[MoE][Refactor] Make select_experts a non-static method (#29067)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-11-24 13:38:04 -05:00
committed by GitHub
parent cec418b5df
commit 8f066146c3
18 changed files with 163 additions and 472 deletions
@@ -696,7 +696,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
def apply(
self,
layer: torch.nn.Module,
layer: FusedMoE,
x: torch.Tensor,
router_logits: torch.Tensor,
top_k: int,
@@ -717,12 +717,11 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
logical_to_physical_map: torch.Tensor | None = None,
logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ModelOptFp8MoEMethod` yet."
)
if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
if layer.enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ModelOptFp8MoEMethod` yet."
)
assert activation == "silu", (
f"Expected 'silu' activation but got {activation}"
)
@@ -740,19 +739,9 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
)
# Expert selection
topk_weights, topk_ids, _ = FusedMoE.select_experts(
topk_weights, topk_ids, _ = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
use_grouped_topk=use_grouped_topk,
top_k=top_k,
renormalize=renormalize,
topk_group=topk_group,
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
routed_scaling_factor=routed_scaling_factor,
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
)
if self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS:
@@ -1459,7 +1448,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
def apply(
self,
layer: torch.nn.Module,
layer: FusedMoE,
x: torch.Tensor,
router_logits: torch.Tensor,
top_k: int,
@@ -1480,16 +1469,16 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
logical_to_physical_map: torch.Tensor | None = None,
logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ModelOptNvFp4FusedMoE` yet."
)
assert activation == "silu", "Only SiLU activation is supported."
if (
self.allow_flashinfer
and self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM
):
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ModelOptNvFp4FusedMoE` yet."
)
return flashinfer_trtllm_fp4_moe(
layer=layer,
x=x,
@@ -1502,19 +1491,9 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
e_score_correction_bias=e_score_correction_bias,
)
topk_weights, topk_ids, _ = FusedMoE.select_experts(
topk_weights, topk_ids, _ = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
use_grouped_topk=use_grouped_topk,
top_k=top_k,
renormalize=renormalize,
topk_group=topk_group,
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
routed_scaling_factor=routed_scaling_factor,
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
)
if self.use_marlin: