Merge branch 'v0.9.2-dev-wm-1' into 'v0.9.2-dev'

暂时去掉profilling标志位，避免影响其他模型

See merge request dcutoolkit/deeplearing/vllm!196

vllm/v1/worker/gpu_model_runner.py

@@ -2224,8 +2224,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
 
     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.
         # TODO: handle encoder-decoder models once we support them.
@@ -2310,7 +2310,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         del hidden_states, output
         self.encoder_cache.clear()
         gc.collect()
-        set_profilling(False)
+        #set_profilling(False)
 
     def capture_model(self) -> None:
         if not self.use_cuda_graph: