[Feature] Expert Parallelism Load Balancer (EPLB) (#18343)

Signed-off-by: Bowen Wang <abmfy@icloud.com>
This commit is contained in:
Bowen Wang
2025-06-26 15:30:21 -07:00
committed by GitHub
parent 07b8fae219
commit e9fd658a73
24 changed files with 2446 additions and 54 deletions
@@ -825,7 +825,16 @@ class Fp8MoEMethod(FusedMoEMethodBase):
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
if enable_eplb:
assert expert_load_view is not None
assert logical_to_physical_map is not None
assert logical_replica_count is not None
assert isinstance(layer, FusedMoE)
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
@@ -839,6 +848,11 @@ class Fp8MoEMethod(FusedMoEMethodBase):
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
enable_eplb=enable_eplb,
expert_map=expert_map,
expert_load_view=expert_load_view,
logical_to_physical_map=logical_to_physical_map,
logical_replica_count=logical_replica_count,
)
if self.rocm_aiter_moe_enabled: