All pull requests

Pull requests

• docs: update coding-agent example to the revised MiMo dataset

  #2513·Syycc-code opened Nov 1, 2026·

  ·

• fix(launcher): quote Ray runtime environment JSON

  #2512·chaoworks opened Oct 31, 2026Contributor·

  ·

• fix(rollout): release timestamps for consumed buffer groups

  #2511·GoldArowana opened Oct 31, 2026Contributor·

  ·

• fix(reward): preserve large integer precision

  #2510·lll-hhh opened Oct 30, 2026·

  ·
  1

  4

• fix(ppo): reduce cancellation in k3 KL estimates

  #2507·LingYi-Liang opened Oct 29, 2026·

  ·

• [ROCm] Support ROCm 7.15 on MI300/MI325

  #2506·jun-amd opened Oct 29, 2026·

  ·

• fix: retain GPU locks while canceling command groups

  #2505·bvolpato opened Oct 28, 2026·

  ·

• test: reuse the single-rank whitening reference

  #2504·bvolpato opened Oct 28, 2026·

  ·

• ci: batch CPU tests to reuse dependency setup

  #2503·bvolpato opened Oct 28, 2026·

  ·

• test: reuse Gloo workers across metric reductions

  #2502·bvolpato opened Oct 28, 2026·

  ·
  1

• test: reuse Gloo workers in score-centering checks

  #2501·bvolpato opened Oct 28, 2026·

  ·
  1

• Send the generated response when continuing a multimodal partial rollout

  #2500·apex-mochen opened Oct 28, 2026·

  ·

• perf(codec): 4.6x faster SampleCodec.prepare token-list encoding

  #2499·dexhunter opened Oct 27, 2026·

  ·

• [Fix] Keep DFLASH draft weights out of colocated weight updates

  #2498·0z5a opened Oct 27, 2026·

  ·

• Add Trustabl agent reliability check to CI

  #2496·Trustabl-Je opened Oct 27, 2026·

  ·

• fix(data): give uniform reward groups exactly zero advantage

  #2494·saloni-garg opened Oct 23, 2026·

  ·
  2

• fix(rm_hub): score the GPQA answer the reply states

  #2492·salma2vec opened Oct 23, 2026·

  ·
  1

  1

• fix(ppo): mask env/tool tokens out of the reward KL penalty

  #2490·Domin-Focus opened Oct 23, 2026·

  ·

• perf: cache DeepGEMM FP8 weights within training steps

  #2488·ggboooy opened Oct 21, 2026·

  ·

• perf(ppo): share sampled-token and score-centering normalization

  #2487·ActuallyEdward opened Oct 20, 2026·

  ·

• fix(rm_hub): accept numeric zero ground truth in grade_answer_verl

  #2486·Domin-Focus opened Oct 20, 2026·

  ·

• Add P2P weight synchronization for disaggregated Megatron and SGLang

  #2485·zhuzilin opened Oct 19, 2026Contributor·

  ·

• feat: add local training performance benchmarks

  #2481·Shomvel opened Oct 19, 2026·

  ·

• Add Qwen3-30B FP8-experts deterministic stack (DeepGEMM, residual checkpoint, sampling seeds)

  #2472·fy1214 opened Oct 15, 2026Contributor·4 tasks·

  ·

• feat: support NVIDIA Nemotron 3.5 Super VL (GRPO on GEO3K, Megatron backend)

  #2471·shifangx opened Oct 14, 2026·

  ·