All pull requests

Pull requests

• Add audio RL fine-tuning skill and ASCEND example

  #16480·vmendelev opened Oct 28, 2026Collaborator·4 tasks done·

  ·
  1

• fix(speechlm2): SALMWithAsrDecoder loads model.init_from_checkpoint

  #16478·albertz opened Oct 28, 2026Member·4 of 8 tasks·

  ·
  1

• fix(speechlm2): instantiate the class named by perception.target

  #16477·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  1

• fix(speechlm2): drop the unread no_rttm_to_ones key from salm_nemotron.yaml

  #16476·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  2

• fix(speechlm2): read inference_guidance_enabled for the EAR-TTS and voice chat inference

  #16475·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  5

• docs(core): yaml example of setup_optimization as a literal block

  core
  #16474·albertz opened Oct 28, 2026Member·4 of 8 tasks·

  ·
  2

• fix(speechlm2): spell the two dataset keys where duplex_eartts_train.py reads them

  #16473·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  2

• fix(speechlm2): voice chat evaluation overrides reach the STT and TTS sub-models

  #16472·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  1

• fix(speechlm2): DuplexEARTTS evaluation rejects a batch with different system prompts

  #16471·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  6

• fix(speechlm2): keep every frame of the teacher-forced EAR-TTS generation

  #16470·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  4

• fix(asr): ParallelExpertEncoder routes all-sentinel speaker targets like None under the online scope

  ASR
  #16469·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  8

• fix(asr): ParallelExpertEncoder cuts external speaker targets past the ASR grid instead of pooling them

  ASR
  #16468·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  4

• fix(speechlm2): per-row audio lengths from DuplexEARTTS.offline_inference

  #16467·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  6

• fix(speechlm2): per-row audio lengths from NemotronVoiceChat.offline_inference

  #16466·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  2

• fix(speechlm2): QformerConnector masks padded encoder frames and returns the prompt length

  #16465·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  2

• fix(speechlm2): decode duplex STT text outputs within each row's own length

  #16464·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  2

• feat(asr): option to skip the final LayerNorm of TransformerEncoder

  ASR
  #16463·albertz opened Oct 28, 2026Member·4 of 8 tasks·

  ·
  2

• fix(speechlm2): write position 0 of prompt-less rows in a mixed-prompt batch

  #16462·albertz opened Oct 28, 2026Member·3 of 8 tasks·

  ·
  2

• Add NemotronTTS v2 for training and serving

  TTS
  #16458·vklimkov-nvidia opened Oct 27, 2026Member·5 of 8 tasks·

  ·
  9

• fix(speechlm2): avoid placeholder token IDs in packed embedding forwards

  community-request
  #16456·BillVan1945 opened Oct 27, 2026·

  ·
  1

  1

• fix(docs): add --shm-size and --ipc=host to Docker run examples

  community-request
  #16453·nouraellm opened Oct 26, 2026·

  ·
  1

  1

• fix(core): set_optim builds the sched config without SchedConfig.min_lr

  core
  #16452·albertz opened Oct 26, 2026Member·3 of 8 tasks·

  ·
  20

• feat(core): allow config targets from packages the user trusts

  core
  #16449·albertz opened Oct 26, 2026Member·4 of 8 tasks·

  ·
  3

• Turbo Transducer: Speed-of-Light Transducer Training with (Nearly) Zero Memory

  ASRcore
  #16447·artbataev opened Oct 25, 2026Collaborator·4 of 8 tasks·

  ·
  6

• Use in-memory waveforms for online PO reward scoring

  TTS
  #16443·blisc opened Oct 23, 2026Collaborator·2 of 5 tasks·

  ·
  1