[2/N] Elastic EP Milestone 2: Integrating NIXL-EP (#35627)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Co-authored-by: Yongji Wu <wuyongji317@gmail.com>
Co-authored-by: Ron Tourgeman <rtourgeman@nvidia.com>
This commit is contained in:
Itay Alroy
2026-03-13 09:25:33 -04:00
committed by GitHub
co-authored by Yongji Wu Ron Tourgeman
parent 82f836d976
commit d5af196c18
14 changed files with 635 additions and 11 deletions
@@ -177,10 +177,11 @@ def determine_expert_placement_strategy(
if (
moe_parallel_config.use_all2all_kernels
and not moe_parallel_config.use_deepep_ll_kernels
and not moe_parallel_config.use_nixl_ep_kernels
):
logger.warning(
"Round-robin expert placement currently only supports "
"the DeepEP low-latency backend, but '%s' was configured. "
"the DeepEP low-latency or NIXL EP backend, but '%s' was configured. "
"Falling back to linear expert placement.",
moe_parallel_config.all2all_backend,
)
@@ -745,10 +746,10 @@ class FusedMoE(CustomOp):
self,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None:
# Currently routing_tables only needed for round-robin expert placement
# with DeepEP-ll all2all backend.
if (
self.expert_placement_strategy != "round_robin"
or not self.moe_parallel_config.use_deepep_ll_kernels
# with DeepEP-ll or NIXL EP all2all backends.
if self.expert_placement_strategy != "round_robin" or (
not self.moe_parallel_config.use_deepep_ll_kernels
and not self.moe_parallel_config.use_nixl_ep_kernels
):
return None