1 Release published by 1 person

• v2.2.1 NVIDIA Neural Modules 2.2.1

  published Mar 31, 2025

25 Pull requests merged by 15 people

• Update base container in Dockerfile.speech

  #12859 merged Apr 2, 2025

• ci: Increase prune time

  #12860 merged Apr 2, 2025

• [automodel] qlora peft

  #12817 merged Apr 2, 2025

• Update resiliency example notebook readme and add links to the brev launchable

  #12843 merged Apr 2, 2025

• [audio] Adding tests for predictive models

  #12823 merged Apr 2, 2025

• Fix timestamps when cuda graphs enabled

  #12808 merged Apr 2, 2025

• Fix trt-llm install

  #12827 merged Apr 1, 2025

• Prune docker images in GHA older than 8hrs

  #12838 merged Apr 1, 2025

• add missing call to _apply_liger_kernel_to_instance

  #12806 merged Apr 1, 2025

• [automodel] add cpu:gloo to backend

  #12832 merged Apr 1, 2025

• TP optimization and code cleanup for llama3 70b LoRA

  #12488 merged Apr 1, 2025

• add finetune support for Auto Configurator

  #12770 merged Apr 1, 2025

• transcribe fix for new hypotheses

  #12801 merged Mar 31, 2025

• Cherry pick Update changelog for r2.2.1 (12818) into r2.2.0

  #12820 merged Mar 31, 2025

• Update changelog for r2.2.1

  #12818 merged Mar 31, 2025

• Alit/hyena recipes

  #12582 merged Mar 31, 2025

• ci: Exclude nlp, mm, vision collections

  #12816 merged Mar 31, 2025

• add __init__.py to make this a package

  #12814 merged Mar 31, 2025

• ci: Move scripts fully down to files

  #12802 merged Mar 31, 2025

• ci: Remove --branch

  #12809 merged Mar 31, 2025

• Fix TransformerBlock cuda_graphs compatibility with MCore

  #12779 merged Mar 31, 2025

• Add BERT/Qwen2.5 Unit test and Refactor all GHA Conversion Tests

  #12785 merged Mar 31, 2025

• ci: Fix flaky LLM tests

  #12807 merged Mar 30, 2025

• ci: Measure multiprocessing

  #12778 merged Mar 29, 2025

• [doc] Fixes for audio doc warnings

  #12736 merged Mar 27, 2025

27 Pull requests opened by 22 people

• test

  #12803 opened Mar 28, 2025

• Non-blocking checkpoint cleanup failure

  #12804 opened Mar 28, 2025

• handle identical duration bins

  #12810 opened Mar 31, 2025

• ci: Horizontal fail-fast

  #12811 opened Mar 31, 2025

• Fix vad param confusion

  #12812 opened Mar 31, 2025

• ci: Bump dependencies

  #12819 opened Mar 31, 2025

• [automodel] Add FSDPv2-compatible context parallelism support.

  #12821 opened Mar 31, 2025

• Allow configuring signal for signal handling

  #12824 opened Mar 31, 2025

• [automodel] Add linear ce loss support

  #12825 opened Mar 31, 2025

• DeepseekV3 SFT finetuning perf config

  #12829 opened Apr 1, 2025

• Adds support for SFT and Llama + Qwen converters

  #12830 opened Apr 1, 2025

• Add energon dataset support for Qwen2VL

  #12831 opened Apr 1, 2025

• Handle CUDA_DEVICE_MAX_CONNECTIONS before job launch

  #12833 opened Apr 1, 2025

• [fault tolerance] Add local checkpointing support

  #12839 opened Apr 1, 2025

• Update LLaVA's next HF exporter to load ViT checkpoint from YAML

  #12841 opened Apr 1, 2025

• Fine tuning data module restructure for huggingface datasets

  #12842 opened Apr 1, 2025

• update streaming conformer

  #12846 opened Apr 1, 2025

• [not ready] Saves checkpoints at specified steps

  #12847 opened Apr 1, 2025

• [automodel] hsdp

  #12851 opened Apr 2, 2025

• Fix qwen2.5 1.5b configuration inheritance bug

  #12852 opened Apr 2, 2025

• Fix pack error when using custom prompt_template

  #12854 opened Apr 2, 2025

• Remove Unexecuted Hyena Code Paths

  #12856 opened Apr 2, 2025

• Improve evo2 dataset test and testability

  #12857 opened Apr 2, 2025

• Guard decord import and update nvidia-resiliency-ext

  #12861 opened Apr 2, 2025

• Add vocab size as attr to GPT and T5 Configs, use file name based logger in llm.gpt.data

  #12862 opened Apr 2, 2025

• Implement Speculative transform script for GPT models

  #12863 opened Apr 3, 2025

• [automodel] fix hellaswag tokenizer call

  #12864 opened Apr 3, 2025

3 Issues closed by 3 people

• HFAutoModelForCausalLM not compatible with latest nvidia-modelopt release. TypeError: _new__load_pretrained_model() missing 1 required positional argument: 'pretrained_model_name_or_path'

  #12836 closed Apr 3, 2025

• Model creates duplicate transcriptions

  #12442 closed Mar 30, 2025

• Broken offline mode of NeMo

  #11899 closed Mar 30, 2025

14 Issues opened by 14 people

• No signal.SIGKILL attribute error under MS Windows 11

  #12858 opened Apr 2, 2025

• generating manifest file for training titanet-large

  #12853 opened Apr 2, 2025

• Configuraion inheritance bug with Qwen2.5 1.5B

  #12849 opened Apr 2, 2025

• canary-1b-flash not generalizing

  #12845 opened Apr 1, 2025

• canary-1b-flash missing punctuation when timestamps are enabled

  #12844 opened Apr 1, 2025

• Cache Aware Streaming script yields different results for different batch_sizes

  #12840 opened Apr 1, 2025

• speechllm TP bug

  #12837 opened Apr 1, 2025

• Bug Report: Failed to Build causal-conv1d During NeMo Installation

  #12835 opened Apr 1, 2025

• Bug after computing timestamps and it's not supported for canary_model?

  #12834 opened Apr 1, 2025

• Deepseek-v2 pretrain recipe doesn't work

  #12828 opened Apr 1, 2025

• Is there any plan for supporting Qwen2.5 VL?

  #12813 opened Mar 31, 2025

• torch.distributed.DistNetworkError

  #12805 opened Mar 29, 2025

• Discrepancy in custom transcribe pipeline vs. `model.transcribe()` for QuartzNet model

  #12800 opened Mar 27, 2025

• BUG - ASR - Finetuned hybrid model timestamps

  #12799 opened Mar 27, 2025

66 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Batched beam search for transducers (RNN-T and TDT)

  #12729 commented on Apr 2, 2025 • 23 new comments

• add nemotron5

  #12660 commented on Apr 2, 2025 • 9 new comments

• WIP: S2S collection

  #12617 commented on Apr 2, 2025 • 7 new comments

• [automodel] add linear loss function

  #12772 commented on Apr 1, 2025 • 7 new comments

• Enable in-fw deployment for eval with OAI compatible server

  #12101 commented on Apr 2, 2025 • 6 new comments

• feat: add support for checkpointing with multistorageclient for Exp M…

  #12747 commented on Apr 2, 2025 • 4 new comments

• Adds Standard transformer flops formula to flops callback

  #12774 commented on Apr 2, 2025 • 3 new comments

• Add PerceiverLayer implementation

  #12618 commented on Mar 31, 2025 • 3 new comments

• Updated preprocess_text.py to fix NeMo-text-processing :: WARNING :: Your input is too long and could take a long time to normalize. Use split_text_into_sentences() to make the input shorter and then call normalize_list().

  #11780 commented on Apr 2, 2025 • 2 new comments

• TRT-LLM tests

  #12671 commented on Apr 3, 2025 • 2 new comments

• Mingyuanm/GitHub ci flux

  #12761 commented on Apr 2, 2025 • 2 new comments

• Commiting changes for nemo-run asr

  #12748 commented on Mar 31, 2025 • 2 new comments

• ONNX and TRT Export Test for LLM Embedding Models

  #12692 commented on Apr 3, 2025 • 1 new comment

• Add Blockwise FP8 to PTQ & EP to modelopt resume

  #12670 commented on Apr 2, 2025 • 1 new comment

• Add unit tests for evaluation

  #12705 commented on Mar 31, 2025 • 1 new comment

• RNNT Timestamps Fix & Timestamps in transcribe_speech_parallel

  #12707 commented on Apr 2, 2025 • 1 new comment

• Add packed sequences for HFDatasetDataModule

  #12551 commented on Mar 31, 2025 • 1 new comment

• Testing deploy_inframework and query_inframework scripts

  #12667 commented on Apr 3, 2025 • 1 new comment

• [Automodel] Add TP/SP support

  #12796 commented on Apr 2, 2025 • 1 new comment

• Fix GPT HF Exporter dtype and head_dim

  #12792 commented on Mar 31, 2025 • 1 new comment

• Fix bugs in `AudioToMelSpectrogramPreprocessor.input_example`

  #12063 commented on Mar 30, 2025 • 1 new comment

• Add safetensor option when saving and restoring models

  #11549 commented on Mar 30, 2025 • 1 new comment

• make a outer forward function for FluxControlnet

  #12662 commented on Apr 2, 2025 • 0 new comments

• sft nemo2.0 checkin

  #12794 commented on Mar 27, 2025 • 0 new comments

• Expose NCCL timeout via existing env var

  #12669 commented on Mar 31, 2025 • 0 new comments

• GB200 LLM performance scripts tuning

  #12791 commented on Apr 2, 2025 • 0 new comments

• Update ssm.py

  #12687 commented on Apr 3, 2025 • 0 new comments

• Update modelopt upperbound to 0.27

  #12788 commented on Apr 3, 2025 • 0 new comments

• Adding more doc-strings to megatron_parallel.py

  #12767 commented on Mar 31, 2025 • 0 new comments

• Bugfix eval test

  #12704 commented on Apr 1, 2025 • 0 new comments

• feat: add support for nemo 2.0 checkpointing with multistorageclient

  #12746 commented on Apr 1, 2025 • 0 new comments

• Adding tests for more coverage

  #12713 commented on Mar 28, 2025 • 0 new comments

• Expand test converage neva / mllama

  #12715 commented on Apr 2, 2025 • 0 new comments

• Export and Deploy Unit Tests

  #12717 commented on Apr 3, 2025 • 0 new comments

• Test TE-free path for peft

  #12733 commented on Apr 2, 2025 • 0 new comments

• Remove code related to canonical adapters since it's not supported in NeMo 2.0

  #12724 commented on Apr 2, 2025 • 0 new comments

• AED Decoding with N-Gram LM

  #12730 commented on Apr 2, 2025 • 0 new comments

• NeMo is not friendly to HF compatibility.

  #12166 commented on Mar 28, 2025 • 0 new comments

• when i use container to do sft for any model, it has context not found error

  #11825 commented on Mar 30, 2025 • 0 new comments

• russian optimal vocab_size

  #12740 commented on Apr 1, 2025 • 0 new comments

• Unable to find JitConfig & JitTransform in nemo/lightning/pytorch/callbacks

  #12710 commented on Apr 2, 2025 • 0 new comments

• Fix checkpoint loading when lm_head is on separate pipeline stage

  #10769 commented on Apr 1, 2025 • 0 new comments

• Add global state cleanup function

  #11172 commented on Apr 1, 2025 • 0 new comments

• Add nemo1 to nemo2 conversion for neva

  #11860 commented on Mar 30, 2025 • 0 new comments

• fix(huggingface-hub): allow offline mode

  #11901 commented on Mar 30, 2025 • 0 new comments

• Make TETransformerLayerAutocast Support Cuda Graph

  #12075 commented on Mar 31, 2025 • 0 new comments

• Remove getattr_proxy to avoid problematic edge cases

  #12176 commented on Apr 3, 2025 • 0 new comments

• Fix: 'IterableDatasetWrapper' has no len() when using Lhotse datasets

  #12190 commented on Mar 29, 2025 • 0 new comments

• Update L2_NeMo_2_NeMo_Mcore_Mixtral_bitexact to reenable failure on mismatch

  #12233 commented on Mar 30, 2025 • 0 new comments

• change loss return format so that it can work with calculate_per_token_loss

  #12459 commented on Mar 29, 2025 • 0 new comments

• Nemo run ipl

  #12470 commented on Apr 1, 2025 • 0 new comments

• first commit

  #12477 commented on Mar 30, 2025 • 0 new comments

• initialize model with metadata

  #12496 commented on Mar 28, 2025 • 0 new comments

• Remove adapter_path from base AutoResume and refactor PEFT checkpoint handling

  #12565 commented on Apr 2, 2025 • 0 new comments

• [resiliency] Add in process integration for Nemo2

  #12589 commented on Apr 1, 2025 • 0 new comments

• Minimize overhead from asynchronous checkpointing

  #12590 commented on Mar 29, 2025 • 0 new comments

• Heh/speech conv dev

  #12598 commented on Apr 1, 2025 • 0 new comments

• ci: pip-installable automodels examples

  #12609 commented on Mar 30, 2025 • 0 new comments

• Feature/wsd scheduler

  #12611 commented on Apr 2, 2025 • 0 new comments

• Hugging Face model deployment support

  #12628 commented on Apr 3, 2025 • 0 new comments

• Ptq parallel config dataclass

  #12632 commented on Apr 1, 2025 • 0 new comments

• Enable use_sharp argument

  #12636 commented on Apr 1, 2025 • 0 new comments

• Variable global and micro batch sizes for different GPUs

  #12640 commented on Mar 29, 2025 • 0 new comments

• [automodel] expose gradient clip value for automodel recipe

  #12648 commented on Mar 31, 2025 • 0 new comments

• Alit/fw eval nm5 ux

  #12658 commented on Apr 2, 2025 • 0 new comments

• FSDP2 support for MultiTask AED models (Canary)

  #12661 commented on Apr 3, 2025 • 0 new comments