Version 5.15.1 of the Transformers library is now available as a patch release, addressing several bugs that affected self-hosted deployments. The most notable fixes target DFlash and MTP candidate generators, where token placement on the correct device was corrected when using device_map=”auto”, and the logit distributions for candidate generators that rely on sampling have been aligned to behave consistently.
Additional generator-related improvements include a fix for MTP configurations that previously failed when the mlp_layer_types field was not present, ensuring smoother initialization across different model variants.
On the image processing side, the library will now automatically fall back from the Lanczos resampling filter to bicubic when running on CUDA, preventing cases where images could not be processed on accelerator hardware. The patch also fixes a device assignment issue for Gemma4 video inputs, so video data is sent to the correct device during inference.
These changes are intended to reduce runtime errors and improve stability for users running models locally on GPU or multi-device setups.
