This patch most notably fixes an issue with the new dtype argument (replacing torch_dtype) in pipelines!
Bug Fixes & Improvements
- Fix broken Llama4 accuracy in MoE part (#40609)
- fix pipeline dtype (#40638)
- Fix self.dropout_p is not defined for SamAttention/Sam2Attention (#40667)
- Fix backward compatibility with accelerate in Trainer (#40668)
- fix broken offline mode when loading tokenizer from hub (#40669)
- [Glm4.5V] fix vLLM support (#40696)