All pull requests

Pull requests

• [NPU] Restore Qwen3.5-35B-A3B test coverage

  #455·CalvinXKY opened Sep 27, 2026Collaborator·

  ·
  5

• [Feature][Training] Collect versioned target features for external draft

  #454·0z5a opened Sep 27, 2026·

  ·
  6

• [Feature] Define TransferQueue group and lease contract

  #453·0z5a opened Sep 27, 2026·

  ·
  5

• [Feature] Define versioned external draft feature contract

  #452·0z5a opened Sep 27, 2026·

  ·
  4

• [Training] Add parameter-wise gradient diagnostics to parallel replay tests

  #451·Levius-Fubuki opened Sep 26, 2026·

  ·
  3

• [Training] Reuse chunked DSpark probability distance for loss and confidence

  #450·Levius-Fubuki opened Sep 26, 2026·

  ·
  4

• [Training] Add NVMe optimizer streaming and variable global batches

  #449·inaniloquentee opened Sep 24, 2026·

  ·
  11

• feat(ascend): add MXFP8 rollout support

  #448·PierceZhou opened Sep 24, 2026·

  ·
  5

• [Feature][Ascend] Support sparse HCCL weight synchronization from Megatron training workers

  #447·wangx700 opened Sep 22, 2026·

  ·
  11

• fix(npu): restore RoPE cos/sin cache after colocate weight update

  #446·Bobfan opened Sep 19, 2026·

  ·
  3

• [Docs] --offload-train is not forced by --colocate, and skipping it took a step from 18.85 s to 3.32 s

  #445·0z5a opened Sep 19, 2026·

  ·
  1

• [Docs] Record the measured cost of the colocate CPU backup and link the upstream report

  #444·0z5a opened Sep 19, 2026·

  ·
  1

• [Observability] Add opt-in spans for the colocate offload/onload boundary

  #443·0z5a opened Sep 19, 2026·

  ·
  4

• [Sync] Update from Slime through #2407

  #441·aoshen02 opened Sep 19, 2026Collaborator·

  ·
  6

• [Bugfix][Rollout] Report engine drain state after an abort

  #440·0z5a opened Sep 18, 2026·

  ·
  5

• [Bugfix][Rollout] Preserve per-request transfer metrics through MultiConnector

  #439·0z5a opened Sep 18, 2026·

  ·
  1

• [Bugfix][Examples] Handle zero NVLink matches without masking topology query errors

  #438·0z5a opened Sep 18, 2026·

  ·
  3

• [Observability] Add opt-in spans for the RolloutManager post-rollout boundary

  #437·0z5a opened Sep 18, 2026·

  ·
  3

• [CI/Build][Hardware][AMD] Rebuild the ROCm image on vllm/vllm-openai-rocm:nightly

  #432·indianspeedster opened Sep 16, 2026Collaborator·

  ·
  5

• feat(ascend): add disk checkpoint serving patches

  #431·wangx700 opened Sep 16, 2026·

  ·
  7

• [Bugfix][Training] Fix native Adam checkpoint resume and add parity checks

  #427·0z5a opened Sep 12, 2026·

  ·
  5

• [Bugfix][Rollout] Reject missing or misaligned sampled-token log probabilities

  #426·0z5a opened Sep 12, 2026·

  ·
  3

• feat(modelexpress): add S3 delta weight update backend

  #425·nv-hwoo opened Sep 12, 2026·

  ·
  11

• [Bugfix][Training] Preserve DSpark rotary embedding dtype

  #420·LOGO127 opened Sep 10, 2026Contributor·

  ·
  7

• [Bugfix][Training] Preserve FP32 PPO entropy reduction under CUDA autocast

  #419·LOGO127 opened Sep 10, 2026Contributor·

  ·
  3