EngineX/xc-llm-kunlun
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[fix]bias bug in kunlun_scale_mm (#126)

Browse Source
This commit is contained in:
Li Wei
2026-01-20 13:24:52 +08:00
committed by GitHub
parent f2019b145f
commit 2a2d773ad0
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm_kunlun/ops/quantization/kernels/kunlun_scale_mm.py
View File

@@ -76,7 +76,7 @@ class KunlunScaledMMLinearKernel(CutlassScaledMMLinearKernel):
out_dtype=x.dtype,
azp_adj=azp_adj,
azp=azp,
bias=bias.to(torch.float32).contiguous() if bias else None,
bias=bias.to(torch.float32).contiguous() if bias is not None else None,
)
else: # symmetric
return torch.ops._C.matmul(
@@ -85,7 +85,7 @@ class KunlunScaledMMLinearKernel(CutlassScaledMMLinearKernel):
out_dtype=x.dtype,
x_pc_max=x_s * 127.0 if static else x_s,
w_pc_max=w_s,
bias=bias.to(torch.float32).contiguous() if bias else None,
bias=bias.to(torch.float32).contiguous() if bias is not None else None,
)
# backup option: lower performance
@@ -95,7 +95,7 @@ class KunlunScaledMMLinearKernel(CutlassScaledMMLinearKernel):
# scale_a=x_s / 127.0 if not static else x_s,
# scale_b=(w_s / 127.0).transpose(0, 1),
# out_dtype=x.dtype,
# bias=bias.to(torch.float32).contiguous() if bias else None,
# bias=bias.to(torch.float32).contiguous() if bias is not None else None,
# )