[Misc] Configurable timeout for execute_model RPC calls via env var (#19544)

Signed-off-by: jinqinn <goodqinjin@163.com>

vllm/envs.py

@@ -130,6 +130,7 @@ if TYPE_CHECKING:
     VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
     VLLM_SLEEP_WHEN_IDLE: bool = False
     VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
     VLLM_KV_CACHE_LAYOUT: Optional[str] = None
     VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
 
@@ -897,6 +898,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_MQ_MAX_CHUNK_BYTES_MB":
     lambda: int(os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")),
 
+    # Timeout in seconds for execute_model RPC calls in multiprocessing
+    # executor (only applies when TP > 1).
+    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS":
+    lambda: int(os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")),
+
     # KV Cache layout used throughout vllm.
     # Some common values are:
     # - NHD

vllm/v1/executor/multiproc_executor.py

@@ -37,11 +37,6 @@ from vllm.worker.worker_base import WorkerWrapperBase
 
 logger = init_logger(__name__)
 
-POLLING_TIMEOUT_MS = 5000
-POLLING_TIMEOUT_S = POLLING_TIMEOUT_MS // 1000
-
-EXECUTE_MODEL_TIMEOUT_S = 300
-
 
 class MultiprocExecutor(Executor):
 
@@ -160,12 +155,12 @@ class MultiprocExecutor(Executor):
         self,
         scheduler_output,
     ) -> Union[ModelRunnerOutput, Future[ModelRunnerOutput]]:
-        (output, ) = self.collective_rpc("execute_model",
-                                         args=(scheduler_output, ),
-                                         unique_reply_rank=self.output_rank,
-                                         non_block=self.max_concurrent_batches
-                                         > 1,
-                                         timeout=EXECUTE_MODEL_TIMEOUT_S)
+        (output, ) = self.collective_rpc(
+            "execute_model",
+            args=(scheduler_output, ),
+            unique_reply_rank=self.output_rank,
+            non_block=self.max_concurrent_batches > 1,
+            timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS)
         return output
 
     def collective_rpc(self,