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 14, 05:30 UTC
[Core] Change `execute_model_with_error_logging()` to be a ctx manager (#27060)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Oct 16, 2025 at 20:45 UTC
fe3b9372adfe968360c6754c15058380ab5e87f1
1 file changed
+11
-15
vllm/v1/engine/core.py
+11
-15
@@ -290,14 +290,11 @@ class EngineCore:
290
# (i.e. client-aborted vs stop criteria met).
291
self.scheduler.finish_requests(request_ids, RequestStatus.FINISHED_ABORTED)
292
293
- def execute_model_with_error_logging(
294
- self,
295
- model_fn: Callable[[SchedulerOutput], ModelRunnerOutput],
296
- scheduler_output: SchedulerOutput,
297
- ) -> ModelRunnerOutput:
293
+ @contextmanager
294
+ def log_error_detail(self, scheduler_output: SchedulerOutput):
295
"""Execute the model and log detailed info on failure."""
296
try:
300
- return model_fn(scheduler_output)
297
+ yield
298
except Exception as err:
299
# We do not want to catch BaseException here since we're only
300
# interested in dumping info when the exception is due to an
@@ -321,15 +318,16 @@ class EngineCore:
318
if not self.scheduler.has_requests():
319
return {}, False
320
scheduler_output = self.scheduler.schedule()
324
- model_output = self.execute_model_with_error_logging(
325
- self.model_executor.execute_model, # type: ignore
326
- scheduler_output,
327
- )
321
+
322
+ with self.log_error_detail(scheduler_output):
323
+ model_output = self.model_executor.execute_model(scheduler_output)
324
+
325
+ assert isinstance(model_output, ModelRunnerOutput)
326
engine_core_outputs = self.scheduler.update_from_output(
327
scheduler_output, model_output
328
)
329
332
- return (engine_core_outputs, scheduler_output.total_num_scheduled_tokens > 0)
330
+ return engine_core_outputs, scheduler_output.total_num_scheduled_tokens > 0
331
332
def post_step(self, model_executed: bool) -> None:
333
if self.use_spec_decode and model_executed:
@@ -386,14 +384,12 @@ class EngineCore:
384
385
# Block until the next result is available.
386
future, scheduler_output = batch_queue.pop()
389
- model_output = self.execute_model_with_error_logging(
390
- lambda _: future.result(), scheduler_output
391
- )
387
+ with self.log_error_detail(scheduler_output):
388
+ model_output = future.result()
389
390
engine_core_outputs = self.scheduler.update_from_output(
391
scheduler_output, model_output
392
)
396
-
393
return engine_core_outputs, model_executed
394
395
def shutdown(self):