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

[Bug] Fix V4-Pro NaN on Blackwell by converting fp8_einsum input scale to ue8m0 (#25733)

Browse Source
This commit is contained in:
Yuhao Yang
2026-05-18 23:48:34 -07:00
committed by GitHub
parent 7c3f614e23
commit 79ea30d1f1
1 changed files with 1 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_v4.py
+1
View File
@@ -623,6 +623,7 @@ class MQALayer(nn.Module):
o.reshape(T * G, D).contiguous(), o.reshape(T * G, D).contiguous(),
group_size=128, group_size=128,
) )
o_s = deep_gemm.ceil_to_ue8m0(o_s)
output = torch.empty(T, G, R, device=o.device, dtype=torch.bfloat16) output = torch.empty(T, G, R, device=o.device, dtype=torch.bfloat16)
deep_gemm.fp8_einsum( deep_gemm.fp8_einsum(
"bhr,hdr->bhd", "bhr,hdr->bhd",