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

fix: change qwen 3.5 linear attention a_log to fp32 (#19961)

Browse Source 
Co-authored-by: sunqi.7 <sunqi.7@bytedance.com>
This commit is contained in:
shiyu7
2026-03-17 22:42:06 -07:00
committed by GitHub
co-authored by sunqi.7
parent 88c40ec16d
commit 0acc1d3c9a
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_5.py
+1 -1
View File
@@ -193,7 +193,7 @@ class Qwen3_5GatedDeltaNet(nn.Module):
torch.ones(self.num_v_heads // self.attn_tp_size), torch.ones(self.num_v_heads // self.attn_tp_size),
) )
self.A_log = nn.Parameter( self.A_log = nn.Parameter(
torch.empty(self.num_v_heads // self.attn_tp_size), torch.empty(self.num_v_heads // self.attn_tp_size, dtype=torch.float32),
) )
set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)}) set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})