暂时去掉profilling标志位，避免影响其他模型

vllm/v1/worker/gpu_model_runner.py

@@ -2611,8 +2611,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
 
     def profile_run(self) -> None:
         # set profiling flag to avoid torch compile
-        set_profilling(True)
-        self._sync_device()
+        # set_profilling(True)
+        # self._sync_device()
 
         # Profile with multimodal encoder & encoder cache.
         if self.supports_mm_inputs:
@@ -2681,7 +2681,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         del hidden_states, output
         self.encoder_cache.clear()
         gc.collect()
-        set_profilling(False)
+        # set_profilling(False)
 
     def capture_model(self) -> None:
         if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE: