修复pd分离开cp引起的校检问题

update VLLM_USE_PD_SPLIT=0 (for dspk)and  VLLM_USE_PD_SPLIT=1 (for others)

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -481,7 +481,9 @@ class P2pNcclConnector(KVConnectorBase_V1):
                 num_scheduled_tokens = (
                     scheduler_output.num_scheduled_tokens)[req_id]
                 num_tokens = (num_scheduled_tokens + num_computed_tokens)
-                assert req_id in self.chunked_prefill
+                # assert req_id in self.chunked_prefill
+                if  req_id not in self.chunked_prefill:
+                    continue
                 block_ids = new_block_ids[0]
                 if not resumed_from_preemption:
                     block_ids = (self.chunked_prefill[req_id][0] + block_ids)

vllm/envs.py

@@ -1642,7 +1642,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
              ("true", "1")),
     # vLLM will split prefill and decode, not mix up
     "VLLM_USE_PD_SPLIT":
-        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in
+        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "False").lower() in
                  ("true", "1")), 
     # vLLM will sync to avoid pp vmfault
     "VLLM_USE_PP_SYNC":

vllm/model_executor/model_loader/utils.py

@@ -198,6 +198,9 @@ def _get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
+            else:
+                if not envs.is_set("VLLM_USE_PD_SPLIT"):
+                    os.environ['VLLM_USE_PD_SPLIT'] = '1'
 
             if os.getenv('GEMM_PAD') != '1': 
                 os.environ['GEMM_PAD'] = '0'
@@ -209,6 +212,9 @@ def _get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
+            else:
+                if not envs.is_set("VLLM_USE_PD_SPLIT"):
+                    os.environ['VLLM_USE_PD_SPLIT'] = '1'
                     
         # awq相关配置
         try: