All pull requests

Pull requests

• docs(review): add approval guardrails and distill the PR review rubric

  #16589·pzelasko opened Nov 10, 2026Collaborator·3 of 7 tasks·

  ·
  1

• fix(speechlm2): support vLLM 0.29 multimodal processing and keep refit padding on device

  #16588·vmendelev opened Nov 10, 2026Collaborator·5 of 8 tasks·

  ·
  1

• Lhotse: filter HQ metadata before source repetition

  common
  #16587·nithinraok opened Nov 10, 2026Member·4 tasks done·

  ·
  1

• fix(deps): AUT-2832 update security floors and Python runtime

  CIdependenciesforce-run-allpython:uv
  #16586·svcnemo-autobot opened Nov 10, 2026Contributor·

  ·
  4

• fix(utils): keep torch.autocast's parameters and the training step's arguments in CUDAGraphCallback

  #16585·albertz opened Nov 9, 2026Member·4 of 7 tasks·

  ·
  2

• Restore FusedAdam DCP checkpoints into live GPU buffers

  #16584·pzelasko opened Nov 9, 2026Collaborator·3 tasks done·

  ·
  3

• Use indexed fixtures for the updated Lhotse checkpoint contract

  #16583·pzelasko opened Nov 9, 2026Collaborator·5 tasks done·

  ·
  1

• Add efficient paired CFG to EasyMagpie v2 serving

  TTS
  #16582·Edresson opened Nov 9, 2026Member·

  ·
  1

• Adapt indexed dataloading to the updated Lhotse checkpoint lifecycle

  common
  #16581·pzelasko opened Nov 9, 2026Collaborator·5 tasks done·

  ·
  1

• Trim consumed tokens from the carried RNNT hypothesis in cache-aware streaming

  ASRcommunity-request
  #16579·ovuruska opened Nov 9, 2026·3 of 7 tasks·

  ·
  1

• cp: Apply runtime loading policy to nested model configurations (16561) into r3.1.0

  cherry-pickRun CICD
  #16577·svcnvidia-nemo-ci opened Nov 9, 2026Contributor·

  ·
  3

• feat(asr): capacity mode of the sequence-packed path for CUDA-graph capture

  ASR
  #16575·albertz opened Nov 9, 2026Member·4 of 6 tasks·

  ·
  3

• fix(common): build CompositeLoss's scales tensor once per device

  common
  #16574·albertz opened Nov 9, 2026Member·4 of 6 tasks·

  ·
  3

• feat(core): multi-tensor Muon step with batched Newton-Schulz iterations

  core
  #16573·albertz opened Nov 9, 2026Member·4 of 8 tasks·

  ·
  6

• feat(asr): autocast_residual, the Conformer residual stream in the autocast dtype

  ASR
  #16572·albertz opened Nov 9, 2026Member·4 of 8 tasks·

  ·
  4

• fix(asr): RotaryPositionalEncoding builds its tables for max_len at construction

  ASR
  #16571·albertz opened Nov 9, 2026Member·3 of 8 tasks·

  ·
  4

• feat(asr): fused activation, dropout and second linear layer of the Conformer feed-forward

  ASR
  #16570·albertz opened Nov 9, 2026Member·6 of 8 tasks·

  ·
  6

• feat(asr): Triton depthwise convolution over sequence-packed utterances

  ASR
  #16569·albertz opened Nov 9, 2026Member·6 of 8 tasks·

  ·
  4

• fix(common): SquaredReLU saves one half-precision tensor under autocast

  common
  #16568·albertz opened Nov 9, 2026Member·3 of 8 tasks·

  ·
  3

• feat(asr): Triton kernel for the sequence-packed relative-position attention

  ASR
  #16567·albertz opened Nov 9, 2026Member·6 of 8 tasks·

  ·
  5

• Skip blank manifest lines before numbering transcriptions

  ASRcommunity-requestwaiting-on-maintainers
  #16566·0z5a opened Nov 8, 2026·3 of 8 tasks·

  ·
  1

  1

• Fix word confidence for character CTC models

  ASRcommunity-requestwaiting-on-maintainers
  #16565·0z5a opened Nov 8, 2026·3 of 8 tasks·

  ·
  1

  1

• cp: Improve artifact path resolution (16492) into r3.1.0

  cherry-pickcoreRun CICD
  #16560·svcnvidia-nemo-ci opened Nov 6, 2026Contributor·

  ·
  3

• cp: Improve Canary prompt parsing (16412) into r3.1.0

  cherry-pickcommonRun CICD
  #16559·svcnvidia-nemo-ci opened Nov 6, 2026Contributor·

  ·
  3

• cp: Improve model archive probing (16411) into r3.1.0

  ASRcherry-pickRun CICD
  #16558·svcnvidia-nemo-ci opened Nov 6, 2026Contributor·

  ·
  3