xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-13 13:35:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model Runner V2][BugFix] Keep reference to GPU tensors in AsyncOutput (#29623)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon 2025-11-27 12:38:53 -08:00 committed by GitHub
parent a5345bf49d
commit ae0ce1be27
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/worker/gpu/async_utils.py
View File

@ -21,6 +21,9 @@ class AsyncOutput(AsyncModelRunnerOutput):
copy_stream: torch.cuda.Stream, copy_stream: torch.cuda.Stream,
copy_event: torch.cuda.Event, 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
# the one where the tensors were created.
self.model_runner_output = model_runner_output self.model_runner_output = model_runner_output
self.sampler_output = sampler_output self.sampler_output = sampler_output
self.num_sampled_tokens = num_sampled_tokens self.num_sampled_tokens = num_sampled_tokens
@ -51,7 +54,9 @@ class AsyncOutput(AsyncModelRunnerOutput):
) )
else: else:
self.logprobs_tensors = None self.logprobs_tensors = None
self.num_sampled_tokens = num_sampled_tokens.to("cpu", non_blocking=True) self.num_sampled_tokens_cpu = num_sampled_tokens.to(
"cpu", non_blocking=True
)
self.prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {} self.prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
if self.model_runner_output.prompt_logprobs_dict: if self.model_runner_output.prompt_logprobs_dict:
for k, v in self.model_runner_output.prompt_logprobs_dict.items(): for k, v in self.model_runner_output.prompt_logprobs_dict.items():
@ -63,7 +68,7 @@ class AsyncOutput(AsyncModelRunnerOutput):
def get_output(self) -> ModelRunnerOutput: def get_output(self) -> ModelRunnerOutput:
self.copy_event.synchronize() self.copy_event.synchronize()
num_sampled_tokens_np = self.num_sampled_tokens.numpy() num_sampled_tokens_np = self.num_sampled_tokens_cpu.numpy()
# NOTE(woosuk): The following code is to ensure compatibility with # NOTE(woosuk): The following code is to ensure compatibility with
# the existing model runner. # the existing model runner.