[feat] FlashInfer CuteDSL MegaMoE integration (#54049)

Signed-off-by: jdebache <jdebache@nvidia.com>
This commit is contained in:
Julien Debache
2026-10-02 15:43:33 -04:00
committed by GitHub
parent c4973d94fe
commit 111f71a6db
33 changed files with 1663 additions and 1036 deletions
@@ -49,6 +49,20 @@ if has_flashinfer_nvlink_one_sided():
logger = init_logger(__name__)
class PassThroughAll2AllManager(All2AllManagerBase):
"""Placeholder for ``all2all_backend="passthrough"``.
The MoE backend dispatches and combines itself, so there is no all2all to
manage.
"""
def get_handle(self, kwargs):
raise RuntimeError(
"passthrough has no all2all handle: dispatch and combine run "
"inside the MoE backend."
)
class AgRsAll2AllManager(All2AllManagerBase):
"""An implementation of all2all communication based on
all-gather (dispatch) and reduce-scatter (combine).
@@ -255,6 +255,12 @@ class CudaCommunicator(DeviceCommunicatorBase):
from .all2all import FlashInferNVLinkOneSidedManager
self.all2all_manager = FlashInferNVLinkOneSidedManager(self.cpu_group)
elif self.all2all_backend == "passthrough":
from .all2all import PassThroughAll2AllManager
self.all2all_manager = PassThroughAll2AllManager(
self.cpu_group, tcp_store_group
)
else:
raise ValueError(f"Unknown all2all backend: {self.all2all_backend}")