dummy_run时使用随机输入

vllm/v1/worker/gpu_model_runner.py

@@ -2054,6 +2054,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 input_ids = None
                 inputs_embeds = self.inputs_embeds[:num_tokens]
             else:
+                self.input_ids[:num_tokens] = torch.randint(0, self.model_config.get_vocab_size(), (num_tokens,),
+                                                            dtype=torch.int32)
                 input_ids = self.input_ids[:num_tokens]
                 inputs_embeds = None
             if self.uses_mrope: