Merge branch 'v0.9.2-dev-ds-wm-1225' into 'v0.9.2-dev-ds'

[fix]解决EP不启用mtp时某些size报错

See merge request dcutoolkit/deeplearing/vllm!322

vllm/v1/worker/gpu_model_runner.py

@@ -2076,10 +2076,14 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
                 num_actual_tokens = round_down(num_tokens, 1 + self.speculative_config.num_lookahead_slots)
                 num_reqs = num_actual_tokens // min_tokens_per_req
         num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
+        
         if not self.ep_sp:
             num_scheduled_tokens_list[-1] += num_tokens % num_reqs
         else:
-            num_scheduled_tokens_list[-1] += num_tokens % min_tokens_per_req
+            if self.speculative_config is not None:
+                num_scheduled_tokens_list[-1] += num_tokens % min_tokens_per_req
+            else:
+                num_scheduled_tokens_list[-1] += num_tokens % num_reqs
 
         assert sum(num_scheduled_tokens_list) == num_tokens
         assert len(num_scheduled_tokens_list) == num_reqs