Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -111,7 +111,8 @@ def load_pipeline() -> FluxPipeline:
                                         vae=vae,
                                         text_encoder=text_encoder,
                                         text_encoder_2=text_encoder_2,
-                                        torch_dtype=torch.bfloat16).to("cuda")
+                                        torch_dtype=torch.bfloat16)
+    pipe.enable_model_cpu_offload()
     empty_cache()
     pipe("cat", guidance_scale=0., max_sequence_length=256, num_inference_steps=4)
     return pipe