All pull requests

Pull requests

• [ci] fix: set explicit model paths for Qwen3-30B nightly jobs

  #8165·lxb007981 opened Oct 9, 2026Contributor·2 tasks done·

  ·

• [doc, ci] feat: add Ascend tutorial doc translation pipeline

  #8164·Geofeng12138 opened Oct 9, 2026·

  ·
  1

• [ci] fix: fix ascend ci

  Ascend
  #8162·wucong25 opened Oct 9, 2026Collaborator·8 tasks·

  ·
  1

• [ci] chore: use a3-8 runner for nightly ppo qwen3-8b fsdp vllm test

  Ascend
  #8161·lxb007981 opened Oct 9, 2026Contributor·2 tasks done·

  ·

• verl supprot rollout trace

  #8158·mengchengTang opened Oct 9, 2026Contributor·8 tasks·

  ·

• [tool, rollout] fix: create BaseTool instances once per trajectory and release them when it ends

  #8156·nataliekung opened Oct 8, 2026·5 of 7 tasks·

  ·

• [hardware] feat: enable TPU GRPO with TorchTitan and vLLM via verl-hardware-plugin

  #8155·lixali opened Oct 8, 2026·2 tasks done·

  ·
  3

• [tool] test: cover reasoning-bearing chat-template trajectories

  #8153·JiahaoTanXX opened Oct 8, 2026·9 of 10 tasks·

  ·
  1

• [fsdp] fix: fail fast when LoRA parameter collection is empty

  #8152·lxb007981 opened Oct 8, 2026Contributor·2 tasks done·

  ·
  2

• [ci, fsdp] fix: make the FSDP rmpad checkpoint test deterministic and actually verify the load

  #8147·MincongZhou opened Oct 8, 2026·5 of 6 tasks·

  ·
  1

• [draft] [worker, hardware] fix: pin each Ray worker's accelerator in both vis…

  #8145·kahlun opened Oct 8, 2026Contributor·6 of 8 tasks·

  ·

• [hardware, model, cfg, tests] feat: Ascend CANN fused linear cross-entropy

  Ascend
  #8143·RitianD opened Oct 8, 2026·5 of 8 tasks·

  ·
  1

  4

• fix(data): make val max_samples subsample follow validation_shuffle

  #8141·stanleyIl opened Oct 8, 2026·

  ·
  1

  1

• [vllm, hardware] fix: only skip causal mask for genuine decode on vLLM-Ascend 0.23

  #8140·fangweii1 opened Oct 8, 2026Contributor·

  ·
  1

• [training_utils, perf] refactor: cache Karmarkar-Karp state spread

  #8137·IsaacLi74 opened Oct 7, 2026·7 of 8 tasks·

  ·
  1

• [rollout, vllm, sglang, perf] feat: prioritize resumed agent requests

  #8136·jiacao-amd opened Oct 7, 2026·5 of 7 tasks·

  ·
  1

• [7/7] [rollout] feat: audit that training and rollout quantize the same layers

  #8132·wengeezhang opened Oct 7, 2026Contributor·6 of 7 tasks·

  ·

• [6/7] [rollout] feat: self-check MXFP8 kernel layouts after every weight sync

  #8131·wengeezhang opened Oct 7, 2026Contributor·6 of 7 tasks·

  ·

• [5/7] [rollout, sglang] feat: MXFP8 SGLang rollout with a refit loader

  #8130·wengeezhang opened Oct 7, 2026Contributor·6 of 7 tasks·

  ·

• [4/7] [rollout, vllm] feat: MXFP8 vLLM rollout on the learner's weight grid

  #8129·wengeezhang opened Oct 7, 2026Contributor·6 of 7 tasks·

  ·

• [2/7] [trainer] feat: fail loudly when a quantized rollout produces garbage

  #8128·wengeezhang opened Oct 7, 2026Contributor·6 of 7 tasks·

  ·

• [1/7] [megatron] feat: MXFP8 training recipe and shared TE weight quantizer

  #8127·wengeezhang opened Oct 7, 2026Contributor·6 of 7 tasks·

  ·

• fix: guard against re-conversion of already-nested teacher_logprobs in rmpad preprocessing

  #8125·NGDBZ opened Oct 7, 2026·

  ·

• [reward] fix: normalize the boxed fallback answer in math_dapo

  #8124·KardeniaPoyu opened Oct 6, 2026·5 of 7 tasks·

  ·
  1

  2

• [draft] Feat/xpu real fp8 training

  #8121·kahlun opened Oct 6, 2026Contributor·

  ·