From b245d8e73f67f1fffed3b46b6152078dc1683fe6 Mon Sep 17 00:00:00 2001 From: Harry Mellor <19981378+hmellor@users.noreply.github.com> Date: Thu, 13 Aug 2026 17:21:55 +0100 Subject: [PATCH] Apply logit softcapping in Transformers modelling backend (#52173) Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> --- vllm/model_executor/models/transformers/causal.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/models/transformers/causal.py b/vllm/model_executor/models/transformers/causal.py index b637df1164..a32e2b54ba 100644 --- a/vllm/model_executor/models/transformers/causal.py +++ b/vllm/model_executor/models/transformers/causal.py @@ -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()