Update src/pipeline.py

src/pipeline.py

@@ -1,4 +1,4 @@
-import gc
+# partial 
 import os
 from typing import TypeAlias
 
@@ -10,7 +10,9 @@ from pipelines.models import TextToImageRequest
 from torch import Generator
 from torchao.quantization import quantize_, int8_weight_only
 from transformers import T5EncoderModel, CLIPTextModel, logging 
+from functools import partial 
 
+my_partial_compile = partial(torch.compile, mode="max-autotune")
 
 Pipeline: TypeAlias = FluxPipeline
 torch.backends.cudnn.benchmark = True
@@ -51,7 +53,7 @@ def load_pipeline() -> Pipeline:
 
     pipeline.to(memory_format=torch.channels_last)
     quantize_(pipeline.vae, int8_weight_only()) 
-    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune", fullgraph=True)
+    pipeline.vae = my_partial_compile(pipeline.vae)
     with torch.inference_mode():
         for _ in range(2):
             pipeline("cat", num_inference_steps=4)