transformers

New Models and Features

• Added Inkling2 (Thinking Machines Lab, 594B/17B active MoE), JEV (decision head), LFM3 (Liquid AI, gated short conv + GQA), and Dots 3 Note Preview (MoE, multimodal, Whisper-style audio encoder, BF16/FP8 support).
• Parallelization: Added ep_implementation selection feature (all-reduce for inference, all-to-all for training) and optimized FSDP/HSDP gradient reduction.
• vLLM: Added support for Omni token counting and text replacement offsets.
• Gemma4: Added prefill optimization and support for passing num_frames_per_video.

Performance Improvements

• Exporters: Integrated by returning ExportedModel, strengthening support for Dynamo, ONNX Runtime, OpenVINO, and ExecuTorch.
• OpenVINO: Decode speed improved by 2.4–3.9x when fused CPU kernels are applied. Based on Xeon Platinum 8570, improved by 4.1x compared to eager (4.9x for int8) and 2.2x compared to torch.compile (2.6x for int8).
• ONNX Runtime: Based on EmbeddingGemma2(A100), improved by 7.6x compared to eager PyTorch and 1.4x compared to torch.compile.

Breaking Changes and Removals

• Removed ep_router; MXFP4 GPT-OSS uses ep_dispatch_experts.
• Removed audio dependencies (librosa, phonemizer, etc.). Wav2Vec2ProcessorWithLM and pyctcdecode/kenlm LM decoding are deprecated.
• Exporters API changes: Changed export() return value and removed output_path parameter.

Key Bug Fixes and Other

• Distributed: Fixed HSDP expert parallelism, ZeRO-3 rank 0 weight reading and composite model initialization, and FSDP2 + cpu_ram_efficient_loading OOM issues.
• Vision/Generation: Fixed bugs in MiniMax-M3-VL, Qwen VL, and PaddleOCR-VL; improved RecurrentGemma padded generation and DeepSeek-V4 beam search cache.
• Environment/CI: Adjusted compatibility for torch 2.15, cu132, and Pillow 13.x; updated Docker images (added NNCF, Nvidia images with torch 2.15/cu132).