Got it

src/pipeline.py

@@ -14,6 +14,11 @@ from transformers import T5EncoderModel, CLIPTextModel
 
 Pipeline: TypeAlias = FluxPipeline
 torch.backends.cudnn.benchmark = True
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
 REVISION = "5ef0012f11a863e5111ec56540302a023bc8587b"
@@ -28,7 +33,7 @@ def load_pipeline() -> Pipeline:
             path, 
             use_safetensors=False,
             local_files_only=True,
-            torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+            torch_dtype=torch.bfloat16)
     vae = AutoencoderTiny.from_pretrained(
             TinyVAE,
             revision=TinyVAE_REV,
@@ -44,10 +49,8 @@ def load_pipeline() -> Pipeline:
         torch_dtype=torch.bfloat16,
     )
 
-    pipeline.to(memory_format=torch.channels_last)
-    pipeline.enable_vae_slicing()
+    pipeline.transformer.to(memory_format=torch.channels_last)
     pipeline.to("cuda")
-    # quantize_(pipeline.vae, int8_weight_only())
     for _ in range(4):
         pipeline("cat", num_inference_steps=4)