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

Set num_fused_shared_experts as num_shared_experts when shared_experts fusion is not disabled (#6736)

Browse Source
This commit is contained in:
Cheng Wan
2025-06-04 15:53:22 -07:00
committed by GitHub
parent f0f84975f4
commit 81964328b7
22 changed files with 381 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
sgl-kernel/python/sgl_kernel/moe.py
View File

@@ -51,8 +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 expert will be replaced with a round-robin shared expert
# routed_scaling_factor: if > 0, the last expert will be scaled by this 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,