Pull requests

• 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·

  ·

• Keep a router's existing registration when an engine registers again

  #2483·key4ng opened Oct 19, 2026·

  ·
  1

• feat: add local training performance benchmarks

  #2481·Shomvel opened Oct 19, 2026·

  ·

• fix(coding_agent_rl): snapshot fails when repo .gitignore ignores .harness

  #2473·dextermayhewjd opened Oct 15, 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·

  ·

• test: disable git auto-maintenance in swe workspace diff test

  #2469·gavinqii opened Oct 13, 2026·

  ·

• fix(tau-bench): train on sampled token ids instead of retokenized chat text

  #2467·PlanckChang opened Oct 12, 2026·6 tasks done·

  ·

• Add opt-in R3 route-mismatch metric and targeted R3 CI checks

  #2463·kar137 opened Oct 11, 2026·

  ·

• fix(data): preserve empty datasets during prompt length filtering

  #2462·GoldArowana opened Oct 11, 2026Contributor·

  ·

• perf: fuse FP32 CUDA updates in StatelessAdam

  #2461·CharlesXu-HQ opened Oct 11, 2026·

  ·

• fix(ray): bind trainer CPU affinity to the selected GPU

  #2460·DongDongJu opened Oct 11, 2026·

  ·

• test: prevent Git maintenance race in SWE workspace fixture

  #2459·GoldArowana opened Oct 10, 2026Contributor·

  ·

• fix(qwen3.5): average TP-replicated linear-attention gradients across TP

  #2451·gavinqii opened Oct 8, 2026·

  ·
  1

• fix: filter_long_prompt crashes on chat-templated multimodal prompts

  #2449·annika-modal opened Oct 7, 2026Contributor·

  ·

• feat(agent): select the rollout sandbox backend by name or import path

  #2447·brandonrjacobs opened Oct 7, 2026·

  ·

• fix(grpo): normalize rewards per query and unique rollout

  #2446·huanhuan6666 opened Oct 6, 2026·

  ·
  1

• Fix Qwen3.5 image positions in SGLang patch bundles

  #2441·1fanwang opened Oct 3, 2026·1 task·

  ·
  1

• Add Megatron LoRA support for actor training with dense and MoE models

  #2440·Easily2 opened Oct 3, 2026·

  ·

• docs: fix Qwen3-4B example link in rollout_buffer READMEs

  #2439·pratikgx opened Oct 2, 2026·

  ·

• fix(rollout): preserve weight versions on stream abort

  #2438·Beverly621 opened Oct 2, 2026Contributor·

  ·

• fix(data): address optional Straw and trusted top-p validation regressions

  #2436·Beverly621 opened Oct 2, 2026Contributor·

  ·
  1