feat: 兼容MTP零消耗和主模型+MTP零消耗(VLLM_ZERO_OVERHEAD_ENHANCE=1)开启

vllm/zero_overhead/v1/gpu_model_runner.py

@@ -703,7 +703,8 @@ class V1ZeroModelRunner(GPUModelRunner):
         is_output_valid = False
         # Get the valid generated tokens.
         sampled_token_ids = sampler_output.sampled_token_ids
-        over_head_enhance = envs.VLLM_ZERO_OVERHEAD_ENHANCE and self.speculative_config
+        over_head_enhance = (envs.VLLM_ZERO_OVERHEAD_ENHANCE and
+                             self.speculative_config is not None)
         if over_head_enhance:
             # if not self.speculative_config:
             #     self.fix_req_ids = self.last_sampled_req_ids