A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 23:00 UTC
[Bugfix] Surface exceptions from non-blocking execute_model in UniProcExecutor to avoid DP deadlocks (#35194)
Signed-off-by: fangyuchu <fangyuchu@qq.com>
fangyuchu committed
Mar 11, 2026 at 11:22 UTC
fa0d353acfa3de7610fdcdb6d23b3b34109c5749
2 files changed
+10
-4
vllm/v1/engine/core.py
+4
-3
@@ -443,9 +443,10 @@ class EngineCore:
443
deferred_scheduler_output = None
444
if self.scheduler.has_requests():
445
scheduler_output = self.scheduler.schedule()
446
- exec_future = self.model_executor.execute_model(
447
- scheduler_output, non_block=True
448
- )
446
+ with self.log_error_detail(scheduler_output):
447
+ exec_future = self.model_executor.execute_model(
448
+ scheduler_output, non_block=True
449
+ )
450
if self.is_ec_consumer:
451
model_executed = scheduler_output.total_num_scheduled_tokens > 0
452
vllm/v1/executor/uniproc_executor.py
+6
-1
@@ -100,12 +100,17 @@ class UniProcExecutor(Executor):
100
def execute_model( # type: ignore[override]
101
self, scheduler_output: SchedulerOutput, non_block: bool = False
102
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
103
- return self.collective_rpc(
103
+ output = self.collective_rpc(
104
"execute_model",
105
args=(scheduler_output,),
106
non_block=non_block,
107
single_value=True,
108
)
109
+ # In non-blocking mode, surface any exception as early as possible.
110
+ if non_block and output.done():
111
+ # Raise the exception in-line if the task failed.
112
+ output.result()
113
+ return output
114
115
def sample_tokens( # type: ignore[override]
116
self, grammar_output: GrammarOutput | None, non_block: bool = False