[BugFix][MRV2] Fix cuda event reuse race (#39115)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-04-07 00:21:09 +00:00
committed by GitHub
parent b2b2c5239e
commit 62095e82c1
2 changed files with 2 additions and 7 deletions
-3
View File
@@ -130,7 +130,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.use_async_scheduling = self.scheduler_config.async_scheduling
self.output_copy_stream = torch.cuda.Stream(self.device)
self.output_copy_event = torch.cuda.Event()
# Pipeline parallelism.
self.use_pp = self.parallel_config.pipeline_parallel_size > 1
@@ -1180,7 +1179,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
num_sampled_tokens=num_sampled,
main_stream=self.main_stream,
copy_stream=self.output_copy_stream,
copy_event=self.output_copy_event,
)
mm_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None
@@ -1270,7 +1268,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
is_valid=is_valid,
main_stream=self.main_stream,
copy_stream=self.output_copy_stream,
copy_event=self.output_copy_event,
)
self.postprocess_pool(input_batch)