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

[BugFix] Async scheduling and PP compatibility with DP (#23770)

Browse Source 
Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill
2025-08-29 08:17:27 -07:00
committed by GitHub
parent 0a2f4c0793
commit d90d8eb674
7 changed files with 105 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
vllm/v1/worker/gpu_worker.py
View File

@@ -354,36 +354,37 @@ class Worker(WorkerBase):
scheduler_output: "SchedulerOutput",
) -> Optional[ModelRunnerOutput]:
intermediate_tensors = None
if not get_pp_group().is_first_rank:
forward_pass = scheduler_output.total_num_scheduled_tokens > 0
if forward_pass and not get_pp_group().is_first_rank:
intermediate_tensors = IntermediateTensors(
get_pp_group().recv_tensor_dict(
all_gather_group=get_tp_group()))
output = self.model_runner.execute_model(scheduler_output,
intermediate_tensors)
parallel_config = self.vllm_config.parallel_config
if parallel_config.distributed_executor_backend != "external_launcher" \
and not get_pp_group().is_last_rank:
assert isinstance(output, IntermediateTensors)
get_pp_group().send_tensor_dict(output.tensors,
all_gather_group=get_tp_group())
kv_connector_output = output.kv_connector_output
if not kv_connector_output:
return None
# In case of PP with kv transfer, we need to pass through the
# kv_connector_output
if (not kv_connector_output.finished_sending
and not kv_connector_output.finished_recving):
return EMPTY_MODEL_RUNNER_OUTPUT
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.kv_connector_output = kv_connector_output
if isinstance(output, ModelRunnerOutput):
return output
assert isinstance(output, ModelRunnerOutput)
assert isinstance(output, IntermediateTensors)
parallel_config = self.vllm_config.parallel_config
assert parallel_config.distributed_executor_backend != (
"external_launcher") and not get_pp_group().is_last_rank
get_pp_group().send_tensor_dict(output.tensors,
all_gather_group=get_tp_group())
kv_connector_output = output.kv_connector_output
if not kv_connector_output:
return None
# In case of PP with kv transfer, we need to pass through the
# kv_connector_output
if (not kv_connector_output.finished_sending
and not kv_connector_output.finished_recving):
return EMPTY_MODEL_RUNNER_OUTPUT
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.kv_connector_output = kv_connector_output
return output
def take_draft_token_ids(self) -> Optional[DraftTokenIds]: