解决大EP mtp报错问题

vllm/v1/worker/gpu_model_runner.py

@@ -2005,7 +2005,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         num_reqs = min(num_tokens, max_num_reqs)
         min_tokens_per_req = num_tokens // num_reqs
 
-        if not is_profile and self.speculative_config is not None and self.speculative_config.num_lookahead_slots > 0:
+        if not is_profile and self.speculative_config is not None \
+            and self.speculative_config.num_lookahead_slots > 0 \
+            and num_tokens >= (1 + self.speculative_config.num_lookahead_slots):
             min_tokens_per_req = (1 + self.speculative_config.num_lookahead_slots)
             num_reqs = num_tokens // min_tokens_per_req
         num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs