feat: pp mtp加入零消耗调度，加入环境变量VLLM_USE_ZERO_MTP，默认打开

vllm/envs.py

@@ -182,6 +182,7 @@ if TYPE_CHECKING:
     VLLM_USE_LIGHTOP_FILL_MOE_ALIGN: bool = False
     USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT: bool = False
     VLLM_USE_PP_BALANCE: bool = False
+    VLLM_USE_ZERO_MTP: bool = False
     VLLM_USE_CUDA_GRAPH_SIZES: bool = False
 
 def get_default_cache_root():
@@ -1152,40 +1153,44 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # vllm pd separation will be used async
     "VLLM_P2P_ASYNC":
     lambda: bool(int(os.getenv("VLLM_P2P_ASYNC", "0"))),
-    
+
     # pd separation p2p async buf tokens
     "VLLM_P2P_BUF_TOKENS":
     lambda: int(os.getenv("VLLM_P2P_BUF_TOKENS", "30000")),
-    
+
     # vllm will enable minimal injection for pipeline parallel scheduling
     "VLLM_SCHED_ENABLE_MINIMAL_INJECTION":
         lambda: (os.getenv("VLLM_SCHED_ENABLE_MINIMAL_INJECTION", "0").lower() in
                  ("true", "1")),
-        
+
     # vLLM will split prefill and decode, not mix up
     "VLLM_USE_PD_SPLIT":
         lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "False").lower() in
                  ("true", "1")), 
-        
+
     # vLLM will sync to avoid pp vmfault
     "VLLM_USE_PP_SYNC":
         lambda: (os.environ.get("VLLM_USE_PP_SYNC", "False").lower() in
                  ("true", "1")),
-        
+
     # vLLM will use lightop to fuse fill and moe align
     "VLLM_USE_LIGHTOP_FILL_MOE_ALIGN":
         lambda: (os.environ.get("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN", "False").lower() in
                  ("true", "1")), 
-        
+
     # vllm will use custom-allreduce rmsquant fused op
     "USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT": 
     lambda: (os.getenv('USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT', '0').lower() in
              ("true", "1")),
-    
+
     "VLLM_USE_PP_BALANCE":
         lambda: (os.getenv('VLLM_USE_PP_BALANCE', '1').lower() in
                  ("true", "1")),
-    
+
+    "VLLM_USE_ZERO_MTP":
+        lambda: (os.getenv('VLLM_USE_ZERO_MTP', '1').lower() in
+                 ("true", "1")),
+
     # vllm will use 1-18... (not only 1 2 4 8 16)
     "VLLM_USE_CUDA_GRAPH_SIZES":
         lambda: (os.getenv('VLLM_USE_CUDA_GRAPH_SIZES', 'False').lower() in