[V1] Improve TP>1 Error Handling + Stack Trace

vllm/v1/engine/async_llm.py

@@ -1,6 +1,5 @@
 import asyncio
 import os
-import signal
 from typing import AsyncGenerator, Dict, List, Mapping, Optional, Type, Union
 
 from vllm.config import ModelConfig, VllmConfig
@@ -42,21 +41,6 @@ def __init__(
         start_engine_loop: bool = True,
     ) -> None:
 
-        # The child processes will send SIGQUIT when unrecoverable
-        # errors happen. We kill the process tree here so that the
-        # stack trace is very evident.
-        # TODO: rather than killing the main process, we should
-        # figure out how to raise an AsyncEngineDeadError and
-        # handle at the API server level so we can return a better
-        # error code to the clients calling VLLM.
-        def sigquit_handler(signum, frame):
-            logger.fatal(
-                "AsyncLLM got SIGQUIT from worker processes, shutting "
-                "down. See stack trace above for root cause issue.")
-            kill_process_tree(os.getpid())
-
-        signal.signal(signal.SIGQUIT, sigquit_handler)
-
         assert start_engine_loop
 
         self.log_requests = log_requests

vllm/v1/engine/core.py

@@ -198,7 +198,7 @@ def signal_handler(signum, frame):
         except Exception:
             traceback = get_exception_traceback()
             logger.error("EngineCore hit an exception: %s", traceback)
-            parent_process.send_signal(signal.SIGQUIT)
+            parent_process.send_signal(signal.SIGUSR1)
 
         finally:
             if engine_core is not None:

vllm/v1/engine/core_client.py

@@ -1,3 +1,5 @@
+import os
+import signal
 import weakref
 from abc import ABC, abstractmethod
 from typing import List, Type
@@ -8,7 +10,8 @@
 
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
-from vllm.utils import get_open_zmq_ipc_path, make_zmq_socket
+from vllm.utils import (get_open_zmq_ipc_path, make_zmq_socket,
+                        kill_process_tree)
 from vllm.v1.engine import (EngineCoreOutput, EngineCoreOutputs,
                             EngineCoreProfile, EngineCoreRequest,
                             EngineCoreRequestType, EngineCoreRequestUnion)
@@ -134,6 +137,20 @@ def __init__(
         executor_class: Type[Executor],
         log_stats: bool = False,
     ):
+        # The child processes will send SIGUSR1 when unrecoverable
+        # errors happen. We kill the process tree here so that the
+        # stack trace is very evident.
+        # TODO(rob): rather than killing the main process, we should
+        # figure out how to raise an AsyncEngineDeadError and
+        # handle at the API server level so we can return a better
+        # error code to the clients calling VLLM.
+        def sigusr1_handler(signum, frame):
+            logger.fatal(
+                "Got SIGUSR1 from worker processes, shutting "
+                "down. See stack trace above for root cause issue.")
+            kill_process_tree(os.getpid())
+        signal.signal(signal.SIGUSR1, sigusr1_handler)
+
         # Serialization setup.
         self.encoder = PickleEncoder()
         self.decoder = msgspec.msgpack.Decoder(EngineCoreOutputs)

vllm/v1/executor/multiproc_executor.py

@@ -9,6 +9,7 @@
 from multiprocessing.process import BaseProcess
 from typing import Any, Dict, List, Optional, Tuple
 
+import psutil
 import zmq
 
 from vllm.config import VllmConfig
@@ -38,6 +39,19 @@ def __init__(self, vllm_config: VllmConfig) -> None:
         # and ensure workers will be terminated.
         self._finalizer = weakref.finalize(self, self.shutdown)
 
+        # The child processes will send SIGUSR1 when unrecoverable
+        # errors happen.
+        def sigusr1_handler(signum, frame):
+            logger.fatal(
+                "MulitprocExecutor got SIGUSR1 from worker processes, shutting "
+                "down. See stack trace above for root cause issue.")
+            # Propagate error up to parent process.
+            parent_process = psutil.Process().parent()
+            parent_process.send_signal(signal.SIGUSR1)
+            self.shutdown()
+
+        signal.signal(signal.SIGUSR1, sigusr1_handler)
+
         self.vllm_config = vllm_config
         self.parallel_config = vllm_config.parallel_config
 
@@ -335,8 +349,11 @@ def signal_handler(signum, frame):
         except SystemExit:
             logger.debug("Worker interrupted.")
 
-        except BaseException as e:
-            logger.exception(e)
+        except Exception:
+            # worker_busy_loop sends exceptions exceptons to Executor
+            # for shutdown, but if there is an error in startup or an
+            # error with IPC itself, we need to alert the parent.
+            psutil.Process().parent().send_signal(signal.SIGUSR1)
             raise
 
         finally:
@@ -377,9 +394,10 @@ def worker_busy_loop(self):
 
             try:
                 output = getattr(self.worker, method)(*args, **kwargs)
-            except BaseException as e:
+            except Exception as e:
                 self.worker_response_mq.enqueue(
                     (WorkerProc.ResponseStatus.FAILURE, e))
+                logger.exception("WorkerProc hit an exception: %s", exc_info=e)
                 continue
 
             self.worker_response_mq.enqueue(

vllm/v1/worker/gpu_worker.py

@@ -35,6 +35,8 @@ def __init__(
         distributed_init_method: str,
     ):
 
+        self.i = 0
+
         # TODO: use WorkerBase.__init__(self, vllm_config=vllm_config)
         self.vllm_config = vllm_config
         self.model_config = vllm_config.model_config
@@ -201,6 +203,10 @@ def execute_model(
         self,
         scheduler_output: "SchedulerOutput",
     ) -> ModelRunnerOutput:
+        if self.rank == 0 and self.i == 10:
+            raise ValueError("ERROR FROM HERE :)")
+        self.i += 1
+
         output = self.model_runner.execute_model(scheduler_output)
         return output if self.rank == 0 else None