All pull requests

Pull requests

• Decode graphs: grow an eager KV cache by the eager increment beside a kept graph store

  #16191·danielhanchen opened Oct 30, 2026Member·

  ·
  2

• Studio Images: encode gallery PNGs about 2x faster with byte-identical output

  #16190·danielhanchen opened Oct 30, 2026Member·

  ·

• Import: scan installed distributions once while transformers and TRL import

  #16189·danielhanchen opened Oct 30, 2026Member·

  ·

• Studio: stop the false "Could not sync pinned chats" toast during a desktop update

  #16188·oobabooga opened Oct 30, 2026Member·

  ·
  1

• Packing: do not serve compiled inference_mode metadata to a training step

  #16183·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• RoPE: compiled stock-model calls index the cache without a host read

  #16182·danielhanchen opened Oct 30, 2026Member·

  ·
  7

• FSDP2 save: one writer for a tensor parallel x FSDP2 model

  #16181·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• FSDP2: average tp gradients over dp when every trainable param is ignored

  #16180·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• Studio: run cheap dataset formatting maps in process for in-memory datasets

  #16179·danielhanchen opened Oct 30, 2026Member·

  ·

• Studio: train on CPU when no GPU is found, and offer Install Training on GPU-less Chat only installs

  #16175·oobabooga opened Oct 30, 2026Member·

  ·
  2

• Studio: overlap web image lookup with text search and independent URL reads in one tool batch

  #16174·danielhanchen opened Oct 30, 2026Member·

  ·

• FSDP2: sum fp16 replicated gradients over unsynced micro-batches in float32

  #16173·danielhanchen opened Oct 30, 2026Member·

  ·
  4

• Flex small-smem tiles: keep float32 calls on the default tiles

  #16172·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• Auto checkpoint skip: count FSDP2 gradient and parameter bytes per rank

  #16171·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• Pipeline parallel: end a non-last stage without a graph break under torch.compile

  #16167·danielhanchen opened Oct 30, 2026Member·

  ·
  4

• FSDP2 QAT export: count the converted copy in the writer's host RAM check

  #16166·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• CLI: audit dataset supervision before loading model weights

  #16165·alkinun opened Oct 30, 2026Collaborator·

  ·
  1

• Studio: stream durable chat events without re-parsing stored payloads

  #16159·danielhanchen opened Oct 30, 2026Member·

  ·

• Studio: carry FP8 / NVFP4 compile cache tags into render and compile threads

  #16158·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• Studio installer: pin the resolved unsloth on the no-torch --no-deps path

  #16157·danielhanchen opened Oct 30, 2026Member·

  ·
  7

• FP8: strip activation_scale sources when matching the 16-bit save conversion

  #16156·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• compressed-tensors: keep offload wrappers when the model spans several devices

  #16155·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• Studio diffusion prequant cache: do not count cgroup shmem as reclaimable

  #16154·danielhanchen opened Oct 30, 2026Member·

  ·
  3

• Studio fine-tuning API: refuse OpenAI training files with per-message weights

  #16153·danielhanchen opened Oct 30, 2026Member·

  ·
  8

• Studio video: load cached weights offline while the Hub is unreachable

  #16152·danielhanchen opened Oct 30, 2026Member·

  ·
  9