fix: gemma 3 not use softcap (#5622)

2025-04-22 01:16:08 -07:00
parent 506be6b892
commit 04f2abcb34
4 changed files with 17 additions and 2 deletions
--- a/python/sglang/srt/models/gemma3_causal.py
+++ b/python/sglang/srt/models/gemma3_causal.py
@@ -189,7 +189,7 @@ class Gemma3Attention(nn.Module):
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            layer_id=layer_id,
-            logit_cap=getattr(self.config, "attn_logit_softcapping", None),
+            logit_cap=0.0,
            # Module must also define `get_attention_sliding_window_size` to correctly initialize
            # attention backend in `ForwardBatch`.
            sliding_window_size=self.sliding_window,