EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Revert "[1/2] sgl-kernel: Fuse routed scaling factor into select_experts" (#8706)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-08-02 20:14:30 +08:00
committed by GitHub
parent ac6962ccd6
commit f9f0138f80
5 changed files with 12 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
sgl-kernel/python/sgl_kernel/moe.py
View File

@@ -44,7 +44,6 @@ def moe_fused_gate(
topk,
num_fused_shared_experts=0,
routed_scaling_factor=0,
apply_routed_scaling_factor_on_output=False,
):
# This fused kernel function is used to select topk expert in a hierarchical 2-layer fashion
# it split group of expert into num_expert_group, and use top2 expert weight sum in each group
@@ -52,13 +51,8 @@ def moe_fused_gate(
# the #experts is decided by the input tensor shape and we currently only support power of 2 #experts
# and #experts should be divisible by num_expert_group. #expert/num_expert_group <= 32 is limited for now.
# for non-supported case, we suggest to use the biased_grouped_topk func in sglang.srt.layers.moe.topk
# num_fused_shared_experts: if > 0, the last several experts will be
# replaced with shared experts. the shared experts will be divided by the
# routed_scaling_factor - this is intended to cancel out later when routed+shared
# output is scaled so that shared experts are not scaled.
# routed_scaling_factor: if > 0, the experts will be scaled by this factor
# apply_routed_scaling_factor_on_output: if true, output will be
# scaled by the routed_scaling_factor
# num_fused_shared_experts: if > 0, the last several experts will be replaced with shared experts
# routed_scaling_factor: if > 0, the shared experts will be scaled by this factor
return torch.ops.sgl_kernel.moe_fused_gate.default(
input_tensor,
bias,
@@ -67,7 +61,6 @@ def moe_fused_gate(
topk,
num_fused_shared_experts,
routed_scaling_factor,
apply_routed_scaling_factor_on_output,
)