Apply logit softcapping in Transformers modelling backend (#52173)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
2026-08-13 16:21:55 +00:00
parent 6014f9e672
commit b245d8e73f
1 changed files with 3 additions and 2 deletions
@@ -59,9 +59,10 @@ class CausalMixin(VllmModelForTextGeneration):
                        self.lm_head = self.lm_head.tie_weights(module)
                        break

-            logit_scale = getattr(self.text_config, "logit_scale", 1.0)
            self.logits_processor = LogitsProcessor(
-                self.text_config.vocab_size, scale=logit_scale
+                self.text_config.vocab_size,
+                scale=getattr(self.text_config, "logit_scale", 1.0),
+                soft_cap=getattr(self.text_config, "final_logit_softcapping", None),
            )
        else:
            self.lm_head = PPMissingLayer()