All pull requests

Pull requests

• [CI] Improve run-slow PR comments and remove CI recap

  #49782·ydshieh opened Oct 20, 2026Collaborator·

  ·
  1

• MXFP8 on the fine-grained machinery (Mxfp8Config)

  #49780·IlyasMoutawwakil opened Oct 20, 2026Member·10/10·

  ·
  3

• NVFP4 on the fine-grained machinery, with the producers' NVFP4 exports

  #49779·IlyasMoutawwakil opened Oct 20, 2026Member·9/10·

  ·
  3

• MXFP4 on the fine-grained machinery, with DeepSeek-V4's FP8 + MXFP4 plan

  #49778·IlyasMoutawwakil opened Oct 20, 2026Member·8/10·

  ·
  3

• FP8 on the fine-grained machinery (Fp8Config)

  #49777·IlyasMoutawwakil opened Oct 20, 2026Member·7/10·

  ·
  3

• Fine-grained block-scaled quantization machinery, with no format

  #49776·IlyasMoutawwakil opened Oct 20, 2026Member·6/10·

  ·
  3

• Experts and TP: per-instance experts flags, experts requirements, a public local-tensor marker, intra-expert styles

  #49775·IlyasMoutawwakil opened Oct 20, 2026Member·5/10·

  ·
  2

• Quantizer API: model-aware conversions, checkpoint renames, config readers and adapters, skip-list resolution, dequantized saves

  #49774·IlyasMoutawwakil opened Oct 20, 2026Member·4/10·

  ·
  2

• Weight loading: many-to-many ops declare themselves, plus four small loader fixes

  #49773·IlyasMoutawwakil opened Oct 20, 2026Member·3/10·

  ·
  2

• Keep a capturing group's escapes when reversing a conversion pattern

  #49772·IlyasMoutawwakil opened Oct 20, 2026Member·2/10·

  ·
  2

• Match a skip-list entry from the start only up to a module boundary

  #49771·IlyasMoutawwakil opened Oct 20, 2026Member·1/10·

  ·
  3

• CTC loss refactor

  #49770·ebezzam opened Oct 20, 2026Contributor·

  ·
  1

• Fix MoE auxiliary loss scaling in Trainer

  #49769·priyanshu0463 opened Oct 20, 2026·

  ·
  1

  1

• Fix background_color handling across DeepSeek-VL and Janus image processors

  Code agent slop
  #49767·tritsystem opened Oct 20, 2026·

  ·
  1

  4

• feat(tokenizers): allow install of v1.0.0-rc.x

  #49764·SBrandeis opened Oct 20, 2026Contributor·

  ·
  4

• fix(parsing): decode escaped string delimiters in json content parser

  #49763·zoechen281 opened Oct 20, 2026·

  ·
  1

  1

• [Atlas Inference] Add GB10/SM121 Hub-kernel path for NVFP4Config

  #49756·AzeezIsh opened Oct 20, 2026Contributor·5 of 6 tasks·

  ·
  1

• Modular suffix overlap - allow non-model classes

  #49754·zucchini-nlp opened Oct 20, 2026Member·

  ·
  4

• Keep image inputs on the first chunk of a chunked prefill

  #49753·digit50 opened Oct 20, 2026Contributor·3 of 6 tasks·

  ·
  1

  3

• fix(cache): clear request position_ids on reset (fixes #49745)

  #49749·Omarinho12345 opened Oct 19, 2026·

  ·
  1

  2

• Fix library import failing without torch when we have PIL installed

  #49748·Sai-Suraj-27 opened Oct 19, 2026Contributor·6 tasks·

  ·
  8

• [jev-likes] have the decision heads set the number of options the model can score

  #49747·apolinario opened Oct 19, 2026·

  ·
  4

• [docs] Quantizer integration

  #49746·stevhliu opened Oct 19, 2026Member·

  ·
  2

• Fix grouped RMSNorm in Mamba2's non-fused path

  #49744·taking-lying-flat opened Oct 19, 2026·4 tasks done·

  ·
  1

  2

• Fold GQA query heads onto their KV head when decoding instead of repeating K/V

  #49743·Yooniel opened Oct 19, 2026Contributor·3 of 6 tasks·

  ·
  1