All pull requests

Pull requests

• Fix remaining modalities loss for causallm

  #50445·SunMarc opened Nov 6, 2026Member·3/3·

  ·
  2

• [FSDP tests] Fork CPU ranks from a forkserver instead of spawning them

  #50444·ydshieh opened Nov 6, 2026Collaborator·

  ·
  2

• Bump the actions group across 1 directory with 14 updates

  dependenciesgithub_actions
  #50442·dependabot[bot] opened Nov 6, 2026Bot·

  ·
  2

• Loss cleaning for CausalLM

  #50441·SunMarc opened Nov 6, 2026Member·2/3·

  ·
  8

• Fix tokenizer.backend being overwritten by backend kwarg

  #50439·samyuktahegde opened Nov 6, 2026Contributor·4 of 6 tasks·

  ·
  1

  1

• [Decision] Add native D1 Omni with modular components and training loss

  #50438·kashif opened Nov 6, 2026Contributor·

  ·
  2

• chore(ci): refresh workflow references (#50436)

  #50437·hf-security-analysis[bot] opened Nov 6, 2026Bot·

  ·
  1

• Fix Mimi encoder frame alignment for non-aligned inputs

  #50429·kasheeesh08 opened Nov 6, 2026·

  ·
  1

  2

• 🚨 Rework slightly subprocessor loading

  #50428·zucchini-nlp opened Nov 6, 2026Member·

  ·
  3

• Fix Qwen VL positions when a text cache is reused with a new image

  #50427·Arnav66692 opened Nov 6, 2026Contributor·

  ·
  1

  3

• Tests: skip audio model tests when torchaudio or librosa is missing

  #50423·IlyasMoutawwakil opened Nov 5, 2026Member·

  ·
  2

• CI: install OpenVINO and NNCF in the daily GPU image

  #50422·IlyasMoutawwakil opened Nov 5, 2026Member·

  ·
  1

• fix(masking): preserve packed isolation with empty KV caches

  #50421·BlazeSR10hawk opened Nov 5, 2026Contributor·

  ·
  1

• Doge: note that the MoE router split is not causal in a full forward pass

  #50420·Andreas-Hum opened Nov 5, 2026Contributor·3 of 5 tasks·

  ·
  1

• [Trainer] Support python dataclass objects in _prepare_input

  #50416·aarushdubey opened Nov 5, 2026·4 of 5 tasks·

  ·
  1

  4

• fix(attention): enforce packed isolation after blockwise overlays

  #50415·BlazeSR10hawk opened Nov 5, 2026Contributor·

  ·
  1

• Keep the GLM-5.3-Flash mHC base and scale in float32

  #50413·albertvillanova opened Nov 5, 2026Member·3 of 5 tasks·

  ·
  1

  7

• Fix Mimi encode last frame when length is not a multiple of frame_size

  #50412·its-amann opened Nov 5, 2026·

  ·
  1

  2

• [Decision] Add native model-specific training losses

  #50411·kashif opened Nov 5, 2026Contributor·2 of 6 tasks·

  ·
  3

• Deprecate Wav2Vec2ProcessorWithLM and pyctcdecode/kenlm LM decoding

  #50408·eustlb opened Nov 5, 2026Collaborator·

  ·
  1

  2

• QA / Add a mic runner in Make

  #50407·tarekziade opened Nov 5, 2026Collaborator·

  ·
  7

• test(fastspeech2_conformer): change set_seed from 42 to 0 in test_training_integration

  #50405·Ridham2509 opened Nov 5, 2026·

  ·
  4

  

• Add support for compiled transformers

  #50401·cselkiraly opened Nov 5, 2026·

  ·
  1

  1

• [Seq2SeqTrainer] Support optimizer_cls_and_kwargs in Seq2SeqTrainer

  #50398·GautamKumarOffical opened Nov 5, 2026·4 of 6 tasks·

  ·
  1

  1

• Title: Optimize Cache Memory Operations and Tensor Reuse

  #50397·Behdad-kanaani opened Nov 5, 2026·

  ·
  3