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

[bugfix] Apply routed scaling factor to cutlass_fused_experts_fp8 (#8688)

Browse Source
This commit is contained in:
Trevor Morris
2025-08-01 19:00:24 -07:00
committed by GitHub
parent b27b11919c
commit 89caf7a3c6
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/quantization/fp8.py
View File

@@ -1039,7 +1039,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
from sglang.srt.layers.moe.cutlass_moe import cutlass_fused_experts_fp8 from sglang.srt.layers.moe.cutlass_moe import cutlass_fused_experts_fp8
topk_weights, topk_ids, _ = topk_output topk_weights, topk_ids, _ = topk_output
return cutlass_fused_experts_fp8( output = cutlass_fused_experts_fp8(
x, x,
layer.w13_weight.transpose(1, 2), layer.w13_weight.transpose(1, 2),
layer.w2_weight.transpose(1, 2), layer.w2_weight.transpose(1, 2),
@@ -1062,6 +1062,10 @@ class Fp8MoEMethod(FusedMoEMethodBase):
self.problem_sizes2, self.problem_sizes2,
use_fp8_blockscale=True, use_fp8_blockscale=True,
) )
# TODO: Fuse into select_experts
if routed_scaling_factor is not None:
output *= routed_scaling_factor
return output
# Expert fusion with FP8 quantization # Expert fusion with FP8 quantization
return fused_experts( return fused_experts(
x, x,