All pull requests

Pull requests

• feat(sft): support single-GPU training resume

  #648·fusang991 opened Oct 31, 2026·6 of 13 tasks·

  ·

• feat(optim): add GaLore AdamW with optional deferred gradient projection

  #647·ymping666 opened Oct 31, 2026·7 of 14 tasks·

  ·
  2

• feat(engine): add opt-in adaptive compression of saved activations

  #646·EricChan2028 opened Oct 30, 2026·7 of 13 tasks·

  ·
  1

• feat(engine): add opt-in KIVI 2/4-bit KV cache for CUDA rollout

  #641·EricChan2028 opened Oct 30, 2026·7 of 13 tasks·

  ·
  1

• feat(optim): add opt-in GaLore AdamW for single-GPU full-parameter training

  #640·EricChan2028 opened Oct 30, 2026·7 of 13 tasks·

  ·
  1

• feat(runtime): share prompt KV across rollout samples

  #639·fusang991 opened Oct 29, 2026·6 of 13 tasks·

  ·
  1

  1

• docs: use Huawei Ascend in README

  #627·p40www opened Oct 23, 2026·2 of 7 tasks·

  ·

• feat(lora): add opt-in LoRA-FA to reduce adapter activation memory

  #622·Dayuxiaoshui opened Oct 22, 2026Contributor·7 of 13 tasks·

  ·
  1

• Feat/opd_implementation

  #609·yuetongz417 opened Oct 20, 2026·

  ·

• perf(engine): project MTP loss logits in recomputed chunks

  #600·Dayuxiaoshui opened Oct 18, 2026Contributor·7 of 13 tasks·

  ·

• feat(dashboard): visualize training execution and learning signals

  #599·xsuler opened Oct 14, 2026Collaborator·5 of 13 tasks·

  ·
  

• feat(classify): support JevForge-style decision scoring end to end

  #575·tangxinyao opened Sep 26, 2026·6 of 7 tasks·

  ·

• feat(serve): add token-level streaming and reasoning_content to /v1/chat/completions

  #574·tangxinyao opened Sep 22, 2026·6 of 13 tasks·

  ·

• feat(engine): FP8 decode weight quantization for train & serve

  #562·CAICAIIs opened Sep 15, 2026Contributor·

  ·

• fix(mlx): support mlx_lm 0.31 stop-sequence API and add response to math SFT loader

  #561·Kelei0215 opened Sep 15, 2026·

  ·
  4

• feat(lora): support Flash V3 and expand native model compatibility

  #556·indevn opened Sep 10, 2026Collaborator·

  ·

• feat(experimental): add world-model VLA workflow

  #554·matianyun opened Sep 8, 2026·5 tasks done·

  ·

• feat: add competition agentic feedback example

  #501·feizhenxuan opened Aug 21, 2026·13 tasks·

  ·

• feat(experimental): add DAPO training support

  #486·CharlesXu-HQ opened Aug 15, 2026·8 of 13 tasks·

  ·
  1

  2

• feat(experimental): add Dr. GRPO

  #484·CharlesXu-HQ opened Aug 14, 2026·7 of 13 tasks·

  ·
  1

  2

• feat: add async policy trainer primitives

  #480·VLa1111 opened Aug 14, 2026·7 of 13 tasks·

  ·
  1

  1

• docs(agentic): define trajectory observability contract

  #472·YAO-001 opened Aug 12, 2026Contributor·7 tasks done·

  ·
  1

  4

• build: adopt file-backed dynamic package version

  #471·YAO-001 opened Aug 12, 2026Contributor·6 of 13 tasks·

  ·
  1

  1

• docs: define greppability conventions for new seams

  #469·remember993 opened Aug 12, 2026·

  ·
  1

• fix(sft): expose filtered row reasons

  #468·SRiftX opened Aug 11, 2026·

  ·
  1