Upload handler.py

handler.py

@@ -216,7 +216,6 @@ def load_pipeline_fast(repo_id: str, dtype: torch.dtype) -> Any:
     pipe.vae.fuse_qkv_projections()
     pipe.transformer.to(memory_format=torch.channels_last)
     pipe.vae.to(memory_format=torch.channels_last)
-    apply_cache_on_pipe(pipe, residual_diff_threshold=0.12)
     if IS_QUANT and not IS_AUTOQ:
         quantize_(pipe.text_encoder, int8_dynamic_activation_int8_weight())
         quantize_(pipe.text_encoder_2, int8_dynamic_activation_int8_weight())
@@ -240,12 +239,11 @@ class EndpointHandler:
         elif IS_COMPILE: self.pipeline = load_pipeline_fast(repo_id, dtype)
         elif IS_LVRAM and IS_CC89: self.pipeline = load_pipeline_lowvram(repo_id, dtype)
         else: self.pipeline = load_pipeline_stable(repo_id, dtype)
-        if IS_PARA: apply_cache_on_pipe(self.pipeline, residual_diff_threshold=0.12)
+        self.pipeline.enable_vae_slicing()
+        self.pipeline.enable_vae_tiling()
         self.pipeline.to("cuda")
-        if not IS_COMPILE:
-            self.pipeline.enable_vae_slicing()
-            self.pipeline.enable_vae_tiling()
-        else:
+        if IS_PARA: apply_cache_on_pipe(self.pipeline, residual_diff_threshold=0.12)
+        if IS_COMPILE:
             print("Compiling pipeline...")
             self.pipeline.transformer = torch.compile(self.pipeline.transformer, mode="max-autotune-no-cudagraphs")
             self.pipeline.vae = torch.compile(self.pipeline.vae, mode="max-autotune-no-cudagraphs")