biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-04-06 17:21:09 -07:00
committed by GitHub
parent b2b2c5239e
commit 62095e82c1
2 changed files with 2 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/v1/worker/gpu/async_utils.py
View File

@@ -17,7 +17,6 @@ class AsyncOutput(AsyncModelRunnerOutput):
num_sampled_tokens: torch.Tensor,
main_stream: torch.cuda.Stream,
copy_stream: torch.cuda.Stream,
copy_event: torch.cuda.Event,
):
# NOTE(woosuk): We must retain references to the GPU tensors,
# as the copy operations are performed on a different CUDA stream than
@@ -25,7 +24,7 @@ class AsyncOutput(AsyncModelRunnerOutput):
self.model_runner_output = model_runner_output
self.sampler_output = sampler_output
self.num_sampled_tokens = num_sampled_tokens
self.copy_event = copy_event
self.copy_event = torch.cuda.Event()
with stream(copy_stream, main_stream):
copy_stream.wait_stream(main_stream)
@@ -78,12 +77,11 @@ class AsyncPoolingOutput(AsyncModelRunnerOutput):
is_valid: torch.Tensor | None,
main_stream: torch.cuda.Stream,
copy_stream: torch.cuda.Stream,
copy_event: torch.cuda.Event,
):
self.model_runner_output = model_runner_output
self.pooler_output = pooler_output
self.is_valid = is_valid
self.copy_event = copy_event
self.copy_event = torch.cuda.Event()
with stream(copy_stream, main_stream):
copy_stream.wait_stream(main_stream)

3
vllm/v1/worker/gpu/model_runner.py
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)