All pull requests

Pull requests

• preserve checkpoint publication and resumed dataloader state

  #2346·coreyhu opened Sep 30, 2026·

  ·
  1

• fix(skycap): bridge token-equivalent assistant replays

  #2345·cbanbury-lila opened Sep 30, 2026·

  ·

• [train] Fold the VLM generator into SkyRLGymGenerator's token-in-token-out loop

  #2344·xinyuangui2 opened Sep 30, 2026Contributor·

  ·

• [skycap] R3 through skycap: step-wise rows carry their own routes

  #2343·kouroshHakha opened Sep 30, 2026Collaborator·

  ·
  4

• [skycap] Harbor: per-step trial phase and retry metrics

  #2342·kouroshHakha opened Sep 29, 2026Collaborator·

  ·
  6

• [sft] Enable GLM-5.3-Flash full fine-tuning in SFT (Megatron)

  #2341·avigyabb opened Sep 29, 2026Collaborator·

  ·

• [skyrl-train] Fix RL loss/gradient normalization across token-packed microbatches

  #2340·dyurk-lila opened Sep 29, 2026Contributor·

  ·
  6

• fix(megatron): pad per-sample TensorList fields when padding token-based microbatches

  #2339·dyurk-lila opened Sep 29, 2026Contributor·

  ·
  3

• fix(megatron): make overlap_grad_reduce work when the microbatch count changes between steps

  #2338·dyurk-lila opened Sep 29, 2026Contributor·

  ·
  4

• fix(megatron): gather updated parameters before weight transfer, HF export and checkpoint save

  #2337·dyurk-lila opened Sep 29, 2026Contributor·

  ·
  3

• fix(rdt): re-apply composed-loader transforms on replay (Mamba2 A = -exp(A_log))

  #2336·dyurk-lila opened Sep 29, 2026Contributor·

  ·
  6

• [skycap] Upload each step's skycap documents to W&B

  #2333·jlee-lila opened Sep 29, 2026Contributor·

  ·
  8

• [r3] Allow router replay on the vLLM ray executor; refuse inference pipeline parallelism

  #2332·erictang000 opened Sep 29, 2026Collaborator·

  ·
  4

• [megatron] Align mixed-dtype distributed-optimizer checkpoint indices

  #2331·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Gather updated params before weight sync, HF export and checkpoint save

  #2330·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Publish mHC projection weights before functional reads under overlap_param_gather

  #2329·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Add freeze_dsa_indexer; freeze through HyperConnection-wrapped layers

  #2328·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Align the cuDNN DSA backward's top-k axis to its 64-wide tile

  #2327·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Compute mixed-dtype gradient norms per dtype group

  #2326·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Pad ragged TensorList fields in _pad_microbatch_to_size

  #2325·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [megatron] Resolve the MoE decoder inside multimodal wrappers in freeze_moe_router

  #2324·erictang000 opened Sep 29, 2026Collaborator·

  ·

• [fsdp] Add lora.base_dtype to store a frozen LoRA base in bf16

  #2323·gc-anyscale opened Sep 29, 2026·

  ·
  8

• [vlm 1/n] Fix silent VLM correctness bugs (eos doubling, LoRA on vision tower, mixed batches, SFT image_url, num_epochs, ckpt_interval)

  #2322·xinyuangui2 opened Sep 29, 2026Contributor·

  ·
  9

• [fix][fsdp] Make manual model CPU offload blocking

  #2320·AlirezAkbary opened Sep 29, 2026·

  ·
  1

• fix(InferenceEngineClient): Close leaking async sessions after asyncio run

  #2319·kyuds opened Sep 28, 2026Collaborator·

  ·
  3