暂时去掉profilling标志位，避免影响其他模型

vllm/v1/worker/gpu_model_runner.py

@@ -2051,8 +2051,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 input_ids = None
                 inputs_embeds = self.inputs_embeds[:num_tokens]
             else:
-                #self.input_ids[:num_tokens] = torch.randint(0, 120000, (num_tokens,), dtype=torch.int32)
-                #self.input_ids[:num_tokens] = torch.arange(num_tokens, dtype=torch.int32, device=self.input_ids.device)
                 input_ids = self.input_ids[:num_tokens]
                 inputs_embeds = None
             if self.uses_mrope:
@@ -2226,8 +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.
@@ -2312,7 +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: