Fullgraph

src/pipeline.py

@@ -10,8 +10,6 @@ from pipelines.models import TextToImageRequest
 from torch import Generator
 from torchao.quantization import quantize_, int8_weight_only
 from transformers import T5EncoderModel, CLIPTextModel, logging 
-import torch._dynamo
-torch._dynamo.config.suppress_errors = True
 
 Pipeline: TypeAlias = FluxPipeline
 
@@ -48,7 +46,7 @@ def load_pipeline() -> Pipeline:
     ).to("cuda")
 
     pipeline.to(memory_format=torch.channels_last)
-    pipeline.transformer = torch.compile(pipeline.transformer)
+    pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
     quantize_(pipeline.vae, int8_weight_only())
         
     PROMPT = 'semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'