[Kernel][MoE] Add GELU_TANH to CPU, CUTLASS, and WNA16 MoE backends (#42027)

Signed-off-by: lesj0610 <lesj0610@users.noreply.github.com>
Co-authored-by: lesj0610 <lesj0610@users.noreply.github.com>
This commit is contained in:
SeongJun Lee
2026-06-02 17:12:08 -04:00
committed by GitHub
co-authored by lesj0610
parent e15f20258b
commit 3099de3617
7 changed files with 119 additions and 7 deletions
@@ -53,6 +53,10 @@ _CPU_MOE_ACT_FN: dict[MoEActivation, Callable[[torch.Tensor], torch.Tensor]] = {
MoEActivation.SILU: lambda x: SiluAndMul(compile_native=False).forward_native(x),
MoEActivation.SWIGLUOAI: _swigluoai_forward_native,
MoEActivation.GELU: _gelu_and_mul,
MoEActivation.GELU_TANH: (
lambda x: F.gelu(x[..., : x.shape[-1] // 2], approximate="tanh")
* x[..., x.shape[-1] // 2 :]
),
}
@@ -322,6 +322,7 @@ class CutlassExpertsFp8Base(mk.FusedMoEExpertsModular):
return activation in [
MoEActivation.SILU,
MoEActivation.GELU,
MoEActivation.GELU_TANH,
MoEActivation.SWIGLUOAI,
]
@@ -719,10 +720,12 @@ class CutlassExpertsFp4(mk.FusedMoEExpertsModular):
return activation in [
MoEActivation.SILU,
MoEActivation.GELU,
MoEActivation.GELU_TANH,
MoEActivation.SWIGLUOAI,
MoEActivation.SWIGLUSTEP,
MoEActivation.SILU_NO_MUL,
MoEActivation.GELU_NO_MUL,
MoEActivation.GELU_TANH_NO_MUL,
MoEActivation.RELU2_NO_MUL,
]
@@ -13,7 +13,6 @@ from vllm.model_executor.layers.fused_moe import (
RoutedExperts,
SharedExperts,
)
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
int4_w4a16_moe_quant_config,
@@ -367,16 +366,13 @@ class MoeWNA16Method(FusedMoEMethodBase):
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
assert layer.activation == MoEActivation.SILU, (
f"Only SiLU activation is supported, not {layer.activation}."
)
return fused_experts(
x,
layer.w13_qweight,
layer.w2_qweight,
topk_weights=topk_weights,
topk_ids=topk_ids,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,