xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 04:15:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Minor] Group async_scheduling related fields in model runner init (#26736)

Browse Source 
Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill 2025-10-14 14:46:37 -07:00 committed by GitHub
parent 9d6964926e
commit ff4810ba73
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
vllm/v1/worker/gpu_model_runner.py
View File

@ -375,9 +375,15 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
) )
self.use_async_scheduling = self.scheduler_config.async_scheduling self.use_async_scheduling = self.scheduler_config.async_scheduling
self.async_output_copy_stream = ( # Separate cuda stream for overlapping transfer of sampled token ids from
torch.cuda.Stream() if self.use_async_scheduling else None # GPU to CPU when async scheduling is enabled.
) self.async_output_copy_stream: torch.cuda.Stream | None = None
# cuda event to synchronize use of reused CPU tensors between steps
# when async scheduling is enabled.
self.prepare_inputs_event: torch.cuda.Event | None = None
if self.use_async_scheduling:
self.async_output_copy_stream = torch.cuda.Stream()
self.prepare_inputs_event = torch.cuda.Event()
# TODO(woosuk): Provide an option to tune the max cudagraph batch size. # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
# The convention is different. # The convention is different.
@ -444,14 +450,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
(3, self.max_num_tokens + 1), dtype=torch.int64 (3, self.max_num_tokens + 1), dtype=torch.int64
) )
# CUDA event to synchronize use of reused CPU tensors between steps
# when async scheduling is enabled.
self.prepare_inputs_event: torch.cuda.Event | None = None
if self.use_async_scheduling:
self.prepare_inputs_event = torch.cuda.Event()
# Start in a completed state.
self.prepare_inputs_event.record(torch.cuda.default_stream())
# None in the first PP rank. The rest are set after load_model. # None in the first PP rank. The rest are set after load_model.
self.intermediate_tensors: IntermediateTensors | None = None self.intermediate_tensors: IntermediateTensors | None = None