Disable neft on eval

fish_speech/models/text2semantic/llama.py

@@ -166,7 +166,7 @@ class Transformer(nn.Module):
         x = torch.stack(vocab_embeds, dim=3)
         x = x.sum(dim=3)
 
-        if self.config.neft_alpha > 0:
+        if self.config.neft_alpha > 0 and self.training:
             # alpha / sqrt(L * D)
             scaled_alpha = self.config.neft_alpha / math.sqrt(
                 self.config.dim * x.shape[2]