[Core] Support async scheduling with uniproc executor (#24219)

Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Nick Hill
2025-09-12 16:34:28 -07:00
committed by GitHub
co-authored by Ronald1995 Robert Shaw
parent 8226dd56bf
commit 4fdd6f5cbf
9 changed files with 103 additions and 55 deletions
+4 -3
View File
@@ -66,11 +66,13 @@ class RayDistributedExecutor(RayDistributedExecutorV0, Executor):
def execute_model(
self,
scheduler_output: SchedulerOutput,
non_block: bool = False,
) -> Union[ModelRunnerOutput, Future[ModelRunnerOutput]]:
"""Execute the model on the Ray workers.
Args:
scheduler_output: The scheduler output to execute.
non_block: If True, the method will return a Future.
Returns:
The model runner output.
@@ -84,7 +86,7 @@ class RayDistributedExecutor(RayDistributedExecutorV0, Executor):
if not self.has_connector:
# Get output only from a single worker (output_rank)
# When PP is not used, we block here until the result is available.
if self.max_concurrent_batches == 1:
if not non_block:
return refs[0].get()
# When PP is used, we return a FutureWrapper immediately so that
@@ -92,7 +94,7 @@ class RayDistributedExecutor(RayDistributedExecutorV0, Executor):
return FutureWrapper(refs)
# Get output from all workers when connector is present
if self.max_concurrent_batches == 1:
if not non_block:
# Block and get results from all workers
outputs = [ref.get() for ref in refs]
return self.kv_output_aggregator.aggregate(outputs)
@@ -106,4 +108,3 @@ class RayDistributedExecutor(RayDistributedExecutorV0, Executor):
if reconfig_request.new_data_parallel_rank == \
ReconfigureRankType.SHUTDOWN_CURRENT_RANK:
self.shutdown()
return