All pull requests

Pull requests

• fix: preserve per-sample training metadata across DP transport

  #2403·heyxingxing opened Sep 22, 2026·

  ·
  1

• fix: propagate fully async worker failures to training

  #2401·heyxingxing opened Sep 22, 2026·

  ·
  1

• refactor: migrate Slime MUSA bootstrap to musa_titan

  #2398·ForAxel opened Sep 21, 2026·

  ·

• feat(profile): safer torch.profiler defaults (rank-0 only + env-gated metadata)

  #2397·leofan-i opened Sep 21, 2026·

  ·

• fix(args): reject a non-positive --eval-temperature

  #2396·PerryLink opened Sep 20, 2026·

  ·

• fix: isolate dataset shuffle random state

  #2395·LZC-BELIEVER opened Sep 19, 2026·

  ·

• fix(vlm): preserve token IDs in multi-turn Qwen rollouts

  #2389·zupengwang opened Sep 16, 2026·

  ·
  1

• fix(R3): reduce host memory usage for rollout routing replay

  #2385·JayPritchet opened Sep 14, 2026·

  ·
  1

• feat: add fused Triton log-probability backend for Megatron

  #2383·YolandaLyj opened Sep 14, 2026·

  ·

• fix: add an opt-out for final checkpoint saves

  #2382·NOAH1ARK opened Sep 13, 2026·

  ·
  1

• fix(rollout): handle fan-out dynamic filter groups

  #2381·QIU-Guanzong opened Sep 13, 2026·

  ·
  1

• fix: restore actor weights before initial rollout sync

  #2380·tengeel opened Sep 12, 2026·

  ·
  1

• feat(nemotron_h): native support for the Nemotron-3 hybrid stack

  #2378·shifangx opened Sep 11, 2026·

  ·

• feat(sft): a VL SFT rollout, so a -vl model config can be fine-tuned

  #2377·shifangx opened Sep 11, 2026·

  ·

• fix(qwen3.5-vl): skip the CP index mapping when there is no context parallelism

  #2376·shifangx opened Sep 10, 2026·

  ·

• fix: place PPO terminal rewards on the owning CP rank

  #2375·MrCapricornLiu opened Sep 9, 2026·

  ·

• fix(rollout): measure each engine's host instead of deriving it from its rank

  #2373·kaysonyu opened Sep 9, 2026·

  ·

• fix: make GenerateState.semaphore a lazy property to avoid event loop mismatch between train and eval

  #2372·LucQueen opened Sep 9, 2026·

  ·

• fix(update-weight): release rollout engine lock when distributed transfer fails

  #2371·wangxianyu0825 opened Sep 9, 2026·

  ·

• [ROCm] Update and unify ROCm support for gfx942 and gfx950

  #2369·LZ-QWQ opened Sep 8, 2026·

  ·
  1

• fix(args): reject --kl-coef with GRPO-family advantage estimators

  #2368·QJSQJS opened Sep 8, 2026·

  ·

• fix: use group-level concurrency in fully-async rollout

  #2367·JayPritchet opened Sep 7, 2026·

  ·

• fix(train): raise clear error when computed train_iters is 0

  #2366·Xuan-1998 opened Sep 6, 2026·

  ·

• fix(data): fill sample requests across multiple epochs

  #2365·feng1201 opened Sep 6, 2026·

  ·

• [docs] debug guide: tell a cold Triton compile from a colocate hang (en + zh)

  #2364·David553 opened Sep 6, 2026·

  ·