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

[BugFix] fix gpt-oss-120b launch failure with --enable-piecewise-cuda-graph (#16757)

Browse Source
This commit is contained in:
Minglei Zhu
2026-01-11 09:19:59 +08:00
committed by GitHub
parent 2a9344d320
commit a2c2c09d7d
1 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/mxfp4.py
+8 -2
View File
@@ -565,10 +565,16 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
from triton_kernels.numerics_details.mxfp import upcast_from_mxfp from triton_kernels.numerics_details.mxfp import upcast_from_mxfp
w13_weight = upcast_from_mxfp( w13_weight = upcast_from_mxfp(
layer.w13_weight, layer.w13_weight_scale, dtype=torch.bfloat16, axis=-1 layer.w13_weight,
layer.w13_weight_scale,
target_dtype=torch.bfloat16,
axis=-1,
) )
w2_weight = upcast_from_mxfp( w2_weight = upcast_from_mxfp(
layer.w2_weight, layer.w2_weight_scale, dtype=torch.bfloat16, axis=-1 layer.w2_weight,
layer.w2_weight_scale,
target_dtype=torch.bfloat16,
axis=-1,
) )
del layer.w13_weight del layer.w13_weight
del layer.w2_weight del layer.w2_weight