All pull requests

Pull requests

• [PR3] [Modular Clean-Up] RoPE canonical parents

  #50333·itazap opened Nov 3, 2026Collaborator·

  ·
  1

• docs(watermarking): fix WatermarkDetector.__call__ docstring types and formatting

  #50332·marlonferreira1800-hue opened Nov 3, 2026·3 tasks done·

  ·
  1

• fix(pipelines): respect audio classification problem type

  #50329·BlazeSR10hawk opened Nov 3, 2026Contributor·

  ·
  1

• Import torchcodec lazily in processing_utils

  #50325·eustlb opened Nov 3, 2026Collaborator·

  ·
  2

• test: dummy change to vit/llama/mistral to verify runtime dep recorder

  #50324·ydshieh opened Nov 3, 2026Collaborator·

  ·
  2

• [Model] LFM3 Family: LFM3-Text

  #50322·paulpak58 opened Nov 3, 2026Contributor·

  ·
  1

• [PR2] [Modular Clean-Up] *Norm Canonical Update and Renaming

  #50321·itazap opened Nov 3, 2026Collaborator·

  ·
  3

• Fix CLAP loss double-counting the text-to-audio direction

  #50319·0xABCD01 opened Nov 3, 2026·3 of 6 tasks·

  ·
  1

  2

• fix(pipelines): validate zero-shot detection batch labels

  #50318·BlazeSR10hawk opened Nov 3, 2026Contributor·

  ·
  1

• [PR1] [Modular Clean-Up] Remove No-op Overrides

  #50317·itazap opened Nov 3, 2026Collaborator·

  ·
  2

• Reject assisted generation with stateful logits processors

  #50314·jeevan1098 opened Nov 3, 2026·

  ·
  1

  1

• Fix Moe aux loss scaling

  #50313·SunMarc opened Nov 3, 2026Member·

  ·
  1

  2

• OpenVINO exporter: let OpenVINO fuse attention, RoPE, the KV cache and compressed matmuls

  #50311·IlyasMoutawwakil opened Nov 3, 2026Member·2 tasks done·

  ·
  1

• Exporters: ONNX Runtime fusions at export time (OnnxConfig(ort_fusions=...))

  #50309·IlyasMoutawwakil opened Nov 3, 2026Member·

  ·
  1

• [kernels] Bump version

  #50307·vasqu opened Nov 3, 2026Collaborator·

  ·
  1

• [Beam] Fix TP on attn

  #50302·vasqu opened Nov 3, 2026Collaborator·

  ·
  3

• Quantize on load from the checkpoint tensor, not its cast to the quantized dtype

  #50301·qgallouedec opened Nov 3, 2026Member·

  ·
  1

  2

• Install kernels via transformers[kernels] in Intel CPU Dockerfile

  #50295·jiqing-feng opened Nov 3, 2026Contributor·

  ·
  1

• [distributed] Support HSDP with expert parallelism

  #50293·3outeille opened Nov 3, 2026Member·

  ·
  1

• [distributed] Divide the expert gradients once when efsdp_size == 1

  #50292·3outeille opened Nov 3, 2026Member·

  ·
  1

• [GraniteSpeech] Fix logits_scaling not applied to logits

  #50290·pszemraj opened Nov 3, 2026Contributor·3 of 6 tasks·

  ·
  1

  1

• Fix QuantizedCache.crop() IndexError breaking speculative decoding

  #50287·siddhant-rajhans opened Nov 3, 2026·

  ·
  1

  3

• [distributed] Trainer: reduce FSDP/HSDP gradients on the last micro-batch only

  #50286·3outeille opened Nov 3, 2026Member·

  ·
  1

• [distributed] Add HSDP (dp_replicate_size) to DistributedConfig

  #50285·3outeille opened Nov 3, 2026Member·

  ·
  1

• ZeRO-3: read weights on rank 0 only and convert them one group at a time

  #50284·3outeille opened Nov 3, 2026Member·

  ·
  1

  4