transformers releases

Details

Breaking Changes & Deprecations

The minimum supported Python version has been raised to 3.11.
Deprecation warnings have been added for existing input names and rope_scaling assignment.
API compatibility changes have been made, including legacy RoPE key normalization and the thinking -> reasoning_content alias configuration.

New Features & Models

MiniCPM-OCR: Added a document understanding model based on SigLIP+Qwen3.5.
Recursive Flint: Support for an MoE model (routing 6 out of 256 experts) that combines a 512-token sliding window with full attention.
UltraVBert: An extension of ModernVBert that supports 512px tiling and 4x4 patch merging (max_image_tiles 27).
TTS models: Added Gradium Phonon TTS, which uses 12.5 Hz, 32-dimensional latents, and the 100M parameter Kyutai Pocket TTS, which supports real-time speed on CPU.
vLLM integration: Added support for Qwen-Omni Thinker base models, Gemma 3n per-layer inputs, and Flint models.

Performance & Infrastructure

Tensor Parallelism(TP): Introduced an embedding-only style and support for widths not divisible by TP size. Made all expert sharding mandatory in MoE TP/EP plans.
Quantization: Added fp8 backend (torch.float8_e4m3fn) to QuantizedCache. Improved performance by removing integer weight dtype casting for DeepSeek-V4.1 FP4 expert processing and TimesFM 2.5 bitsandbytes 4/8-bit.
Cache/Generation: Improved Flash Attention static cache compatibility and added support for chunked prefill cache continuation.
CI/Infra: Added Intel XPU CI and upgraded to torch 2.14.1. Removed SAM module splitting and improved various test infrastructure.

Bug Fixes

Gemma4: Fixed Processor import errors.
Vision: Fixed PNG transparency white background compositing and DeepSeek-VL/Janus background_color handling errors.
Generation: Fixed BLOOM static cache/compiled generate crashes and HunYuan-VL/Idefics2/3 batch/beam search image allocation errors.
Compatibility: Fixed torch.float64 crashes in MPS environments and torch 2.6/2.7 DtensorShardOperation errors.

Details

Breaking Changes & Deprecations

The minimum supported Python version has been raised to 3.11.
Deprecation warnings have been added for existing input names and rope_scaling assignment.
API compatibility changes have been made, including legacy RoPE key normalization and the thinking -> reasoning_content alias configuration.

New Features & Models

MiniCPM-OCR: Added a document understanding model based on SigLIP+Qwen3.5.
Recursive Flint: Support for an MoE model (routing 6 out of 256 experts) that combines a 512-token sliding window with full attention.
UltraVBert: An extension of ModernVBert that supports 512px tiling and 4x4 patch merging (max_image_tiles 27).
TTS models: Added Gradium Phonon TTS, which uses 12.5 Hz, 32-dimensional latents, and the 100M parameter Kyutai Pocket TTS, which supports real-time speed on CPU.
vLLM integration: Added support for Qwen-Omni Thinker base models, Gemma 3n per-layer inputs, and Flint models.

Performance & Infrastructure

Tensor Parallelism(TP): Introduced an embedding-only style and support for widths not divisible by TP size. Made all expert sharding mandatory in MoE TP/EP plans.
Quantization: Added fp8 backend (torch.float8_e4m3fn) to QuantizedCache. Improved performance by removing integer weight dtype casting for DeepSeek-V4.1 FP4 expert processing and TimesFM 2.5 bitsandbytes 4/8-bit.
Cache/Generation: Improved Flash Attention static cache compatibility and added support for chunked prefill cache continuation.
CI/Infra: Added Intel XPU CI and upgraded to torch 2.14.1. Removed SAM module splitting and improved various test infrastructure.

Bug Fixes

Gemma4: Fixed Processor import errors.
Vision: Fixed PNG transparency white background compositing and DeepSeek-VL/Janus background_color handling errors.
Generation: Fixed BLOOM static cache/compiled generate crashes and HunYuan-VL/Idefics2/3 batch/beam search image allocation errors.
Compatibility: Fixed torch.float64 crashes in MPS environments and torch 2.6/2.7 DtensorShardOperation errors.

transformers