[Core] Release FlashInfer allreduce workspace with --enable-nccl-comm-suspend (#59360)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
aoshen02
2026-10-10 02:26:40 +00:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 193922d6ff
commit 1dfe9fbdae
3 changed files with 98 additions and 4 deletions
@@ -814,12 +814,19 @@ class CudaCommunicator(DeviceCommunicatorBase):
return bd is not None and bd.uniform
def suspend(self) -> None:
from .flashinfer_all_reduce import checkpoint_prepare_fi_ar_workspaces
# FlashInfer AR syncs over the gloo cpu_group, so order vs. NCCL is free.
checkpoint_prepare_fi_ar_workspaces(self.cpu_group, skip_unsupported=True)
if self.pynccl_comm is not None:
self.pynccl_comm.suspend()
def resume(self) -> None:
from .flashinfer_all_reduce import checkpoint_restore_fi_ar_workspaces
if self.pynccl_comm is not None:
self.pynccl_comm.resume()
checkpoint_restore_fi_ar_workspaces(self.cpu_group, skip_unsupported=True)
def checkpoint_prepare(self) -> None:
# Only FlashInfer all-reduce and FlashInfer all2all are supported for now.
@@ -360,14 +360,30 @@ def _fi_ar_workspaces_for_group(group: ProcessGroup) -> list[Any]:
return group_workspaces
def checkpoint_prepare_fi_ar_workspaces(group: ProcessGroup) -> None:
def checkpoint_prepare_fi_ar_workspaces(
group: ProcessGroup, skip_unsupported: bool = False
) -> None:
for workspace in _fi_ar_workspaces_for_group(group):
workspace.checkpoint_prepare()
try:
workspace.checkpoint_prepare()
except NotImplementedError:
if not skip_unsupported:
raise
logger.info_once(
"FlashInfer all-reduce workspace backed by torch symmetric memory "
"stays resident during suspend."
)
def checkpoint_restore_fi_ar_workspaces(group: ProcessGroup) -> None:
def checkpoint_restore_fi_ar_workspaces(
group: ProcessGroup, skip_unsupported: bool = False
) -> None:
for workspace in _fi_ar_workspaces_for_group(group):
workspace.checkpoint_restore(TorchDistBackend(group=group))
try:
workspace.checkpoint_restore(TorchDistBackend(group=group))
except NotImplementedError:
if not skip_unsupported:
raise
atexit.register(destroy_fi_ar_workspace)