All pull requests

Pull requests

• [megatron] Default fused_lm_head_logprob to auto

  #2609·erictang000 opened Nov 3, 2026Collaborator·

  ·
  5

• [train] Default enable_score_centering to auto

  #2608·erictang000 opened Nov 3, 2026Collaborator·

  ·
  4

• [megatron] Qwen3.8-Flash-Next context parallelism up to 262K

  #2607·xinyuangui2 opened Nov 2, 2026Collaborator·

  ·
  1

• [megatron] Opt-in MoE all-to-all that sends each token once per destination EP rank

  #2606·avigyabb opened Nov 2, 2026Collaborator·

  ·
  10

• [chore] Bump vLLM to 0.33.0 (torch held at 2.13); WeightSource/NCCL/NIXL fixes; skyrl-vllm-upgrade skill

  #2604·SumanthRH opened Nov 2, 2026Member·

  ·

• [fix][arctic_rl] detect limit at the top level only when capping bird reward queries

  #2603·saloni-garg opened Nov 2, 2026·

  ·
  1

  15

• [1/n][integration] Move Daytona labelling and the sandbox quota guard to skyrl/integration/sandboxes

  #2602·xinyuangui2 opened Nov 2, 2026Collaborator·

  ·
  6

• [megatron] Opt-in head-full DSA sparse attention (Ulysses-style all-to-all over TP)

  #2601·avigyabb opened Nov 2, 2026Collaborator·

  ·
  7

• perf(megatron): copy-free grouped_mm forward for per-expert LoRA adapters

  #2599·avigyabb opened Nov 2, 2026Collaborator·

  ·
  9

• refactor(weight-sync): add reusable provider controls [1/8]

  #2596·biswapanda opened Nov 1, 2026·

  ·

• [examples] Add SDPO and SRPO self-distillation example

  #2595·devpatelio opened Nov 1, 2026Collaborator·

  ·

• fix(harbor_skycap): validate rewards and honor bind hosts

  #2594·githubergsx opened Oct 31, 2026·

  ·
  1

  5

• [megatron][vllm] Routed-expert NVFP4 in trainer and sampler (humans& recipe) with serialized NVFP4 weight sync

  #2593·xinyuangui2 opened Oct 31, 2026Collaborator·

  ·

• feat(skyrl_train): native TorchNeuron (Trainium2) inference engine + FSDP trainer port

  #2592·susmit-holbox opened Oct 30, 2026·

  ·
  15

• [skyrl-train] Run the FSDP2 trainer on Cloud TPU through TorchTPU

  #2591·debjyotictr-sudo opened Oct 30, 2026·

  ·
  20

• [fsdp] Enable single-XPU training through a shared torch.accelerator wrapper

  #2590·wanhakim opened Oct 30, 2026·

  ·
  25

• [docs] New model support: check upstream before patching and match its shape

  #2588·avigyabb opened Oct 30, 2026Collaborator·

  ·
  3

• [vllm] Clamp sleep to level 1 for online NVFP4; NVFP4 rollout recipe + findings

  #2583·xinyuangui2 opened Oct 29, 2026Collaborator·

  ·

• perf(trainer): trainer.balance_dp_tokens balances tokens across DP ranks per mini-batch

  #2580·aoshen02 opened Oct 29, 2026·

  ·

• perf(lora-sync): faster colocated adapter sync and frozen-weight offload

  #2579·aoshen02 opened Oct 29, 2026·

  ·

• fix(megatron): old logprobs from the update forward, per-rank RNG, expert LoRA import and memory sync

  #2578·aoshen02 opened Oct 29, 2026·

  ·

• feat(megatron): ML4 bridge, layer spec and modules with NVFP4/FP8 QAT

  #2577·aoshen02 opened Oct 29, 2026·

  ·

• feat(fsdp): ML4 (Mistral Large 4 NVFP4) LoRA RL on FSDP2

  #2576·aoshen02 opened Oct 29, 2026·

  ·

• [DO NOT MERGE] DeepSeek-V4 / V4.1-Flash: full working branch and best tested recipes

  #2573·avigyabb opened Oct 28, 2026Collaborator·

  ·

• [tests] SKYRL_TEST_LOG_PATH overrides the GPU test helpers' log path

  #2570·avigyabb opened Oct 27, 2026Collaborator·

  ·