transformer only

src/pipeline.py

@@ -13,6 +13,7 @@ from transformers import T5EncoderModel, CLIPTextModel
 
 
 Pipeline: TypeAlias = FluxPipeline
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 torch.backends.cudnn.benchmark = True
 
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
@@ -38,11 +39,11 @@ def load_pipeline() -> Pipeline:
         torch_dtype=torch.bfloat16,
     )
 
-    pipeline.to(memory_format=torch.channels_last)
+    pipeline.transformer.to(memory_format=torch.channels_last)
     pipeline.to("cuda")
     for _ in range(4):
         pipeline("cat", num_inference_steps=4)
-
+    torch.cuda.empty_cache()
     return pipeline
 
 @torch.inference_mode()