("op_1486_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1486_cast_fp16 = slice_by_index(begin = var_1486_begin_0, end = var_1486_end_0, end_mask = var_1486_end_mask_0, squeeze_mask = var_1486_squeeze_mask_0, x = coreml_update_state_47)[name = string("op_1486_cast_fp16")];
            tensor<int32, [4]> var_1489_begin_0 = const()[name = string("op_1489_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1489_end_mask_0 = const()[name = string("op_1489_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1489_cast_fp16 = slice_by_index(begin = var_1489_begin_0, end = concat_11, end_mask = var_1489_end_mask_0, x = var_1486_cast_fp16)[name = string("op_1489_cast_fp16")];
            tensor<int32, [4]> var_1491_shape_cast_fp16 = shape(x = var_1484_cast_fp16)[name = string("op_1491_shape_cast_fp16")];
            int32 gather_139 = const()[name = string("gather_139"), val = int32(1)];
            int32 gather_140 = const()[name = string("gather_140"), val = int32(8)];
            int32 gather_141_axis_0 = const()[name = string("gather_141_axis_0"), val = int32(0)];
            int32 gather_141_batch_dims_0 = const()[name = string("gather_141_batch_dims_0"), val = int32(0)];
            bool gather_141_validate_indices_0 = const()[name = string("gather_141_validate_indices_0"), val = bool(false)];
            string var_1491_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1491_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_141_to_uint16 = const()[name = string("select_141_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1491_shape_cast_fp16_to_uint16 = cast(dtype = var_1491_shape_cast_fp16_to_uint16_dtype_0, x = var_1491_shape_cast_fp16)[name = string("cast_70")];
            uint16 gather_141_cast_uint16 = gather(axis = gather_141_axis_0, batch_dims = gather_141_batch_dims_0, indices = select_141_to_uint16, validate_indices = gather_141_validate_indices_0, x = var_1491_shape_cast_fp16_to_uint16)[name = string("gather_141_cast_uint16")];
            string gather_141_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_141_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_142 = const()[name = string("gather_142"), val = int32(64)];
            tensor<int32, [1]> var_1498_axes_0 = const()[name = string("op_1498_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1498_cast_fp16 = expand_dims(axes = var_1498_axes_0, x = var_1484_cast_fp16)[name = string("op_1498_cast_fp16")];
            tensor<int32, [5]> shape_157_cast_fp16 = shape(x = var_1498_cast_fp16)[name = string("shape_157_cast_fp16")];
            int32 concat_146_axis_0 = const()[name = string("concat_146_axis_0"), val = int32(0)];
            bool concat_146_interleave_0 = const()[name = string("concat_146_interleave_0"), val = bool(false)];
            int32 gather_141_cast_uint16_to_int32 = cast(dtype = gather_141_cast_uint16_to_int32_dtype_0, x = gather_141_cast_uint16)[name = string("cast_69")];
            tensor<int32, [5]> concat_146 = concat(axis = concat_146_axis_0, interleave = concat_146_interleave_0, values = (gather_139, gather_140, var_59, gather_141_cast_uint16_to_int32, gather_142))[name = string("concat_146")];
            tensor<int32, [5]> real_div_14 = real_div(x = concat_146, y = shape_157_cast_fp16)[name = string("real_div_14")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_221_cast_fp16 = tile(reps = real_div_14, x = var_1498_cast_fp16)[name = string("hidden_states_221_cast_fp16")];
            tensor<int32, [4]> concat_147x = const()[name = string("concat_147x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_31_cast_fp16 = reshape(shape = concat_147x, x = hidden_states_221_cast_fp16)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [4]> var_1508_shape_cast_fp16 = shape(x = var_1489_cast_fp16)[name = string("op_1508_shape_cast_fp16")];
            int32 gather_143 = const()[name = string("gather_143"), val = int32(1)];
            int32 gather_144 = const()[name = string("gather_144"), val = int32(8)];
            int32 gather_145_axis_0 = const()[name = string("gather_145_axis_0"), val = int32(0)];
            int32 gather_145_batch_dims_0 = const()[name = string("gather_145_batch_dims_0"), val = int32(0)];
            bool gather_145_validate_indices_0 = const()[name = string("gather_145_validate_indices_0"), val = bool(false)];
            string var_1508_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1508_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_145_to_uint16 = const()[name = string("select_145_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1508_shape_cast_fp16_to_uint16 = cast(dtype = var_1508_shape_cast_fp16_to_uint16_dtype_0, x = var_1508_shape_cast_fp16)[name = string("cast_68")];
            uint16 gather_145_cast_uint16 = gather(axis = gather_145_axis_0, batch_dims = gather_145_batch_dims_0, indices = select_145_to_uint16, validate_indices = gather_145_validate_indices_0, x = var_1508_shape_cast_fp16_to_uint16)[name = string("gather_145_cast_uint16")];
            string gather_145_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_145_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_146 = const()[name = string("gather_146"), val = int32(64)];
            tensor<int32, [1]> var_1515_axes_0 = const()[name = string("op_1515_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1515_cast_fp16 = expand_dims(axes = var_1515_axes_0, x = var_1489_cast_fp16)[name = string("op_1515_cast_fp16")];
            tensor<int32, [5]> shape_162_cast_fp16 = shape(x = var_1515_cast_fp16)[name = string("shape_162_cast_fp16")];
            int32 concat_148_axis_0 = const()[name = string("concat_148_axis_0"), val = int32(0)];
            bool concat_148_interleave_0 = const()[name = string("concat_148_interleave_0"), val = bool(false)];
            int32 gather_145_cast_uint16_to_int32 = cast(dtype = gather_145_cast_uint16_to_int32_dtype_0, x = gather_145_cast_uint16)[name = string("cast_67")];
            tensor<int32, [5]> concat_148 = concat(axis = concat_148_axis_0, interleave = concat_148_interleave_0, values = (gather_143, gather_144, var_59, gather_145_cast_uint16_to_int32, gather_146))[name = string("concat_148")];
            tensor<int32, [5]> real_div_15 = real_div(x = concat_148, y = shape_162_cast_fp16)[name = string("real_div_15")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_225_cast_fp16 = tile(reps = real_div_15, x = var_1515_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<int32, [4]> concat_149x = const()[name = string("concat_149x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_31_cast_fp16 = reshape(shape = concat_149x, x = hidden_states_225_cast_fp16)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> var_1525_shape_cast_fp16 = shape(x = key_states_31_cast_fp16)[name = string("op_1525_shape_cast_fp16")];
            int32 gather_147_axis_0 = const()[name = string("gather_147_axis_0"), val = int32(0)];
            int32 gather_147_batch_dims_0 = const()[name = string("gather_147_batch_dims_0"), val = int32(0)];
            bool gather_147_validate_indices_0 = const()[name = string("gather_147_validate_indices_0"), val = bool(false)];
            string var_1525_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1525_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_147_to_uint16 = const()[name = string("select_147_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1525_shape_cast_fp16_to_uint16 = cast(dtype = var_1525_shape_cast_fp16_to_uint16_dtype_0, x = var_1525_shape_cast_fp16)[name = string("cast_66")];
            uint16 gather_147_cast_uint16 = gather(axis = gather_147_axis_0, batch_dims = gather_147_batch_dims_0, indices = select_147_to_uint16, validate_indices = gather_147_validate_indices_0, x = var_1525_shape_cast_fp16_to_uint16)[name = string("gather_147_cast_uint16")];
            string gather_147_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_147_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_150_values0_0 = const()[name = string("concat_150_values0_0"), val = int32(1)];
            int32 concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = int32(1)];
            int32 concat_150_values2_0 = const()[name = string("concat_150_values2_0"), val = int32(0)];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            int32 gather_147_cast_uint16_to_int32 = cast(dtype = gather_147_cast_uint16_to_int32_dtype_0, x = gather_147_cast_uint16)[name = string("cast_65")];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (concat_150_values0_0, concat_150_values1_0, concat_150_values2_0, gather_147_cast_uint16_to_int32))[name = string("concat_150")];
            tensor<int32, [4]> causal_mask_17_begin_0 = const()[name = string("causal_mask_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_17_end_mask_0 = const()[name = string("causal_mask_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_17_cast_fp16 = slice_by_index(begin = causal_mask_17_begin_0, end = concat_150, end_mask = causal_mask_17_end_mask_0, x = causal_mask)[name = string("causal_mask_17_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_29_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_17_cast_fp16, key = key_states_31_cast_fp16, query = query_states_31_cast_fp16, value = value_states_31_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1531_perm_0 = const()[name = string("op_1531_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_151_axis_0 = const()[name = string("concat_151_axis_0"), val = int32(0)];
            bool concat_151_interleave_0 = const()[name = string("concat_151_interleave_0"), val = bool(false)];
            int32 gather_131_cast_uint16_to_int32 = cast(dtype = gather_131_cast_uint16_to_int32_dtype_0, x = gather_131_cast_uint16)[name = string("cast_64")];
            tensor<int32, [3]> concat_151 = concat(axis = concat_151_axis_0, interleave = concat_151_interleave_0, values = (gather_130, gather_131_cast_uint16_to_int32, var_48))[name = string("concat_151")];
            tensor<fp16, [1, ?, 32, 64]> var_1531_cast_fp16 = transpose(perm = var_1531_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, ?, ?]> input_57_cast_fp16 = reshape(shape = concat_151, x = var_1531_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390849472))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392946688))))[name = string("model_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_229_cast_fp16 = add(x = hidden_states_209_cast_fp16, y = linear_52_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            fp16 var_54_promoted_15_to_fp16 = const()[name = string("op_54_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1540_cast_fp16 = pow(x = hidden_states_229_cast_fp16, y = var_54_promoted_15_to_fp16)[name = string("op_1540_cast_fp16")];
            tensor<int32, [1]> variance_31_axes_0 = const()[name = string("variance_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_31_keep_dims_0 = const()[name = string("variance_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_31_cast_fp16 = reduce_mean(axes = variance_31_axes_0, keep_dims = variance_31_keep_dims_0, x = var_1540_cast_fp16)[name = string("variance_31_cast_fp16")];
            fp16 var_1543_to_fp16 = const()[name = string("op_1543_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1544_cast_fp16 = add(x = variance_31_cast_fp16, y = var_1543_to_fp16)[name = string("op_1544_cast_fp16")];
            fp32 var_1545_epsilon_0 = const()[name = string("op_1545_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1545_cast_fp16 = rsqrt(epsilon = var_1545_epsilon_0, x = var_1544_cast_fp16)[name = string("op_1545_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_233_cast_fp16 = mul(x = hidden_states_229_cast_fp16, y = var_1545_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393208896)))];
            tensor<fp16, [1, ?, 2048]> input_59_cast_fp16 = mul(x = model_model_layers_7_post_attention_layernorm_weight_to_fp16, y = hidden_states_233_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393213056))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401601728))))[name = string("model_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_1557_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1557_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402650368))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411039040))))[name = string("model_model_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_63_cast_fp16 = mul(x = var_1557_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412087680))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420476352))))[name = string("model_model_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_55_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_239_cast_fp16 = add(x = hidden_states_229_cast_fp16, y = linear_55_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            fp16 var_54_promoted_16_to_fp16 = const()[name = string("op_54_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1570_cast_fp16 = pow(x = hidden_states_239_cast_fp16, y = var_54_promoted_16_to_fp16)[name = string("op_1570_cast_fp16")];
            tensor<int32, [1]> variance_33_axes_0 = const()[name = string("variance_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_33_keep_dims_0 = const()[name = string("variance_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_33_cast_fp16 = reduce_mean(axes = variance_33_axes_0, keep_dims = variance_33_keep_dims_0, x = var_1570_cast_fp16)[name = string("variance_33_cast_fp16")];
            fp16 var_1573_to_fp16 = const()[name = string("op_1573_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1574_cast_fp16 = add(x = variance_33_cast_fp16, y = var_1573_to_fp16)[name = string("op_1574_cast_fp16")];
            fp32 var_1575_epsilon_0 = const()[name = string("op_1575_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1575_cast_fp16 = rsqrt(epsilon = var_1575_epsilon_0, x = var_1574_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_243_cast_fp16 = mul(x = hidden_states_239_cast_fp16, y = var_1575_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421524992)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_247_cast_fp16 = mul(x = model_model_layers_8_input_layernorm_weight_to_fp16, y = hidden_states_243_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [3]> var_1586_shape_cast_fp16 = shape(x = hidden_states_247_cast_fp16)[name = string("op_1586_shape_cast_fp16")];
            int32 gather_148 = const()[name = string("gather_148"), val = int32(1)];
            int32 gather_149_axis_0 = const()[name = string("gather_149_axis_0"), val = int32(0)];
            int32 gather_149_batch_dims_0 = const()[name = string("gather_149_batch_dims_0"), val = int32(0)];
            bool gather_149_validate_indices_0 = const()[name = string("gather_149_validate_indices_0"), val = bool(false)];
            string var_1586_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1586_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_149_to_uint16 = const()[name = string("select_149_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1586_shape_cast_fp16_to_uint16 = cast(dtype = var_1586_shape_cast_fp16_to_uint16_dtype_0, x = var_1586_shape_cast_fp16)[name = string("cast_63")];
            uint16 gather_149_cast_uint16 = gather(axis = gather_149_axis_0, batch_dims = gather_149_batch_dims_0, indices = select_149_to_uint16, validate_indices = gather_149_validate_indices_0, x = var_1586_shape_cast_fp16_to_uint16)[name = string("gather_149_cast_uint16")];
            string gather_149_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_149_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421529152))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423626368))))[name = string("model_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_247_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423888576))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424412928))))[name = string("model_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_247_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424478528))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425002880))))[name = string("model_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_247_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> concat_152x = const()[name = string("concat_152x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_1595_cast_fp16 = reshape(shape = concat_152x, x = linear_56_cast_fp16)[name = string("op_1595_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_153x = const()[name = string("concat_153x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1598_cast_fp16 = reshape(shape = concat_153x, x = linear_57_cast_fp16)[name = string("op_1598_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_154x = const()[name = string("concat_154x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1601_cast_fp16 = reshape(shape = concat_154x, x = linear_58_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [4]> v_state_17_perm_0 = const()[name = string("v_state_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1595_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 32, ?, 64]> var_1605_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1605_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_1616_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_1616_cast_fp16")];
            bool var_1618_interleave_0 = const()[name = string("op_1618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_1618_cast_fp16 = concat(axis = var_48, interleave = var_1618_interleave_0, values = (var_1616_cast_fp16, x1_33_cast_fp16))[name = string("op_1618_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_1619_cast_fp16 = mul(x = var_1618_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1619_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_35_cast_fp16 = add(x = var_1605_cast_fp16, y = var_1619_cast_fp16)[name = string("query_states_35_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1598_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 8, ?, 64]> var_1621_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1621_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_1632_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1632_cast_fp16")];
            bool var_1634_interleave_0 = const()[name = string("op_1634_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_1634_cast_fp16 = concat(axis = var_48, interleave = var_1634_interleave_0, values = (var_1632_cast_fp16, x1_35_cast_fp16))[name = string("op_1634_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1635_cast_fp16 = mul(x = var_1634_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1635_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_17_cast_fp16 = add(x = var_1621_cast_fp16, y = var_1635_cast_fp16)[name = string("k_state_17_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_157_values0_0 = const()[name = string("concat_157_values0_0"), val = tensor<int32, [1]>([8])];
            int32 concat_157_axis_0 = const()[name = string("concat_157_axis_0"), val = int32(0)];
            bool concat_157_interleave_0 = const()[name = string("concat_157_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_157 = concat(axis = concat_157_axis_0, interleave = concat_157_interleave_0, values = (concat_157_values0_0, expand_dims_96, expand_dims_97, expand_dims_2, expand_dims_99))[name = string("concat_157")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_157, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = k_state_17_cast_fp16, x = coreml_update_state_46)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_48_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_48 = read_state(input = key_cache)[name = string("coreml_update_state_48")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_17_cast_fp16 = transpose(perm = v_state_17_perm_0, x = var_1601_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_157, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = v_state_17_cast_fp16, x = coreml_update_state_47)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_49_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_49 = read_state(input = value_cache)[name = string("coreml_update_state_49")];
            tensor<int32, [5]> var_1658_begin_0 = const()[name = string("op_1658_begin_0"), val = tensor<int32, [5]>([8, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1658_end_0 = const()[name = string("op_1658_end_0"), val = tensor<int32, [5]>([9, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1658_end_mask_0 = const()[name = string("op_1658_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1658_squeeze_mask_0 = const()[name = string("op_1658_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1658_cast_fp16 = slice_by_index(begin = var_1658_begin_0, end = var_1658_end_0, end_mask = var_1658_end_mask_0, squeeze_mask = var_1658_squeeze_mask_0, x = coreml_update_state_48)[name = string("op_1658_cast_fp16")];
            tensor<int32, [4]> var_1661_begin_0 = const()[name = string("op_1661_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1661_end_mask_0 = const()[name = string("op_1661_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1661_cast_fp16 = slice_by_index(begin = var_1661_begin_0, end = concat_11, end_mask = var_1661_end_mask_0, x = var_1658_cast_fp16)[name = string("op_1661_cast_fp16")];
            tensor<int32, [5]> var_1663_begin_0 = const()[name = string("op_1663_begin_0"), val = tensor<int32, [5]>([8, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1663_end_0 = const()[name = string("op_1663_end_0"), val = tensor<int32, [5]>([9, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1663_end_mask_0 = const()[name = string("op_1663_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1663_squeeze_mask_0 = const()[name = string("op_1663_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1663_cast_fp16 = slice_by_index(begin = var_1663_begin_0, end = var_1663_end_0, end_mask = var_1663_end_mask_0, squeeze_mask = var_1663_squeeze_mask_0, x = coreml_update_state_49)[name = string("op_1663_cast_fp16")];
            tensor<int32, [4]> var_1666_begin_0 = const()[name = string("op_1666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1666_end_mask_0 = const()[name = string("op_1666_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1666_cast_fp16 = slice_by_index(begin = var_1666_begin_0, end = concat_11, end_mask = var_1666_end_mask_0, x = var_1663_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<int32, [4]> var_1668_shape_cast_fp16 = shape(x = var_1661_cast_fp16)[name = string("op_1668_shape_cast_fp16")];
            int32 gather_157 = const()[name = string("gather_157"), val = int32(1)];
            int32 gather_158 = const()[name = string("gather_158"), val = int32(8)];
            int32 gather_159_axis_0 = const()[name = string("gather_159_axis_0"), val = int32(0)];
            int32 gather_159_batch_dims_0 = const()[name = string("gather_159_batch_dims_0"), val = int32(0)];
            bool gather_159_validate_indices_0 = const()[name = string("gather_159_validate_indices_0"), val = bool(false)];
            string var_1668_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1668_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_159_to_uint16 = const()[name = string("select_159_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1668_shape_cast_fp16_to_uint16 = cast(dtype = var_1668_shape_cast_fp16_to_uint16_dtype_0, x = var_1668_shape_cast_fp16)[name = string("cast_62")];
            uint16 gather_159_cast_uint16 = gather(axis = gather_159_axis_0, batch_dims = gather_159_batch_dims_0, indices = select_159_to_uint16, validate_indices = gather_159_validate_indices_0, x = var_1668_shape_cast_fp16_to_uint16)[name = string("gather_159_cast_uint16")];
            string gather_159_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_159_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_160 = const()[name = string("gather_160"), val = int32(64)];
            tensor<int32, [1]> var_1675_axes_0 = const()[name = string("op_1675_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1675_cast_fp16 = expand_dims(axes = var_1675_axes_0, x = var_1661_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [5]> shape_177_cast_fp16 = shape(x = var_1675_cast_fp16)[name = string("shape_177_cast_fp16")];
            int32 concat_165_axis_0 = const()[name = string("concat_165_axis_0"), val = int32(0)];
            bool concat_165_interleave_0 = const()[name = string("concat_165_interleave_0"), val = bool(false)];
            int32 gather_159_cast_uint16_to_int32 = cast(dtype = gather_159_cast_uint16_to_int32_dtype_0, x = gather_159_cast_uint16)[name = string("cast_61")];
            tensor<int32, [5]> concat_165 = concat(axis = concat_165_axis_0, interleave = concat_165_interleave_0, values = (gather_157, gather_158, var_59, gather_159_cast_uint16_to_int32, gather_160))[name = string("concat_165")];
            tensor<int32, [5]> real_div_16 = real_div(x = concat_165, y = shape_177_cast_fp16)[name = string("real_div_16")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_251_cast_fp16 = tile(reps = real_div_16, x = var_1675_cast_fp16)[name = string("hidden_states_251_cast_fp16")];
            tensor<int32, [4]> concat_166x = const()[name = string("concat_166x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_35_cast_fp16 = reshape(shape = concat_166x, x = hidden_states_251_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [4]> var_1685_shape_cast_fp16 = shape(x = var_1666_cast_fp16)[name = string("op_1685_shape_cast_fp16")];
            int32 gather_161 = const()[name = string("gather_161"), val = int32(1)];
            int32 gather_162 = const()[name = string("gather_162"), val = int32(8)];
            int32 gather_163_axis_0 = const()[name = string("gather_163_axis_0"), val = int32(0)];
            int32 gather_163_batch_dims_0 = const()[name = string("gather_163_batch_dims_0"), val = int32(0)];
            bool gather_163_validate_indices_0 = const()[name = string("gather_163_validate_indices_0"), val = bool(false)];
            string var_1685_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1685_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_163_to_uint16 = const()[name = string("select_163_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1685_shape_cast_fp16_to_uint16 = cast(dtype = var_1685_shape_cast_fp16_to_uint16_dtype_0, x = var_1685_shape_cast_fp16)[name = string("cast_60")];
            uint16 gather_163_cast_uint16 = gather(axis = gather_163_axis_0, batch_dims = gather_163_batch_dims_0, indices = select_163_to_uint16, validate_indices = gather_163_validate_indices_0, x = var_1685_shape_cast_fp16_to_uint16)[name = string("gather_163_cast_uint16")];
            string gather_163_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_163_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_164 = const()[name = string("gather_164"), val = int32(64)];
            tensor<int32, [1]> var_1692_axes_0 = const()[name = string("op_1692_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1692_cast_fp16 = expand_dims(axes = var_1692_axes_0, x = var_1666_cast_fp16)[name = string("op_1692_cast_fp16")];
            tensor<int32, [5]> shape_182_cast_fp16 = shape(x = var_1692_cast_fp16)[name = string("shape_182_cast_fp16")];
            int32 concat_167_axis_0 = const()[name = string("concat_167_axis_0"), val = int32(0)];
            bool concat_167_interleave_0 = const()[name = string("concat_167_interleave_0"), val = bool(false)];
            int32 gather_163_cast_uint16_to_int32 = cast(dtype = gather_163_cast_uint16_to_int32_dtype_0, x = gather_163_cast_uint16)[name = string("cast_59")];
            tensor<int32, [5]> concat_167 = concat(axis = concat_167_axis_0, interleave = concat_167_interleave_0, values = (gather_161, gather_162, var_59, gather_163_cast_uint16_to_int32, gather_164))[name = string("concat_167")];
            tensor<int32, [5]> real_div_17 = real_div(x = concat_167, y = shape_182_cast_fp16)[name = string("real_div_17")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_255_cast_fp16 = tile(reps = real_div_17, x = var_1692_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [4]> concat_168x = const()[name = string("concat_168x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_35_cast_fp16 = reshape(shape = concat_168x, x = hidden_states_255_cast_fp16)[name = string("value_states_35_cast_fp16")];
            tensor<int32, [4]> var_1702_shape_cast_fp16 = shape(x = key_states_35_cast_fp16)[name = string("op_1702_shape_cast_fp16")];
            int32 gather_165_axis_0 = const()[name = string("gather_165_axis_0"), val = int32(0)];
            int32 gather_165_batch_dims_0 = const()[name = string("gather_165_batch_dims_0"), val = int32(0)];
            bool gather_165_validate_indices_0 = const()[name = string("gather_165_validate_indices_0"), val = bool(false)];
            string var_1702_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1702_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_165_to_uint16 = const()[name = string("select_165_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1702_shape_cast_fp16_to_uint16 = cast(dtype = var_1702_shape_cast_fp16_to_uint16_dtype_0, x = var_1702_shape_cast_fp16)[name = string("cast_58")];
            uint16 gather_165_cast_uint16 = gather(axis = gather_165_axis_0, batch_dims = gather_165_batch_dims_0, indices = select_165_to_uint16, validate_indices = gather_165_validate_indices_0, x = var_1702_shape_cast_fp16_to_uint16)[name = string("gather_165_cast_uint16")];
            string gather_165_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_165_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_169_values0_0 = const()[name = string("concat_169_values0_0"), val = int32(1)];
            int32 concat_169_values1_0 = const()[name = string("concat_169_values1_0"), val = int32(1)];
            int32 concat_169_values2_0 = const()[name = string("concat_169_values2_0"), val = int32(0)];
            int32 concat_169_axis_0 = const()[name = string("concat_169_axis_0"), val = int32(0)];
            bool concat_169_interleave_0 = const()[name = string("concat_169_interleave_0"), val = bool(false)];
            int32 gather_165_cast_uint16_to_int32 = cast(dtype = gather_165_cast_uint16_to_int32_dtype_0, x = gather_165_cast_uint16)[name = string("cast_57")];
            tensor<int32, [4]> concat_169 = concat(axis = concat_169_axis_0, interleave = concat_169_interleave_0, values = (concat_169_values0_0, concat_169_values1_0, concat_169_values2_0, gather_165_cast_uint16_to_int32))[name = string("concat_169")];
            tensor<int32, [4]> causal_mask_19_begin_0 = const()[name = string("causal_mask_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_19_end_mask_0 = const()[name = string("causal_mask_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_19_cast_fp16 = slice_by_index(begin = causal_mask_19_begin_0, end = concat_169, end_mask = causal_mask_19_end_mask_0, x = causal_mask)[name = string("causal_mask_19_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_33_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_19_cast_fp16, key = key_states_35_cast_fp16, query = query_states_35_cast_fp16, value = value_states_35_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1708_perm_0 = const()[name = string("op_1708_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_170_axis_0 = const()[name = string("concat_170_axis_0"), val = int32(0)];
            bool concat_170_interleave_0 = const()[name = string("concat_170_interleave_0"), val = bool(false)];
            int32 gather_149_cast_uint16_to_int32 = cast(dtype = gather_149_cast_uint16_to_int32_dtype_0, x = gather_149_cast_uint16)[name = string("cast_56")];
            tensor<int32, [3]> concat_170 = concat(axis = concat_170_axis_0, interleave = concat_170_interleave_0, values = (gather_148, gather_149_cast_uint16_to_int32, var_48))[name = string("concat_170")];
            tensor<fp16, [1, ?, 32, 64]> var_1708_cast_fp16 = transpose(perm = var_1708_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, ?, ?]> input_65_cast_fp16 = reshape(shape = concat_170, x = var_1708_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425068480))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427165696))))[name = string("model_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_59_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_259_cast_fp16 = add(x = hidden_states_239_cast_fp16, y = linear_59_cast_fp16)[name = string("hidden_states_259_cast_fp16")];
            fp16 var_54_promoted_17_to_fp16 = const()[name = string("op_54_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1717_cast_fp16 = pow(x = hidden_states_259_cast_fp16, y = var_54_promoted_17_to_fp16)[name = string("op_1717_cast_fp16")];
            tensor<int32, [1]> variance_35_axes_0 = const()[name = string("variance_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_35_keep_dims_0 = const()[name = string("variance_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_35_cast_fp16 = reduce_mean(axes = variance_35_axes_0, keep_dims = variance_35_keep_dims_0, x = var_1717_cast_fp16)[name = string("variance_35_cast_fp16")];
            fp16 var_1720_to_fp16 = const()[name = string("op_1720_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1721_cast_fp16 = add(x = variance_35_cast_fp16, y = var_1720_to_fp16)[name = string("op_1721_cast_fp16")];
            fp32 var_1722_epsilon_0 = const()[name = string("op_1722_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1722_cast_fp16 = rsqrt(epsilon = var_1722_epsilon_0, x = var_1721_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_263_cast_fp16 = mul(x = hidden_states_259_cast_fp16, y = var_1722_cast_fp16)[name = string("hidden_states_263_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427427904)))];
            tensor<fp16, [1, ?, 2048]> input_67_cast_fp16 = mul(x = model_model_layers_8_post_attention_layernorm_weight_to_fp16, y = hidden_states_263_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427432064))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435820736))))[name = string("model_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_1734_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1734_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436869376))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445258048))))[name = string("model_model_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_71_cast_fp16 = mul(x = var_1734_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446306688))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454695360))))[name = string("model_model_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_269_cast_fp16 = add(x = hidden_states_259_cast_fp16, y = linear_62_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            fp16 var_54_promoted_18_to_fp16 = const()[name = string("op_54_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1747_cast_fp16 = pow(x = hidden_states_269_cast_fp16, y = var_54_promoted_18_to_fp16)[name = string("op_1747_cast_fp16")];
            tensor<int32, [1]> variance_37_axes_0 = const()[name = string("variance_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_37_keep_dims_0 = const()[name = string("variance_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_37_cast_fp16 = reduce_mean(axes = variance_37_axes_0, keep_dims = variance_37_keep_dims_0, x = var_1747_cast_fp16)[name = string("variance_37_cast_fp16")];
            fp16 var_1750_to_fp16 = const()[name = string("op_1750_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1751_cast_fp16 = add(x = variance_37_cast_fp16, y = var_1750_to_fp16)[name = string("op_1751_cast_fp16")];
            fp32 var_1752_epsilon_0 = const()[name = string("op_1752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1752_cast_fp16 = rsqrt(epsilon = var_1752_epsilon_0, x = var_1751_cast_fp16)[name = string("op_1752_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_273_cast_fp16 = mul(x = hidden_states_269_cast_fp16, y = var_1752_cast_fp16)[name = string("hidden_states_273_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455744000)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_277_cast_fp16 = mul(x = model_model_layers_9_input_layernorm_weight_to_fp16, y = hidden_states_273_cast_fp16)[name = string("hidden_states_277_cast_fp16")];
            tensor<int32, [3]> var_1763_shape_cast_fp16 = shape(x = hidden_states_277_cast_fp16)[name = string("op_1763_shape_cast_fp16")];
            int32 gather_166 = const()[name = string("gather_166"), val = int32(1)];
            int32 gather_167_axis_0 = const()[name = string("gather_167_axis_0"), val = int32(0)];
            int32 gather_167_batch_dims_0 = const()[name = string("gather_167_batch_dims_0"), val = int32(0)];
            bool gather_167_validate_indices_0 = const()[name = string("gather_167_validate_indices_0"), val = bool(false)];
            string var_1763_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1763_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_167_to_uint16 = const()[name = string("select_167_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1763_shape_cast_fp16_to_uint16 = cast(dtype = var_1763_shape_cast_fp16_to_uint16_dtype_0, x = var_1763_shape_cast_fp16)[name = string("cast_55")];
            uint16 gather_167_cast_uint16 = gather(axis = gather_167_axis_0, batch_dims = gather_167_batch_dims_0, indices = select_167_to_uint16, validate_indices = gather_167_validate_indices_0, x = var_1763_shape_cast_fp16_to_uint16)[name = string("gather_167_cast_uint16")];
            string gather_167_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_167_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455748160))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457845376))))[name = string("model_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_277_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458107584))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458631936))))[name = string("model_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_277_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458697536))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459221888))))[name = string("model_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_277_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> concat_171x = const()[name = string("concat_171x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_1772_cast_fp16 = reshape(shape = concat_171x, x = linear_63_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_172x = const()[name = string("concat_172x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1775_cast_fp16 = reshape(shape = concat_172x, x = linear_64_cast_fp16)[name = string("op_1775_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_173x = const()[name = string("concat_173x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1778_cast_fp16 = reshape(shape = concat_173x, x = linear_65_cast_fp16)[name = string("op_1778_cast_fp16")];
            tensor<int32, [4]> v_state_19_perm_0 = const()[name = string("v_state_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1772_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 32, ?, 64]> var_1782_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_1793_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_1793_cast_fp16")];
            bool var_1795_interleave_0 = const()[name = string("op_1795_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_1795_cast_fp16 = concat(axis = var_48, interleave = var_1795_interleave_0, values = (var_1793_cast_fp16, x1_37_cast_fp16))[name = string("op_1795_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_1796_cast_fp16 = mul(x = var_1795_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_39_cast_fp16 = add(x = var_1782_cast_fp16, y = var_1796_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1775_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, ?, 64]> var_1798_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1798_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_1809_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1809_cast_fp16")];
            bool var_1811_interleave_0 = const()[name = string("op_1811_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_1811_cast_fp16 = concat(axis = var_48, interleave = var_1811_interleave_0, values = (var_1809_cast_fp16, x1_39_cast_fp16))[name = string("op_1811_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1812_cast_fp16 = mul(x = var_1811_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1812_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_19_cast_fp16 = add(x = var_1798_cast_fp16, y = var_1812_cast_fp16)[name = string("k_state_19_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_176_values0_0 = const()[name = string("concat_176_values0_0"), val = tensor<int32, [1]>([9])];
            int32 concat_176_axis_0 = const()[name = string("concat_176_axis_0"), val = int32(0)];
            bool concat_176_interleave_0 = const()[name = string("concat_176_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_176 = concat(axis = concat_176_axis_0, interleave = concat_176_interleave_0, values = (concat_176_values0_0, expand_dims_108, expand_dims_109, expand_dims_2, expand_dims_111))[name = string("concat_176")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_176, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = k_state_19_cast_fp16, x = coreml_update_state_48)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_50_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_50 = read_state(input = key_cache)[name = string("coreml_update_state_50")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_19_cast_fp16 = transpose(perm = v_state_19_perm_0, x = var_1778_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_176, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = v_state_19_cast_fp16, x = coreml_update_state_49)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_51_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_51 = read_state(input = value_cache)[name = string("coreml_update_state_51")];
            tensor<int32, [5]> var_1835_begin_0 = const()[name = string("op_1835_begin_0"), val = tensor<int32, [5]>([9, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1835_end_0 = const()[name = string("op_1835_end_0"), val = tensor<int32, [5]>([10, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1835_end_mask_0 = const()[name = string("op_1835_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1835_squeeze_mask_0 = const()[name = string("op_1835_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1835_cast_fp16 = slice_by_index(begin = var_1835_begin_0, end = var_1835_end_0, end_mask = var_1835_end_mask_0, squeeze_mask = var_1835_squeeze_mask_0, x = coreml_update_state_50)[name = string("op_1835_cast_fp16")];
            tensor<int32, [4]> var_1838_begin_0 = const()[name = string("op_1838_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1838_end_mask_0 = const()[name = string("op_1838_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1838_cast_fp16 = slice_by_index(begin = var_1838_begin_0, end = concat_11, end_mask = var_1838_end_mask_0, x = var_1835_cast_fp16)[name = string("op_1838_cast_fp16")];
            tensor<int32, [5]> var_1840_begin_0 = const()[name = string("op_1840_begin_0"), val = tensor<int32, [5]>([9, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1840_end_0 = const()[name = string("op_1840_end_0"), val = tensor<int32, [5]>([10, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_1840_end_mask_0 = const()[name = string("op_1840_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1840_squeeze_mask_0 = const()[name = string("op_1840_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_1840_cast_fp16 = slice_by_index(begin = var_1840_begin_0, end = var_1840_end_0, end_mask = var_1840_end_mask_0, squeeze_mask = var_1840_squeeze_mask_0, x = coreml_update_state_51)[name = string("op_1840_cast_fp16")];
            tensor<int32, [4]> var_1843_begin_0 = const()[name = string("op_1843_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1843_end_mask_0 = const()[name = string("op_1843_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_1843_cast_fp16 = slice_by_index(begin = var_1843_begin_0, end = concat_11, end_mask = var_1843_end_mask_0, x = var_1840_cast_fp16)[name = string("op_1843_cast_fp16")];
            tensor<int32, [4]> var_1845_shape_cast_fp16 = shape(x = var_1838_cast_fp16)[name = string("op_1845_shape_cast_fp16")];
            int32 gather_175 = const()[name = string("gather_175"), val = int32(1)];
            int32 gather_176 = const()[name = string("gather_176"), val = int32(8)];
            int32 gather_177_axis_0 = const()[name = string("gather_177_axis_0"), val = int32(0)];
            int32 gather_177_batch_dims_0 = const()[name = string("gather_177_batch_dims_0"), val = int32(0)];
            bool gather_177_validate_indices_0 = const()[name = string("gather_177_validate_indices_0"), val = bool(false)];
            string var_1845_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1845_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_177_to_uint16 = const()[name = string("select_177_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1845_shape_cast_fp16_to_uint16 = cast(dtype = var_1845_shape_cast_fp16_to_uint16_dtype_0, x = var_1845_shape_cast_fp16)[name = string("cast_54")];
            uint16 gather_177_cast_uint16 = gather(axis = gather_177_axis_0, batch_dims = gather_177_batch_dims_0, indices = select_177_to_uint16, validate_indices = gather_177_validate_indices_0, x = var_1845_shape_cast_fp16_to_uint16)[name = string("gather_177_cast_uint16")];
            string gather_177_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_177_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_178 = const()[name = string("gather_178"), val = int32(64)];
            tensor<int32, [1]> var_1852_axes_0 = const()[name = string("op_1852_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1852_cast_fp16 = expand_dims(axes = var_1852_axes_0, x = var_1838_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [5]> shape_197_cast_fp16 = shape(x = var_1852_cast_fp16)[name = string("shape_197_cast_fp16")];
            int32 concat_184_axis_0 = const()[name = string("concat_184_axis_0"), val = int32(0)];
            bool concat_184_interleave_0 = const()[name = string("concat_184_interleave_0"), val = bool(false)];
            int32 gather_177_cast_uint16_to_int32 = cast(dtype = gather_177_cast_uint16_to_int32_dtype_0, x = gather_177_cast_uint16)[name = string("cast_53")];
            tensor<int32, [5]> concat_184 = concat(axis = concat_184_axis_0, interleave = concat_184_interleave_0, values = (gather_175, gather_176, var_59, gather_177_cast_uint16_to_int32, gather_178))[name = string("concat_184")];
            tensor<int32, [5]> real_div_18 = real_div(x = concat_184, y = shape_197_cast_fp16)[name = string("real_div_18")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_281_cast_fp16 = tile(reps = real_div_18, x = var_1852_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<int32, [4]> concat_185x = const()[name = string("concat_185x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_39_cast_fp16 = reshape(shape = concat_185x, x = hidden_states_281_cast_fp16)[name = string("key_states_39_cast_fp16")];
            tensor<int32, [4]> var_1862_shape_cast_fp16 = shape(x = var_1843_cast_fp16)[name = string("op_1862_shape_cast_fp16")];
            int32 gather_179 = const()[name = string("gather_179"), val = int32(1)];
            int32 gather_180 = const()[name = string("gather_180"), val = int32(8)];
            int32 gather_181_axis_0 = const()[name = string("gather_181_axis_0"), val = int32(0)];
            int32 gather_181_batch_dims_0 = const()[name = string("gather_181_batch_dims_0"), val = int32(0)];
            bool gather_181_validate_indices_0 = const()[name = string("gather_181_validate_indices_0"), val = bool(false)];
            string var_1862_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1862_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_181_to_uint16 = const()[name = string("select_181_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1862_shape_cast_fp16_to_uint16 = cast(dtype = var_1862_shape_cast_fp16_to_uint16_dtype_0, x = var_1862_shape_cast_fp16)[name = string("cast_52")];
            uint16 gather_181_cast_uint16 = gather(axis = gather_181_axis_0, batch_dims = gather_181_batch_dims_0, indices = select_181_to_uint16, validate_indices = gather_181_validate_indices_0, x = var_1862_shape_cast_fp16_to_uint16)[name = string("gather_181_cast_uint16")];
            string gather_181_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_181_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_182 = const()[name = string("gather_182"), val = int32(64)];
            tensor<int32, [1]> var_1869_axes_0 = const()[name = string("op_1869_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_1869_cast_fp16 = expand_dims(axes = var_1869_axes_0, x = var_1843_cast_fp16)[name = string("op_1869_cast_fp16")];
            tensor<int32, [5]> shape_202_cast_fp16 = shape(x = var_1869_cast_fp16)[name = string("shape_202_cast_fp16")];
            int32 concat_186_axis_0 = const()[name = string("concat_186_axis_0"), val = int32(0)];
            bool concat_186_interleave_0 = const()[name = string("concat_186_interleave_0"), val = bool(false)];
            int32 gather_181_cast_uint16_to_int32 = cast(dtype = gather_181_cast_uint16_to_int32_dtype_0, x = gather_181_cast_uint16)[name = string("cast_51")];
            tensor<int32, [5]> concat_186 = concat(axis = concat_186_axis_0, interleave = concat_186_interleave_0, values = (gather_179, gather_180, var_59, gather_181_cast_uint16_to_int32, gather_182))[name = string("concat_186")];
            tensor<int32, [5]> real_div_19 = real_div(x = concat_186, y = shape_202_cast_fp16)[name = string("real_div_19")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_285_cast_fp16 = tile(reps = real_div_19, x = var_1869_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> concat_187x = const()[name = string("concat_187x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_39_cast_fp16 = reshape(shape = concat_187x, x = hidden_states_285_cast_fp16)[name = string("value_states_39_cast_fp16")];
            tensor<int32, [4]> var_1879_shape_cast_fp16 = shape(x = key_states_39_cast_fp16)[name = string("op_1879_shape_cast_fp16")];
            int32 gather_183_axis_0 = const()[name = string("gather_183_axis_0"), val = int32(0)];
            int32 gather_183_batch_dims_0 = const()[name = string("gather_183_batch_dims_0"), val = int32(0)];
            bool gather_183_validate_indices_0 = const()[name = string("gather_183_validate_indices_0"), val = bool(false)];
            string var_1879_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1879_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_183_to_uint16 = const()[name = string("select_183_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1879_shape_cast_fp16_to_uint16 = cast(dtype = var_1879_shape_cast_fp16_to_uint16_dtype_0, x = var_1879_shape_cast_fp16)[name = string("cast_50")];
            uint16 gather_183_cast_uint16 = gather(axis = gather_183_axis_0, batch_dims = gather_183_batch_dims_0, indices = select_183_to_uint16, validate_indices = gather_183_validate_indices_0, x = var_1879_shape_cast_fp16_to_uint16)[name = string("gather_183_cast_uint16")];
            string gather_183_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_183_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_188_values0_0 = const()[name = string("concat_188_values0_0"), val = int32(1)];
            int32 concat_188_values1_0 = const()[name = string("concat_188_values1_0"), val = int32(1)];
            int32 concat_188_values2_0 = const()[name = string("concat_188_values2_0"), val = int32(0)];
            int32 concat_188_axis_0 = const()[name = string("concat_188_axis_0"), val = int32(0)];
            bool concat_188_interleave_0 = const()[name = string("concat_188_interleave_0"), val = bool(false)];
            int32 gather_183_cast_uint16_to_int32 = cast(dtype = gather_183_cast_uint16_to_int32_dtype_0, x = gather_183_cast_uint16)[name = string("cast_49")];
            tensor<int32, [4]> concat_188 = concat(axis = concat_188_axis_0, interleave = concat_188_interleave_0, values = (concat_188_values0_0, concat_188_values1_0, concat_188_values2_0, gather_183_cast_uint16_to_int32))[name = string("concat_188")];
            tensor<int32, [4]> causal_mask_21_begin_0 = const()[name = string("causal_mask_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_21_end_mask_0 = const()[name = string("causal_mask_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_21_cast_fp16 = slice_by_index(begin = causal_mask_21_begin_0, end = concat_188, end_mask = causal_mask_21_end_mask_0, x = causal_mask)[name = string("causal_mask_21_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_37_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_21_cast_fp16, key = key_states_39_cast_fp16, query = query_states_39_cast_fp16, value = value_states_39_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1885_perm_0 = const()[name = string("op_1885_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_189_axis_0 = const()[name = string("concat_189_axis_0"), val = int32(0)];
            bool concat_189_interleave_0 = const()[name = string("concat_189_interleave_0"), val = bool(false)];
            int32 gather_167_cast_uint16_to_int32 = cast(dtype = gather_167_cast_uint16_to_int32_dtype_0, x = gather_167_cast_uint16)[name = string("cast_48")];
            tensor<int32, [3]> concat_189 = concat(axis = concat_189_axis_0, interleave = concat_189_interleave_0, values = (gather_166, gather_167_cast_uint16_to_int32, var_48))[name = string("concat_189")];
            tensor<fp16, [1, ?, 32, 64]> var_1885_cast_fp16 = transpose(perm = var_1885_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, ?, ?]> input_73_cast_fp16 = reshape(shape = concat_189, x = var_1885_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459287488))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461384704))))[name = string("model_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_289_cast_fp16 = add(x = hidden_states_269_cast_fp16, y = linear_66_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            fp16 var_54_promoted_19_to_fp16 = const()[name = string("op_54_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1894_cast_fp16 = pow(x = hidden_states_289_cast_fp16, y = var_54_promoted_19_to_fp16)[name = string("op_1894_cast_fp16")];
            tensor<int32, [1]> variance_39_axes_0 = const()[name = string("variance_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_39_keep_dims_0 = const()[name = string("variance_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_39_cast_fp16 = reduce_mean(axes = variance_39_axes_0, keep_dims = variance_39_keep_dims_0, x = var_1894_cast_fp16)[name = string("variance_39_cast_fp16")];
            fp16 var_1897_to_fp16 = const()[name = string("op_1897_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1898_cast_fp16 = add(x = variance_39_cast_fp16, y = var_1897_to_fp16)[name = string("op_1898_cast_fp16")];
            fp32 var_1899_epsilon_0 = const()[name = string("op_1899_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1899_cast_fp16 = rsqrt(epsilon = var_1899_epsilon_0, x = var_1898_cast_fp16)[name = string("op_1899_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_293_cast_fp16 = mul(x = hidden_states_289_cast_fp16, y = var_1899_cast_fp16)[name = string("hidden_states_293_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461646912)))];
            tensor<fp16, [1, ?, 2048]> input_75_cast_fp16 = mul(x = model_model_layers_9_post_attention_layernorm_weight_to_fp16, y = hidden_states_293_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461651072))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470039744))))[name = string("model_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_1911_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1911_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471088384))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(479477056))))[name = string("model_model_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_79_cast_fp16 = mul(x = var_1911_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480525696))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488914368))))[name = string("model_model_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_69_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_299_cast_fp16 = add(x = hidden_states_289_cast_fp16, y = linear_69_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            fp16 var_54_promoted_20_to_fp16 = const()[name = string("op_54_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_1924_cast_fp16 = pow(x = hidden_states_299_cast_fp16, y = var_54_promoted_20_to_fp16)[name = string("op_1924_cast_fp16")];
            tensor<int32, [1]> variance_41_axes_0 = const()[name = string("variance_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_41_keep_dims_0 = const()[name = string("variance_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_41_cast_fp16 = reduce_mean(axes = variance_41_axes_0, keep_dims = variance_41_keep_dims_0, x = var_1924_cast_fp16)[name = string("variance_41_cast_fp16")];
            fp16 var_1927_to_fp16 = const()[name = string("op_1927_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_1928_cast_fp16 = add(x = variance_41_cast_fp16, y = var_1927_to_fp16)[name = string("op_1928_cast_fp16")];
            fp32 var_1929_epsilon_0 = const()[name = string("op_1929_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1929_cast_fp16 = rsqrt(epsilon = var_1929_epsilon_0, x = var_1928_cast_fp16)[name = string("op_1929_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_303_cast_fp16 = mul(x = hidden_states_299_cast_fp16, y = var_1929_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489963008)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_307_cast_fp16 = mul(x = model_model_layers_10_input_layernorm_weight_to_fp16, y = hidden_states_303_cast_fp16)[name = string("hidden_states_307_cast_fp16")];
            tensor<int32, [3]> var_1940_shape_cast_fp16 = shape(x = hidden_states_307_cast_fp16)[name = string("op_1940_shape_cast_fp16")];
            int32 gather_184 = const()[name = string("gather_184"), val = int32(1)];
            int32 gather_185_axis_0 = const()[name = string("gather_185_axis_0"), val = int32(0)];
            int32 gather_185_batch_dims_0 = const()[name = string("gather_185_batch_dims_0"), val = int32(0)];
            bool gather_185_validate_indices_0 = const()[name = string("gather_185_validate_indices_0"), val = bool(false)];
            string var_1940_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1940_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_185_to_uint16 = const()[name = string("select_185_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1940_shape_cast_fp16_to_uint16 = cast(dtype = var_1940_shape_cast_fp16_to_uint16_dtype_0, x = var_1940_shape_cast_fp16)[name = string("cast_47")];
            uint16 gather_185_cast_uint16 = gather(axis = gather_185_axis_0, batch_dims = gather_185_batch_dims_0, indices = select_185_to_uint16, validate_indices = gather_185_validate_indices_0, x = var_1940_shape_cast_fp16_to_uint16)[name = string("gather_185_cast_uint16")];
            string gather_185_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_185_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489967168))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492064384))))[name = string("model_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_307_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492326592))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492850944))))[name = string("model_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_307_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492916544))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493440896))))[name = string("model_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_307_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> concat_190x = const()[name = string("concat_190x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_1949_cast_fp16 = reshape(shape = concat_190x, x = linear_70_cast_fp16)[name = string("op_1949_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_191x = const()[name = string("concat_191x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1952_cast_fp16 = reshape(shape = concat_191x, x = linear_71_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_192x = const()[name = string("concat_192x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_1955_cast_fp16 = reshape(shape = concat_192x, x = linear_72_cast_fp16)[name = string("op_1955_cast_fp16")];
            tensor<int32, [4]> v_state_21_perm_0 = const()[name = string("v_state_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1949_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 32, ?, 64]> var_1959_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1959_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_1970_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_1970_cast_fp16")];
            bool var_1972_interleave_0 = const()[name = string("op_1972_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_1972_cast_fp16 = concat(axis = var_48, interleave = var_1972_interleave_0, values = (var_1970_cast_fp16, x1_41_cast_fp16))[name = string("op_1972_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_1973_cast_fp16 = mul(x = var_1972_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1973_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_43_cast_fp16 = add(x = var_1959_cast_fp16, y = var_1973_cast_fp16)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1952_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, ?, 64]> var_1975_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1975_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_1986_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_1986_cast_fp16")];
            bool var_1988_interleave_0 = const()[name = string("op_1988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_1988_cast_fp16 = concat(axis = var_48, interleave = var_1988_interleave_0, values = (var_1986_cast_fp16, x1_43_cast_fp16))[name = string("op_1988_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1989_cast_fp16 = mul(x = var_1988_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1989_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_21_cast_fp16 = add(x = var_1975_cast_fp16, y = var_1989_cast_fp16)[name = string("k_state_21_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_195_values0_0 = const()[name = string("concat_195_values0_0"), val = tensor<int32, [1]>([10])];
            int32 concat_195_axis_0 = const()[name = string("concat_195_axis_0"), val = int32(0)];
            bool concat_195_interleave_0 = const()[name = string("concat_195_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_195 = concat(axis = concat_195_axis_0, interleave = concat_195_interleave_0, values = (concat_195_values0_0, expand_dims_120, expand_dims_121, expand_dims_2, expand_dims_123))[name = string("concat_195")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_195, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = k_state_21_cast_fp16, x = coreml_update_state_50)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_52_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_52 = read_state(input = key_cache)[name = string("coreml_update_state_52")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_21_cast_fp16 = transpose(perm = v_state_21_perm_0, x = var_1955_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_195, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = v_state_21_cast_fp16, x = coreml_update_state_51)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_53_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_53 = read_state(input = value_cache)[name = string("coreml_update_state_53")];
            tensor<int32, [5]> var_2012_begin_0 = const()[name = string("op_2012_begin_0"), val = tensor<int32, [5]>([10, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2012_end_0 = const()[name = string("op_2012_end_0"), val = tensor<int32, [5]>([11, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2012_end_mask_0 = const()[name = string("op_2012_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2012_squeeze_mask_0 = const()[name = string("op_2012_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2012_cast_fp16 = slice_by_index(begin = var_2012_begin_0, end = var_2012_end_0, end_mask = var_2012_end_mask_0, squeeze_mask = var_2012_squeeze_mask_0, x = coreml_update_state_52)[name = string("op_2012_cast_fp16")];
            tensor<int32, [4]> var_2015_begin_0 = const()[name = string("op_2015_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2015_end_mask_0 = const()[name = string("op_2015_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2015_cast_fp16 = slice_by_index(begin = var_2015_begin_0, end = concat_11, end_mask = var_2015_end_mask_0, x = var_2012_cast_fp16)[name = string("op_2015_cast_fp16")];
            tensor<int32, [5]> var_2017_begin_0 = const()[name = string("op_2017_begin_0"), val = tensor<int32, [5]>([10, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2017_end_0 = const()[name = string("op_2017_end_0"), val = tensor<int32, [5]>([11, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2017_end_mask_0 = const()[name = string("op_2017_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2017_squeeze_mask_0 = const()[name = string("op_2017_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2017_cast_fp16 = slice_by_index(begin = var_2017_begin_0, end = var_2017_end_0, end_mask = var_2017_end_mask_0, squeeze_mask = var_2017_squeeze_mask_0, x = coreml_update_state_53)[name = string("op_2017_cast_fp16")];
            tensor<int32, [4]> var_2020_begin_0 = const()[name = string("op_2020_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2020_end_mask_0 = const()[name = string("op_2020_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = concat_11, end_mask = var_2020_end_mask_0, x = var_2017_cast_fp16)[name = string("op_2020_cast_fp16")];
            tensor<int32, [4]> var_2022_shape_cast_fp16 = shape(x = var_2015_cast_fp16)[name = string("op_2022_shape_cast_fp16")];
            int32 gather_193 = const()[name = string("gather_193"), val = int32(1)];
            int32 gather_194 = const()[name = string("gather_194"), val = int32(8)];
            int32 gather_195_axis_0 = const()[name = string("gather_195_axis_0"), val = int32(0)];
            int32 gather_195_batch_dims_0 = const()[name = string("gather_195_batch_dims_0"), val = int32(0)];
            bool gather_195_validate_indices_0 = const()[name = string("gather_195_validate_indices_0"), val = bool(false)];
            string var_2022_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2022_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_195_to_uint16 = const()[name = string("select_195_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2022_shape_cast_fp16_to_uint16 = cast(dtype = var_2022_shape_cast_fp16_to_uint16_dtype_0, x = var_2022_shape_cast_fp16)[name = string("cast_46")];
            uint16 gather_195_cast_uint16 = gather(axis = gather_195_axis_0, batch_dims = gather_195_batch_dims_0, indices = select_195_to_uint16, validate_indices = gather_195_validate_indices_0, x = var_2022_shape_cast_fp16_to_uint16)[name = string("gather_195_cast_uint16")];
            string gather_195_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_195_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_196 = const()[name = string("gather_196"), val = int32(64)];
            tensor<int32, [1]> var_2029_axes_0 = const()[name = string("op_2029_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2029_cast_fp16 = expand_dims(axes = var_2029_axes_0, x = var_2015_cast_fp16)[name = string("op_2029_cast_fp16")];
            tensor<int32, [5]> shape_217_cast_fp16 = shape(x = var_2029_cast_fp16)[name = string("shape_217_cast_fp16")];
            int32 concat_203_axis_0 = const()[name = string("concat_203_axis_0"), val = int32(0)];
            bool concat_203_interleave_0 = const()[name = string("concat_203_interleave_0"), val = bool(false)];
            int32 gather_195_cast_uint16_to_int32 = cast(dtype = gather_195_cast_uint16_to_int32_dtype_0, x = gather_195_cast_uint16)[name = string("cast_45")];
            tensor<int32, [5]> concat_203 = concat(axis = concat_203_axis_0, interleave = concat_203_interleave_0, values = (gather_193, gather_194, var_59, gather_195_cast_uint16_to_int32, gather_196))[name = string("concat_203")];
            tensor<int32, [5]> real_div_20 = real_div(x = concat_203, y = shape_217_cast_fp16)[name = string("real_div_20")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_311_cast_fp16 = tile(reps = real_div_20, x = var_2029_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [4]> concat_204x = const()[name = string("concat_204x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_43_cast_fp16 = reshape(shape = concat_204x, x = hidden_states_311_cast_fp16)[name = string("key_states_43_cast_fp16")];
            tensor<int32, [4]> var_2039_shape_cast_fp16 = shape(x = var_2020_cast_fp16)[name = string("op_2039_shape_cast_fp16")];
            int32 gather_197 = const()[name = string("gather_197"), val = int32(1)];
            int32 gather_198 = const()[name = string("gather_198"), val = int32(8)];
            int32 gather_199_axis_0 = const()[name = string("gather_199_axis_0"), val = int32(0)];
            int32 gather_199_batch_dims_0 = const()[name = string("gather_199_batch_dims_0"), val = int32(0)];
            bool gather_199_validate_indices_0 = const()[name = string("gather_199_validate_indices_0"), val = bool(false)];
            string var_2039_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2039_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_199_to_uint16 = const()[name = string("select_199_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2039_shape_cast_fp16_to_uint16 = cast(dtype = var_2039_shape_cast_fp16_to_uint16_dtype_0, x = var_2039_shape_cast_fp16)[name = string("cast_44")];
            uint16 gather_199_cast_uint16 = gather(axis = gather_199_axis_0, batch_dims = gather_199_batch_dims_0, indices = select_199_to_uint16, validate_indices = gather_199_validate_indices_0, x = var_2039_shape_cast_fp16_to_uint16)[name = string("gather_199_cast_uint16")];
            string gather_199_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_199_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_200 = const()[name = string("gather_200"), val = int32(64)];
            tensor<int32, [1]> var_2046_axes_0 = const()[name = string("op_2046_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2046_cast_fp16 = expand_dims(axes = var_2046_axes_0, x = var_2020_cast_fp16)[name = string("op_2046_cast_fp16")];
            tensor<int32, [5]> shape_222_cast_fp16 = shape(x = var_2046_cast_fp16)[name = string("shape_222_cast_fp16")];
            int32 concat_205_axis_0 = const()[name = string("concat_205_axis_0"), val = int32(0)];
            bool concat_205_interleave_0 = const()[name = string("concat_205_interleave_0"), val = bool(false)];
            int32 gather_199_cast_uint16_to_int32 = cast(dtype = gather_199_cast_uint16_to_int32_dtype_0, x = gather_199_cast_uint16)[name = string("cast_43")];
            tensor<int32, [5]> concat_205 = concat(axis = concat_205_axis_0, interleave = concat_205_interleave_0, values = (gather_197, gather_198, var_59, gather_199_cast_uint16_to_int32, gather_200))[name = string("concat_205")];
            tensor<int32, [5]> real_div_21 = real_div(x = concat_205, y = shape_222_cast_fp16)[name = string("real_div_21")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_315_cast_fp16 = tile(reps = real_div_21, x = var_2046_cast_fp16)[name = string("hidden_states_315_cast_fp16")];
            tensor<int32, [4]> concat_206x = const()[name = string("concat_206x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_43_cast_fp16 = reshape(shape = concat_206x, x = hidden_states_315_cast_fp16)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> var_2056_shape_cast_fp16 = shape(x = key_states_43_cast_fp16)[name = string("op_2056_shape_cast_fp16")];
            int32 gather_201_axis_0 = const()[name = string("gather_201_axis_0"), val = int32(0)];
            int32 gather_201_batch_dims_0 = const()[name = string("gather_201_batch_dims_0"), val = int32(0)];
            bool gather_201_validate_indices_0 = const()[name = string("gather_201_validate_indices_0"), val = bool(false)];
            string var_2056_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2056_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_201_to_uint16 = const()[name = string("select_201_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2056_shape_cast_fp16_to_uint16 = cast(dtype = var_2056_shape_cast_fp16_to_uint16_dtype_0, x = var_2056_shape_cast_fp16)[name = string("cast_42")];
            uint16 gather_201_cast_uint16 = gather(axis = gather_201_axis_0, batch_dims = gather_201_batch_dims_0, indices = select_201_to_uint16, validate_indices = gather_201_validate_indices_0, x = var_2056_shape_cast_fp16_to_uint16)[name = string("gather_201_cast_uint16")];
            string gather_201_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_201_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_207_values0_0 = const()[name = string("concat_207_values0_0"), val = int32(1)];
            int32 concat_207_values1_0 = const()[name = string("concat_207_values1_0"), val = int32(1)];
            int32 concat_207_values2_0 = const()[name = string("concat_207_values2_0"), val = int32(0)];
            int32 concat_207_axis_0 = const()[name = string("concat_207_axis_0"), val = int32(0)];
            bool concat_207_interleave_0 = const()[name = string("concat_207_interleave_0"), val = bool(false)];
            int32 gather_201_cast_uint16_to_int32 = cast(dtype = gather_201_cast_uint16_to_int32_dtype_0, x = gather_201_cast_uint16)[name = string("cast_41")];
            tensor<int32, [4]> concat_207 = concat(axis = concat_207_axis_0, interleave = concat_207_interleave_0, values = (concat_207_values0_0, concat_207_values1_0, concat_207_values2_0, gather_201_cast_uint16_to_int32))[name = string("concat_207")];
            tensor<int32, [4]> causal_mask_23_begin_0 = const()[name = string("causal_mask_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_23_end_mask_0 = const()[name = string("causal_mask_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_23_cast_fp16 = slice_by_index(begin = causal_mask_23_begin_0, end = concat_207, end_mask = causal_mask_23_end_mask_0, x = causal_mask)[name = string("causal_mask_23_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_41_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_23_cast_fp16, key = key_states_43_cast_fp16, query = query_states_43_cast_fp16, value = value_states_43_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_2062_perm_0 = const()[name = string("op_2062_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_208_axis_0 = const()[name = string("concat_208_axis_0"), val = int32(0)];
            bool concat_208_interleave_0 = const()[name = string("concat_208_interleave_0"), val = bool(false)];
            int32 gather_185_cast_uint16_to_int32 = cast(dtype = gather_185_cast_uint16_to_int32_dtype_0, x = gather_185_cast_uint16)[name = string("cast_40")];
            tensor<int32, [3]> concat_208 = concat(axis = concat_208_axis_0, interleave = concat_208_interleave_0, values = (gather_184, gather_185_cast_uint16_to_int32, var_48))[name = string("concat_208")];
            tensor<fp16, [1, ?, 32, 64]> var_2062_cast_fp16 = transpose(perm = var_2062_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, ?, ?]> input_81_cast_fp16 = reshape(shape = concat_208, x = var_2062_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493506496))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495603712))))[name = string("model_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_73_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_319_cast_fp16 = add(x = hidden_states_299_cast_fp16, y = linear_73_cast_fp16)[name = string("hidden_states_319_cast_fp16")];
            fp16 var_54_promoted_21_to_fp16 = const()[name = string("op_54_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2071_cast_fp16 = pow(x = hidden_states_319_cast_fp16, y = var_54_promoted_21_to_fp16)[name = string("op_2071_cast_fp16")];
            tensor<int32, [1]> variance_43_axes_0 = const()[name = string("variance_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_43_keep_dims_0 = const()[name = string("variance_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_43_cast_fp16 = reduce_mean(axes = variance_43_axes_0, keep_dims = variance_43_keep_dims_0, x = var_2071_cast_fp16)[name = string("variance_43_cast_fp16")];
            fp16 var_2074_to_fp16 = const()[name = string("op_2074_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2075_cast_fp16 = add(x = variance_43_cast_fp16, y = var_2074_to_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_323_cast_fp16 = mul(x = hidden_states_319_cast_fp16, y = var_2076_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495865920)))];
            tensor<fp16, [1, ?, 2048]> input_83_cast_fp16 = mul(x = model_model_layers_10_post_attention_layernorm_weight_to_fp16, y = hidden_states_323_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495870080))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504258752))))[name = string("model_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_2088_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_2088_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505307392))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513696064))))[name = string("model_model_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_87_cast_fp16 = mul(x = var_2088_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514744704))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523133376))))[name = string("model_model_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_329_cast_fp16 = add(x = hidden_states_319_cast_fp16, y = linear_76_cast_fp16)[name = string("hidden_states_329_cast_fp16")];
            fp16 var_54_promoted_22_to_fp16 = const()[name = string("op_54_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2101_cast_fp16 = pow(x = hidden_states_329_cast_fp16, y = var_54_promoted_22_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> variance_45_axes_0 = const()[name = string("variance_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_45_keep_dims_0 = const()[name = string("variance_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_45_cast_fp16 = reduce_mean(axes = variance_45_axes_0, keep_dims = variance_45_keep_dims_0, x = var_2101_cast_fp16)[name = string("variance_45_cast_fp16")];
            fp16 var_2104_to_fp16 = const()[name = string("op_2104_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2105_cast_fp16 = add(x = variance_45_cast_fp16, y = var_2104_to_fp16)[name = string("op_2105_cast_fp16")];
            fp32 var_2106_epsilon_0 = const()[name = string("op_2106_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2106_cast_fp16 = rsqrt(epsilon = var_2106_epsilon_0, x = var_2105_cast_fp16)[name = string("op_2106_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_333_cast_fp16 = mul(x = hidden_states_329_cast_fp16, y = var_2106_cast_fp16)[name = string("hidden_states_333_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524182016)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_337_cast_fp16 = mul(x = model_model_layers_11_input_layernorm_weight_to_fp16, y = hidden_states_333_cast_fp16)[name = string("hidden_states_337_cast_fp16")];
            tensor<int32, [3]> var_2117_shape_cast_fp16 = shape(x = hidden_states_337_cast_fp16)[name = string("op_2117_shape_cast_fp16")];
            int32 gather_202 = const()[name = string("gather_202"), val = int32(1)];
            int32 gather_203_axis_0 = const()[name = string("gather_203_axis_0"), val = int32(0)];
            int32 gather_203_batch_dims_0 = const()[name = string("gather_203_batch_dims_0"), val = int32(0)];
            bool gather_203_validate_indices_0 = const()[name = string("gather_203_validate_indices_0"), val = bool(false)];
            string var_2117_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2117_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_203_to_uint16 = const()[name = string("select_203_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2117_shape_cast_fp16_to_uint16 = cast(dtype = var_2117_shape_cast_fp16_to_uint16_dtype_0, x = var_2117_shape_cast_fp16)[name = string("cast_39")];
            uint16 gather_203_cast_uint16 = gather(axis = gather_203_axis_0, batch_dims = gather_203_batch_dims_0, indices = select_203_to_uint16, validate_indices = gather_203_validate_indices_0, x = var_2117_shape_cast_fp16_to_uint16)[name = string("gather_203_cast_uint16")];
            string gather_203_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_203_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524186176))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526283392))))[name = string("model_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_337_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526545600))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527069952))))[name = string("model_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_337_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527135552))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527659904))))[name = string("model_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_337_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> concat_209x = const()[name = string("concat_209x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_2126_cast_fp16 = reshape(shape = concat_209x, x = linear_77_cast_fp16)[name = string("op_2126_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_210x = const()[name = string("concat_210x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2129_cast_fp16 = reshape(shape = concat_210x, x = linear_78_cast_fp16)[name = string("op_2129_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_211x = const()[name = string("concat_211x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2132_cast_fp16 = reshape(shape = concat_211x, x = linear_79_cast_fp16)[name = string("op_2132_cast_fp16")];
            tensor<int32, [4]> v_state_23_perm_0 = const()[name = string("v_state_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_2126_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 32, ?, 64]> var_2136_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_2147_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_2147_cast_fp16")];
            bool var_2149_interleave_0 = const()[name = string("op_2149_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_2149_cast_fp16 = concat(axis = var_48, interleave = var_2149_interleave_0, values = (var_2147_cast_fp16, x1_45_cast_fp16))[name = string("op_2149_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_2150_cast_fp16 = mul(x = var_2149_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_47_cast_fp16 = add(x = var_2136_cast_fp16, y = var_2150_cast_fp16)[name = string("query_states_47_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_2129_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, ?, 64]> var_2152_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_2163_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_2163_cast_fp16")];
            bool var_2165_interleave_0 = const()[name = string("op_2165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_2165_cast_fp16 = concat(axis = var_48, interleave = var_2165_interleave_0, values = (var_2163_cast_fp16, x1_47_cast_fp16))[name = string("op_2165_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2166_cast_fp16 = mul(x = var_2165_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2166_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_23_cast_fp16 = add(x = var_2152_cast_fp16, y = var_2166_cast_fp16)[name = string("k_state_23_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_214_values0_0 = const()[name = string("concat_214_values0_0"), val = tensor<int32, [1]>([11])];
            int32 concat_214_axis_0 = const()[name = string("concat_214_axis_0"), val = int32(0)];
            bool concat_214_interleave_0 = const()[name = string("concat_214_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_214 = concat(axis = concat_214_axis_0, interleave = concat_214_interleave_0, values = (concat_214_values0_0, expand_dims_132, expand_dims_133, expand_dims_2, expand_dims_135))[name = string("concat_214")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_214, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = k_state_23_cast_fp16, x = coreml_update_state_52)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_54_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_54 = read_state(input = key_cache)[name = string("coreml_update_state_54")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_23_cast_fp16 = transpose(perm = v_state_23_perm_0, x = var_2132_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_214, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = v_state_23_cast_fp16, x = coreml_update_state_53)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_55_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_55 = read_state(input = value_cache)[name = string("coreml_update_state_55")];
            tensor<int32, [5]> var_2189_begin_0 = const()[name = string("op_2189_begin_0"), val = tensor<int32, [5]>([11, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2189_end_0 = const()[name = string("op_2189_end_0"), val = tensor<int32, [5]>([12, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2189_end_mask_0 = const()[name = string("op_2189_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2189_squeeze_mask_0 = const()[name = string("op_2189_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, squeeze_mask = var_2189_squeeze_mask_0, x = coreml_update_state_54)[name = string("op_2189_cast_fp16")];
            tensor<int32, [4]> var_2192_begin_0 = const()[name = string("op_2192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2192_end_mask_0 = const()[name = string("op_2192_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2192_cast_fp16 = slice_by_index(begin = var_2192_begin_0, end = concat_11, end_mask = var_2192_end_mask_0, x = var_2189_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<int32, [5]> var_2194_begin_0 = const()[name = string("op_2194_begin_0"), val = tensor<int32, [5]>([11, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2194_end_0 = const()[name = string("op_2194_end_0"), val = tensor<int32, [5]>([12, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2194_end_mask_0 = const()[name = string("op_2194_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2194_squeeze_mask_0 = const()[name = string("op_2194_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2194_cast_fp16 = slice_by_index(begin = var_2194_begin_0, end = var_2194_end_0, end_mask = var_2194_end_mask_0, squeeze_mask = var_2194_squeeze_mask_0, x = coreml_update_state_55)[name = string("op_2194_cast_fp16")];
            tensor<int32, [4]> var_2197_begin_0 = const()[name = string("op_2197_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2197_end_mask_0 = const()[name = string("op_2197_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2197_cast_fp16 = slice_by_index(begin = var_2197_begin_0, end = concat_11, end_mask = var_2197_end_mask_0, x = var_2194_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<int32, [4]> var_2199_shape_cast_fp16 = shape(x = var_2192_cast_fp16)[name = string("op_2199_shape_cast_fp16")];
            int32 gather_211 = const()[name = string("gather_211"), val = int32(1)];
            int32 gather_212 = const()[name = string("gather_212"), val = int32(8)];
            int32 gather_213_axis_0 = const()[name = string("gather_213_axis_0"), val = int32(0)];
            int32 gather_213_batch_dims_0 = const()[name = string("gather_213_batch_dims_0"), val = int32(0)];
            bool gather_213_validate_indices_0 = const()[name = string("gather_213_validate_indices_0"), val = bool(false)];
            string var_2199_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2199_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_213_to_uint16 = const()[name = string("select_213_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2199_shape_cast_fp16_to_uint16 = cast(dtype = var_2199_shape_cast_fp16_to_uint16_dtype_0, x = var_2199_shape_cast_fp16)[name = string("cast_38")];
            uint16 gather_213_cast_uint16 = gather(axis = gather_213_axis_0, batch_dims = gather_213_batch_dims_0, indices = select_213_to_uint16, validate_indices = gather_213_validate_indices_0, x = var_2199_shape_cast_fp16_to_uint16)[name = string("gather_213_cast_uint16")];
            string gather_213_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_213_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_214 = const()[name = string("gather_214"), val = int32(64)];
            tensor<int32, [1]> var_2206_axes_0 = const()[name = string("op_2206_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2206_cast_fp16 = expand_dims(axes = var_2206_axes_0, x = var_2192_cast_fp16)[name = string("op_2206_cast_fp16")];
            tensor<int32, [5]> shape_237_cast_fp16 = shape(x = var_2206_cast_fp16)[name = string("shape_237_cast_fp16")];
            int32 concat_222_axis_0 = const()[name = string("concat_222_axis_0"), val = int32(0)];
            bool concat_222_interleave_0 = const()[name = string("concat_222_interleave_0"), val = bool(false)];
            int32 gather_213_cast_uint16_to_int32 = cast(dtype = gather_213_cast_uint16_to_int32_dtype_0, x = gather_213_cast_uint16)[name = string("cast_37")];
            tensor<int32, [5]> concat_222 = concat(axis = concat_222_axis_0, interleave = concat_222_interleave_0, values = (gather_211, gather_212, var_59, gather_213_cast_uint16_to_int32, gather_214))[name = string("concat_222")];
            tensor<int32, [5]> real_div_22 = real_div(x = concat_222, y = shape_237_cast_fp16)[name = string("real_div_22")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_341_cast_fp16 = tile(reps = real_div_22, x = var_2206_cast_fp16)[name = string("hidden_states_341_cast_fp16")];
            tensor<int32, [4]> concat_223x = const()[name = string("concat_223x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_47_cast_fp16 = reshape(shape = concat_223x, x = hidden_states_341_cast_fp16)[name = string("key_states_47_cast_fp16")];
            tensor<int32, [4]> var_2216_shape_cast_fp16 = shape(x = var_2197_cast_fp16)[name = string("op_2216_shape_cast_fp16")];
            int32 gather_215 = const()[name = string("gather_215"), val = int32(1)];
            int32 gather_216 = const()[name = string("gather_216"), val = int32(8)];
            int32 gather_217_axis_0 = const()[name = string("gather_217_axis_0"), val = int32(0)];
            int32 gather_217_batch_dims_0 = const()[name = string("gather_217_batch_dims_0"), val = int32(0)];
            bool gather_217_validate_indices_0 = const()[name = string("gather_217_validate_indices_0"), val = bool(false)];
            string var_2216_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2216_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_217_to_uint16 = const()[name = string("select_217_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2216_shape_cast_fp16_to_uint16 = cast(dtype = var_2216_shape_cast_fp16_to_uint16_dtype_0, x = var_2216_shape_cast_fp16)[name = string("cast_36")];
            uint16 gather_217_cast_uint16 = gather(axis = gather_217_axis_0, batch_dims = gather_217_batch_dims_0, indices = select_217_to_uint16, validate_indices = gather_217_validate_indices_0, x = var_2216_shape_cast_fp16_to_uint16)[name = string("gather_217_cast_uint16")];
            string gather_217_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_217_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_218 = const()[name = string("gather_218"), val = int32(64)];
            tensor<int32, [1]> var_2223_axes_0 = const()[name = string("op_2223_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2223_cast_fp16 = expand_dims(axes = var_2223_axes_0, x = var_2197_cast_fp16)[name = string("op_2223_cast_fp16")];
            tensor<int32, [5]> shape_242_cast_fp16 = shape(x = var_2223_cast_fp16)[name = string("shape_242_cast_fp16")];
            int32 concat_224_axis_0 = const()[name = string("concat_224_axis_0"), val = int32(0)];
            bool concat_224_interleave_0 = const()[name = string("concat_224_interleave_0"), val = bool(false)];
            int32 gather_217_cast_uint16_to_int32 = cast(dtype = gather_217_cast_uint16_to_int32_dtype_0, x = gather_217_cast_uint16)[name = string("cast_35")];
            tensor<int32, [5]> concat_224 = concat(axis = concat_224_axis_0, interleave = concat_224_interleave_0, values = (gather_215, gather_216, var_59, gather_217_cast_uint16_to_int32, gather_218))[name = string("concat_224")];
            tensor<int32, [5]> real_div_23 = real_div(x = concat_224, y = shape_242_cast_fp16)[name = string("real_div_23")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_345_cast_fp16 = tile(reps = real_div_23, x = var_2223_cast_fp16)[name = string("hidden_states_345_cast_fp16")];
            tensor<int32, [4]> concat_225x = const()[name = string("concat_225x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_47_cast_fp16 = reshape(shape = concat_225x, x = hidden_states_345_cast_fp16)[name = string("value_states_47_cast_fp16")];
            tensor<int32, [4]> var_2233_shape_cast_fp16 = shape(x = key_states_47_cast_fp16)[name = string("op_2233_shape_cast_fp16")];
            int32 gather_219_axis_0 = const()[name = string("gather_219_axis_0"), val = int32(0)];
            int32 gather_219_batch_dims_0 = const()[name = string("gather_219_batch_dims_0"), val = int32(0)];
            bool gather_219_validate_indices_0 = const()[name = string("gather_219_validate_indices_0"), val = bool(false)];
            string var_2233_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2233_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_219_to_uint16 = const()[name = string("select_219_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2233_shape_cast_fp16_to_uint16 = cast(dtype = var_2233_shape_cast_fp16_to_uint16_dtype_0, x = var_2233_shape_cast_fp16)[name = string("cast_34")];
            uint16 gather_219_cast_uint16 = gather(axis = gather_219_axis_0, batch_dims = gather_219_batch_dims_0, indices = select_219_to_uint16, validate_indices = gather_219_validate_indices_0, x = var_2233_shape_cast_fp16_to_uint16)[name = string("gather_219_cast_uint16")];
            string gather_219_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_219_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_226_values0_0 = const()[name = string("concat_226_values0_0"), val = int32(1)];
            int32 concat_226_values1_0 = const()[name = string("concat_226_values1_0"), val = int32(1)];
            int32 concat_226_values2_0 = const()[name = string("concat_226_values2_0"), val = int32(0)];
            int32 concat_226_axis_0 = const()[name = string("concat_226_axis_0"), val = int32(0)];
            bool concat_226_interleave_0 = const()[name = string("concat_226_interleave_0"), val = bool(false)];
            int32 gather_219_cast_uint16_to_int32 = cast(dtype = gather_219_cast_uint16_to_int32_dtype_0, x = gather_219_cast_uint16)[name = string("cast_33")];
            tensor<int32, [4]> concat_226 = concat(axis = concat_226_axis_0, interleave = concat_226_interleave_0, values = (concat_226_values0_0, concat_226_values1_0, concat_226_values2_0, gather_219_cast_uint16_to_int32))[name = string("concat_226")];
            tensor<int32, [4]> causal_mask_25_begin_0 = const()[name = string("causal_mask_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_25_end_mask_0 = const()[name = string("causal_mask_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_25_cast_fp16 = slice_by_index(begin = causal_mask_25_begin_0, end = concat_226, end_mask = causal_mask_25_end_mask_0, x = causal_mask)[name = string("causal_mask_25_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_45_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_25_cast_fp16, key = key_states_47_cast_fp16, query = query_states_47_cast_fp16, value = value_states_47_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_2239_perm_0 = const()[name = string("op_2239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_227_axis_0 = const()[name = string("concat_227_axis_0"), val = int32(0)];
            bool concat_227_interleave_0 = const()[name = string("concat_227_interleave_0"), val = bool(false)];
            int32 gather_203_cast_uint16_to_int32 = cast(dtype = gather_203_cast_uint16_to_int32_dtype_0, x = gather_203_cast_uint16)[name = string("cast_32")];
            tensor<int32, [3]> concat_227 = concat(axis = concat_227_axis_0, interleave = concat_227_interleave_0, values = (gather_202, gather_203_cast_uint16_to_int32, var_48))[name = string("concat_227")];
            tensor<fp16, [1, ?, 32, 64]> var_2239_cast_fp16 = transpose(perm = var_2239_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, ?, ?]> input_89_cast_fp16 = reshape(shape = concat_227, x = var_2239_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527725504))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529822720))))[name = string("model_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_349_cast_fp16 = add(x = hidden_states_329_cast_fp16, y = linear_80_cast_fp16)[name = string("hidden_states_349_cast_fp16")];
            fp16 var_54_promoted_23_to_fp16 = const()[name = string("op_54_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2248_cast_fp16 = pow(x = hidden_states_349_cast_fp16, y = var_54_promoted_23_to_fp16)[name = string("op_2248_cast_fp16")];
            tensor<int32, [1]> variance_47_axes_0 = const()[name = string("variance_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_47_keep_dims_0 = const()[name = string("variance_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_47_cast_fp16 = reduce_mean(axes = variance_47_axes_0, keep_dims = variance_47_keep_dims_0, x = var_2248_cast_fp16)[name = string("variance_47_cast_fp16")];
            fp16 var_2251_to_fp16 = const()[name = string("op_2251_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2252_cast_fp16 = add(x = variance_47_cast_fp16, y = var_2251_to_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2253_cast_fp16 = rsqrt(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_353_cast_fp16 = mul(x = hidden_states_349_cast_fp16, y = var_2253_cast_fp16)[name = string("hidden_states_353_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530084928)))];
            tensor<fp16, [1, ?, 2048]> input_91_cast_fp16 = mul(x = model_model_layers_11_post_attention_layernorm_weight_to_fp16, y = hidden_states_353_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530089088))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538477760))))[name = string("model_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_2265_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_2265_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539526400))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547915072))))[name = string("model_model_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_95_cast_fp16 = mul(x = var_2265_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548963712))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557352384))))[name = string("model_model_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_83_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_359_cast_fp16 = add(x = hidden_states_349_cast_fp16, y = linear_83_cast_fp16)[name = string("hidden_states_359_cast_fp16")];
            fp16 var_54_promoted_24_to_fp16 = const()[name = string("op_54_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2278_cast_fp16 = pow(x = hidden_states_359_cast_fp16, y = var_54_promoted_24_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<int32, [1]> variance_49_axes_0 = const()[name = string("variance_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_49_keep_dims_0 = const()[name = string("variance_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_49_cast_fp16 = reduce_mean(axes = variance_49_axes_0, keep_dims = variance_49_keep_dims_0, x = var_2278_cast_fp16)[name = string("variance_49_cast_fp16")];
            fp16 var_2281_to_fp16 = const()[name = string("op_2281_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2282_cast_fp16 = add(x = variance_49_cast_fp16, y = var_2281_to_fp16)[name = string("op_2282_cast_fp16")];
            fp32 var_2283_epsilon_0 = const()[name = string("op_2283_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2283_cast_fp16 = rsqrt(epsilon = var_2283_epsilon_0, x = var_2282_cast_fp16)[name = string("op_2283_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_363_cast_fp16 = mul(x = hidden_states_359_cast_fp16, y = var_2283_cast_fp16)[name = string("hidden_states_363_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558401024)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_367_cast_fp16 = mul(x = model_model_layers_12_input_layernorm_weight_to_fp16, y = hidden_states_363_cast_fp16)[name = string("hidden_states_367_cast_fp16")];
            tensor<int32, [3]> var_2294_shape_cast_fp16 = shape(x = hidden_states_367_cast_fp16)[name = string("op_2294_shape_cast_fp16")];
            int32 gather_220 = const()[name = string("gather_220"), val = int32(1)];
            int32 gather_221_axis_0 = const()[name = string("gather_221_axis_0"), val = int32(0)];
            int32 gather_221_batch_dims_0 = const()[name = string("gather_221_batch_dims_0"), val = int32(0)];
            bool gather_221_validate_indices_0 = const()[name = string("gather_221_validate_indices_0"), val = bool(false)];
            string var_2294_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2294_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_221_to_uint16 = const()[name = string("select_221_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2294_shape_cast_fp16_to_uint16 = cast(dtype = var_2294_shape_cast_fp16_to_uint16_dtype_0, x = var_2294_shape_cast_fp16)[name = string("cast_31")];
            uint16 gather_221_cast_uint16 = gather(axis = gather_221_axis_0, batch_dims = gather_221_batch_dims_0, indices = select_221_to_uint16, validate_indices = gather_221_validate_indices_0, x = var_2294_shape_cast_fp16_to_uint16)[name = string("gather_221_cast_uint16")];
            string gather_221_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_221_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558405184))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560502400))))[name = string("model_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_367_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560764608))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561288960))))[name = string("model_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_367_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561354560))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561878912))))[name = string("model_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_367_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> concat_228x = const()[name = string("concat_228x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_2303_cast_fp16 = reshape(shape = concat_228x, x = linear_84_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_229x = const()[name = string("concat_229x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2306_cast_fp16 = reshape(shape = concat_229x, x = linear_85_cast_fp16)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_230x = const()[name = string("concat_230x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2309_cast_fp16 = reshape(shape = concat_230x, x = linear_86_cast_fp16)[name = string("op_2309_cast_fp16")];
            tensor<int32, [4]> v_state_25_perm_0 = const()[name = string("v_state_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_2303_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 32, ?, 64]> var_2313_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2313_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_2324_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_2324_cast_fp16")];
            bool var_2326_interleave_0 = const()[name = string("op_2326_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_2326_cast_fp16 = concat(axis = var_48, interleave = var_2326_interleave_0, values = (var_2324_cast_fp16, x1_49_cast_fp16))[name = string("op_2326_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_2327_cast_fp16 = mul(x = var_2326_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2327_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_51_cast_fp16 = add(x = var_2313_cast_fp16, y = var_2327_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_2306_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 8, ?, 64]> var_2329_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_2340_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_2340_cast_fp16")];
            bool var_2342_interleave_0 = const()[name = string("op_2342_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_2342_cast_fp16 = concat(axis = var_48, interleave = var_2342_interleave_0, values = (var_2340_cast_fp16, x1_51_cast_fp16))[name = string("op_2342_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2343_cast_fp16 = mul(x = var_2342_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2343_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_25_cast_fp16 = add(x = var_2329_cast_fp16, y = var_2343_cast_fp16)[name = string("k_state_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_233_values0_0 = const()[name = string("concat_233_values0_0"), val = tensor<int32, [1]>([12])];
            int32 concat_233_axis_0 = const()[name = string("concat_233_axis_0"), val = int32(0)];
            bool concat_233_interleave_0 = const()[name = string("concat_233_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_233 = concat(axis = concat_233_axis_0, interleave = concat_233_interleave_0, values = (concat_233_values0_0, expand_dims_144, expand_dims_145, expand_dims_2, expand_dims_147))[name = string("concat_233")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_233, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = k_state_25_cast_fp16, x = coreml_update_state_54)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_56_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_25_cast_fp16 = transpose(perm = v_state_25_perm_0, x = var_2309_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_233, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = v_state_25_cast_fp16, x = coreml_update_state_55)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_57_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [5]> var_2366_begin_0 = const()[name = string("op_2366_begin_0"), val = tensor<int32, [5]>([12, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2366_end_0 = const()[name = string("op_2366_end_0"), val = tensor<int32, [5]>([13, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2366_end_mask_0 = const()[name = string("op_2366_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2366_squeeze_mask_0 = const()[name = string("op_2366_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2366_cast_fp16 = slice_by_index(begin = var_2366_begin_0, end = var_2366_end_0, end_mask = var_2366_end_mask_0, squeeze_mask = var_2366_squeeze_mask_0, x = coreml_update_state_56)[name = string("op_2366_cast_fp16")];
            tensor<int32, [4]> var_2369_begin_0 = const()[name = string("op_2369_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2369_end_mask_0 = const()[name = string("op_2369_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2369_cast_fp16 = slice_by_index(begin = var_2369_begin_0, end = concat_11, end_mask = var_2369_end_mask_0, x = var_2366_cast_fp16)[name = string("op_2369_cast_fp16")];
            tensor<int32, [5]> var_2371_begin_0 = const()[name = string("op_2371_begin_0"), val = tensor<int32, [5]>([12, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2371_end_0 = const()[name = string("op_2371_end_0"), val = tensor<int32, [5]>([13, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2371_end_mask_0 = const()[name = string("op_2371_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2371_squeeze_mask_0 = const()[name = string("op_2371_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2371_cast_fp16 = slice_by_index(begin = var_2371_begin_0, end = var_2371_end_0, end_mask = var_2371_end_mask_0, squeeze_mask = var_2371_squeeze_mask_0, x = coreml_update_state_57)[name = string("op_2371_cast_fp16")];
            tensor<int32, [4]> var_2374_begin_0 = const()[name = string("op_2374_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2374_end_mask_0 = const()[name = string("op_2374_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2374_cast_fp16 = slice_by_index(begin = var_2374_begin_0, end = concat_11, end_mask = var_2374_end_mask_0, x = var_2371_cast_fp16)[name = string("op_2374_cast_fp16")];
            tensor<int32, [4]> var_2376_shape_cast_fp16 = shape(x = var_2369_cast_fp16)[name = string("op_2376_shape_cast_fp16")];
            int32 gather_229 = const()[name = string("gather_229"), val = int32(1)];
            int32 gather_230 = const()[name = string("gather_230"), val = int32(8)];
            int32 gather_231_axis_0 = const()[name = string("gather_231_axis_0"), val = int32(0)];
            int32 gather_231_batch_dims_0 = const()[name = string("gather_231_batch_dims_0"), val = int32(0)];
            bool gather_231_validate_indices_0 = const()[name = string("gather_231_validate_indices_0"), val = bool(false)];
            string var_2376_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2376_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_231_to_uint16 = const()[name = string("select_231_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2376_shape_cast_fp16_to_uint16 = cast(dtype = var_2376_shape_cast_fp16_to_uint16_dtype_0, x = var_2376_shape_cast_fp16)[name = string("cast_30")];
            uint16 gather_231_cast_uint16 = gather(axis = gather_231_axis_0, batch_dims = gather_231_batch_dims_0, indices = select_231_to_uint16, validate_indices = gather_231_validate_indices_0, x = var_2376_shape_cast_fp16_to_uint16)[name = string("gather_231_cast_uint16")];
            string gather_231_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_231_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_232 = const()[name = string("gather_232"), val = int32(64)];
            tensor<int32, [1]> var_2383_axes_0 = const()[name = string("op_2383_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2383_cast_fp16 = expand_dims(axes = var_2383_axes_0, x = var_2369_cast_fp16)[name = string("op_2383_cast_fp16")];
            tensor<int32, [5]> shape_257_cast_fp16 = shape(x = var_2383_cast_fp16)[name = string("shape_257_cast_fp16")];
            int32 concat_241_axis_0 = const()[name = string("concat_241_axis_0"), val = int32(0)];
            bool concat_241_interleave_0 = const()[name = string("concat_241_interleave_0"), val = bool(false)];
            int32 gather_231_cast_uint16_to_int32 = cast(dtype = gather_231_cast_uint16_to_int32_dtype_0, x = gather_231_cast_uint16)[name = string("cast_29")];
            tensor<int32, [5]> concat_241 = concat(axis = concat_241_axis_0, interleave = concat_241_interleave_0, values = (gather_229, gather_230, var_59, gather_231_cast_uint16_to_int32, gather_232))[name = string("concat_241")];
            tensor<int32, [5]> real_div_24 = real_div(x = concat_241, y = shape_257_cast_fp16)[name = string("real_div_24")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_371_cast_fp16 = tile(reps = real_div_24, x = var_2383_cast_fp16)[name = string("hidden_states_371_cast_fp16")];
            tensor<int32, [4]> concat_242x = const()[name = string("concat_242x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_51_cast_fp16 = reshape(shape = concat_242x, x = hidden_states_371_cast_fp16)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [4]> var_2393_shape_cast_fp16 = shape(x = var_2374_cast_fp16)[name = string("op_2393_shape_cast_fp16")];
            int32 gather_233 = const()[name = string("gather_233"), val = int32(1)];
            int32 gather_234 = const()[name = string("gather_234"), val = int32(8)];
            int32 gather_235_axis_0 = const()[name = string("gather_235_axis_0"), val = int32(0)];
            int32 gather_235_batch_dims_0 = const()[name = string("gather_235_batch_dims_0"), val = int32(0)];
            bool gather_235_validate_indices_0 = const()[name = string("gather_235_validate_indices_0"), val = bool(false)];
            string var_2393_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2393_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_235_to_uint16 = const()[name = string("select_235_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2393_shape_cast_fp16_to_uint16 = cast(dtype = var_2393_shape_cast_fp16_to_uint16_dtype_0, x = var_2393_shape_cast_fp16)[name = string("cast_28")];
            uint16 gather_235_cast_uint16 = gather(axis = gather_235_axis_0, batch_dims = gather_235_batch_dims_0, indices = select_235_to_uint16, validate_indices = gather_235_validate_indices_0, x = var_2393_shape_cast_fp16_to_uint16)[name = string("gather_235_cast_uint16")];
            string gather_235_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_235_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_236 = const()[name = string("gather_236"), val = int32(64)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2400_cast_fp16 = expand_dims(axes = var_2400_axes_0, x = var_2374_cast_fp16)[name = string("op_2400_cast_fp16")];
            tensor<int32, [5]> shape_262_cast_fp16 = shape(x = var_2400_cast_fp16)[name = string("shape_262_cast_fp16")];
            int32 concat_243_axis_0 = const()[name = string("concat_243_axis_0"), val = int32(0)];
            bool concat_243_interleave_0 = const()[name = string("concat_243_interleave_0"), val = bool(false)];
            int32 gather_235_cast_uint16_to_int32 = cast(dtype = gather_235_cast_uint16_to_int32_dtype_0, x = gather_235_cast_uint16)[name = string("cast_27")];
            tensor<int32, [5]> concat_243 = concat(axis = concat_243_axis_0, interleave = concat_243_interleave_0, values = (gather_233, gather_234, var_59, gather_235_cast_uint16_to_int32, gather_236))[name = string("concat_243")];
            tensor<int32, [5]> real_div_25 = real_div(x = concat_243, y = shape_262_cast_fp16)[name = string("real_div_25")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_375_cast_fp16 = tile(reps = real_div_25, x = var_2400_cast_fp16)[name = string("hidden_states_375_cast_fp16")];
            tensor<int32, [4]> concat_244x = const()[name = string("concat_244x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_51_cast_fp16 = reshape(shape = concat_244x, x = hidden_states_375_cast_fp16)[name = string("value_states_51_cast_fp16")];
            tensor<int32, [4]> var_2410_shape_cast_fp16 = shape(x = key_states_51_cast_fp16)[name = string("op_2410_shape_cast_fp16")];
            int32 gather_237_axis_0 = const()[name = string("gather_237_axis_0"), val = int32(0)];
            int32 gather_237_batch_dims_0 = const()[name = string("gather_237_batch_dims_0"), val = int32(0)];
            bool gather_237_validate_indices_0 = const()[name = string("gather_237_validate_indices_0"), val = bool(false)];
            string var_2410_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2410_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_237_to_uint16 = const()[name = string("select_237_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2410_shape_cast_fp16_to_uint16 = cast(dtype = var_2410_shape_cast_fp16_to_uint16_dtype_0, x = var_2410_shape_cast_fp16)[name = string("cast_26")];
            uint16 gather_237_cast_uint16 = gather(axis = gather_237_axis_0, batch_dims = gather_237_batch_dims_0, indices = select_237_to_uint16, validate_indices = gather_237_validate_indices_0, x = var_2410_shape_cast_fp16_to_uint16)[name = string("gather_237_cast_uint16")];
            string gather_237_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_237_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_245_values0_0 = const()[name = string("concat_245_values0_0"), val = int32(1)];
            int32 concat_245_values1_0 = const()[name = string("concat_245_values1_0"), val = int32(1)];
            int32 concat_245_values2_0 = const()[name = string("concat_245_values2_0"), val = int32(0)];
            int32 concat_245_axis_0 = const()[name = string("concat_245_axis_0"), val = int32(0)];
            bool concat_245_interleave_0 = const()[name = string("concat_245_interleave_0"), val = bool(false)];
            int32 gather_237_cast_uint16_to_int32 = cast(dtype = gather_237_cast_uint16_to_int32_dtype_0, x = gather_237_cast_uint16)[name = string("cast_25")];
            tensor<int32, [4]> concat_245 = concat(axis = concat_245_axis_0, interleave = concat_245_interleave_0, values = (concat_245_values0_0, concat_245_values1_0, concat_245_values2_0, gather_237_cast_uint16_to_int32))[name = string("concat_245")];
            tensor<int32, [4]> causal_mask_27_begin_0 = const()[name = string("causal_mask_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_27_end_mask_0 = const()[name = string("causal_mask_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_27_cast_fp16 = slice_by_index(begin = causal_mask_27_begin_0, end = concat_245, end_mask = causal_mask_27_end_mask_0, x = causal_mask)[name = string("causal_mask_27_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_49_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_27_cast_fp16, key = key_states_51_cast_fp16, query = query_states_51_cast_fp16, value = value_states_51_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2416_perm_0 = const()[name = string("op_2416_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_246_axis_0 = const()[name = string("concat_246_axis_0"), val = int32(0)];
            bool concat_246_interleave_0 = const()[name = string("concat_246_interleave_0"), val = bool(false)];
            int32 gather_221_cast_uint16_to_int32 = cast(dtype = gather_221_cast_uint16_to_int32_dtype_0, x = gather_221_cast_uint16)[name = string("cast_24")];
            tensor<int32, [3]> concat_246 = concat(axis = concat_246_axis_0, interleave = concat_246_interleave_0, values = (gather_220, gather_221_cast_uint16_to_int32, var_48))[name = string("concat_246")];
            tensor<fp16, [1, ?, 32, 64]> var_2416_cast_fp16 = transpose(perm = var_2416_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, ?, ?]> input_97_cast_fp16 = reshape(shape = concat_246, x = var_2416_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561944512))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564041728))))[name = string("model_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_87_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_379_cast_fp16 = add(x = hidden_states_359_cast_fp16, y = linear_87_cast_fp16)[name = string("hidden_states_379_cast_fp16")];
            fp16 var_54_promoted_25_to_fp16 = const()[name = string("op_54_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2425_cast_fp16 = pow(x = hidden_states_379_cast_fp16, y = var_54_promoted_25_to_fp16)[name = string("op_2425_cast_fp16")];
            tensor<int32, [1]> variance_51_axes_0 = const()[name = string("variance_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_51_keep_dims_0 = const()[name = string("variance_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_51_cast_fp16 = reduce_mean(axes = variance_51_axes_0, keep_dims = variance_51_keep_dims_0, x = var_2425_cast_fp16)[name = string("variance_51_cast_fp16")];
            fp16 var_2428_to_fp16 = const()[name = string("op_2428_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2429_cast_fp16 = add(x = variance_51_cast_fp16, y = var_2428_to_fp16)[name = string("op_2429_cast_fp16")];
            fp32 var_2430_epsilon_0 = const()[name = string("op_2430_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2430_cast_fp16 = rsqrt(epsilon = var_2430_epsilon_0, x = var_2429_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_383_cast_fp16 = mul(x = hidden_states_379_cast_fp16, y = var_2430_cast_fp16)[name = string("hidden_states_383_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564303936)))];
            tensor<fp16, [1, ?, 2048]> input_99_cast_fp16 = mul(x = model_model_layers_12_post_attention_layernorm_weight_to_fp16, y = hidden_states_383_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564308096))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572696768))))[name = string("model_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_2442_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2442_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573745408))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582134080))))[name = string("model_model_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_103_cast_fp16 = mul(x = var_2442_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583182720))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591571392))))[name = string("model_model_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_389_cast_fp16 = add(x = hidden_states_379_cast_fp16, y = linear_90_cast_fp16)[name = string("hidden_states_389_cast_fp16")];
            fp16 var_54_promoted_26_to_fp16 = const()[name = string("op_54_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2455_cast_fp16 = pow(x = hidden_states_389_cast_fp16, y = var_54_promoted_26_to_fp16)[name = string("op_2455_cast_fp16")];
            tensor<int32, [1]> variance_53_axes_0 = const()[name = string("variance_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_53_keep_dims_0 = const()[name = string("variance_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_53_cast_fp16 = reduce_mean(axes = variance_53_axes_0, keep_dims = variance_53_keep_dims_0, x = var_2455_cast_fp16)[name = string("variance_53_cast_fp16")];
            fp16 var_2458_to_fp16 = const()[name = string("op_2458_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2459_cast_fp16 = add(x = variance_53_cast_fp16, y = var_2458_to_fp16)[name = string("op_2459_cast_fp16")];
            fp32 var_2460_epsilon_0 = const()[name = string("op_2460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2460_cast_fp16 = rsqrt(epsilon = var_2460_epsilon_0, x = var_2459_cast_fp16)[name = string("op_2460_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_393_cast_fp16 = mul(x = hidden_states_389_cast_fp16, y = var_2460_cast_fp16)[name = string("hidden_states_393_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592620032)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_397_cast_fp16 = mul(x = model_model_layers_13_input_layernorm_weight_to_fp16, y = hidden_states_393_cast_fp16)[name = string("hidden_states_397_cast_fp16")];
            tensor<int32, [3]> var_2471_shape_cast_fp16 = shape(x = hidden_states_397_cast_fp16)[name = string("op_2471_shape_cast_fp16")];
            int32 gather_238 = const()[name = string("gather_238"), val = int32(1)];
            int32 gather_239_axis_0 = const()[name = string("gather_239_axis_0"), val = int32(0)];
            int32 gather_239_batch_dims_0 = const()[name = string("gather_239_batch_dims_0"), val = int32(0)];
            bool gather_239_validate_indices_0 = const()[name = string("gather_239_validate_indices_0"), val = bool(false)];
            string var_2471_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2471_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_239_to_uint16 = const()[name = string("select_239_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2471_shape_cast_fp16_to_uint16 = cast(dtype = var_2471_shape_cast_fp16_to_uint16_dtype_0, x = var_2471_shape_cast_fp16)[name = string("cast_23")];
            uint16 gather_239_cast_uint16 = gather(axis = gather_239_axis_0, batch_dims = gather_239_batch_dims_0, indices = select_239_to_uint16, validate_indices = gather_239_validate_indices_0, x = var_2471_shape_cast_fp16_to_uint16)[name = string("gather_239_cast_uint16")];
            string gather_239_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_239_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592624192))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594721408))))[name = string("model_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_397_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594983616))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595507968))))[name = string("model_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_397_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595573568))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596097920))))[name = string("model_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_397_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> concat_247x = const()[name = string("concat_247x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_2480_cast_fp16 = reshape(shape = concat_247x, x = linear_91_cast_fp16)[name = string("op_2480_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_248x = const()[name = string("concat_248x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2483_cast_fp16 = reshape(shape = concat_248x, x = linear_92_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_249x = const()[name = string("concat_249x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2486_cast_fp16 = reshape(shape = concat_249x, x = linear_93_cast_fp16)[name = string("op_2486_cast_fp16")];
            tensor<int32, [4]> v_state_27_perm_0 = const()[name = string("v_state_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2480_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 32, ?, 64]> var_2490_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_2501_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_2501_cast_fp16")];
            bool var_2503_interleave_0 = const()[name = string("op_2503_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_2503_cast_fp16 = concat(axis = var_48, interleave = var_2503_interleave_0, values = (var_2501_cast_fp16, x1_53_cast_fp16))[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_2504_cast_fp16 = mul(x = var_2503_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_55_cast_fp16 = add(x = var_2490_cast_fp16, y = var_2504_cast_fp16)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2483_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 8, ?, 64]> var_2506_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2506_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_2517_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_2517_cast_fp16")];
            bool var_2519_interleave_0 = const()[name = string("op_2519_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_2519_cast_fp16 = concat(axis = var_48, interleave = var_2519_interleave_0, values = (var_2517_cast_fp16, x1_55_cast_fp16))[name = string("op_2519_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2520_cast_fp16 = mul(x = var_2519_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_27_cast_fp16 = add(x = var_2506_cast_fp16, y = var_2520_cast_fp16)[name = string("k_state_27_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_252_values0_0 = const()[name = string("concat_252_values0_0"), val = tensor<int32, [1]>([13])];
            int32 concat_252_axis_0 = const()[name = string("concat_252_axis_0"), val = int32(0)];
            bool concat_252_interleave_0 = const()[name = string("concat_252_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_252 = concat(axis = concat_252_axis_0, interleave = concat_252_interleave_0, values = (concat_252_values0_0, expand_dims_156, expand_dims_157, expand_dims_2, expand_dims_159))[name = string("concat_252")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_252, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = k_state_27_cast_fp16, x = coreml_update_state_56)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_58_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_27_cast_fp16 = transpose(perm = v_state_27_perm_0, x = var_2486_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_252, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = v_state_27_cast_fp16, x = coreml_update_state_57)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_59_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [5]> var_2543_begin_0 = const()[name = string("op_2543_begin_0"), val = tensor<int32, [5]>([13, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2543_end_0 = const()[name = string("op_2543_end_0"), val = tensor<int32, [5]>([14, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2543_end_mask_0 = const()[name = string("op_2543_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2543_squeeze_mask_0 = const()[name = string("op_2543_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2543_cast_fp16 = slice_by_index(begin = var_2543_begin_0, end = var_2543_end_0, end_mask = var_2543_end_mask_0, squeeze_mask = var_2543_squeeze_mask_0, x = coreml_update_state_58)[name = string("op_2543_cast_fp16")];
            tensor<int32, [4]> var_2546_begin_0 = const()[name = string("op_2546_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2546_end_mask_0 = const()[name = string("op_2546_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2546_cast_fp16 = slice_by_index(begin = var_2546_begin_0, end = concat_11, end_mask = var_2546_end_mask_0, x = var_2543_cast_fp16)[name = string("op_2546_cast_fp16")];
            tensor<int32, [5]> var_2548_begin_0 = const()[name = string("op_2548_begin_0"), val = tensor<int32, [5]>([13, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2548_end_0 = const()[name = string("op_2548_end_0"), val = tensor<int32, [5]>([14, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2548_end_mask_0 = const()[name = string("op_2548_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2548_squeeze_mask_0 = const()[name = string("op_2548_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2548_cast_fp16 = slice_by_index(begin = var_2548_begin_0, end = var_2548_end_0, end_mask = var_2548_end_mask_0, squeeze_mask = var_2548_squeeze_mask_0, x = coreml_update_state_59)[name = string("op_2548_cast_fp16")];
            tensor<int32, [4]> var_2551_begin_0 = const()[name = string("op_2551_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2551_end_mask_0 = const()[name = string("op_2551_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2551_cast_fp16 = slice_by_index(begin = var_2551_begin_0, end = concat_11, end_mask = var_2551_end_mask_0, x = var_2548_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<int32, [4]> var_2553_shape_cast_fp16 = shape(x = var_2546_cast_fp16)[name = string("op_2553_shape_cast_fp16")];
            int32 gather_247 = const()[name = string("gather_247"), val = int32(1)];
            int32 gather_248 = const()[name = string("gather_248"), val = int32(8)];
            int32 gather_249_axis_0 = const()[name = string("gather_249_axis_0"), val = int32(0)];
            int32 gather_249_batch_dims_0 = const()[name = string("gather_249_batch_dims_0"), val = int32(0)];
            bool gather_249_validate_indices_0 = const()[name = string("gather_249_validate_indices_0"), val = bool(false)];
            string var_2553_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2553_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_249_to_uint16 = const()[name = string("select_249_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2553_shape_cast_fp16_to_uint16 = cast(dtype = var_2553_shape_cast_fp16_to_uint16_dtype_0, x = var_2553_shape_cast_fp16)[name = string("cast_22")];
            uint16 gather_249_cast_uint16 = gather(axis = gather_249_axis_0, batch_dims = gather_249_batch_dims_0, indices = select_249_to_uint16, validate_indices = gather_249_validate_indices_0, x = var_2553_shape_cast_fp16_to_uint16)[name = string("gather_249_cast_uint16")];
            string gather_249_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_249_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_250 = const()[name = string("gather_250"), val = int32(64)];
            tensor<int32, [1]> var_2560_axes_0 = const()[name = string("op_2560_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2560_cast_fp16 = expand_dims(axes = var_2560_axes_0, x = var_2546_cast_fp16)[name = string("op_2560_cast_fp16")];
            tensor<int32, [5]> shape_277_cast_fp16 = shape(x = var_2560_cast_fp16)[name = string("shape_277_cast_fp16")];
            int32 concat_260_axis_0 = const()[name = string("concat_260_axis_0"), val = int32(0)];
            bool concat_260_interleave_0 = const()[name = string("concat_260_interleave_0"), val = bool(false)];
            int32 gather_249_cast_uint16_to_int32 = cast(dtype = gather_249_cast_uint16_to_int32_dtype_0, x = gather_249_cast_uint16)[name = string("cast_21")];
            tensor<int32, [5]> concat_260 = concat(axis = concat_260_axis_0, interleave = concat_260_interleave_0, values = (gather_247, gather_248, var_59, gather_249_cast_uint16_to_int32, gather_250))[name = string("concat_260")];
            tensor<int32, [5]> real_div_26 = real_div(x = concat_260, y = shape_277_cast_fp16)[name = string("real_div_26")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_401_cast_fp16 = tile(reps = real_div_26, x = var_2560_cast_fp16)[name = string("hidden_states_401_cast_fp16")];
            tensor<int32, [4]> concat_261x = const()[name = string("concat_261x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_55_cast_fp16 = reshape(shape = concat_261x, x = hidden_states_401_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [4]> var_2570_shape_cast_fp16 = shape(x = var_2551_cast_fp16)[name = string("op_2570_shape_cast_fp16")];
            int32 gather_251 = const()[name = string("gather_251"), val = int32(1)];
            int32 gather_252 = const()[name = string("gather_252"), val = int32(8)];
            int32 gather_253_axis_0 = const()[name = string("gather_253_axis_0"), val = int32(0)];
            int32 gather_253_batch_dims_0 = const()[name = string("gather_253_batch_dims_0"), val = int32(0)];
            bool gather_253_validate_indices_0 = const()[name = string("gather_253_validate_indices_0"), val = bool(false)];
            string var_2570_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2570_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_253_to_uint16 = const()[name = string("select_253_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2570_shape_cast_fp16_to_uint16 = cast(dtype = var_2570_shape_cast_fp16_to_uint16_dtype_0, x = var_2570_shape_cast_fp16)[name = string("cast_20")];
            uint16 gather_253_cast_uint16 = gather(axis = gather_253_axis_0, batch_dims = gather_253_batch_dims_0, indices = select_253_to_uint16, validate_indices = gather_253_validate_indices_0, x = var_2570_shape_cast_fp16_to_uint16)[name = string("gather_253_cast_uint16")];
            string gather_253_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_253_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_254 = const()[name = string("gather_254"), val = int32(64)];
            tensor<int32, [1]> var_2577_axes_0 = const()[name = string("op_2577_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2577_cast_fp16 = expand_dims(axes = var_2577_axes_0, x = var_2551_cast_fp16)[name = string("op_2577_cast_fp16")];
            tensor<int32, [5]> shape_282_cast_fp16 = shape(x = var_2577_cast_fp16)[name = string("shape_282_cast_fp16")];
            int32 concat_262_axis_0 = const()[name = string("concat_262_axis_0"), val = int32(0)];
            bool concat_262_interleave_0 = const()[name = string("concat_262_interleave_0"), val = bool(false)];
            int32 gather_253_cast_uint16_to_int32 = cast(dtype = gather_253_cast_uint16_to_int32_dtype_0, x = gather_253_cast_uint16)[name = string("cast_19")];
            tensor<int32, [5]> concat_262 = concat(axis = concat_262_axis_0, interleave = concat_262_interleave_0, values = (gather_251, gather_252, var_59, gather_253_cast_uint16_to_int32, gather_254))[name = string("concat_262")];
            tensor<int32, [5]> real_div_27 = real_div(x = concat_262, y = shape_282_cast_fp16)[name = string("real_div_27")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_405_cast_fp16 = tile(reps = real_div_27, x = var_2577_cast_fp16)[name = string("hidden_states_405_cast_fp16")];
            tensor<int32, [4]> concat_263x = const()[name = string("concat_263x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_55_cast_fp16 = reshape(shape = concat_263x, x = hidden_states_405_cast_fp16)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> var_2587_shape_cast_fp16 = shape(x = key_states_55_cast_fp16)[name = string("op_2587_shape_cast_fp16")];
            int32 gather_255_axis_0 = const()[name = string("gather_255_axis_0"), val = int32(0)];
            int32 gather_255_batch_dims_0 = const()[name = string("gather_255_batch_dims_0"), val = int32(0)];
            bool gather_255_validate_indices_0 = const()[name = string("gather_255_validate_indices_0"), val = bool(false)];
            string var_2587_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2587_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_255_to_uint16 = const()[name = string("select_255_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2587_shape_cast_fp16_to_uint16 = cast(dtype = var_2587_shape_cast_fp16_to_uint16_dtype_0, x = var_2587_shape_cast_fp16)[name = string("cast_18")];
            uint16 gather_255_cast_uint16 = gather(axis = gather_255_axis_0, batch_dims = gather_255_batch_dims_0, indices = select_255_to_uint16, validate_indices = gather_255_validate_indices_0, x = var_2587_shape_cast_fp16_to_uint16)[name = string("gather_255_cast_uint16")];
            string gather_255_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_255_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_264_values0_0 = const()[name = string("concat_264_values0_0"), val = int32(1)];
            int32 concat_264_values1_0 = const()[name = string("concat_264_values1_0"), val = int32(1)];
            int32 concat_264_values2_0 = const()[name = string("concat_264_values2_0"), val = int32(0)];
            int32 concat_264_axis_0 = const()[name = string("concat_264_axis_0"), val = int32(0)];
            bool concat_264_interleave_0 = const()[name = string("concat_264_interleave_0"), val = bool(false)];
            int32 gather_255_cast_uint16_to_int32 = cast(dtype = gather_255_cast_uint16_to_int32_dtype_0, x = gather_255_cast_uint16)[name = string("cast_17")];
            tensor<int32, [4]> concat_264 = concat(axis = concat_264_axis_0, interleave = concat_264_interleave_0, values = (concat_264_values0_0, concat_264_values1_0, concat_264_values2_0, gather_255_cast_uint16_to_int32))[name = string("concat_264")];
            tensor<int32, [4]> causal_mask_29_begin_0 = const()[name = string("causal_mask_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_29_end_mask_0 = const()[name = string("causal_mask_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_29_cast_fp16 = slice_by_index(begin = causal_mask_29_begin_0, end = concat_264, end_mask = causal_mask_29_end_mask_0, x = causal_mask)[name = string("causal_mask_29_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_53_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_29_cast_fp16, key = key_states_55_cast_fp16, query = query_states_55_cast_fp16, value = value_states_55_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2593_perm_0 = const()[name = string("op_2593_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_265_axis_0 = const()[name = string("concat_265_axis_0"), val = int32(0)];
            bool concat_265_interleave_0 = const()[name = string("concat_265_interleave_0"), val = bool(false)];
            int32 gather_239_cast_uint16_to_int32 = cast(dtype = gather_239_cast_uint16_to_int32_dtype_0, x = gather_239_cast_uint16)[name = string("cast_16")];
            tensor<int32, [3]> concat_265 = concat(axis = concat_265_axis_0, interleave = concat_265_interleave_0, values = (gather_238, gather_239_cast_uint16_to_int32, var_48))[name = string("concat_265")];
            tensor<fp16, [1, ?, 32, 64]> var_2593_cast_fp16 = transpose(perm = var_2593_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, ?, ?]> input_105_cast_fp16 = reshape(shape = concat_265, x = var_2593_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596163520))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598260736))))[name = string("model_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_409_cast_fp16 = add(x = hidden_states_389_cast_fp16, y = linear_94_cast_fp16)[name = string("hidden_states_409_cast_fp16")];
            fp16 var_54_promoted_27_to_fp16 = const()[name = string("op_54_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2602_cast_fp16 = pow(x = hidden_states_409_cast_fp16, y = var_54_promoted_27_to_fp16)[name = string("op_2602_cast_fp16")];
            tensor<int32, [1]> variance_55_axes_0 = const()[name = string("variance_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_55_keep_dims_0 = const()[name = string("variance_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_55_cast_fp16 = reduce_mean(axes = variance_55_axes_0, keep_dims = variance_55_keep_dims_0, x = var_2602_cast_fp16)[name = string("variance_55_cast_fp16")];
            fp16 var_2605_to_fp16 = const()[name = string("op_2605_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2606_cast_fp16 = add(x = variance_55_cast_fp16, y = var_2605_to_fp16)[name = string("op_2606_cast_fp16")];
            fp32 var_2607_epsilon_0 = const()[name = string("op_2607_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2607_cast_fp16 = rsqrt(epsilon = var_2607_epsilon_0, x = var_2606_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_413_cast_fp16 = mul(x = hidden_states_409_cast_fp16, y = var_2607_cast_fp16)[name = string("hidden_states_413_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598522944)))];
            tensor<fp16, [1, ?, 2048]> input_107_cast_fp16 = mul(x = model_model_layers_13_post_attention_layernorm_weight_to_fp16, y = hidden_states_413_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598527104))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606915776))))[name = string("model_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_2619_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2619_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607964416))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616353088))))[name = string("model_model_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_111_cast_fp16 = mul(x = var_2619_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617401728))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(625790400))))[name = string("model_model_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_97_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_419_cast_fp16 = add(x = hidden_states_409_cast_fp16, y = linear_97_cast_fp16)[name = string("hidden_states_419_cast_fp16")];
            fp16 var_54_promoted_28_to_fp16 = const()[name = string("op_54_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2632_cast_fp16 = pow(x = hidden_states_419_cast_fp16, y = var_54_promoted_28_to_fp16)[name = string("op_2632_cast_fp16")];
            tensor<int32, [1]> variance_57_axes_0 = const()[name = string("variance_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_57_keep_dims_0 = const()[name = string("variance_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_57_cast_fp16 = reduce_mean(axes = variance_57_axes_0, keep_dims = variance_57_keep_dims_0, x = var_2632_cast_fp16)[name = string("variance_57_cast_fp16")];
            fp16 var_2635_to_fp16 = const()[name = string("op_2635_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2636_cast_fp16 = add(x = variance_57_cast_fp16, y = var_2635_to_fp16)[name = string("op_2636_cast_fp16")];
            fp32 var_2637_epsilon_0 = const()[name = string("op_2637_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2637_cast_fp16 = rsqrt(epsilon = var_2637_epsilon_0, x = var_2636_cast_fp16)[name = string("op_2637_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_423_cast_fp16 = mul(x = hidden_states_419_cast_fp16, y = var_2637_cast_fp16)[name = string("hidden_states_423_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626839040)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_427_cast_fp16 = mul(x = model_model_layers_14_input_layernorm_weight_to_fp16, y = hidden_states_423_cast_fp16)[name = string("hidden_states_427_cast_fp16")];
            tensor<int32, [3]> var_2648_shape_cast_fp16 = shape(x = hidden_states_427_cast_fp16)[name = string("op_2648_shape_cast_fp16")];
            int32 gather_256 = const()[name = string("gather_256"), val = int32(1)];
            int32 gather_257_axis_0 = const()[name = string("gather_257_axis_0"), val = int32(0)];
            int32 gather_257_batch_dims_0 = const()[name = string("gather_257_batch_dims_0"), val = int32(0)];
            bool gather_257_validate_indices_0 = const()[name = string("gather_257_validate_indices_0"), val = bool(false)];
            string var_2648_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2648_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_257_to_uint16 = const()[name = string("select_257_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2648_shape_cast_fp16_to_uint16 = cast(dtype = var_2648_shape_cast_fp16_to_uint16_dtype_0, x = var_2648_shape_cast_fp16)[name = string("cast_15")];
            uint16 gather_257_cast_uint16 = gather(axis = gather_257_axis_0, batch_dims = gather_257_batch_dims_0, indices = select_257_to_uint16, validate_indices = gather_257_validate_indices_0, x = var_2648_shape_cast_fp16_to_uint16)[name = string("gather_257_cast_uint16")];
            string gather_257_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_257_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626843200))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628940416))))[name = string("model_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_427_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629202624))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629726976))))[name = string("model_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_427_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629792576))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630316928))))[name = string("model_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_427_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> concat_266x = const()[name = string("concat_266x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_2657_cast_fp16 = reshape(shape = concat_266x, x = linear_98_cast_fp16)[name = string("op_2657_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_267x = const()[name = string("concat_267x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2660_cast_fp16 = reshape(shape = concat_267x, x = linear_99_cast_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_268x = const()[name = string("concat_268x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2663_cast_fp16 = reshape(shape = concat_268x, x = linear_100_cast_fp16)[name = string("op_2663_cast_fp16")];
            tensor<int32, [4]> v_state_29_perm_0 = const()[name = string("v_state_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2657_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 32, ?, 64]> var_2667_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_2678_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_2678_cast_fp16")];
            bool var_2680_interleave_0 = const()[name = string("op_2680_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_2680_cast_fp16 = concat(axis = var_48, interleave = var_2680_interleave_0, values = (var_2678_cast_fp16, x1_57_cast_fp16))[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_2681_cast_fp16 = mul(x = var_2680_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2681_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_59_cast_fp16 = add(x = var_2667_cast_fp16, y = var_2681_cast_fp16)[name = string("query_states_59_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2660_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, ?, 64]> var_2683_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_2694_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_2694_cast_fp16")];
            bool var_2696_interleave_0 = const()[name = string("op_2696_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_2696_cast_fp16 = concat(axis = var_48, interleave = var_2696_interleave_0, values = (var_2694_cast_fp16, x1_59_cast_fp16))[name = string("op_2696_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2697_cast_fp16 = mul(x = var_2696_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2697_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_29_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2697_cast_fp16)[name = string("k_state_29_cast_fp16")];
            tensor<int32, [1]> expand_dims_168 = const()[name = string("expand_dims_168"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_169 = const()[name = string("expand_dims_169"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_171 = const()[name = string("expand_dims_171"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_271_values0_0 = const()[name = string("concat_271_values0_0"), val = tensor<int32, [1]>([14])];
            int32 concat_271_axis_0 = const()[name = string("concat_271_axis_0"), val = int32(0)];
            bool concat_271_interleave_0 = const()[name = string("concat_271_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_271 = concat(axis = concat_271_axis_0, interleave = concat_271_interleave_0, values = (concat_271_values0_0, expand_dims_168, expand_dims_169, expand_dims_2, expand_dims_171))[name = string("concat_271")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_15_stride_0 = const()[name = string("key_cache_internal_tensor_assign_15_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_15_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_15_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_271, begin_mask = key_cache_internal_tensor_assign_15_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_15_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_15_squeeze_mask_0, stride = key_cache_internal_tensor_assign_15_stride_0, update = k_state_29_cast_fp16, x = coreml_update_state_58)[name = string("key_cache_internal_tensor_assign_15_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_15_cast_fp16, input = key_cache)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_15_stride_0 = const()[name = string("value_cache_internal_tensor_assign_15_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_15_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_15_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_29_cast_fp16 = transpose(perm = v_state_29_perm_0, x = var_2663_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_271, begin_mask = value_cache_internal_tensor_assign_15_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_15_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_15_squeeze_mask_0, stride = value_cache_internal_tensor_assign_15_stride_0, update = v_state_29_cast_fp16, x = coreml_update_state_59)[name = string("value_cache_internal_tensor_assign_15_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_15_cast_fp16, input = value_cache)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [5]> var_2720_begin_0 = const()[name = string("op_2720_begin_0"), val = tensor<int32, [5]>([14, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2720_end_0 = const()[name = string("op_2720_end_0"), val = tensor<int32, [5]>([15, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2720_end_mask_0 = const()[name = string("op_2720_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2720_squeeze_mask_0 = const()[name = string("op_2720_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2720_cast_fp16 = slice_by_index(begin = var_2720_begin_0, end = var_2720_end_0, end_mask = var_2720_end_mask_0, squeeze_mask = var_2720_squeeze_mask_0, x = coreml_update_state_60)[name = string("op_2720_cast_fp16")];
            tensor<int32, [4]> var_2723_begin_0 = const()[name = string("op_2723_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2723_end_mask_0 = const()[name = string("op_2723_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2723_cast_fp16 = slice_by_index(begin = var_2723_begin_0, end = concat_11, end_mask = var_2723_end_mask_0, x = var_2720_cast_fp16)[name = string("op_2723_cast_fp16")];
            tensor<int32, [5]> var_2725_begin_0 = const()[name = string("op_2725_begin_0"), val = tensor<int32, [5]>([14, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2725_end_0 = const()[name = string("op_2725_end_0"), val = tensor<int32, [5]>([15, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2725_end_mask_0 = const()[name = string("op_2725_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2725_squeeze_mask_0 = const()[name = string("op_2725_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2725_cast_fp16 = slice_by_index(begin = var_2725_begin_0, end = var_2725_end_0, end_mask = var_2725_end_mask_0, squeeze_mask = var_2725_squeeze_mask_0, x = coreml_update_state_61)[name = string("op_2725_cast_fp16")];
            tensor<int32, [4]> var_2728_begin_0 = const()[name = string("op_2728_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2728_end_mask_0 = const()[name = string("op_2728_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2728_cast_fp16 = slice_by_index(begin = var_2728_begin_0, end = concat_11, end_mask = var_2728_end_mask_0, x = var_2725_cast_fp16)[name = string("op_2728_cast_fp16")];
            tensor<int32, [4]> var_2730_shape_cast_fp16 = shape(x = var_2723_cast_fp16)[name = string("op_2730_shape_cast_fp16")];
            int32 gather_265 = const()[name = string("gather_265"), val = int32(1)];
            int32 gather_266 = const()[name = string("gather_266"), val = int32(8)];
            int32 gather_267_axis_0 = const()[name = string("gather_267_axis_0"), val = int32(0)];
            int32 gather_267_batch_dims_0 = const()[name = string("gather_267_batch_dims_0"), val = int32(0)];
            bool gather_267_validate_indices_0 = const()[name = string("gather_267_validate_indices_0"), val = bool(false)];
            string var_2730_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2730_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_267_to_uint16 = const()[name = string("select_267_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2730_shape_cast_fp16_to_uint16 = cast(dtype = var_2730_shape_cast_fp16_to_uint16_dtype_0, x = var_2730_shape_cast_fp16)[name = string("cast_14")];
            uint16 gather_267_cast_uint16 = gather(axis = gather_267_axis_0, batch_dims = gather_267_batch_dims_0, indices = select_267_to_uint16, validate_indices = gather_267_validate_indices_0, x = var_2730_shape_cast_fp16_to_uint16)[name = string("gather_267_cast_uint16")];
            string gather_267_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_267_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_268 = const()[name = string("gather_268"), val = int32(64)];
            tensor<int32, [1]> var_2737_axes_0 = const()[name = string("op_2737_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2737_cast_fp16 = expand_dims(axes = var_2737_axes_0, x = var_2723_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<int32, [5]> shape_297_cast_fp16 = shape(x = var_2737_cast_fp16)[name = string("shape_297_cast_fp16")];
            int32 concat_279_axis_0 = const()[name = string("concat_279_axis_0"), val = int32(0)];
            bool concat_279_interleave_0 = const()[name = string("concat_279_interleave_0"), val = bool(false)];
            int32 gather_267_cast_uint16_to_int32 = cast(dtype = gather_267_cast_uint16_to_int32_dtype_0, x = gather_267_cast_uint16)[name = string("cast_13")];
            tensor<int32, [5]> concat_279 = concat(axis = concat_279_axis_0, interleave = concat_279_interleave_0, values = (gather_265, gather_266, var_59, gather_267_cast_uint16_to_int32, gather_268))[name = string("concat_279")];
            tensor<int32, [5]> real_div_28 = real_div(x = concat_279, y = shape_297_cast_fp16)[name = string("real_div_28")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_431_cast_fp16 = tile(reps = real_div_28, x = var_2737_cast_fp16)[name = string("hidden_states_431_cast_fp16")];
            tensor<int32, [4]> concat_280x = const()[name = string("concat_280x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_59_cast_fp16 = reshape(shape = concat_280x, x = hidden_states_431_cast_fp16)[name = string("key_states_59_cast_fp16")];
            tensor<int32, [4]> var_2747_shape_cast_fp16 = shape(x = var_2728_cast_fp16)[name = string("op_2747_shape_cast_fp16")];
            int32 gather_269 = const()[name = string("gather_269"), val = int32(1)];
            int32 gather_270 = const()[name = string("gather_270"), val = int32(8)];
            int32 gather_271_axis_0 = const()[name = string("gather_271_axis_0"), val = int32(0)];
            int32 gather_271_batch_dims_0 = const()[name = string("gather_271_batch_dims_0"), val = int32(0)];
            bool gather_271_validate_indices_0 = const()[name = string("gather_271_validate_indices_0"), val = bool(false)];
            string var_2747_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2747_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_271_to_uint16 = const()[name = string("select_271_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2747_shape_cast_fp16_to_uint16 = cast(dtype = var_2747_shape_cast_fp16_to_uint16_dtype_0, x = var_2747_shape_cast_fp16)[name = string("cast_12")];
            uint16 gather_271_cast_uint16 = gather(axis = gather_271_axis_0, batch_dims = gather_271_batch_dims_0, indices = select_271_to_uint16, validate_indices = gather_271_validate_indices_0, x = var_2747_shape_cast_fp16_to_uint16)[name = string("gather_271_cast_uint16")];
            string gather_271_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_271_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_272 = const()[name = string("gather_272"), val = int32(64)];
            tensor<int32, [1]> var_2754_axes_0 = const()[name = string("op_2754_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2754_cast_fp16 = expand_dims(axes = var_2754_axes_0, x = var_2728_cast_fp16)[name = string("op_2754_cast_fp16")];
            tensor<int32, [5]> shape_302_cast_fp16 = shape(x = var_2754_cast_fp16)[name = string("shape_302_cast_fp16")];
            int32 concat_281_axis_0 = const()[name = string("concat_281_axis_0"), val = int32(0)];
            bool concat_281_interleave_0 = const()[name = string("concat_281_interleave_0"), val = bool(false)];
            int32 gather_271_cast_uint16_to_int32 = cast(dtype = gather_271_cast_uint16_to_int32_dtype_0, x = gather_271_cast_uint16)[name = string("cast_11")];
            tensor<int32, [5]> concat_281 = concat(axis = concat_281_axis_0, interleave = concat_281_interleave_0, values = (gather_269, gather_270, var_59, gather_271_cast_uint16_to_int32, gather_272))[name = string("concat_281")];
            tensor<int32, [5]> real_div_29 = real_div(x = concat_281, y = shape_302_cast_fp16)[name = string("real_div_29")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_435_cast_fp16 = tile(reps = real_div_29, x = var_2754_cast_fp16)[name = string("hidden_states_435_cast_fp16")];
            tensor<int32, [4]> concat_282x = const()[name = string("concat_282x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_59_cast_fp16 = reshape(shape = concat_282x, x = hidden_states_435_cast_fp16)[name = string("value_states_59_cast_fp16")];
            tensor<int32, [4]> var_2764_shape_cast_fp16 = shape(x = key_states_59_cast_fp16)[name = string("op_2764_shape_cast_fp16")];
            int32 gather_273_axis_0 = const()[name = string("gather_273_axis_0"), val = int32(0)];
            int32 gather_273_batch_dims_0 = const()[name = string("gather_273_batch_dims_0"), val = int32(0)];
            bool gather_273_validate_indices_0 = const()[name = string("gather_273_validate_indices_0"), val = bool(false)];
            string var_2764_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2764_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_273_to_uint16 = const()[name = string("select_273_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2764_shape_cast_fp16_to_uint16 = cast(dtype = var_2764_shape_cast_fp16_to_uint16_dtype_0, x = var_2764_shape_cast_fp16)[name = string("cast_10")];
            uint16 gather_273_cast_uint16 = gather(axis = gather_273_axis_0, batch_dims = gather_273_batch_dims_0, indices = select_273_to_uint16, validate_indices = gather_273_validate_indices_0, x = var_2764_shape_cast_fp16_to_uint16)[name = string("gather_273_cast_uint16")];
            string gather_273_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_273_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_283_values0_0 = const()[name = string("concat_283_values0_0"), val = int32(1)];
            int32 concat_283_values1_0 = const()[name = string("concat_283_values1_0"), val = int32(1)];
            int32 concat_283_values2_0 = const()[name = string("concat_283_values2_0"), val = int32(0)];
            int32 concat_283_axis_0 = const()[name = string("concat_283_axis_0"), val = int32(0)];
            bool concat_283_interleave_0 = const()[name = string("concat_283_interleave_0"), val = bool(false)];
            int32 gather_273_cast_uint16_to_int32 = cast(dtype = gather_273_cast_uint16_to_int32_dtype_0, x = gather_273_cast_uint16)[name = string("cast_9")];
            tensor<int32, [4]> concat_283 = concat(axis = concat_283_axis_0, interleave = concat_283_interleave_0, values = (concat_283_values0_0, concat_283_values1_0, concat_283_values2_0, gather_273_cast_uint16_to_int32))[name = string("concat_283")];
            tensor<int32, [4]> causal_mask_31_begin_0 = const()[name = string("causal_mask_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_31_end_mask_0 = const()[name = string("causal_mask_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_31_cast_fp16 = slice_by_index(begin = causal_mask_31_begin_0, end = concat_283, end_mask = causal_mask_31_end_mask_0, x = causal_mask)[name = string("causal_mask_31_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_57_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_31_cast_fp16, key = key_states_59_cast_fp16, query = query_states_59_cast_fp16, value = value_states_59_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2770_perm_0 = const()[name = string("op_2770_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_284_axis_0 = const()[name = string("concat_284_axis_0"), val = int32(0)];
            bool concat_284_interleave_0 = const()[name = string("concat_284_interleave_0"), val = bool(false)];
            int32 gather_257_cast_uint16_to_int32 = cast(dtype = gather_257_cast_uint16_to_int32_dtype_0, x = gather_257_cast_uint16)[name = string("cast_8")];
            tensor<int32, [3]> concat_284 = concat(axis = concat_284_axis_0, interleave = concat_284_interleave_0, values = (gather_256, gather_257_cast_uint16_to_int32, var_48))[name = string("concat_284")];
            tensor<fp16, [1, ?, 32, 64]> var_2770_cast_fp16 = transpose(perm = var_2770_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, ?, ?]> input_113_cast_fp16 = reshape(shape = concat_284, x = var_2770_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630382528))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632479744))))[name = string("model_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_101_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_439_cast_fp16 = add(x = hidden_states_419_cast_fp16, y = linear_101_cast_fp16)[name = string("hidden_states_439_cast_fp16")];
            fp16 var_54_promoted_29_to_fp16 = const()[name = string("op_54_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2779_cast_fp16 = pow(x = hidden_states_439_cast_fp16, y = var_54_promoted_29_to_fp16)[name = string("op_2779_cast_fp16")];
            tensor<int32, [1]> variance_59_axes_0 = const()[name = string("variance_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_59_keep_dims_0 = const()[name = string("variance_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_59_cast_fp16 = reduce_mean(axes = variance_59_axes_0, keep_dims = variance_59_keep_dims_0, x = var_2779_cast_fp16)[name = string("variance_59_cast_fp16")];
            fp16 var_2782_to_fp16 = const()[name = string("op_2782_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2783_cast_fp16 = add(x = variance_59_cast_fp16, y = var_2782_to_fp16)[name = string("op_2783_cast_fp16")];
            fp32 var_2784_epsilon_0 = const()[name = string("op_2784_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2784_cast_fp16 = rsqrt(epsilon = var_2784_epsilon_0, x = var_2783_cast_fp16)[name = string("op_2784_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_443_cast_fp16 = mul(x = hidden_states_439_cast_fp16, y = var_2784_cast_fp16)[name = string("hidden_states_443_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632741952)))];
            tensor<fp16, [1, ?, 2048]> input_115_cast_fp16 = mul(x = model_model_layers_14_post_attention_layernorm_weight_to_fp16, y = hidden_states_443_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632746112))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(641134784))))[name = string("model_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_2796_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2796_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642183424))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650572096))))[name = string("model_model_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_119_cast_fp16 = mul(x = var_2796_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651620736))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660009408))))[name = string("model_model_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_104_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_449_cast_fp16 = add(x = hidden_states_439_cast_fp16, y = linear_104_cast_fp16)[name = string("hidden_states_449_cast_fp16")];
            fp16 var_54_promoted_30_to_fp16 = const()[name = string("op_54_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2809_cast_fp16 = pow(x = hidden_states_449_cast_fp16, y = var_54_promoted_30_to_fp16)[name = string("op_2809_cast_fp16")];
            tensor<int32, [1]> variance_61_axes_0 = const()[name = string("variance_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_61_keep_dims_0 = const()[name = string("variance_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_61_cast_fp16 = reduce_mean(axes = variance_61_axes_0, keep_dims = variance_61_keep_dims_0, x = var_2809_cast_fp16)[name = string("variance_61_cast_fp16")];
            fp16 var_2812_to_fp16 = const()[name = string("op_2812_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2813_cast_fp16 = add(x = variance_61_cast_fp16, y = var_2812_to_fp16)[name = string("op_2813_cast_fp16")];
            fp32 var_2814_epsilon_0 = const()[name = string("op_2814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2814_cast_fp16 = rsqrt(epsilon = var_2814_epsilon_0, x = var_2813_cast_fp16)[name = string("op_2814_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_453_cast_fp16 = mul(x = hidden_states_449_cast_fp16, y = var_2814_cast_fp16)[name = string("hidden_states_453_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661058048)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_457_cast_fp16 = mul(x = model_model_layers_15_input_layernorm_weight_to_fp16, y = hidden_states_453_cast_fp16)[name = string("hidden_states_457_cast_fp16")];
            tensor<int32, [3]> var_2825_shape_cast_fp16 = shape(x = hidden_states_457_cast_fp16)[name = string("op_2825_shape_cast_fp16")];
            int32 gather_274 = const()[name = string("gather_274"), val = int32(1)];
            int32 gather_275_axis_0 = const()[name = string("gather_275_axis_0"), val = int32(0)];
            int32 gather_275_batch_dims_0 = const()[name = string("gather_275_batch_dims_0"), val = int32(0)];
            bool gather_275_validate_indices_0 = const()[name = string("gather_275_validate_indices_0"), val = bool(false)];
            string var_2825_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2825_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_275_to_uint16 = const()[name = string("select_275_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2825_shape_cast_fp16_to_uint16 = cast(dtype = var_2825_shape_cast_fp16_to_uint16_dtype_0, x = var_2825_shape_cast_fp16)[name = string("cast_7")];
            uint16 gather_275_cast_uint16 = gather(axis = gather_275_axis_0, batch_dims = gather_275_batch_dims_0, indices = select_275_to_uint16, validate_indices = gather_275_validate_indices_0, x = var_2825_shape_cast_fp16_to_uint16)[name = string("gather_275_cast_uint16")];
            string gather_275_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_275_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 2048]> model_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661062208))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663159424))))[name = string("model_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_457_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663421632))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663945984))))[name = string("model_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_457_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [512, 2048]> model_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664011584))), scale = tensor<fp16, [512, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664535936))))[name = string("model_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 512]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_457_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> concat_285x = const()[name = string("concat_285x"), val = tensor<int32, [4]>([1, -1, 32, 64])];
            tensor<fp16, [1, ?, 32, 64]> var_2834_cast_fp16 = reshape(shape = concat_285x, x = linear_105_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_286x = const()[name = string("concat_286x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2837_cast_fp16 = reshape(shape = concat_286x, x = linear_106_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_287x = const()[name = string("concat_287x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
            tensor<fp16, [1, ?, 8, 64]> var_2840_cast_fp16 = reshape(shape = concat_287x, x = linear_107_cast_fp16)[name = string("op_2840_cast_fp16")];
            tensor<int32, [4]> v_state_perm_0 = const()[name = string("v_state_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 32, ?, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_2834_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 32, ?, 64]> var_2844_cast_fp16 = mul(x = q_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2844_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 32, 0, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, ?, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 32, 0, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, ?, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, ?, 32]> var_2855_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            bool var_2857_interleave_0 = const()[name = string("op_2857_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, ?, 64]> var_2857_cast_fp16 = concat(axis = var_48, interleave = var_2857_interleave_0, values = (var_2855_cast_fp16, x1_61_cast_fp16))[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> var_2858_cast_fp16 = mul(x = var_2857_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> query_states_cast_fp16 = add(x = var_2844_cast_fp16, y = var_2858_cast_fp16)[name = string("query_states_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_2837_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, ?, 64]> var_2860_cast_fp16 = mul(x = k_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2860_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 8, 0, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, ?, 32]> var_2871_cast_fp16 = mul(x = x2_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_2871_cast_fp16")];
            bool var_2873_interleave_0 = const()[name = string("op_2873_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 64]> var_2873_cast_fp16 = concat(axis = var_48, interleave = var_2873_interleave_0, values = (var_2871_cast_fp16, x1_cast_fp16))[name = string("op_2873_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2874_cast_fp16 = mul(x = var_2873_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2874_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> k_state_cast_fp16 = add(x = var_2860_cast_fp16, y = var_2874_cast_fp16)[name = string("k_state_cast_fp16")];
            tensor<int32, [1]> expand_dims_180 = const()[name = string("expand_dims_180"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_181 = const()[name = string("expand_dims_181"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_183 = const()[name = string("expand_dims_183"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_290_values0_0 = const()[name = string("concat_290_values0_0"), val = tensor<int32, [1]>([15])];
            int32 concat_290_axis_0 = const()[name = string("concat_290_axis_0"), val = int32(0)];
            bool concat_290_interleave_0 = const()[name = string("concat_290_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_290 = concat(axis = concat_290_axis_0, interleave = concat_290_interleave_0, values = (concat_290_values0_0, expand_dims_180, expand_dims_181, expand_dims_2, expand_dims_183))[name = string("concat_290")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_16_stride_0 = const()[name = string("key_cache_internal_tensor_assign_16_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_16_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_16_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [16, 1, 8, 2048, 64]> key_cache_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_290, begin_mask = key_cache_internal_tensor_assign_16_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_16_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_16_squeeze_mask_0, stride = key_cache_internal_tensor_assign_16_stride_0, update = k_state_cast_fp16, x = coreml_update_state_60)[name = string("key_cache_internal_tensor_assign_16_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_16_cast_fp16, input = key_cache)[name = string("coreml_update_state_62_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_16_stride_0 = const()[name = string("value_cache_internal_tensor_assign_16_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_16_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_16_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, ?, 64]> v_state_cast_fp16 = transpose(perm = v_state_perm_0, x = var_2840_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [16, 1, 8, 2048, 64]> value_cache_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_290, begin_mask = value_cache_internal_tensor_assign_16_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_16_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_16_squeeze_mask_0, stride = value_cache_internal_tensor_assign_16_stride_0, update = v_state_cast_fp16, x = coreml_update_state_61)[name = string("value_cache_internal_tensor_assign_16_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_16_cast_fp16, input = value_cache)[name = string("coreml_update_state_63_write_state")];
            tensor<fp16, [16, 1, 8, 2048, 64]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [5]> var_2897_begin_0 = const()[name = string("op_2897_begin_0"), val = tensor<int32, [5]>([15, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2897_end_0 = const()[name = string("op_2897_end_0"), val = tensor<int32, [5]>([16, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2897_end_mask_0 = const()[name = string("op_2897_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2897_squeeze_mask_0 = const()[name = string("op_2897_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2897_cast_fp16 = slice_by_index(begin = var_2897_begin_0, end = var_2897_end_0, end_mask = var_2897_end_mask_0, squeeze_mask = var_2897_squeeze_mask_0, x = coreml_update_state_62)[name = string("op_2897_cast_fp16")];
            tensor<int32, [4]> var_2900_begin_0 = const()[name = string("op_2900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2900_end_mask_0 = const()[name = string("op_2900_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2900_cast_fp16 = slice_by_index(begin = var_2900_begin_0, end = concat_11, end_mask = var_2900_end_mask_0, x = var_2897_cast_fp16)[name = string("op_2900_cast_fp16")];
            tensor<int32, [5]> var_2902_begin_0 = const()[name = string("op_2902_begin_0"), val = tensor<int32, [5]>([15, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2902_end_0 = const()[name = string("op_2902_end_0"), val = tensor<int32, [5]>([16, 1, 8, 2048, 64])];
            tensor<bool, [5]> var_2902_end_mask_0 = const()[name = string("op_2902_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2902_squeeze_mask_0 = const()[name = string("op_2902_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 8, 2048, 64]> var_2902_cast_fp16 = slice_by_index(begin = var_2902_begin_0, end = var_2902_end_0, end_mask = var_2902_end_mask_0, squeeze_mask = var_2902_squeeze_mask_0, x = coreml_update_state_63)[name = string("op_2902_cast_fp16")];
            tensor<int32, [4]> var_2905_begin_0 = const()[name = string("op_2905_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2905_end_mask_0 = const()[name = string("op_2905_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, ?, 64]> var_2905_cast_fp16 = slice_by_index(begin = var_2905_begin_0, end = concat_11, end_mask = var_2905_end_mask_0, x = var_2902_cast_fp16)[name = string("op_2905_cast_fp16")];
            tensor<int32, [4]> var_2907_shape_cast_fp16 = shape(x = var_2900_cast_fp16)[name = string("op_2907_shape_cast_fp16")];
            int32 gather_283 = const()[name = string("gather_283"), val = int32(1)];
            int32 gather_284 = const()[name = string("gather_284"), val = int32(8)];
            int32 gather_285_axis_0 = const()[name = string("gather_285_axis_0"), val = int32(0)];
            int32 gather_285_batch_dims_0 = const()[name = string("gather_285_batch_dims_0"), val = int32(0)];
            bool gather_285_validate_indices_0 = const()[name = string("gather_285_validate_indices_0"), val = bool(false)];
            string var_2907_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2907_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_285_to_uint16 = const()[name = string("select_285_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2907_shape_cast_fp16_to_uint16 = cast(dtype = var_2907_shape_cast_fp16_to_uint16_dtype_0, x = var_2907_shape_cast_fp16)[name = string("cast_6")];
            uint16 gather_285_cast_uint16 = gather(axis = gather_285_axis_0, batch_dims = gather_285_batch_dims_0, indices = select_285_to_uint16, validate_indices = gather_285_validate_indices_0, x = var_2907_shape_cast_fp16_to_uint16)[name = string("gather_285_cast_uint16")];
            string gather_285_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_285_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_286 = const()[name = string("gather_286"), val = int32(64)];
            tensor<int32, [1]> var_2914_axes_0 = const()[name = string("op_2914_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2914_cast_fp16 = expand_dims(axes = var_2914_axes_0, x = var_2900_cast_fp16)[name = string("op_2914_cast_fp16")];
            tensor<int32, [5]> shape_317_cast_fp16 = shape(x = var_2914_cast_fp16)[name = string("shape_317_cast_fp16")];
            int32 concat_298_axis_0 = const()[name = string("concat_298_axis_0"), val = int32(0)];
            bool concat_298_interleave_0 = const()[name = string("concat_298_interleave_0"), val = bool(false)];
            int32 gather_285_cast_uint16_to_int32 = cast(dtype = gather_285_cast_uint16_to_int32_dtype_0, x = gather_285_cast_uint16)[name = string("cast_5")];
            tensor<int32, [5]> concat_298 = concat(axis = concat_298_axis_0, interleave = concat_298_interleave_0, values = (gather_283, gather_284, var_59, gather_285_cast_uint16_to_int32, gather_286))[name = string("concat_298")];
            tensor<int32, [5]> real_div_30 = real_div(x = concat_298, y = shape_317_cast_fp16)[name = string("real_div_30")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_461_cast_fp16 = tile(reps = real_div_30, x = var_2914_cast_fp16)[name = string("hidden_states_461_cast_fp16")];
            tensor<int32, [4]> concat_299x = const()[name = string("concat_299x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> key_states_cast_fp16 = reshape(shape = concat_299x, x = hidden_states_461_cast_fp16)[name = string("key_states_cast_fp16")];
            tensor<int32, [4]> var_2924_shape_cast_fp16 = shape(x = var_2905_cast_fp16)[name = string("op_2924_shape_cast_fp16")];
            int32 gather_287 = const()[name = string("gather_287"), val = int32(1)];
            int32 gather_288 = const()[name = string("gather_288"), val = int32(8)];
            int32 gather_289_axis_0 = const()[name = string("gather_289_axis_0"), val = int32(0)];
            int32 gather_289_batch_dims_0 = const()[name = string("gather_289_batch_dims_0"), val = int32(0)];
            bool gather_289_validate_indices_0 = const()[name = string("gather_289_validate_indices_0"), val = bool(false)];
            string var_2924_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2924_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_289_to_uint16 = const()[name = string("select_289_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2924_shape_cast_fp16_to_uint16 = cast(dtype = var_2924_shape_cast_fp16_to_uint16_dtype_0, x = var_2924_shape_cast_fp16)[name = string("cast_4")];
            uint16 gather_289_cast_uint16 = gather(axis = gather_289_axis_0, batch_dims = gather_289_batch_dims_0, indices = select_289_to_uint16, validate_indices = gather_289_validate_indices_0, x = var_2924_shape_cast_fp16_to_uint16)[name = string("gather_289_cast_uint16")];
            string gather_289_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_289_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_290 = const()[name = string("gather_290"), val = int32(64)];
            tensor<int32, [1]> var_2931_axes_0 = const()[name = string("op_2931_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, ?, 64]> var_2931_cast_fp16 = expand_dims(axes = var_2931_axes_0, x = var_2905_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<int32, [5]> shape_322_cast_fp16 = shape(x = var_2931_cast_fp16)[name = string("shape_322_cast_fp16")];
            int32 concat_300_axis_0 = const()[name = string("concat_300_axis_0"), val = int32(0)];
            bool concat_300_interleave_0 = const()[name = string("concat_300_interleave_0"), val = bool(false)];
            int32 gather_289_cast_uint16_to_int32 = cast(dtype = gather_289_cast_uint16_to_int32_dtype_0, x = gather_289_cast_uint16)[name = string("cast_3")];
            tensor<int32, [5]> concat_300 = concat(axis = concat_300_axis_0, interleave = concat_300_interleave_0, values = (gather_287, gather_288, var_59, gather_289_cast_uint16_to_int32, gather_290))[name = string("concat_300")];
            tensor<int32, [5]> real_div_31 = real_div(x = concat_300, y = shape_322_cast_fp16)[name = string("real_div_31")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_465_cast_fp16 = tile(reps = real_div_31, x = var_2931_cast_fp16)[name = string("hidden_states_465_cast_fp16")];
            tensor<int32, [4]> concat_301x = const()[name = string("concat_301x"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, ?, 64]> value_states_cast_fp16 = reshape(shape = concat_301x, x = hidden_states_465_cast_fp16)[name = string("value_states_cast_fp16")];
            tensor<int32, [4]> var_2941_shape_cast_fp16 = shape(x = key_states_cast_fp16)[name = string("op_2941_shape_cast_fp16")];
            int32 gather_291_axis_0 = const()[name = string("gather_291_axis_0"), val = int32(0)];
            int32 gather_291_batch_dims_0 = const()[name = string("gather_291_batch_dims_0"), val = int32(0)];
            bool gather_291_validate_indices_0 = const()[name = string("gather_291_validate_indices_0"), val = bool(false)];
            string var_2941_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2941_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_291_to_uint16 = const()[name = string("select_291_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2941_shape_cast_fp16_to_uint16 = cast(dtype = var_2941_shape_cast_fp16_to_uint16_dtype_0, x = var_2941_shape_cast_fp16)[name = string("cast_2")];
            uint16 gather_291_cast_uint16 = gather(axis = gather_291_axis_0, batch_dims = gather_291_batch_dims_0, indices = select_291_to_uint16, validate_indices = gather_291_validate_indices_0, x = var_2941_shape_cast_fp16_to_uint16)[name = string("gather_291_cast_uint16")];
            string gather_291_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_291_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_302_values0_0 = const()[name = string("concat_302_values0_0"), val = int32(1)];
            int32 concat_302_values1_0 = const()[name = string("concat_302_values1_0"), val = int32(1)];
            int32 concat_302_values2_0 = const()[name = string("concat_302_values2_0"), val = int32(0)];
            int32 concat_302_axis_0 = const()[name = string("concat_302_axis_0"), val = int32(0)];
            bool concat_302_interleave_0 = const()[name = string("concat_302_interleave_0"), val = bool(false)];
            int32 gather_291_cast_uint16_to_int32 = cast(dtype = gather_291_cast_uint16_to_int32_dtype_0, x = gather_291_cast_uint16)[name = string("cast_1")];
            tensor<int32, [4]> concat_302 = concat(axis = concat_302_axis_0, interleave = concat_302_interleave_0, values = (concat_302_values0_0, concat_302_values1_0, concat_302_values2_0, gather_291_cast_uint16_to_int32))[name = string("concat_302")];
            tensor<int32, [4]> causal_mask_begin_0 = const()[name = string("causal_mask_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_end_mask_0 = const()[name = string("causal_mask_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_cast_fp16 = slice_by_index(begin = causal_mask_begin_0, end = concat_302, end_mask = causal_mask_end_mask_0, x = causal_mask)[name = string("causal_mask_cast_fp16")];
            tensor<fp16, [1, 32, ?, 64]> attn_output_61_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_cast_fp16, key = key_states_cast_fp16, query = query_states_cast_fp16, value = value_states_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2947_perm_0 = const()[name = string("op_2947_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_303_axis_0 = const()[name = string("concat_303_axis_0"), val = int32(0)];
            bool concat_303_interleave_0 = const()[name = string("concat_303_interleave_0"), val = bool(false)];
            int32 gather_275_cast_uint16_to_int32 = cast(dtype = gather_275_cast_uint16_to_int32_dtype_0, x = gather_275_cast_uint16)[name = string("cast_0")];
            tensor<int32, [3]> concat_303 = concat(axis = concat_303_axis_0, interleave = concat_303_interleave_0, values = (gather_274, gather_275_cast_uint16_to_int32, var_48))[name = string("concat_303")];
            tensor<fp16, [1, ?, 32, 64]> var_2947_cast_fp16 = transpose(perm = var_2947_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, ?, ?]> input_121_cast_fp16 = reshape(shape = concat_303, x = var_2947_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [2048, 2048]> model_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664601536))), scale = tensor<fp16, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666698752))))[name = string("model_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_469_cast_fp16 = add(x = hidden_states_449_cast_fp16, y = linear_108_cast_fp16)[name = string("hidden_states_469_cast_fp16")];
            fp16 var_54_promoted_31_to_fp16 = const()[name = string("op_54_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2956_cast_fp16 = pow(x = hidden_states_469_cast_fp16, y = var_54_promoted_31_to_fp16)[name = string("op_2956_cast_fp16")];
            tensor<int32, [1]> variance_63_axes_0 = const()[name = string("variance_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_63_keep_dims_0 = const()[name = string("variance_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_63_cast_fp16 = reduce_mean(axes = variance_63_axes_0, keep_dims = variance_63_keep_dims_0, x = var_2956_cast_fp16)[name = string("variance_63_cast_fp16")];
            fp16 var_2959_to_fp16 = const()[name = string("op_2959_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2960_cast_fp16 = add(x = variance_63_cast_fp16, y = var_2959_to_fp16)[name = string("op_2960_cast_fp16")];
            fp32 var_2961_epsilon_0 = const()[name = string("op_2961_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2961_cast_fp16 = rsqrt(epsilon = var_2961_epsilon_0, x = var_2960_cast_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_473_cast_fp16 = mul(x = hidden_states_469_cast_fp16, y = var_2961_cast_fp16)[name = string("hidden_states_473_cast_fp16")];
            tensor<fp16, [2048]> model_model_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666960960)))];
            tensor<fp16, [1, ?, 2048]> input_123_cast_fp16 = mul(x = model_model_layers_15_post_attention_layernorm_weight_to_fp16, y = hidden_states_473_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666965120))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675353792))))[name = string("model_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> var_2973_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2973_cast_fp16")];
            tensor<fp16, [8192, 2048]> model_model_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8192, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676402432))), scale = tensor<fp16, [8192, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(684791104))))[name = string("model_model_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8192]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, ?, 8192]> input_127_cast_fp16 = mul(x = var_2973_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [2048, 8192]> model_model_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 8192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685839744))), scale = tensor<fp16, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694228416))))[name = string("model_model_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_111_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_model_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_479_cast_fp16 = add(x = hidden_states_469_cast_fp16, y = linear_111_cast_fp16)[name = string("hidden_states_479_cast_fp16")];
            fp16 var_54_promoted_32_to_fp16 = const()[name = string("op_54_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 2048]> var_2982_cast_fp16 = pow(x = hidden_states_479_cast_fp16, y = var_54_promoted_32_to_fp16)[name = string("op_2982_cast_fp16")];
            tensor<int32, [1]> variance_axes_0 = const()[name = string("variance_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_keep_dims_0 = const()[name = string("variance_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_cast_fp16 = reduce_mean(axes = variance_axes_0, keep_dims = variance_keep_dims_0, x = var_2982_cast_fp16)[name = string("variance_cast_fp16")];
            fp16 var_2985_to_fp16 = const()[name = string("op_2985_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, ?, 1]> var_2986_cast_fp16 = add(x = variance_cast_fp16, y = var_2985_to_fp16)[name = string("op_2986_cast_fp16")];
            fp32 var_2987_epsilon_0 = const()[name = string("op_2987_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2987_cast_fp16 = rsqrt(epsilon = var_2987_epsilon_0, x = var_2986_cast_fp16)[name = string("op_2987_cast_fp16")];
            tensor<fp16, [1, ?, 2048]> hidden_states_483_cast_fp16 = mul(x = hidden_states_479_cast_fp16, y = var_2987_cast_fp16)[name = string("hidden_states_483_cast_fp16")];
            tensor<fp16, [2048]> model_model_norm_weight_to_fp16 = const()[name = string("model_model_norm_weight_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695277056)))];
            tensor<fp16, [1, ?, 2048]> hidden_states_cast_fp16 = mul(x = model_model_norm_weight_to_fp16, y = hidden_states_483_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [128256]> linear_112_bias_0_to_fp16 = const()[name = string("linear_112_bias_0_to_fp16"), val = tensor<fp16, [128256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695281216)))];
            tensor<fp16, [1, ?, 128256]> logits = linear(bias = linear_112_bias_0_to_fp16, weight = model_model_embed_tokens_weight_to_fp16_quantized, x = hidden_states_cast_fp16)[name = string("linear_112_cast_fp16")];
        } -> (logits);
}