Upload handler.py

handler.py

@@ -26,17 +26,17 @@ class EndpointHandler:
         self.pipe.enable_vae_tiling()
         #self.pipe.transformer.fuse_qkv_projections()
         #self.pipe.vae.fuse_qkv_projections()
-        self.pipe.transformer.to(memory_format=torch.channels_last)
-        self.pipe.vae.to(memory_format=torch.channels_last)
+        #self.pipe.transformer.to(memory_format=torch.channels_last)
+        #self.pipe.vae.to(memory_format=torch.channels_last)
         apply_cache_on_pipe(self.pipe, residual_diff_threshold=0.12)
-        self.pipe.transformer = torch.compile(
-            self.pipe.transformer, mode="max-autotune-no-cudagraphs",
-        )
-        self.pipe.vae = torch.compile(
-            self.pipe.vae, mode="max-autotune-no-cudagraphs",
-        )
-        #self.pipe.transformer = autoquant(self.pipe.transformer, error_on_unseen=False)
-        #self.pipe.vae = autoquant(self.pipe.vae, error_on_unseen=False)
+        #self.pipe.transformer = torch.compile(
+        #    self.pipe.transformer, mode="max-autotune-no-cudagraphs",
+        #)
+        #self.pipe.vae = torch.compile(
+        #    self.pipe.vae, mode="max-autotune-no-cudagraphs",
+        #)
+        self.pipe.transformer = autoquant(self.pipe.transformer, error_on_unseen=False)
+        self.pipe.vae = autoquant(self.pipe.vae, error_on_unseen=False)
         
         gc.collect()
         torch.cuda.empty_cache()