minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Revert moe sum reduce for marlin moe (#13314)

Browse Source
This commit is contained in:
Ke Bao
2025-11-15 15:57:41 +08:00
committed by GitHub
parent 8a4373405e
commit 2a96e302cb
1 changed files with 3 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
sgl-kernel/python/sgl_kernel/fused_moe.py
+3 -9
View File
@@ -3,7 +3,6 @@ from typing import Optional
import torch import torch
from sgl_kernel.elementwise import silu_and_mul from sgl_kernel.elementwise import silu_and_mul
from sgl_kernel.moe import moe_sum_reduce
def get_scalar_type(num_bits: int, has_zp: bool): def get_scalar_type(num_bits: int, has_zp: bool):
@@ -206,15 +205,10 @@ def fused_marlin_moe(
is_zp_float=False, is_zp_float=False,
).view(-1, topk, K) ).view(-1, topk, K)
if routed_scaling_factor is None:
routed_scaling_factor = 1.0
output = hidden_states if inplace else torch.empty_like(hidden_states) output = hidden_states if inplace else torch.empty_like(hidden_states)
moe_sum_reduce( torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1, out=output)
intermediate_cache3, if routed_scaling_factor is not None:
output, output *= routed_scaling_factor
routed_scaling_factor,
)
return output return output