All pull requests

Pull requests

• ci: guard shared checkpoint caches against incompatible writes

  #3909·yueming-yuan opened Oct 3, 2026Collaborator·

  ·

• Read SGLang output-store references in session sample collection

  #3908·guapisolo opened Oct 3, 2026Collaborator·

  ·

• Resolve SGLang output-store references in rollout generation

  #3907·guapisolo opened Oct 3, 2026Collaborator·

  ·

• [WIP] Fix CP backward dependency on empty response ranks

  #3906·BaiqingL opened Oct 3, 2026Contributor·

  ·

• Support independent projections in MultiLoRA

  run-ci-multi-lora
  #3905·yueming-yuan opened Oct 3, 2026Collaborator·

  ·

• Add distributional PPO critic, critic update ratio, and truncation bootstrap

  #3903·CharlesXu-HQ opened Oct 3, 2026·

  ·

• fix(ci): skip GPU runners for empty test selections

  #3902·guapisolo opened Oct 3, 2026Collaborator·

  ·
  1

• Add full-parameter training to the Tinker backend

  #3901·kailash109 opened Oct 2, 2026·

  ·

• fix: respect FP8 quantization exclusions during weight refit

  #3900·kaixih opened Oct 2, 2026Contributor·

  ·

• Support shared-outer expert LoRA in Tinker

  run-ci-multi-lora
  #3899·yueming-yuan opened Oct 2, 2026Collaborator·

  ·

• [wip]Add grouped NVFP4 QDQ

  #3898·wolfcomos opened Oct 2, 2026·

  ·

• fix(ray): publish weight version before checksum logging

  #3896·Beverly621 opened Oct 2, 2026·

  ·

• [WIP] Add gpu-delta weight publication and in-place receiver coordination

  #3894·zianglih opened Oct 2, 2026Contributor·

  ·

• fix(megatron): stop the MTP loss from training the policy when MTP training is off

  #3893·shawnghu opened Oct 2, 2026·

  ·

• Add FlashREINFORCE Recipe (Single Rollout Critic Free, Batch Mean, Seq-level Trust Region)

  #3892·billxbf opened Oct 2, 2026·

  ·

• feat(dynamo): define sidecar launch configuration (1/4)

  #3890·AndyDai-nv opened Oct 1, 2026·

  ·

• fix(ci): add the missing pytest entry to three fast-gpu tests

  #3889·XinyuJiangCMU opened Oct 1, 2026Contributor·

  ·

• feat(opd): route distillation teachers by LoRA adapter

  #3887·HJSang opened Oct 1, 2026Collaborator·

  ·

• feat: add ModelExpress S3 weight updates

  #3886·nv-hwoo opened Oct 1, 2026·

  ·

• feat(checkpoint): add Hugging Face Hub publishing during megatron training

  #3885·thibautbar opened Oct 1, 2026·

  ·

• Freeze the sparse-attention indexer, drop indexer replay, unify the selection entrypoint

  #3879·HJSang opened Sep 30, 2026Collaborator·

  ·

• fix: preserve disk delta history through the resume version

  #3875·jvmncs opened Sep 30, 2026·

  ·

• Support Tinker image inputs and preserve multimodal training tensors

  #3874·kailash109 opened Sep 30, 2026·

  ·

• feat(weight-sync): load external weight-transfer protocols

  #3868·ajcasagrande opened Sep 29, 2026·

  ·

• Move copyright notices to NOTICE; restore standard Apache-2.0 LICENSE text

  #3865·maocheng23 opened Sep 29, 2026Collaborator·

  ·
  2