All pull requests

Pull requests

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

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

  ·
  12

• [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·

  ·
  13

• [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·

  ·

• [generator] Optional per-trajectory prefix-cache salt (generator.cache_salt_per_trajectory)

  #2568·avigyabb opened Oct 27, 2026Collaborator·

  ·
  10

• [infra] Forward VLLM_FLASHINFER_AUTOTUNE_CACHE_DIR to the Ray runtime env

  #2564·avigyabb opened Oct 27, 2026Collaborator·

  ·
  4

• [r3] Map HybridModel sub-layer indices onto rollout blocks for router replay

  #2562·avigyabb opened Oct 27, 2026Collaborator·

  ·
  10

• [megatron] chunked_mlp: router replay under selective "moe" recompute; sync with Megatron-LM#8463

  #2559·xinyuangui2 opened Oct 27, 2026Collaborator·

  ·

• [megatron] Bump megatron-bridge to 7171584b and megatron-core to ed9c2c8c; retire landed patches

  run_h100_gpu_cirun_megatron_gpu_ci_modelsrun_megatron_gpu_ci
  #2558·erictang000 opened Oct 27, 2026Collaborator·5 of 6 tasks·

  ·
  6

• [skyrl-train] Release GPUs when create_model fails part-way

  #2555·ArturNiederfahrenhorst opened Oct 27, 2026Collaborator·

  ·
  1

  12

  

• [megatron] DeepSeek-V4.1-Flash (text only): Megatron + vLLM training support

  #2547·avigyabb opened Oct 26, 2026Collaborator·

  ·

• [megatron] DeepSeek-V4-Flash: Megatron + vLLM training support

  #2546·avigyabb opened Oct 26, 2026Collaborator·

  ·
  12

• [inference] Survive and count non-finite sampled logprobs

  #2538·avigyabb opened Oct 26, 2026Collaborator·

  ·
  8

• [fix][train] Give uniform reward groups exactly zero GRPO advantage

  #2531·saloni-garg opened Oct 23, 2026·

  ·
  4

• [tinker] Honour train_attn/train_mlp on the Megatron backend

  run_tinker_skyrl_backend_gpu_ci
  #2529·ArturNiederfahrenhorst opened Oct 22, 2026Collaborator·

  ·
  1

  8

  

• [megatron] GLM-5.3-Flash long-context integration branch (all optimizations, not for merge)

  #2528·avigyabb opened Oct 22, 2026Collaborator·

  ·

• feat(train): time optimizer steps separately from forward and backward

  x-lila
  #2527·dyurk-lila opened Oct 22, 2026Collaborator·

  ·
  1