A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Fix idefics3 `tie_word_embeddings` (#25454)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Sep 23, 2025 at 18:06 UTC
231c2c63e4decd0cbf863690dfffe88e1d97a003
1 file changed
+1
-1
vllm/model_executor/models/idefics3.py
+1
-1
@@ -608,7 +608,7 @@ class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal,
608
prefix=maybe_prefix(prefix, "lm_head"),
609
)
610
if self.config.text_config.tie_word_embeddings:
611
- self.lm_head.weight = self.model.text_model.wte.weight
611
+ self.lm_head.weight = self.model.text_model.embed_tokens.weight
612
self.logits_processor = LogitsProcessor(config.text_config.vocab_size)
613
614
def _parse_and_validate_image_input(