[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
@@ -1068,6 +1068,12 @@ class FusedMoEParallelConfig:
and self.all2all_backend == "allgather_reducescatter"
)
@property
def use_passthrough_all2all(self):
# Not gated on use_all2all_kernels: the experts dispatch and combine
# internally in every EP topology, including TP-only.
return self.all2all_backend == "passthrough"
@property
def use_mori_kernels(self):
return self.use_all2all_kernels and self.all2all_backend in (
@@ -1530,6 +1536,10 @@ class FusedMoEConfig:
def use_deepep_ll_kernels(self):
return self.moe_parallel_config.use_deepep_ll_kernels
@property
def use_passthrough_all2all(self):
return self.moe_parallel_config.use_passthrough_all2all
@property
def use_mori_kernels(self):
return self.moe_parallel_config.use_mori_kernels