All pull requests

Pull requests

• build: enable Python 3.15 wheels

  #2109·shaanmajid opened Oct 10, 2026·

  ·
  2

• Add Windows ARM64 build and CI support

  #2099·Harishmcw opened Sep 10, 2026·

  ·
  1

• Register nvidia-cublas-cu12/nvidia-cudnn-cu12 DLL dirs on Windows import

  #2098·si-kui-a opened Sep 9, 2026·

  ·

• Guard dim_t overflow in compute_size, reshape and grow

  #2097·kyo-zzz opened Sep 2, 2026·

  ·
  2

• Fix decoder prefix preprocessing in alternatives mode

  #2087·akira-in-tech opened Aug 21, 2026·

  ·
  1

• Make flash_attention effective for the Whisper encoder

  #2083·nilsreck opened Aug 14, 2026·

  ·

• Add native Apple Metal/MPS backend

  #2077·TBO22 opened Jul 21, 2026·

  ·
  18

• Allow Whisper decoding to use the full decoder context

  #2075·Emre-Akgul opened Jul 11, 2026·

  ·
  1

  22

• Migrate CUDA build to first-class CMake CUDA language

  #2067·drgaborlukacs opened Jun 25, 2026·

  ·

• feat: add experimental runtime LoRA delta injection (ENABLE_LORA_RUNTIME)

  #2046·wonsup-shin opened May 14, 2026·

  ·

• cuda: add simple cudaMalloc/cudaFree allocator as opt-in (workaround for #2038)

  #2044·T0nd3 opened May 12, 2026·4 of 6 tasks·

  ·
  1

• ops: native HIP Flash Attention kernels for AMD RDNA3 (gfx11)

  #2043·T0nd3 opened May 12, 2026·6 tasks done·

  ·

• tests: enable GPU test suites for HIP builds and add ROCm-specific Python tests

  #2042·T0nd3 opened May 11, 2026·

  ·
  2

• docs: add Windows ROCm/HIP build guide

  #2041·T0nd3 opened May 11, 2026·3 tasks done·

  ·

• Add FLASH_ATTN_HDIMS option to limit kernel compilation

  #2029·Caellian opened Apr 3, 2026·

  ·

• Fix ThreadPool shutdown deadlock on Windows with CUDA

  #2027·Deverydoo opened Mar 23, 2026·6 tasks done·

  ·
  4

• Return logits beamsearch

  #2025·cmullovisl opened Mar 8, 2026·

  ·
  1

• aarch64 docker images

  #2023·NicolasBFR opened Mar 4, 2026·

  ·

• Make it possible to use a transformer model as encoder.

  #2022·robinp opened Mar 3, 2026·

  ·

• fix issue with return_logits_vocab option

  #2020·cmullovisl opened Feb 27, 2026·

  ·
  1

• docs: add build instructions for NVIDIA Jetson (aarch64 + CUDA)

  #2019·vimalk78 opened Feb 25, 2026·

  ·
  5

• Fix index calculation for logits_vocab

  #2018·erip opened Feb 23, 2026·

  ·
  1

  1

• Add return_attention support to Generator

  #2010·QuentinFuxa opened Feb 6, 2026·

  ·
  1

  1

• Add a new model: WavLM

  #1966·a2d8a4v opened Jan 2, 2026Contributor·

  ·
  7

• deduplicate and order suppression tokens in apply not add

  #1923·rjames-0 opened Oct 31, 2025·

  ·