All pull requests

Pull requests

• Fix Mask2Former _init_weights to match Meta's original initialization

  #50596·KMaruthi2002 opened Nov 10, 2026·

  ·
  1

  2

• Cast the unquantized weights of a dequantized FP8 checkpoint to the model's dtype

  #50595·qgallouedec opened Nov 10, 2026Member·

  ·
  1

  2

• [WIP] [Reference] Fix quantized model loading

  #50593·kylesayrs opened Nov 10, 2026Contributor·

  ·

• Fix center_crop odd-padding placement to match the torchvision backend

  #50592·InesZenkri opened Nov 10, 2026·

  ·
  1

  1

• fix(hy_v4): use interleaved RoPE to match checkpoint layout

  #50589·Mmy666666 opened Nov 10, 2026·

  ·
  3

• Add native loss support for JEV decision models

  #50586·kashif opened Nov 10, 2026Contributor·

  ·
  2

• 🚨🚨 Rename past_key_values/cache_params to cache

  #50585·Cyrilvallez opened Nov 10, 2026Member·

  ·
  18

• Support Enum in get_json_schema, raise on custom classes

  #50583·Rocketknight1 opened Nov 10, 2026Member·

  ·
  1

  5

• fix(pipelines): respect problem_type in image and video classification pipelines

  #50582·surendra902 opened Nov 10, 2026·

  ·
  2

• Add TeleChat3

  New model
  #50580·hmellor opened Nov 10, 2026Member·

  ·
  5

• Add IQuestCoder

  New model
  #50579·hmellor opened Nov 10, 2026Member·

  ·
  3

• Add InternLM3

  New model
  #50578·hmellor opened Nov 10, 2026Member·

  ·
  5

• docs: remove broken contributor link in UL2 documentation

  #50576·tjt913397-sudo opened Nov 10, 2026·6 tasks·

  ·
  3

• Fix per-layer SwiGLU clipping in BailingHybrid

  #50574·icavan opened Nov 10, 2026Contributor·

  ·
  1

• Follow the SentencePiece model when converting a bare tokenizer.model

  #50555·hmellor opened Nov 10, 2026Member·

  ·
  1

• Fix RecurrentGemma decode-step mask built from a recurrent layer (padded batch generation broken)

  #50554·ydshieh opened Nov 10, 2026Collaborator·

  ·
  4

• Qualcomm AI Engine Direct - QNN Minimal Change to Support IO Alignment

  #50552·winskuo-quic opened Nov 10, 2026·

  ·
  1

• Decision: batch the prompts of a request with slice_logits_to_keep

  #50548·ArthurZucker opened Nov 10, 2026Collaborator·

  ·
  5

• Fix Step3.7 vision RoPE spatial W/H axis order

  #50547·caohch-1 opened Nov 10, 2026·

  ·
  1

  2

• Removing legacy requires_grad workaround in Wav2Vec2FeatureEncoder to fix torch.compile (fullgraph=True)

  #50544·kamboj-sajan opened Nov 10, 2026·4 of 6 tasks·

  ·
  1

  7

• Fix eager torchao import initializing CUDA

  #50542·jaynarayan-yadav opened Nov 10, 2026·

  ·
  1

  4

• [docs] Update index

  #50541·stevhliu opened Nov 9, 2026Member·

  ·
  1

• [PR4] [Modular Cleanup] *MLP Canonicals & Renaming

  #50539·itazap opened Nov 9, 2026Collaborator·

  ·

• [PR3] [Modular Cleanup] *RotaryEmbedding canonical parents and renaming

  #50537·itazap opened Nov 9, 2026Collaborator·

  ·

• fix(wav2vec2): preserve word delimiter around unk token

  #50536·TejasLamba2006 opened Nov 9, 2026·

  ·
  1

  2