[core] Turn off GPU communication overlap for Ray executor (vllm-project#12051)

Signed-off-by: Rui Qiao <[email protected]>
Signed-off-by: Isotr0py <[email protected]>

Author: ruisearch42

vllm/envs.py

@@ -45,7 +45,7 @@
     VLLM_USE_RAY_SPMD_WORKER: bool = False
     VLLM_USE_RAY_COMPILED_DAG: bool = False
     VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL: bool = True
-    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = True
+    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
     VLLM_WORKER_MULTIPROC_METHOD: str = "fork"
     VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
     VLLM_IMAGE_FETCH_TIMEOUT: int = 5
@@ -340,11 +340,11 @@ def get_default_config_root():
     lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL", "1"))
                  ),
 
-    # If the env var is set, it enables GPU communication overlap in
-    # Ray's compiled DAG. This flag is ignored if
+    # If the env var is set, it enables GPU communication overlap
+    # (experimental feature) in Ray's compiled DAG. This flag is ignored if
     # VLLM_USE_RAY_COMPILED_DAG is not set.
     "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
-    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "1"))
+    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
                  ),
 
     # Use dedicated multiprocess context for workers.