All pull requests

Pull requests

• feat(mlx): native GRPO on Apple Silicon, dense baseline (#1976)

  #2182·dryn85-max opened Oct 24, 2026Contributor·

  ·
  1

• test(sft): verify effective vision and audio checkpointing

  ci:full
  #2180·SauravMukherjee44 opened Oct 24, 2026·4 of 5 tasks·

  ·
  1

  2

• test: pin atomic writers to the destination directory

  ci:full
  #2179·SauravMukherjee44 opened Oct 24, 2026·4 of 5 tasks·

  ·
  1

  1

• fix(cloud): support local training input uploads (#1704)

  #2177·Hitesh-XS opened Oct 24, 2026·6 of 10 tasks·

  ·
  1

• test(benchmarks): publish convergence data preparation

  #2172·kok-o opened Oct 24, 2026Contributor·7 tasks done·

  ·

• feat(ship): add versioned JSONL progress

  #2171·kok-o opened Oct 24, 2026Contributor·4 tasks done·

  ·

• fix(monitoring): contain and atomically replace spike recovery hints

  ci:full
  #2169·Nari-Ab opened Oct 24, 2026Contributor·

  ·

• fix(config): refuse vision row formats on kto and reward_model tasks (closes #2078)

  #2168·Tanmay9223 opened Oct 24, 2026Contributor·7 of 10 tasks·

  ·
  1

• fix(export): clean up temporary merge directory when downstream steps fail (Refs #1464)

  #2164·ShivanshShukla opened Oct 23, 2026Contributor·

  ·
  1

• perf: measure selective attention checkpointing for streamed NF4

  ci:full
  #2161·Shutaru opened Oct 23, 2026Contributor·8 of 9 tasks·

  ·
  1

  2

• fix(streaming): save and resume boundary LoRA adapters

  ci:full
  #2159·kok-o opened Oct 23, 2026Contributor·3 tasks done·

  ·

• feat(stream): Read a streamed routed expert twice per step instead of three times

  #2158·Batyrkhantzy111 opened Oct 22, 2026Contributor·8 of 10 tasks·

  ·
  4

• fix(stream): build fused MoE experts on meta without a CPU allocation first (#2154)

  ci:full
  #2157·ybai08 opened Oct 22, 2026Contributor·5 of 10 tasks·

  ·
  1

  3

• fix(bench): run each stream_loss_chunk_vram arm in a process of its own (#1693)

  #2151·ybai08 opened Oct 22, 2026Contributor·6 of 10 tasks·

  ·
  3

• feat(vram): one peak-VRAM estimate for the train pre-flight, soup profile, MCP profile and soup plan

  #2146·abdulwaarith0 opened Oct 22, 2026Contributor·

  ·
  3

• fix(profile): know the RTX 30-series laptop GPUs

  ci:full
  #2144·abdulwaarith0 opened Oct 22, 2026Contributor·

  ·
  2

• Add export behavior verification for merged LoRA models

  #2143·CODING-DARSH opened Oct 22, 2026Contributor·3 of 11 tasks·

  ·

• docs(training): clarify curriculum_dynamic telemetry contract (Refs #808)

  #2142·Nari-Ab opened Oct 22, 2026Contributor·

  ·
  1

• fix(adapters): require numpy and safetensors for adapter commands (#2056)

  #2141·Nari-Ab opened Oct 22, 2026Contributor·

  ·
  1

  4

• feat(mlx): add native token distillation (#1979)

  #2139·Amix29 opened Oct 22, 2026Contributor·

  ·
  1

• feat(config): add alias table and suggestion ranking for foreign parameter names (#2099)

  ci:full
  #2138·Nari-Ab opened Oct 22, 2026Contributor·

  ·
  1

  1

• feat(mlx): add full-state LoRA SFT resume (#1977)

  #2137·Amix29 opened Oct 22, 2026Contributor·

  ·
  1

• feat(mlx): add projected LoRA-FA with complete adapter checkpoints

  #2135·Amix29 opened Oct 22, 2026Contributor·

  ·
  1

• feat(mlx): add native LoRA+ A/B AdamW groups

  #2134·Amix29 opened Oct 21, 2026Contributor·

  ·
  1

• feat(mlx): honor per-module LoRA rank and alpha patterns (#1996)

  #2133·Amix29 opened Oct 21, 2026Contributor·

  ·
  1