All pull requests

Pull requests

• feat(checkpoint): support sharded generation-prefix recovery

  Documentation
  #4323·macandro96 opened Sep 29, 2026Contributor·8/8·

  ·
  3

• feat(checkpoint): support sharded Gym turn recovery

  Documentation
  #4322·macandro96 opened Sep 29, 2026Contributor·7/8·

  ·
  1

• perf(checkpoint): reduce Gym control-plane overhead

  #4321·macandro96 opened Sep 29, 2026Contributor·6/8·

  ·
  2

• fix(checkpoint): harden generation-prefix recovery

  #4320·macandro96 opened Sep 29, 2026Contributor·5/8·

  ·
  1

• feat(checkpoint): add generation-prefix recovery

  #4319·macandro96 opened Sep 29, 2026Contributor·4/8·

  ·
  2

• refactor(dtensor): drop the mesh-derived CP fallback from the x-token path

  CI:Lfast
  #4318·yuki-97 opened Sep 29, 2026Contributor·

  ·
  7

• fix(vlm): add joint MOPD video support

  #4316·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• fix(vlm): load frozen vision weights and disable unused MTP

  #4315·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• fix(vlm): align training media with rollout placeholders

  #4314·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• feat(vlm): port mixed SA-V and CapRL manifests to main

  #4313·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• fix(vllm): align RADIO LayerScale patch with Super loader

  #4312·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• fix(vlm): recognize Super Omni placeholder processor

  #4311·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• fix(vlm): preserve cached video source timing

  #4310·ehosseiniasl opened Sep 28, 2026Contributor·

  ·
  1

• fix(gym): preserve cached video timing from JSONL metadata

  #4309·goelarushi opened Sep 28, 2026·

  ·
  1

• pjin/main-dev-20260928

  Documentation
  #4308·pjin-nvidia opened Sep 28, 2026Contributor·

  ·
  3

• feat(rollout): replay MInf routing indices

  Documentation
  #4307·lauradang opened Sep 28, 2026Contributor·

  ·
  1

  

• test: exercise Gym checkpoint recovery with CPU mocks

  #4306·terrykong opened Sep 28, 2026Collaborator·

  ·
  1

• feat(energon): add external encoder plugins and precomputed masks

  Documentation
  #4296·rohitrango opened Sep 28, 2026Contributor·3/3·

  ·
  1

• feat(data): support preformatted SFT tokenization

  #4295·rohitrango opened Sep 28, 2026Contributor·2/3·

  ·
  1

• feat(sft): add loader and trainer telemetry

  #4294·rohitrango opened Sep 28, 2026Contributor·1/3·

  ·
  1

• feat(megatron): register attention packing metadata and scope scratch

  community-requestDocumentation
  #4293·layalir opened Sep 28, 2026·

  ·
  2

• fix(rewards): do not treat empty alphanumeric answers as matches

  community-requestwaiting-on-customer
  #4292·rwinkelman opened Sep 28, 2026·

  ·
  1

  2

• feat(megatron): Pass VL related config fields to MCore

  CI:Lfast
  #4288·guyueh1 opened Sep 28, 2026Contributor·

  ·
  2

• fix(megatron): match Megatron-LM loss normalization

  #4287·guyueh1 opened Sep 28, 2026Contributor·

  ·
  14

• feat(megatron): share prefixes across packed rows

  #4286·cmunley1 opened Sep 27, 2026Contributor·2/2·

  ·
  1