All pull requests

Pull requests

• Fix typo in quantization documentation, "ly" -> "by"

  #50217·Zyntriptic opened Nov 1, 2026·

  ·
  1

• Fix repetition penalty for masked padding tokens

  #50215·Narendra02053 opened Nov 1, 2026·

  ·
  1

• Fix/trainer parallel token count 50212

  #50214·BlazeSR10hawk opened Nov 1, 2026Contributor·

  ·
  1

• Gguf glm4moe

  #50213·mshayeb opened Nov 1, 2026·6 tasks·

  ·
  2

• fix(pipelines): preserve batched generation metadata

  #50210·BlazeSR10hawk opened Nov 1, 2026Contributor·

  ·
  1

• Clarify import errors when an unsupported PyTorch version is installed

  #50208·ChrisChen667788 opened Nov 1, 2026·1 task·

  ·
  2

• [GGUF] Fix Mixtral routing and expert mapping, Qwen3 head_dim, and Mistral detection

  #50206·RohanQ17 opened Oct 31, 2026·4 of 5 tasks·

  ·
  1

  2

• Fix prompt lookup decoding with stateful logits processors

  #50205·jeevan1098 opened Oct 31, 2026·

  ·
  1

  2

• Add Jeva

  #50203·kashif opened Oct 31, 2026Contributor·3 tasks done·

  ·
  2

• Fix offset mapping padding in tokenizer.pad

  #50202·yuanx749 opened Oct 31, 2026Contributor·3 of 5 tasks·

  ·
  1

• Fix DeBERTa q_bias/v_bias staying float32 when loaded in another dtype

  #50201·tarekziade opened Oct 31, 2026Collaborator·

  ·
  1

  4

• fix: preserve multiple candidates in multimodal pipelines

  #50198·BlazeSR10hawk opened Oct 31, 2026Contributor·

  ·
  1

• [Mamba] Fused kernel at inference without cache or padding

  #50193·ArthurZucker opened Oct 31, 2026Collaborator·

  ·
  2

• AMD CI: add Radeon AI Pro R9700 (gfx1201)

  #50192·dhekimian opened Oct 31, 2026·

  ·
  1

• AMD CI: Strix Halo (gfx1151) on ROCm 10.1

  #50191·dhekimian opened Oct 31, 2026·

  ·
  1

• Detect packed sequences when the cache is still empty

  #50190·ArthurZucker opened Oct 31, 2026Collaborator·

  ·
  1

• Split KV-sharing and rope attention variants into per-layer classes

  #50189·ArthurZucker opened Oct 31, 2026Collaborator·

  ·
  19

• [NemotronH] Use config.time_step_limit for the Mamba2 dt limit

  #50188·ljhnick opened Oct 30, 2026·

  ·
  1

  3

• Run packed rows through the gated delta net layers as a single row

  #50186·qgallouedec opened Oct 30, 2026Member·

  ·
  1

  2

• Replace torch.FloatTensor with torch.Tensor in the model files

  #50183·qgallouedec opened Oct 30, 2026Member·

  ·
  1

  2

• Replace torch.FloatTensor with torch.Tensor outside the model files

  #50182·qgallouedec opened Oct 30, 2026Member·

  ·
  2

• Use repo-wide linguist-generated globs so new models don't touch .gitattributes

  #50181·ArthurZucker opened Oct 30, 2026Collaborator·

  ·
  2

• [VibeVoice] Add sample argument to disable VAE noise in audio features

  #50179·hmellor opened Oct 30, 2026Member·

  ·
  2

• fix(core_model_loading): strip end-anchor suffix when reversing WeightRenaming patterns

  #50178·devtyagi3909 opened Oct 30, 2026·

  ·
  1

  1

• fix(sam3): preserve position_ids buffer in float32 for rotary embeddings

  #50176·devtyagi3909 opened Oct 30, 2026·

  ·
  1

  2