All pull requests

Pull requests

• [megatron] Bump megatron-bridge to 7171584b and megatron-core to ed9c2c8c; drop vendored k-pool DSA indexer

  #2558·erictang000 opened Oct 27, 2026Collaborator·

  ·
  6

• Fix main GPU CI test failures

  run_megatron_gpu_ci
  #2556·SumanthRH opened Oct 27, 2026Member·

  ·
  6

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

  #2555·ArturNiederfahrenhorst opened Oct 27, 2026Collaborator·

  ·
  1

  12

• [vllm] Zero DeepSeek-V4.1's BF16 FlashMLA prefill KV workspace (NaN first tokens)

  #2551·avigyabb opened Oct 27, 2026Collaborator·

  ·
  5

• [trainer] Fire training callbacks from the fully async trainer and the dummy SFT loop

  #2550·SumanthRH opened Oct 26, 2026Member·

  ·
  1

• [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·

  ·

• [vllm] Allocate a pipeline stage's KV cache only for the layers it owns

  #2545·avigyabb opened Oct 26, 2026Collaborator·

  ·
  5

• [megatron] Fix the distributed optimizer for param groups mixing fp32 and bf16

  #2539·avigyabb opened Oct 26, 2026Collaborator·

  ·
  6

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

  #2538·avigyabb opened Oct 26, 2026Collaborator·

  ·

• [ci] SkyRL-JAX-GPU: skip fork PRs on the plain trigger, add run_jax_gpu_ci label run

  #2535·ArturNiederfahrenhorst opened Oct 26, 2026Collaborator·

  ·
  5

• [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

  #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

• [fix] Stop Gated DeltaNet state leaking between packed sequences

  #2526·anthonypetertc opened Oct 22, 2026·

  ·
  17

• [megatron] Fix OOM on native checkpoint save and resume

  x-lila
  #2525·sgabriel92 opened Oct 22, 2026·

  ·
  9

• [skyrl-train][bugfix] Run the sampler weight broadcast off the event loop

  x-lila
  #2524·dyurk-lila opened Oct 21, 2026Collaborator·

  ·
  8

  

• [sft] Count context parallelism in the batch-size check

  x-lila
  #2523·dyurk-lila opened Oct 21, 2026Collaborator·

  ·
  1

• [fix][megatron] Call mark_not_offload in the fused LM head path

  x-lila
  #2522·dyurk-lila opened Oct 21, 2026Collaborator·

  ·
  4

  

• perf(megatron): fuse sample-support candidate projection with Triton

  x-lila
  #2521·dyurk-lila opened Oct 21, 2026Collaborator·5/5·

  ·
  6

• [tinker] Reject backend_config values the server overrides on the SkyRL-Train backends

  #2516·ArturNiederfahrenhorst opened Oct 20, 2026Collaborator·

  ·
  1

  14

  

• Add opt-in policies for unbridged Harbor skycap rollouts

  #2514·jadenfix opened Oct 19, 2026Contributor·

  ·
  1

• [megatron][vllm] Qwen3.8-Flash-Next (qwen4_exp) support: Megatron bridge, QSA sparse attention, GSM8K + DAPO recipes

  #2513·xinyuangui2 opened Oct 17, 2026Collaborator·10 of 13 tasks·

  ·
  5

• [megatron][vlm] freeze_language_model / freeze_vision_model / freeze_vision_projection

  #2509·gc-anyscale opened Oct 16, 2026Contributor·

  ·
  18