[MoE Refactor] EPLB refactoring for FusedMoE (#41055)

Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
bnellnm
2026-05-12 14:16:31 -04:00
committed by GitHub
co-authored by Robert Shaw
parent 4d591db470
commit d9b4990783
17 changed files with 83 additions and 98 deletions
-3
View File
@@ -1004,7 +1004,6 @@ def make_fake_moe_layer(
activation: str = "silu",
indices_type: torch.dtype | None = None,
expert_map: torch.Tensor | None = None,
enable_eplb: bool = False,
expert_load_view: torch.Tensor | None = None,
logical_to_physical_map: torch.Tensor | None = None,
logical_replica_count: torch.Tensor | None = None,
@@ -1022,7 +1021,6 @@ def make_fake_moe_layer(
router = create_fused_moe_router(
top_k=top_k,
global_num_experts=global_num_experts,
# eplb_state=None, # TODO
renormalize=renormalize,
use_grouped_topk=use_grouped_topk,
num_expert_group=num_expert_group,
@@ -1032,7 +1030,6 @@ def make_fake_moe_layer(
routed_scaling_factor=routed_scaling_factor,
e_score_correction_bias=e_score_correction_bias,
num_fused_shared_experts=0, # TODO
enable_eplb=enable_eplb,
# TODO(bnell): once we can construct the MK at init time, we
# can make this a value.
indices_type_getter=lambda: indices_type,