NVIDIA Neural Modules 3.0.0

ASR

Changelog

• Add Per-Stream Phrase Boosting for Cache-Aware RNN-T by @artbataev :: PR: #15344
• docs: clarify audio preprocessing for ASR inference by @stanley1208 :: PR: #15308
• feat(asr): Support Parquet/Arrow datasets with embedded audio bytes via Lhotse by @v4xsh :: PR: #15303
• Add buffered inference support for SALM models by @naymaraq :: PR: #15364
• docs: fix typo and path error in ASR documentation by @stanley1208 :: PR: #15372
• docs: fix multiple typos in ASR bucketing documentation by @stanley1208 :: PR: #15376
• Transformation stage 0: NeMo docs revision by @pzelasko :: PR: #15363
• [TTS] Remove un-supported tts models and related docs by @subhankar-ghosh :: PR: #15368
• Fixing could not find ctc_segmentation. in CTC tutorial by @Jorjeous :: PR: #15403
• Adapt to use env variable for adapter mixin model loading by @nithinraok :: PR: #15406
• fix: add missing docstring for transcript args in EncDecMultiTaskModel.forward() by @stanley1208 :: PR: #15396
• Transformation stage 1: Remove deprecated collections by @pzelasko :: PR: #15378
• Fix cuda-python usage for CUDA graphs by @artbataev :: PR: #15416
• docs: fix broken tutorial links pointing to non-existent stable branch by @stanley1208 :: PR: #15420
• fix: replace assert statements with raise ValueError in EncDecMultiTaskModel by @stanley1208 :: PR: #15444
• update transformers version by @nithinraok :: PR: #15365
• Relax kaldialign pin in ASR requirements to unblock Python 3.14 compatibility by @basnijholt :: PR: #15438
• Verbose fix in AED models by @nune-tadevosyan :: PR: #15509
• Fix numba-cuda and cuda-python installation and usage by @artbataev :: PR: #15506
• Inference Tutorial for Streaming ASR Pipelines by @naymaraq :: PR: #15526
• Add AIS batch loading support to LhotseSpeechToTextBpeDataset by @gaikwadabhishek :: PR: #15538
• Clean ASR older and non-maintained models and its respective modules by @nithinraok :: PR: #15507
• Changed the documentation getting started structure by @Ssofja :: PR: #15460
• Ignore PnC for WER calculation: streaming ASR inference by @artbataev :: PR: #15550
• Add ASR-EOU models and training/eval scripts by @stevehuang52 :: PR: #14740
• Add streaming inference support for Unified RNNT model by @andrusenkoau :: PR: #15522
• [TTS][Docs][MagpieTTS] Add MagpieTTS finetuning docs by @subhankar-ghosh :: PR: #15546
• [Fix] Make cpWER calculation identical to meeteval by @tango4j :: PR: #15573
• Fix [ASR Inference Pipelines]: Minor bug fixes by @naymaraq :: PR: #15576
• fix(asr): mask padded frames after feature normalization by @arushidNV :: PR: #15562
• Remove the PnC flag from the ASR inference pipelines by @naymaraq :: PR: #15591
• fix use_feat_cache feature by @naymaraq :: PR: #15600
• add a parakeet unified model link to readme by @andrusenkoau :: PR: #15604
• Remove tail margin algorithms from all ASR pipelines by @naymaraq :: PR: #15605
• fix: add missing docstrings to data pipeline utility functions by @stanley1208 :: PR: #15560
• Add xatt trimming for multitask beam decoding by @azziko :: PR: #15466
• fix: clamp shift perturbation for short audio instead of silently skipping by @stanley1208 :: PR: #15561
• fix(ci): increase GPU speech unit test timeouts (TTS: 30 min, ASR: 60 min) by @ko3n1g :: PR: #15651
• Remove the use of pickle throughout codebase by @blisc :: PR: #15629
• ci(speech): split L0_Unit_Tests GPU/CPU ASR and CPU SpeechLM2 into parallel buckets by @ko3n1g :: PR: #15654
• SALMAutomodel - long-context support (chunking, AC) and batch of fixes by @pzelasko :: PR: #15648
• handle issue 15598 for remote code fix by @nithinraok :: PR: #15652
• Refactored all ASR collections documentation by @Ssofja :: PR: #15542
• Add Transformer Encoder for ASR by @nithinraok :: PR: #15661
• Reformat files with black. Remove exclusion from black rules. by @artbataev :: PR: #15687
• Causal transformer by @nithinraok :: PR: #15730
• add freeze, unfreeze methods with experimental tag by @nithinraok :: PR: #15729
• Add ASR fine-tuning skill by @pzelasko :: PR: #15733
• Add unified prompt model support for multilingual ASR and streaming inference by @enas-albasiri :: PR: #15666
• ci: Install CI dependencies from uv lock by @chtruong814 :: PR: #15697
• Removing the dependency on Pyannote for Diarization and VAD by @tango4j :: PR: #15632
• Remove stale NeMo dependencies on uv base by @pzelasko :: PR: #15734
• Add prompt RNNT ASR docs by @kingformatty :: PR: #15738
• Adding Conformer encoder I/O-styled Transformer encoder by @tango4j :: PR: #15703
• Fix peak memory in TransformerEncoder rel_pos attention by @ipmedenn :: PR: #15751
• Support CUDA 12 container builds by @pzelasko :: PR: #15747
• Transducer Decoding: 2.4x faster decoding with confidence by @artbataev :: PR: #15315
• Add experimental simulated chunked streaming decoding for transducers by @artbataev :: PR: #15517
• Confidence in RNN-T inference pipeline (buffered and cache-aware) by @artbataev :: PR: #15765
• Update README.md - Nemotron-3.5-ASR-Streaming by @pzelasko :: PR: #15766
• Add batched streaming beam search for RNN-T (mALSD+mAES) and TDT (mALSD) by @lilithgrigoryan :: PR: #15753
• True-bfloat16 inference for cache aware pipelines by @naymaraq :: PR: #15763
• adding community featured models by @Ssofja :: PR: #15744
• Improve NEST GPU Utilization 4/N: 250x Faster Masking on GPU by @MahmoudAshraf97 :: PR: #15739
• add soe fixes in documentation by @Ssofja :: PR: #15774
• refactored Speaker Diarizarion collection's docs by @Ssofja :: PR: #15647
• fix: Move git refs for deep-ep and transformer engine to uv sources by @chtruong814 :: PR: #15758
• cp: fix: Replace direct hydra.utils.instantiate with safe_instantiate (#15802) by @chtruong814 :: PR: #15822
• Adding fix of asr_diar tutorial for r3.0.0 by @tango4j :: PR: #15928
• fix-tutorial-install-line (#15879) by @nithinraok :: PR: #15934
• Fix cpWER imports in ASR diarization tutorial for r3.0.0 by @tango4j :: PR: #15974
• docs: clarify audio preprocessing for ASR inference (#15308) by @stanley1208 :: PR: #0
• feat(asr): Support Parquet/Arrow datasets with embedded audio bytes via Lhotse (#15303) by @v4xsh :: PR: #0
• docs: fix typo and path error in ASR documentation (#15372) by @stanley1208 :: PR: #0
• docs: fix multiple typos in ASR bucketing documentation (#15376) by @stanley1208 :: PR: #0
• switch to soundfile from scipy. added sox with mp3support. added symlinc to ctc_ctc_segmentation command so in colab env wold recognize it. (#15403) by @Jorjeous :: PR: #0
• Relax kaldialign requirement in ASR dependencies (#15438) by @basnijholt :: PR: #0
• Inference Tutorial for Streaming ASR Pipelines (#15526) by @naymaraq :: PR: #0
• Clean ASR older and non-maintained models and its respective modules (#15507) by @nithinraok :: PR: #0
• Ignore PnC for WER calculation: streaming ASR inference (#15550) by @artbataev :: PR: #0
• Add ASR-EOU models and training/eval scripts (#14740) by @stevehuang52 :: PR: #0
• Add streaming inference support for Unified RNNT model (#15522) by @andrusenkoau :: PR: #0
• Fix [ASR Inference Pipelines]: Minor bug fixes (#15576) by @naymaraq :: PR: #0
• Remove the PnC flag from the ASR inference pipelines (#15591) by @naymaraq :: PR: #0
• Remove tail margin algorithms from all ASR pipelines (#15605) by @naymaraq :: PR: #0
• fix(ci): increase GPU speech unit test timeouts (TTS: 30 min, ASR: 60 min) (#15651) by @ko3n1g :: PR: #0
• ci(speech): split L0_Unit_Tests GPU/CPU ASR and CPU SpeechLM2 into parallel buckets (#15654) by @ko3n1g :: PR: #0
• Refactored all ASR collections documentation (#15542) by @Ssofja :: PR: #0
• Add Transformer Encoder for ASR (#15661) by @nithinraok :: PR: #0
• Add ASR fine-tuning skill (#15733) by @pzelasko :: PR: #0
• Add unified prompt model support for multilingual ASR and streaming inference (#15666) by @enas-albasiri :: PR: #0
• Add prompt RNNT ASR docs (#15738) by @kingformatty :: PR: #0
• Transducer Decoding: 2.4x faster decoding with confidence (#15315) by @artbataev :: PR: #0
• Add experimental simulated chunked streaming decoding for transducers (#15517) by @artbataev :: PR: #0
• Fix ASR transducer nemo checkpoint export (#15876) (#15880) by @svcnvidia-nemo-ci :: PR: #0
• Set trusted checkpoint env for ASR tutorial restores (#15875) (#15903) by @svcnvidia-nemo-ci :: PR: #0
• Fix/ctc segmentation tutorial deps (#15901) (#15904) by @svcnvidia-nemo-ci :: PR: #0
• Adding fix of asr_diar tutorial for r3.0.0 (#15928) by @tango4j :: PR: #0
• Fix ASR tutorial notebook restore and setup paths (#15908) (#15935) by @svcnvidia-nemo-ci :: PR: #0
• Disable RNNT CUDA graphs in Multilang ASR notebook (#15937) (#15938) by @svcnvidia-nemo-ci :: PR: #0
• fix Online_ASR_Microphone_Demo_Buffered_Streaming.ipynb failed with ConfigKeyError (#15964) (#15970) by @svcnvidia-nemo-ci :: PR: #0
• fix ASR_with_Subword_Tokenization.ipynb fails: InterpolationKeyError 'model.sample_rate' (#15945) (#15946) by @svcnvidia-nemo-ci :: PR: #0
• Fix cpWER imports in ASR diarization tutorial (#15974) by @tango4j :: PR: #0

TTS

Changelog

• [MagpieTTS][TTS] Support Hi and Ja in MagpieTTS by @subhankar-ghosh :: PR: #15320
• Transformation stage 0: NeMo docs revision by @pzelasko :: PR: #15363
• [MagpieTTS][bugfix] defaults to force_map_dataset=True for validation datast to avoid duplicates by a factor of num_workers. by @XuesongYang :: PR: #15387
• [TTS] Fix bugs in MagpieTTS evaluation by @rlangman :: PR: #15381
• [TTS] Remove un-supported tts models and related docs by @subhankar-ghosh :: PR: #15368
• [MagpieTTS] Mixture-of-Experts by @XuesongYang :: PR: #15370
• CI: increase cer to 0.08 from 0.06 to avoid chance to fail. by @XuesongYang :: PR: #15395
• [voice agent] Add support for nemotron-nano-v3, magpie-tts and misc improvements by @stevehuang52 :: PR: #15318
• Transformation stage 1: Remove deprecated collections by @pzelasko :: PR: #15378
• unblock CI: increase to 0.08 from 0.03 by @XuesongYang :: PR: #15413
• docs: fix broken tutorial links pointing to non-existent stable branch by @stanley1208 :: PR: #15420
• Fix Nanocodec's FSQ codes function: tensor creation misplacement by @Ferdydh :: PR: #15405
• [MagpieTTS] set seed and shard_seed=randomized for val dataloader to ensure deterministic multi-GPU eval by @XuesongYang :: PR: #15427
• [TTS] Infrastructure for parallelization of evaluation (scoring) by @rfejgin :: PR: #15417
• docs: Enable docs self-publishing including nightly update by @chtruong814 :: PR: #15432
• Implement workaround for Japanese transcript normalization by @subhankar-ghosh :: PR: #15448
• Nemotron VoiceChat STT + TTS joint class by @Edresson :: PR: #15456
• [TTS][Magpietts] Unify Longform and Standard Inference logic by @subhankar-ghosh :: PR: #15375
• [magpietts] added multiple validation dataloaders and log metrics per val data by @XuesongYang :: PR: #15348
• End-of-Utterance metric by @rfejgin :: PR: #15462
• update transformers version by @nithinraok :: PR: #15365
• [MagpieTTS] Magpietts longform unify by @subhankar-ghosh :: PR: #15477
• Add Zenodo download URLs to broken links false positives by @XuesongYang :: PR: #15488
• [lhotse] Added support for re-weighting datasets with temperature on the fly. by @XuesongYang :: PR: #15200
• Bugfix: add new EoU metrics names to CSV header by @rfejgin :: PR: #15493
• EoU classifier: better handling of input sampling rate by @rfejgin :: PR: #15497
• [MagpieTTS][TTS] ja longform fix and StrEnum import fix by @subhankar-ghosh :: PR: #15499
• Minor cleanup in EoU classifier unit test by @rfejgin :: PR: #15511
• zero shot disabling magpie by @paarthneekhara :: PR: #15503
• [TTS][MagpieTTS] Add seen speaker test with Feb26 Open Source model by @subhankar-ghosh :: PR: #15510
• TTS Testing Fix: Make Runs more Deterministic by @blisc :: PR: #15508
• MagpieTTS refactor by @paarthneekhara :: PR: #15504
• [TTS][MagpieTTS] Remove HF dependencies from CI Tests by @subhankar-ghosh :: PR: #15544
• [MagpieTTS][TTS] Rename index for attention prior weights by @subhankar-ghosh :: PR: #15551
• bring back zeroshot disabling code in MagpieTTS by @paarthneekhara :: PR: #15564
• MagpieTTS decoder model on top of NeMo main branch by @paarthneekhara :: PR: #15277
• Make TTS L2 Tests More Resilient by @blisc :: PR: #15584
• [TTS][Docs][MagpieTTS] Add MagpieTTS finetuning docs by @subhankar-ghosh :: PR: #15546
• [TTS][Magpietts] Added CFG distillation by @artem-gorodetskii :: PR: #15568
• [TTS] Add code for training semantic codec by @rlangman :: PR: #15524
• [TTS] Added TTS comparison report by @artem-gorodetskii :: PR: #15621
• Add hi-IN , Ko-KR and pt-BR IPA tokenizer support by @quapham :: PR: #15567
• Add Arabic char tokenizer and Japanese-English katakana support by @quapham :: PR: #15614
• fix(ci): increase GPU speech unit test timeouts (TTS: 30 min, ASR: 60 min) by @ko3n1g :: PR: #15651
• [audio codec] Add support for Lhotse training format by @rfejgin :: PR: #15622
• Remove the use of pickle throughout codebase by @blisc :: PR: #15629
• Update text processing in Magpietts eval by @rlangman :: PR: #15608
• GRPO fix for frame stacked models by @shehzeen :: PR: #15643
• fix: count_input_cfg_levels now resolves string file references by @XuesongYang :: PR: #15646
• [MagpieTTS] Load Whisper with torch_dtype="auto" for fp16 inference by @matteolippi :: PR: #15680
• val audio logging fix in EasyMagpieTTS by @paarthneekhara :: PR: #15688
• Easymagpie ZeroShot disabling option using speaker encoder by @paarthneekhara :: PR: #15639
• Add verbose flag to UTMOSv2Calculator by @rlangman :: PR: #15699
• Reformat files with black. Remove exclusion from black rules. by @artbataev :: PR: #15687
• clean commit with EMTTS text embedding disabling and pronunciation co… by @paarthneekhara :: PR: #15717
• ci: Install CI dependencies from uv lock by @chtruong814 :: PR: #15697
• Removing the dependency on Pyannote for Diarization and VAD by @tango4j :: PR: #15632
• Remove stale NeMo dependencies on uv base by @pzelasko :: PR: #15734
• [TTS] add max_duration = 20.0 to lhotse training yaml configs. by @XuesongYang :: PR: #15696
• [TTS] Add comprehensive MoE dispatch unit tests with golden-value regression checks by @XuesongYang :: PR: #15698
• [Audio Codec] Lhotse data loading updates and fixes by @rfejgin :: PR: #15742
• docs: overhaul installation instructions around uv + bring-your-own Python/PyTorch/CUDA by @pzelasko :: PR: #15769
• [TTS][Magpietts] Added local transformer CFG distillation by @artem-gorodetskii :: PR: #15748
• cp: fix: Replace direct hydra.utils.instantiate with safe_instantiate (#15802) by @chtruong814 :: PR: #15822
• fix-tutorial-install-line (#15879) by @nithinraok :: PR: #15934
• Cherry pick tts notebook to release branch by @blisc :: PR: #15960
• CherryPick: Fix loading of MagpieTTS and add test for all 3 versions (#15941) by @blisc :: PR: #15979
• [MagpieTTS][TTS] Support Hi and Ja in MagpieTTS (#15320) by @subhankar-ghosh :: PR: #0
• [TTS] Fix bugs in MagpieTTS evaluation (#15381) by @rlangman :: PR: #0
• [TTS] Remove un-supported tts models and related docs (#15368) by @subhankar-ghosh :: PR: #0
• [MagpieTTS] Mixture-of-Experts (#15370) by @XuesongYang :: PR: #0
• [voice agent] Add support for nemotron-nano-v3, magpie-tts and misc improvements (#15318) by @stevehuang52 :: PR: #0
• [MagpieTTS] fix val dataloader config: set seed and shard_seed=randomized to ensure deterministic multi-GPU eval (#15427) by @XuesongYang :: PR: #0
• [TTS] Infrastructure for parallelization of evaluation (scoring) (#15417) by @rfejgin :: PR: #0
• Nemotron VoiceChat STT + TTS joint class (#15456) by @Edresson :: PR: #0
• [TTS][Magpietts] Unify Longform and Standard Inference logic (#15375) by @subhankar-ghosh :: PR: #0
• [magpietts] added multiple validation dataloaders and log metrics per val data (#15348) by @XuesongYang :: PR: #0
• [MagpieTTS] Magpietts longform unify (#15477) by @subhankar-ghosh :: PR: #0
• [TTS] EoU metric classifier: better handling of input sampling rate (#15497) by @rfejgin :: PR: #0
• TTS Testing Fix: Make Runs more Deterministic (#15508) by @blisc :: PR: #0
• MagpieTTS refactor (#15504) by @paarthneekhara :: PR: #0
• [TTS][MagpieTTS] Remove HF dependencies from CI Tests (#15544) by @subhankar-ghosh :: PR: #0
• MagpieTTS decoder model on top of NeMo main branch (#15277) by @paarthneekhara :: PR: #0
• Make TTS L2 Tests More Resilient (#15584) by @blisc :: PR: #0
• [TTS][Docs][MagpieTTS] Add MagpieTTS finetuning docs (#15546) by @subhankar-ghosh :: PR: #0
• [TTS][Magpietts] Added CFG distillation (#15568) by @artem-gorodetskii :: PR: #0
• [TTS] Add code for training semantic codec (#15524) by @rlangman :: PR: #0
• [TTS] Added TTS comparison report (#15621) by @artem-gorodetskii :: PR: #0
• fix(ci): increase GPU speech unit test timeouts (TTS: 30 min, ASR: 60 min) (#15651) by @ko3n1g :: PR: #0
• Update text processing in Magpietts eval (#15608) by @rlangman :: PR: #0
• [MagpieTTS] Load Whisper with torch_dtype="auto" for fp16 inference (#15680) by @matteolippi :: PR: #0
• clean commit with EMTTS text embedding disabling and pronunciation co… (#15717) by @paarthneekhara :: PR: #0
• [TTS][Magpietts] Added local transformer CFG distillation (#15748) by @artem-gorodetskii :: PR: #0
• [TTS] Whitelist TTS AudioTrimmer and ArtifactGenerator (#15914) (#15916) by @svcnvidia-nemo-ci :: PR: #0
• Cherry pick tts notebook to release branch (#15960) by @blisc :: PR: #0
• CherryPick: Fix loading of MagpieTTS and add test for all 3 versions (#15941) (#15979) by @blisc :: PR: #0

NLP / NMT

Changelog

• [speechlm2] SALM with TDT decoder input; "sharegpt" format support; various improvements by @pzelasko :: PR: #15316
• [speechm2] Support indexed sharegpt JSONL and webdataset formats by @pzelasko :: PR: #15410
• [speechlm2] SALM with TDT decoder input; "sharegpt" format support; various improvements (#15316) by @pzelasko :: PR: #0
• [speechm2] Support indexed sharegpt JSONL and webdataset formats (#15410) by @pzelasko :: PR: #0

NeMo Tools

Changelog

• Updating SDE; s3 read; comparison mode; security fix by @Jorjeous :: PR: #15500
• [r3.0.0] Fix SDE_HowTo_v2 notebook: replace removed QuartzNet15x5Base-En (fix_6411583) by @Jorjeous :: PR: #15898
• Updating SDE; s3 read; comparison mode; security fix (#15500) by @Jorjeous :: PR: #0
• Fix/ctc segmentation tutorial deps (#15901) (#15904) by @svcnvidia-nemo-ci :: PR: #0
• numba compatibility for SDE (#15694) (#15890) by @svcnvidia-nemo-ci :: PR: #0

Export

Changelog

• Fix dynamic_shapes variable naming in ONNX dynamo export by @yashwantbezawada :: PR: #15047
• Support offline SpeechLM2 HF exports by @pzelasko :: PR: #15736
• Fix dynamic_shapes variable naming in ONNX dynamo export (#15047) by @yashwantbezawada :: PR: #0
• Support offline SpeechLM2 HF exports (#15736) by @pzelasko :: PR: #0
• Fix ASR transducer nemo checkpoint export (#15876) (#15880) by @svcnvidia-nemo-ci :: PR: #0

Bugfixes

Changelog

• [TTS] Add comprehensive MoE dispatch unit tests with golden-value regression checks by @XuesongYang :: PR: #15698
• Add comprehensive MoE dispatch unit tests with golden-value regression checks (#15698) by @XuesongYang :: PR: #0
• Pin SALM to portable MoE dispatcher (#15963) (#15969) by @svcnvidia-nemo-ci :: PR: #0

Uncategorized:

Changelog

• Update Imports in Audio Notebook by @blisc :: PR: #15345
• Version bump to 2.8.0rc0.dev0 by @github-actions[bot] :: PR: #15350
• Disable nemo2 functional tests by @chtruong814 :: PR: #15354
• Clarify when to use TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD by @chtruong814 :: PR: #15353
• Add create_all_gather_group configuration option by @jeffnvidia :: PR: #15253
• Remove deprecated LLM, VLM, and diffusion tutorials by @yaoyu-33 :: PR: #15357
• Update changelog for r2.6.0 by @github-actions[bot] :: PR: #15366
• [voice agent] fix dependency for nemo26.02 by @stevehuang52 :: PR: #15380
• Update audio collection yamls to include absolute path to loss functions by @blisc :: PR: #15384
• Fix macro accuracy when changing labels by @stevehuang52 :: PR: #15379
• fixes nemo tutorial for loading non registered classes by @nithinraok :: PR: #15398
• default weights to false by @nithinraok :: PR: #15397
• Handle race condition in exp_manager's timer gracefully by @pzelasko :: PR: #15409
• Fix BNR 2.0 inference alignment error with input signal padding by @ManasiRemane :: PR: #15388
• Remove typing annotation breaking earlier python versions in speechlm2 by @pzelasko :: PR: #15414
• ci: Update release workflow to include generated changelog by @chtruong814 :: PR: #15429
• ci: Remove alerts by @ko3n1g :: PR: #15434
• ci: Add-credentials-for-docs by @ko3n1g :: PR: #15439
• ci: Pass app id to release job by @chtruong814 :: PR: #15442
• Implement Duplex Speech-to-text model and rebase by @kevinhu-nv :: PR: #15092
• Add a functional test of frame stacking by @rfejgin :: PR: #15424
• Auto-detect use_bucketing and validate batch size config by @pzelasko :: PR: #15461
• Script to migrate .nemo checkpoints that used torchaudio preprocessor by @pzelasko :: PR: #15437
• fix for nemotron-speech-streaming-0.6b by @stevehuang52 :: PR: #15473
• Projection connector by @nune-tadevosyan :: PR: #15468
• Update SpeechLM2 collection documentation and fix dataloader custom audio resample issue by @Edresson :: PR: #15478
• ci: Disable selective triggering of tests by @chtruong814 :: PR: #15484
• Add functional tests for init, train step, and inference for every supported released model by @pzelasko :: PR: #15433
• Fix E2E nightly tests: move missing conftest.py by @pzelasko :: PR: #15501
• remove protobuf from reqs by @nithinraok :: PR: #15502
• add claude review by @nithinraok :: PR: #15515
• Fix BlendableDataset crash when torch.distributed is not initialized (#15465) by @v4xsh :: PR: #15505
• add claude answer and fix by @nithinraok :: PR: #15518
• revert code owners by @nithinraok :: PR: #15521
• Fix e2e nightly tests: move shared utils out of conftest.py by @pzelasko :: PR: #15519
• Update README by @blisc :: PR: #15495
• Fix a broken entry in README.md after re-write by @pzelasko :: PR: #15527
• chore: test FW-CI-templates ko3n1g/fix/linkcheck-retry-backoff by @ko3n1g :: PR: #15525
• bump release workflow by @ko3n1g :: PR: #15532
• Script for saving audio augmented by Lhotse samplers/dataloaders by @racoiaws :: PR: #14808
• Fix IsADirectoryError when cleaning up unfinished distributed checkpoints by @pzelasko :: PR: #15541
• Add support for partial transcription prefix in the prompt by @azziko :: PR: #15449
• docs: Fix docs build by setting uv conflicts for cu12 vs cu13 by @chtruong814 :: PR: #15548
• ci: Update docs build job to exclude cu12 extra by @chtruong814 :: PR: #15553
• ci: upgrade GitHub Actions for Node.js 24 compatibility by @ko3n1g :: PR: #15537
• Add VoiceChat to README by @blisc :: PR: #15547
• chore: bump _code_freeze workflow to v0.86.0 by @ko3n1g :: PR: #15569
• Fix Qwen3 SALM LoRA init by @pzelasko :: PR: #15570
• Allow frame stacking with text context by @rfejgin :: PR: #15585
• docs: Fix docs link in README by @chtruong814 :: PR: #15594
• README.md: next and last stable release info by @pzelasko :: PR: #15602
• README: refer to "vanilla" 26.02 container by @pzelasko :: PR: #15606
• fix: typo in readme by @L0RD-ZER0 :: PR: #15539
• Limit to peft<=0.18.0 until CI environment is updated by @pzelasko :: PR: #15610
• add agent skill for debugging distributed training log failures by @gaikwadabhishek :: PR: #15612
• Fix codec e2e tests by @rlangman :: PR: #15607
• Implement formatter to ensure Nemotron VoiceChat speech decoder reproducibility, speed up training and support half precision inference by @Edresson :: PR: #15583
• SALM with NeMo Automodel integration for Nemotron Nano V3 LLM backbone by @pzelasko :: PR: #15447
• CLAUDE.md, PR babysitter, local dev skills + symlinks for Codex/Cursor by @pzelasko :: PR: #15613
• Bump protobufjs from 7.5.3 to 7.5.5 in /examples/voice_agent/client by @dependabot[bot] :: PR: #15619
• Bump mako from 1.3.10 to 1.3.11 by @dependabot[bot] :: PR: #15618
• Bump pillow from 12.1.1 to 12.2.0 by @dependabot[bot] :: PR: #15603
• Bump vite from 6.3.6 to 6.4.2 in /examples/voice_agent/client by @dependabot[bot] :: PR: #15586
• chore(beep boop 🤖): bump FW-CI-templates workflow pins to v0.88.0 by @svcnvidia-nemo-ci :: PR: #15581
• Bump pygments from 2.19.2 to 2.20.0 by @dependabot[bot] :: PR: #15565
• Bump picomatch from 4.0.2 to 4.0.4 in /examples/voice_agent/client by @dependabot[bot] :: PR: #15554
• ci(babysitter): switch from check_run to workflow_run trigger by @pzelasko :: PR: #15627
• ci: restrict testpypi publish trigger to release branches by @pzelasko :: PR: #15628
• ci: onboard copy-pr-bot by @ko3n1g :: PR: #15631
• fix: Remove diskcache from requirements by @chtruong814 :: PR: #15630
• docs: Add SECURITY.md by @chtruong814 :: PR: #15636
• Remove nemo_log_globalrank-N_localrank-M.txt file creation by @pzelasko :: PR: #15626
• [Voice Agent] fix empty tools by @stevehuang52 :: PR: #15634
• to_hf.py + PromptFormatter: produce vLLM-ready SpeechLM checkpoints with backbone-native chat_template by @DongjiGao :: PR: #15623
• [Voice Agent] Update README by @stevehuang52 :: PR: #15642
• Dockerfile: build from the official PyTorch image by @artbataev :: PR: #15638
• feat(ci): migrate from Azure to AWS ephemeral runners by @ko3n1g :: PR: #15620
• ci: switch formatting CI to check-only and enable pre-commit hooks by @chtruong814 :: PR: #15656
• Updated estimate_token_bins.py script; remove deprecated speechlm1 scripts by @pzelasko :: PR: #15655
• [ci, build] fix: unblock testpypi publish — drop direct git URL dep, bump CI template to v0.94.1 by @ko3n1g :: PR: #15659
• Re-add codeowners by @chtruong814 :: PR: #15660
• [build] fix: restore nemo_automodel in requirements_speechlm2.txt by @ko3n1g :: PR: #15662
• Add vLLM support for NeMo SpeechLM by @DongjiGao :: PR: #15520
• Reduce code duplication in audio collection + some small fixes by @racoiaws :: PR: #15587
• ci: Major refactor of release-workflows by @ko3n1g :: PR: #15668
• Preserve SpeechLM perception checkpoint dtype by @DongjiGao :: PR: #15686
• ci: Fix CI approval queue by @chtruong814 :: PR: #15693
• ci: remove build-docs and build-test-publish-wheel workflows by @ko3n1g :: PR: #15685
• ci: validate release branch-rules by @ko3n1g :: PR: #15718
• Fix LazyNeMoIterator for non-tarred S3 audio data by @melllinia :: PR: #15732
• Support encoder input chunking for SALM vLLM inference by @pzelasko :: PR: #15716
• fix for release nightly docs CI by @Ssofja :: PR: #15735
• Docs visual improvements by @Ssofja :: PR: #15745
• Build Automodel compiled dependencies in CI image by @pzelasko :: PR: #15737
• [speechlm2] SALMAutomodel: THD (packed sequence) and context parallel support by @pzelasko :: PR: #15679
• ci: bump _release_library.yml to v1.4.3 by @ko3n1g :: PR: #15749
• Mark legacy docs navigation pages as orphaned by @pzelasko :: PR: #15755
• chore: Update dependencies to address CVEs and also update Dockerfile by @chtruong814 :: PR: #15756
• fix: Move code copy to last docker stage by @chtruong814 :: PR: #15761
• docs: update SpeechLM parallelism strategy docs by @pzelasko :: PR: #15773
• chore: Bump next release to be 3.0 by @chtruong814 :: PR: #15783
• Freeze tags in in r3.0.0 by @github-actions[bot] :: PR: #15788
• fix: Remove weights_only=False (15801) by @svcnvidia-nemo-ci :: PR: #15804
• chore: Bump lxml and gitpython to address CVEs (15785) by @svcnvidia-nemo-ci :: PR: #15789
• Fix reported security vulnerabilities (15803) by @svcnvidia-nemo-ci :: PR: #15806
• fix: Secure tar archive extraction (15811) by @svcnvidia-nemo-ci :: PR: #15821
• fix: voice agent WebSocket URL host configuration (15813) by @svcnvidia-nemo-ci :: PR: #15823
• [codex] Fix NEST SSL model-support training batches (15812) by @svcnvidia-nemo-ci :: PR: #15824
• Fix duplex_stt.yaml format (15829) by @svcnvidia-nemo-ci :: PR: #15833
• Bump mako and nltk dependencies to address CVEs (15832) by @svcnvidia-nemo-ci :: PR: #15836
• fix decorator safe instantiation bug (15848) by @svcnvidia-nemo-ci :: PR: #15852
• add fixes (15850) by @svcnvidia-nemo-ci :: PR: #15851
• Update Libsox Installation in Docker (15834) by @svcnvidia-nemo-ci :: PR: #15854
• fix align_eou (15871) by @svcnvidia-nemo-ci :: PR: #15874
• Update Dockerfile (15866) by @svcnvidia-nemo-ci :: PR: #15873
• fix ssl contrastive loss (15872) by @svcnvidia-nemo-ci :: PR: #15883
• chore: Update Python and pyarrow to address CVEs (15877) by @svcnvidia-nemo-ci :: PR: #15885
• fix publish-hf-notebook-xet (15878) by @svcnvidia-nemo-ci :: PR: #15881
• Fix speechlm2 tutorial config on automodel (15884) by @svcnvidia-nemo-ci :: PR: #15886
• Fix ASR transducer nemo checkpoint export (15876) by @svcnvidia-nemo-ci :: PR: #15880
• fix: NemotronVoiceChat.from_pretrained() crash on huggingface_hub 1.x (15892) by @svcnvidia-nemo-ci :: PR: #15896
• Set trusted checkpoint env for ASR tutorial restores (15875) by @svcnvidia-nemo-ci :: PR: #15903
• Fix/ctc segmentation tutorial deps (15901) by @svcnvidia-nemo-ci :: PR: #15904
• numba compatibility for SDE (15694) by @svcnvidia-nemo-ci :: PR: #15890
• chore: Bump aiohttp to >=3.14.0 (15909) by @svcnvidia-nemo-ci :: PR: #15910
• Fix bug/6408633 and bug/6071397 (15912) by @svcnvidia-nemo-ci :: PR: #15913
• [TTS] Whitelist TTS AudioTrimmer and ArtifactGenerator (15914) by @svcnvidia-nemo-ci :: PR: #15916
• Switch to TinyLlama-1.1B-Chat-v1.0 to resolve auto-conversion exception (bug/6071397) (15924) by @svcnvidia-nemo-ci :: PR: #15927
• chore: Bump pillow to 12.3.0 (15925) by @svcnvidia-nemo-ci :: PR: #15929
• Fix ASR tutorial notebook restore and setup paths (15908) by @svcnvidia-nemo-ci :: PR: #15935
• Disable RNNT CUDA graphs in Multilang ASR notebook (15937) by @svcnvidia-nemo-ci :: PR: #15938
• Fix ModelCheckpoint monitor key on duplex_stt.yaml (bug/6492704) (15939) by @svcnvidia-nemo-ci :: PR: #15942
• fix Online_ASR_Microphone_Demo_Buffered_Streaming.ipynb failed with ConfigKeyError (15964) by @svcnvidia-nemo-ci :: PR: #15970
• fix 01_NeMo_Models.ipynb hit UnsafeTargetError (15944) by @svcnvidia-nemo-ci :: PR: #15958
• fix ASR_with_Subword_Tokenization.ipynb fails: InterpolationKeyError 'model.sample_rate' (15945) by @svcnvidia-nemo-ci :: PR: #15946
• Fix 02_NeMo_Adapters.ipynb fails with UnsafeTargetError for notebook-… by @lilithgrigoryan :: PR: #15972
• Temporarily pin SALM Automodel dispatcher to torch (15963) by @svcnvidia-nemo-ci :: PR: #15969
• docs: qualify SALM Automodel parallelism support (15976) by @svcnvidia-nemo-ci :: PR: #15978
• fix(docker): AUT-1107 upgrade OpenSSL packages (15973) by @svcnvidia-nemo-ci :: PR: #15982
• [speechlm2] Temporarily support native Nemotron-V3 with Automodel 0.4 (15980) by @svcnvidia-nemo-ci :: PR: #15981
• Fix SALM Automodel rc0 compatibility (#15992) by @DongjiGao :: PR: #15997
• docs: Adding audio false positive docs link (16023) by @svcnvidia-nemo-ci :: PR: #16024
• docs: Prep 3.0.0 docs by @chtruong814 :: PR: #16037
• Add SetupToolFix from #15930 by @blisc :: PR: #16034
• update import statements in audio collection (#15345) by @blisc :: PR: #0
• [🤖]: Howdy folks, let's bump NeMo-Toolkit to 2.8.0rc0.dev0 ! (#15350) by @github-actions[bot] :: PR: #0
• Add Per-Stream Phrase Boosting for Cache-Aware RNN-T (#15344) by @artbataev :: PR: #0
• Disable nemo2 functional tests (#15354) by @chtruong814 :: PR: #0
• Clarify when to use TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD (#15353) by @chtruong814 :: PR: #0
• Add create_all_gather_group configuration option (#15253) by @jeffnvidia :: PR: #0
• Remove deprecated LLM, VLM, and diffusion tutorials (#15357) by @yaoyu-33 :: PR: #0
• Update changelog for r2.6.0 (#15366) by @github-actions[bot] :: PR: #0
• Add buffered inference support for SALM models (#15364) by @naymaraq :: PR: #0
• [voice agent] fix dependency for nemo26.02 (#15380) by @stevehuang52 :: PR: #0
• update yaml to include new location of losses that were removed in #15211 (#15384) by @blisc :: PR: #0
• Fix macro accuracy when changing labels (#15379) by @stevehuang52 :: PR: #0
• Transformation stage 0: NeMo docs revision (#15363) by @pzelasko :: PR: #0
• [bugfix] defaults to force_map_dataset=True to avoide duplicates as a (#15387) by @XuesongYang :: PR: #0
• CI: increase cer to 0.08 from 0.06 to avoid chance to fail. (#15395) by @XuesongYang :: PR: #0
• fixes nemo tutorial for loading non registered classes (#15398) by @nithinraok :: PR: #0
• default weights to false (#15397) by @nithinraok :: PR: #0
• Adapt to use env variable for adapter mixin model loading (#15406) by @nithinraok :: PR: #0
• fix: add missing docstring for transcript args in EncDecMultiTaskModel.forward() (#15396) by @stanley1208 :: PR: #0
• Transformation stage 1: Remove deprecated collections (#15378) by @pzelasko :: PR: #0
• Handle race condition in exp_manager's timer gracefully (#15409) by @pzelasko :: PR: #0
• Fix BNR 2.0 inference alignment error with input signal padding (#15388) by @ManasiRemane :: PR: #0
• unblock CI: increase to 0.08 from 0.03 (#15413) by @XuesongYang :: PR: #0
• Remove typing annotation breaking earlier python versions in speechlm2 (#15414) by @pzelasko :: PR: #0
• Fix cuda-python usage for CUDA graphs (#15416) by @artbataev :: PR: #0
• docs: fix broken tutorial links pointing to non-existent stable branch (#15420) by @stanley1208 :: PR: #0
• ci: Update release workflow to include generated changelog (#15429) by @chtruong814 :: PR: #0
• Specify device for codebook indices creation (#15405) by @Ferdydh :: PR: #0
• ci: Remove alerts (#15434) by @ko3n1g :: PR: #0
• ci: Add-credentials-for-docs (#15439) by @ko3n1g :: PR: #0
• ci: Pass app id to release job (#15442) by @chtruong814 :: PR: #0
• Implement Duplex Speech-to-text model (#15092) by @kevinhu-nv :: PR: #0
• Add a functional test of frame stacking (#15424) by @rfejgin :: PR: #0
• docs: Enable docs self-publishing including nightly update (#15432) by @chtruong814 :: PR: #0
• Implement workaround for Japanese transcript normalization (#15448) by @subhankar-ghosh :: PR: #0
• Auto-detect use_bucketing and validate batch size config (#15461) by @pzelasko :: PR: #0
• Script to migrate .nemo checkpoints that used torchaudio preprocessor (#15437) by @pzelasko :: PR: #0
• fix for nemotron-speech-streaming-0.6b (#15473) by @stevehuang52 :: PR: #0
• fix: replace assert statements with raise ValueError in EncDecMultiTaskModel (#15444) by @stanley1208 :: PR: #0
• Projection connector (#15468) by @nune-tadevosyan :: PR: #0
• Update SpeechLM2 collection documentation and fix dataloader custom audio resample issue (#15478) by @Edresson :: PR: #0
• End-of-Utterance metric (#15462) by @rfejgin :: PR: #0
• ci: Disable selective triggering of tests (#15484) by @chtruong814 :: PR: #0
• update transformers version (#15365) by @nithinraok :: PR: #0
• Add Zenodo download URLs to broken links false positives (#15488) by @XuesongYang :: PR: #0
• [lhotse] Added support for re-weighting datasets with temperature on the fly. (#15200) by @XuesongYang :: PR: #0
• Bugfix: add new EoU metric names to metric CSV file's header (#15493) by @rfejgin :: PR: #0
• Add functional tests for init, train step, and inference for every supported released model (#15433) by @pzelasko :: PR: #0
• Fix E2E nightly tests: move missing conftest.py (#15501) by @pzelasko :: PR: #0
• ja longform fix and StrEnum import fix (#15499) by @subhankar-ghosh :: PR: #0
• remove protobuf from reqs (#15502) by @nithinraok :: PR: #0
• Minor cleanup in EoU classifier unit test (#15511) by @rfejgin :: PR: #0
• add claude review (#15515) by @nithinraok :: PR: #0
• Fix BlendableDataset crash in non-distributed envs (#15465) (#15505) by @v4xsh :: PR: #0
• zero shot disabling magpie (#15503) by @paarthneekhara :: PR: #0
• add claude answer and fix (#15518) by @nithinraok :: PR: #0
• revert code owners (#15521) by @nithinraok :: PR: #0
• Add seen speaker test. (#15510) by @subhankar-ghosh :: PR: #0
• Fix e2e nightly tests: move shared utils out of conftest.py (#15519) by @pzelasko :: PR: #0
• Verbose fix in AED models (#15509) by @nune-tadevosyan :: PR: #0
• Update README (#15495) by @blisc :: PR: #0
• Fix a broken entry in README.md after re-write (#15527) by @pzelasko :: PR: #0
• chore: test FW-CI-templates ko3n1g/fix/linkcheck-retry-backoff (#15525) by @ko3n1g :: PR: #0
• bump release workflow (#15532) by @ko3n1g :: PR: #0
• Script for saving audio augmented by Lhotse samplers/dataloaders (#14808) by @racoiaws :: PR: #0
• Fix numba-cuda and cuda-python installation and usage (#15506) by @artbataev :: PR: #0
• Add AIS batch loading support to LhotseSpeechToTextBpeDataset (#15538) by @gaikwadabhishek :: PR: #0
• Fix IsADirectoryError when cleaning up unfinished distributed checkpoints (#15541) by @pzelasko :: PR: #0
• Changed the documentation getting started structure (#15460) by @Ssofja :: PR: #0
• Add support for partial transcription prefix in the prompt (#15449) by @azziko :: PR: #0
• docs: Fix docs build by setting uv conflicts for cu12 vs cu13 (#15548) by @chtruong814 :: PR: #0
• ci: Update docs build job to exclude cu12 extra (#15553) by @chtruong814 :: PR: #0
• Rename index for attention prior weights (#15551) by @subhankar-ghosh :: PR: #0
• ci: upgrade GitHub Actions for Node.js 24 compatibility (#15537) by @ko3n1g :: PR: #0
• Add VoiceChat to README (#15547) by @blisc :: PR: #0
• chore: bump _code_freeze workflow to v0.86.0 (#15569) by @ko3n1g :: PR: #0
• bring back zeroshot disabling code (#15564) by @paarthneekhara :: PR: #0
• Fix Qwen3 SALM LoRA init (#15570) by @pzelasko :: PR: #0
• Permit text context with frame stacking (#15585) by @rfejgin :: PR: #0
• [Fix] Make cpWER calculation identical to meeteval (#15573) by @tango4j :: PR: #0
• docs: Fix docs link in README (#15594) by @chtruong814 :: PR: #0
• README.md: next and last stable release info (#15602) by @pzelasko :: PR: #0
• Mask padded frames after feature normalization (#15562) by @arushidNV :: PR: #0
• fix use_feat_cache feature (#15600) by @naymaraq :: PR: #0
• add a unified model link to readme (#15604) by @andrusenkoau :: PR: #0
• README: refer to "vanilla" 26.02 container (#15606) by @pzelasko :: PR: #0
• fix: typo in readme (#15539) by @L0RD-ZER0 :: PR: #0
• Limit to peft<=0.18.0 until CI environment is updated (#15610) by @pzelasko :: PR: #0
• add agent skill for debugging distributed training log failures (#15612) by @gaikwadabhishek :: PR: #0
• Fix codec e2e tests (#15607) by @rlangman :: PR: #0
• Implement formatter to ensure Nemotron VoiceChat speech decoder reproducibility, speed up training and support half precision inference (#15583) by @Edresson :: PR: #0
• SALM with NeMo Automodel integration for Nemotron Nano V3 LLM backbone (#15447) by @pzelasko :: PR: #0
• CLAUDE.md, PR babysitter, local dev skills + symlinks for Codex/Cursor (#15613) by @pzelasko :: PR: #0
• Bump protobufjs from 7.5.3 to 7.5.5 in /examples/voice_agent/client (#15619) by @dependabot[bot] :: PR: #0
• Bump mako from 1.3.10 to 1.3.11 (#15618) by @dependabot[bot] :: PR: #0
• Bump pillow from 12.1.1 to 12.2.0 (#15603) by @dependabot[bot] :: PR: #0
• Bump vite from 6.3.6 to 6.4.2 in /examples/voice_agent/client (#15586) by @dependabot[bot] :: PR: #0
• chore: bump FW-CI-templates workflow pins to v0.88.0 (#15581) by @svcnvidia-nemo-ci :: PR: #0
• Bump pygments from 2.19.2 to 2.20.0 (#15565) by @dependabot[bot] :: PR: #0
• Bump picomatch from 4.0.2 to 4.0.4 in /examples/voice_agent/client (#15554) by @dependabot[bot] :: PR: #0
• fix: add missing docstrings to data pipeline utility functions (#15560) by @stanley1208 :: PR: #0
• ci(babysitter): switch from check_run to workflow_run trigger (#15627) by @pzelasko :: PR: #0
• ci: restrict testpypi publish trigger to release branches (#15628) by @pzelasko :: PR: #0
• ci: onboard copy-pr-bot (#15631) by @ko3n1g :: PR: #0
• fix: Remove diskcache from requirements (#15630) by @chtruong814 :: PR: #0
• docs: Add SECURITY.md (#15636) by @chtruong814 :: PR: #0
• Add xatt trimming for multitask beam decoding (#15466) by @azziko :: PR: #0
• Remove nemo_log_globalrank-N_localrank-M.txt file creation (#15626) by @pzelasko :: PR: #0
• [Voice Agent] fix empty tools (#15634) by @stevehuang52 :: PR: #0
• fix: clamp shift perturbation for short audio instead of silently skipping (#15561) by @stanley1208 :: PR: #0
• to_hf.py + PromptFormatter: produce vLLM-ready SpeechLM checkpoints with backbone-native chat_template (#15623) by @DongjiGao :: PR: #0
• Add hi-IN , Ko-KR and pt-BR IPA tokenizer support (#15567) by @quapham :: PR: #0
• [Voice Agent] Update README (#15642) by @stevehuang52 :: PR: #0
• Dockerfile: build from the official PyTorch image (#15638) by @artbataev :: PR: #0
• Add Arabic char tokenizer and Japanese-English katakana support (#15614) by @quapham :: PR: #0
• feat(ci): migrate from Azure to AWS ephemeral runners (#15620) by @ko3n1g :: PR: #0
• [audio codec] Add support for Lhotse training format (#15622) by @rfejgin :: PR: #0
• Remove the use of pickle throughout codebase (#15629) by @blisc :: PR: #0
• GRPO fix for frame stacked models (#15643) by @shehzeen :: PR: #0
• SALMAutomodel - long-context support (chunking, AC) and batch of fixes (#15648) by @pzelasko :: PR: #0
• handle issue 15598 for remote code fix (#15652) by @nithinraok :: PR: #0
• fix: count_input_cfg_levels now resolves string file references (#15646) by @XuesongYang :: PR: #0
• ci: switch formatting CI to check-only and enable pre-commit hooks (#15656) by @chtruong814 :: PR: #0
• Updated estimate_token_bins.py script; remove deprecated speechlm1 scripts (#15655) by @pzelasko :: PR: #0
• [ci, build] fix: unblock testpypi publish — drop direct git URL dep, bump CI template to v0.94.1 (#15659) by @ko3n1g :: PR: #0
• Re-add codeowners (#15660) by @chtruong814 :: PR: #0
• [build] fix: restore nemo_automodel in requirements_speechlm2.txt (#15662) by @ko3n1g :: PR: #0
• Add vLLM support for NeMo SpeechLM (#15520) by @DongjiGao :: PR: #0
• Reduce code duplication in audio collection + some small fixes (#15587) by @racoiaws :: PR: #0
• ci: Major refactor of release-workflows (#15668) by @ko3n1g :: PR: #0
• val audio logging fix (#15688) by @paarthneekhara :: PR: #0
• Preserve SpeechLM perception checkpoint dtype (#15686) by @DongjiGao :: PR: #0
• ci: Fix CI approval queue (#15693) by @chtruong814 :: PR: #0
• Easymagpie ZeroShot disabling option using speaker encoder (#15639) by @paarthneekhara :: PR: #0
• Add verbose flag to UTMOSv2Calculator (#15699) by @rlangman :: PR: #0
• ci: remove build-docs and build-test-publish-wheel workflows (#15685) by @ko3n1g :: PR: #0
• Reformat files with black. Remove exclusion from black rules. (#15687) by @artbataev :: PR: #0
• Causal transformer (#15730) by @nithinraok :: PR: #0
• ci: validate release branch-rules (#15718) by @ko3n1g :: PR: #0
• Fix LazyNeMoIterator for non-tarred S3 audio data (#15732) by @melllinia :: PR: #0
• add freeze, unfreeze methods with experimental tag (#15729) by @nithinraok :: PR: #0
• Support encoder input chunking for SALM vLLM inference (#15716) by @pzelasko :: PR: #0
• ci: Install CI dependencies from uv lock (#15697) by @chtruong814 :: PR: #0
• Removing the dependency on Pyannote for Diarization and VAD (#15632) by @tango4j :: PR: #0
• fix for release nightly docs CI (#15735) by @Ssofja :: PR: #0
• Remove stale NeMo dependencies on uv base (#15734) by @pzelasko :: PR: #0
• Docs visual improvements (#15745) by @Ssofja :: PR: #0
• Adding Conformer encoder I/O-styled Transformer encoder (#15703) by @tango4j :: PR: #0
• Build Automodel compiled dependencies in CI image (#15737) by @pzelasko :: PR: #0
• [speechlm2] SALMAutomodel: THD (packed sequence) and context parallel support (#15679) by @pzelasko :: PR: #0
• add max_duration = 20.0 to lhotse training yaml configs. (#15696) by @XuesongYang :: PR: #0
• ci: bump _release_library.yml to v1.4.3 (#15749) by @ko3n1g :: PR: #0
• Fixing memory consumption issues with rel_pos in Transformer encoder (#15751) by @ipmedenn :: PR: #0
• Mark legacy navigation docs as orphaned (#15755) by @pzelasko :: PR: #0
• chore: Update dependencies to address CVEs and also update Dockerfile (#15756) by @chtruong814 :: PR: #0
• [Audio Codec] Lhotse data loading updates and fixes (#15742) by @rfejgin :: PR: #0
• Support CUDA 12 container builds (#15747) by @pzelasko :: PR: #0
• fix: Move code copy to last docker stage (#15761) by @chtruong814 :: PR: #0
• Confidence in RNN-T inference pipeline (buffered and cache-aware) (#15765) by @artbataev :: PR: #0
• Update README.md (#15766) by @pzelasko :: PR: #0
• Add batched streaming beam search for RNN-T (mALSD+mAES) and TDT (mALSD) (#15753) by @lilithgrigoryan :: PR: #0
• fix for true bfloat16 inference (#15763) by @naymaraq :: PR: #0
• adding community featured models (#15744) by @Ssofja :: PR: #0
• Refactor masking methods (#15739) by @MahmoudAshraf97 :: PR: #0
• add soe fixes in documentation (#15774) by @Ssofja :: PR: #0
• docs: update SpeechLM parallelism strategy docs (#15773) by @pzelasko :: PR: #0
• refactored Speaker Diarizarion collection's docs (#15647) by @Ssofja :: PR: #0
• fix: Move git refs for deep-ep and transformer engine to uv sources (#15758) by @chtruong814 :: PR: #0
• docs: overhaul installation instructions around uv + bring-your-own Python/PyTorch/CUDA (#15769) by @pzelasko :: PR: #0
• chore: Bump next release to be 3.0 (#15783) by @chtruong814 :: PR: #0
• [🤠]: Howdy folks, let's release NeMo r3.0.0 ! (#15788) by @github-actions[bot] :: PR: #0
• Remove weights_only=False (#15801) (#15804) by @svcnvidia-nemo-ci :: PR: #0
• chore: Bump lxml and gitpython to address CVEs (#15785) (#15789) by @svcnvidia-nemo-ci :: PR: #0
• Fix reported security vulnerabilities (#15803) (#15806) by @svcnvidia-nemo-ci :: PR: #0
• fix: Secure tar archive extraction (#15811) (#15821) by @svcnvidia-nemo-ci :: PR: #0
• fix: Replace direct hydra.utils.instantiate with safe_instantiate (#15802) (#15822) by @chtruong814 :: PR: #0
• Fix voice agent websocket host handling (#15813) (#15823) by @svcnvidia-nemo-ci :: PR: #0
• [codex] Fix NEST SSL model-support training batches (#15812) (#15824) by @svcnvidia-nemo-ci :: PR: #0
• Fix duplex_stt.yaml format (#15829) (#15833) by @svcnvidia-nemo-ci :: PR: #0
• Bump dependencies for CVE (#15832) (#15836) by @svcnvidia-nemo-ci :: PR: #0
• fix decorator safe instantiation bug (#15848) (#15852) by @svcnvidia-nemo-ci :: PR: #0
• add fixes (#15850) (#15851) by @svcnvidia-nemo-ci :: PR: #0
• Update Libsox Installation in Docker (#15834) (#15854) by @svcnvidia-nemo-ci :: PR: #0
• fix align_eou (#15871) (#15874) by @svcnvidia-nemo-ci :: PR: #0
• update dockerfile (#15866) (#15873) by @svcnvidia-nemo-ci :: PR: #0
• fix ssl contrastive loss (#15872) (#15883) by @svcnvidia-nemo-ci :: PR: #0
• chore: Update Python and pyarrow to address CVEs (#15877) (#15885) by @svcnvidia-nemo-ci :: PR: #0
• fix publish-hf-notebook-xet (#15878) (#15881) by @svcnvidia-nemo-ci :: PR: #0
• fix config on automodel (#15884) (#15886) by @svcnvidia-nemo-ci :: PR: #0
• update model (fix_6411583) (#15898) by @Jorjeous :: PR: #0
• remove deprecated proxies/resume_download from NemotronVoiceChat._from_pretrained (#15892) (#15896) by @svcnvidia-nemo-ci :: PR: #0
• chore: Bump aiohttp to >=3.14.0 (#15909) (#15910) by @svcnvidia-nemo-ci :: PR: #0
• Fix bug/6408633 and bug/6071397 (#15912) (#15913) by @svcnvidia-nemo-ci :: PR: #0
• Switch to TinyLlama-1.1B-Chat-v1.0 to resolve auto-conversion exception (bug/6071397) (#15924) (#15927) by @svcnvidia-nemo-ci :: PR: #0
• chore: Bump pillow to 12.3.0 (#15925) (#15929) by @svcnvidia-nemo-ci :: PR: #0
• fix-tutorial-install-line (#15879) (#15934) by @nithinraok :: PR: #0
• Fix monitor key on duplex_stt.yaml (#15939) (#15942) by @svcnvidia-nemo-ci :: PR: #0
• fix (#15944) (#15958) by @svcnvidia-nemo-ci :: PR: #0
• Fix 02_NeMo_Adapters.ipynb fails with UnsafeTargetError for notebook-local ResidualMLP (#15959) (#15972) by @lilithgrigoryan :: PR: #0
• Qualify SALM Automodel parallelism support (#15976) (#15978) by @svcnvidia-nemo-ci :: PR: #0
• fix(docker): upgrade OpenSSL packages (#15973) (#15982) by @svcnvidia-nemo-ci :: PR: #0
• Hotfix native Nemotron-V3 Automodel parallelization (#15980) (#15981) by @svcnvidia-nemo-ci :: PR: #0
• Fix SALM Automodel rc0 compatibility (#15992) (#15997) by @DongjiGao :: PR: #0
• Adding audio false positive docs link (#16023) (#16024) by @svcnvidia-nemo-ci :: PR: #0
• docs: Prep 3.0.0 docs (#16037) by @chtruong814 :: PR: #0
• Add SetupToolFix from #15930 (#16034) by @blisc :: PR: #0