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

Co-authored-by: sunqi.7 <sunqi.7@bytedance.com>
2026-03-18 13:42:06 +08:00
parent 88c40ec16d
commit 0acc1d3c9a
1 changed files with 1 additions and 1 deletions
--- a/python/sglang/srt/models/qwen3_5.py
+++ b/python/sglang/srt/models/qwen3_5.py
@@ -193,7 +193,7 @@ class Qwen3_5GatedDeltaNet(nn.Module):
            torch.ones(self.num_v_heads // self.attn_tp_size),
        )
        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)})