more compile

src/pipeline.py

@@ -52,9 +52,9 @@ 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)
-    
-    for _ in range(2):
-        pipeline("cat", num_inference_steps=4)
+    with torch.inference_mode():
+        for _ in range(2):
+            pipeline("cat", num_inference_steps=4)
 
     return pipeline