Merge branch 'v0.9.2-dev-ds-new' into 'v0.9.2-dev-ds'

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

See merge request dcutoolkit/deeplearing/vllm!277

vllm/envs.py

@@ -178,6 +178,7 @@ if TYPE_CHECKING:
     VLLM_SCHED_ENABLE_MINIMAL_INJECTION: bool = False
     VLLM_USE_PD_SPLIT: bool = False
     VLLM_USE_PP_BALANCE: bool = False
+    VLLM_USE_ZERO_MTP: bool = False
 
 def get_default_cache_root():
     return os.getenv(
@@ -1171,6 +1172,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "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")),
 }
 
 # --8<-- [end:env-vars-definition]