diff --git a/python/sglang/srt/models/qwen3_5.py b/python/sglang/srt/models/qwen3_5.py
index fd7aaf258..0cdb4dd83 100644
--- 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)})