[PerfFix] Avoid separate thread for MP executor shm spin (#28012)

Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill
2025-11-04 08:33:55 -08:00
committed by GitHub
parent 05cae69f0f
commit c9f66da8fd
9 changed files with 145 additions and 133 deletions
+5 -6
View File
@@ -435,26 +435,25 @@ class RayDistributedExecutor(Executor):
# When PP is used, we return a FutureWrapper immediately so that
# the scheduler can yield to the next batch.
return FutureWrapper(refs)
return FutureWrapper(refs[0])
# Get output from all workers when connector is present
assert self.kv_output_aggregator is not None
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)
return self.kv_output_aggregator.aggregate(ray.get(refs))
# Return a future that will aggregate outputs from all workers
return FutureWrapper(refs, self.kv_output_aggregator)
def collective_rpc(
def collective_rpc( # type: ignore[override]
self,
method: str | Callable,
timeout: float | None = None,
args: tuple = (),
kwargs: dict[str, Any] | None = None,
non_block: bool = False,
) -> list[Any]:
) -> list[Any] | Future[list[Any]]:
"""Runs the given method on all workers."""
sent_method = method if isinstance(method, str) else cloudpickle.dumps(method)
del method
@@ -470,7 +469,7 @@ class RayDistributedExecutor(Executor):
# Get the results of the ray workers.
if non_block:
return [FutureWrapper((output,)) for output in ray_worker_outputs]
return FutureWrapper(ray_worker_outputs)
return ray.get(ray_worker_outputs, timeout=timeout)