All pull requests

Pull requests

• This PR fixes Gemma3n fp16 loading when missing weights need to be in…

  #49147·sywangyi opened Sep 28, 2026Contributor·

  ·
  3

• docs: fix unclosed and mis-indented code fences

  #49146·gonisulaimann opened Sep 28, 2026·

  ·
  1

• Map bare list, tuple and dict annotations to the right JSON schema type in get_json_schema

  #49145·825pranav opened Sep 27, 2026·

  ·
  1

  1

• [serge] Fix 2 integration tests for model jais2 failing with output_mismatch (tensor values differ (2))

  #49143·sergereview[bot] opened Sep 27, 2026Bot·

  ·
  3

• [serge] Fix 2 integration tests for model cohere2 failing with other (other (2))

  #49142·sergereview[bot] opened Sep 27, 2026Bot·

  ·
  3

• Skip the query/key head repeat for grouped value heads in gated DeltaNet

  #49141·manav-simile opened Sep 27, 2026·

  ·
  2

• fix nested set_attn_implementation

  #49138·eustlb opened Sep 27, 2026Collaborator·

  ·
  1

  2

• Fix get_json_schema dropping items/enum for unions of list/dict/Literal types

  #49136·JoeyTan21 opened Sep 27, 2026·

  ·
  1

  1

• Fix UMT5 decoder self-attention not being causal

  #49135·the-cross-art opened Sep 27, 2026·

  ·
  1

  2

• [Gemma] Fix test_model_7b_fp16_static_cache expected value for cuda 8 after #49084

  #49133·ydshieh opened Sep 27, 2026Collaborator·

  ·
  13

• [New model] Add K2 Horizon

  #49132·boyangaaaaa opened Sep 27, 2026·

  ·
  2

• [docs] Add warning about remove_unused_columns silently dropping custom dataset fields

  #49131·qarong opened Sep 27, 2026·

  ·
  1

• Fix DeepSeek-OCR-2 tokenizer loading when model config imports fail

  #49130·forest-david opened Sep 27, 2026·

  ·
  3

• Skip flash tests that fall back to a hub kernel when kernels is missing

  #49129·Abdennacer-Badaoui opened Sep 27, 2026Member·

  ·
  2

• Add Strix Halo (gfx1151) Atlas Inference Hub-kernel path for Qwen3.5/3.6/3.8 Gated DeltaNet

  #49127·AzeezIsh opened Sep 26, 2026Contributor·

  ·
  1

• dummy: trigger CI on gemma3n test_model_4b_batch

  #49125·ydshieh opened Sep 26, 2026Collaborator·

  ·
  15

• Fix Trainer checkpoint resume crashing on CPU with multiple processes

  #49123·neevmodh opened Sep 26, 2026·

  ·
  1

  2

• [HF Exporters] Adds opt-in ExecuTorch off-graph KV cache export

  #49120·metascroy opened Sep 25, 2026Contributor·

  ·
  10

• Fix MusicGen forward(labels=...) when decoder.decoder_start_token_id is None

  #49118·Omarinho12345 opened Sep 25, 2026·

  ·
  1

  2

• Fix SequenceBiasLogitsProcessor edge cases: token id 0 and prefix equal to context

  #49117·lucaluo925 opened Sep 25, 2026·

  ·
  1

  1

• qwen3 models map to wrong tokenizer class on the hub

  #49116·itazap opened Sep 25, 2026Collaborator·

  ·
  1

  8

• Little micro-optimization for FlashAttention

  #49115·Rocketknight1 opened Sep 25, 2026Member·

  ·
  3

• pass pretokenizer regex charsmap to qwen2 tokenizer

  #49114·itazap opened Sep 25, 2026Collaborator·

  ·
  3

• [Attention] Deprecate "paged|" prefix for SDPA and flash

  #49112·remi-or opened Sep 25, 2026Collaborator·5/5·

  ·
  3

• Fix parse_generate_flags to properly handle backslashes in string values

  #49110·dajiaohuang opened Sep 25, 2026Contributor·

  ·
  1

  2