1 Release published by 1 person

• v2.0.0 NVIDIA Neural Modules 2.0.0

  published Nov 14, 2024

37 Pull requests merged by 21 people

• Update import 'pytorch_lightning' -> 'lightning.pytorch'

  #11252 merged Nov 18, 2024

• ci: Fix release workflow

  #11286 merged Nov 17, 2024

• Integrate lm-eval-harness for evaluations in NeMo

  #10621 merged Nov 16, 2024

• Create phi3mini.py

  #11281 merged Nov 16, 2024

• remove redundant docs

  #11302 merged Nov 15, 2024

• Revert "fix(export): GPT models w/ bias=False convert properly"

  #11301 merged Nov 15, 2024

• ci: Exclude CPU machines from scan

  #11300 merged Nov 15, 2024

• Add T5TTS

  #11193 merged Nov 15, 2024

• Fixes per comments

  #11280 merged Nov 15, 2024

• Add openai-gelu in gated activation for TRTLLM export

  #11293 merged Nov 15, 2024

• Fix head_size in NeMo to HF checkpoint converters for width pruned model support

  #11230 merged Nov 15, 2024

• Sync vfm branch with main branch

  #11288 merged Nov 14, 2024

• chore: Add changelog

  #11283 merged Nov 14, 2024

• Remove opencc upperbound

  #10909 merged Nov 14, 2024

• Use MegatronDataSampler in HfDatasetDataModule

  #11274 merged Nov 14, 2024

• Change default ckpt name

  #11277 merged Nov 14, 2024

• Handle _io_unflatten_object when _thread_local.output_dir is not available

  #11199 merged Nov 14, 2024

• Configure no restart validation loop in nl.Trainer

  #11029 merged Nov 13, 2024

• Fix Finetune Recipe

  #11267 merged Nov 13, 2024

• Add llama 3.1 recipes

  #11273 merged Nov 13, 2024

• update nemo1->2 conversion according to changes in main

  #11253 merged Nov 13, 2024

• Beam search algorithm implementation for TDT models

  #10903 merged Nov 13, 2024

• Update pruning and distillation tutorial notebooks

  #11091 merged Nov 13, 2024

• Advanced Diffusion Training Features

  #11246 merged Nov 13, 2024

• fix(export): update API for disabling device reassignment in TRTLLM for Aligner

  #10863 merged Nov 12, 2024

• ci: Run secrets detector on pull_request_target

  #11263 merged Nov 12, 2024

• fix(export): GPT models w/ bias=False convert properly

  #11255 merged Nov 12, 2024

• [Doc fixes] update file names, installation instructions, bad links

  #11045 merged Nov 12, 2024

• chore(beep boop 🤖): Bump MCORE_TAG=aded519... (2024-11-12)

  #11260 merged Nov 12, 2024

• Remove builder_opt param from trtllm-build for TensorRT-LLM >= 0.14.0

  #11259 merged Nov 12, 2024

• ci: Fix secrets detector

  #11205 merged Nov 12, 2024

• ci: Move bump mcore to templates

  #11229 merged Nov 12, 2024

• Fix finetuning datamodule resume

  #11187 merged Nov 12, 2024

• Handling tokenizer in PTQ for Nemo 2.0

  #11237 merged Nov 12, 2024

• Bump Dockerfile.ci (2024-11-12)

  #11254 merged Nov 12, 2024

• Hyena wrapper: Weight decay override function

  #11203 merged Nov 11, 2024

• Bump Dockerfile.ci (2024-11-11)

  #11247 merged Nov 11, 2024

33 Pull requests opened by 23 people

• Lhotse support for transcribe_speech_parallel

  #11249 opened Nov 11, 2024

• PTQ memory optimization

  #11257 opened Nov 12, 2024

• Add option to set cp_comm_type

  #11258 opened Nov 12, 2024

• TE acceleration using callbacks

  #11261 opened Nov 12, 2024

• perf summary on homepage

  #11262 opened Nov 12, 2024

• Aligner/nemotron5

  #11264 opened Nov 12, 2024

• Add checklist for config validations

  #11265 opened Nov 12, 2024

• Introducing TensorRT lazy export and caching option with trt_compile()

  #11266 opened Nov 13, 2024

• Add option to change batch size if needed

  #11268 opened Nov 13, 2024

• Adding alinger export

  #11269 opened Nov 13, 2024

• chore: Add `devN` to semver

  #11271 opened Nov 13, 2024

• add inference

  #11275 opened Nov 13, 2024

• Enable ucc backend for pp

  #11276 opened Nov 13, 2024

• chore(beep boop 🤖): Bump `MCORE_TAG=64cbae5...` (2024-11-14)

  #11278 opened Nov 14, 2024

• Adding multimodal examples

  #11279 opened Nov 14, 2024

• Sortformer Diarizer 4spk v1 model PR Part 1: models, modules and dataloaders

  #11282 opened Nov 14, 2024

• Draft: Add bert Model to NeMo 2.0

  #11285 opened Nov 14, 2024

• AttributeError: module 'signal' has no attribute 'SIGKILL'

  #11287 opened Nov 14, 2024

• Add `attention_bias` argument in transformer block and transformer layer modules, addressing change in MCore

  #11289 opened Nov 14, 2024

• [nemo ux] cudagraph plugin

  #11290 opened Nov 14, 2024

• Huvu/t5 nemo2.0 nemoci

  #11291 opened Nov 14, 2024

• chore(beep boop 🤖): Bump `MCORE_TAG=4c4215f...` (2024-11-15)

  #11292 opened Nov 15, 2024

• Fix broken links

  #11294 opened Nov 15, 2024

• Draft: Interface for asymmetric pipeline schedule

  #11295 opened Nov 15, 2024

• chore: Add experimental directory

  #11296 opened Nov 15, 2024

• ci: Deploy containers

  #11298 opened Nov 15, 2024

• fix perf plugin CUDA_DEVICE_MAX_CONNECTIONS setting

  #11299 opened Nov 15, 2024

• Added clamp to SFT loss denominator

  #11304 opened Nov 15, 2024

• chore(beep boop 🤖): Bump `MCORE_TAG=63b8520...` (2024-11-16)

  #11305 opened Nov 16, 2024

• Remove pytorch-lightning

  #11306 opened Nov 16, 2024

• chore(beep boop 🤖): Bump `MCORE_TAG=ce507ee...` (2024-11-17)

  #11308 opened Nov 17, 2024

• Don't use nvidia_torch_version until after None check

  #11309 opened Nov 17, 2024

• chore(beep boop 🤖): Bump `MCORE_TAG=6c88bfc...` (2024-11-18)

  #11310 opened Nov 18, 2024

8 Issues closed by 2 people

• Unable to merge lora weights: "world_size (1) is not divisible by 4"

  #10782 closed Nov 17, 2024

• Resuming from a checkpoint that ended before the epoch ended and your dataloader is not resumable

  #10797 closed Nov 17, 2024

• [Question] Converting a Megatron-LM ckpt to Nemo

  #10831 closed Nov 17, 2024

• Conversion script of phi3 from HF to Nemo

  #10825 closed Nov 16, 2024

• Converting HF model to Nemo gets an error

  #10264 closed Nov 15, 2024

• Unable to decode using canary 1b model

  #10680 closed Nov 15, 2024

• [NeVa Pretraining] Vision Encoder Created on All GPUs During Pipeline Parallelism

  #10805 closed Nov 15, 2024

• Using MSDD model with a different speaker embedding model

  #10681 closed Nov 11, 2024

7 Issues opened by 7 people

• Japanese model name for Forced Aligner

  #11307 opened Nov 16, 2024

• OOM with RAM with Lhotse

  #11303 opened Nov 15, 2024

• Missing BOS tokens for HF tokenizer

  #11297 opened Nov 15, 2024

• using skip_nan_grad with gradient accumulation for ASR

  #11272 opened Nov 13, 2024

• Fail to convert Llama3 Nemo 2.0 checkpoint to HF

  #11256 opened Nov 12, 2024

• Initialize a Parakeet Cache-Aware Streaming model's encoder from an offline model.

  #11250 opened Nov 11, 2024

• How to convert downloaded local models into Nemo files？

  #11248 opened Nov 11, 2024

57 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.

• NeMo 2.0 SFT PEFT notebooks

  #10874 commented on Nov 17, 2024 • 33 new comments

• Add slimpajama example

  #10671 commented on Nov 15, 2024 • 29 new comments

• Fix: Data from AIStore

  #11241 commented on Nov 15, 2024 • 14 new comments

• Profiling - support Chakra & Kineto trace dumping

  #11115 commented on Nov 17, 2024 • 13 new comments

• Add StragglerDetection and FTlauncher to NeMo2.0

  #11117 commented on Nov 15, 2024 • 9 new comments

• Add support for NeMo Run to ASR

  #10933 commented on Nov 16, 2024 • 9 new comments

• Add PP support in NeVA along with few bug fixes

  #11170 commented on Nov 15, 2024 • 8 new comments

• [NeMo-UX] Support `load_strictness`

  #10612 commented on Nov 15, 2024 • 5 new comments

• [NeMo-UX] Add option to drop optimizer states

  #11089 commented on Nov 15, 2024 • 5 new comments

• nemo2 peft merge

  #11017 commented on Nov 17, 2024 • 3 new comments

• fix: regular torch optims (e.g., sgd) no longer error with closure spec

  #11189 commented on Nov 13, 2024 • 2 new comments

• Add MCore FSDP2 support

  #11216 commented on Nov 14, 2024 • 2 new comments

• Add recipe configs validating

  #10954 commented on Nov 15, 2024 • 1 new comment

• add JitTransform

  #11131 commented on Nov 12, 2024 • 1 new comment

• Fast N-Gram LM on GPU + greedy decoding (RNN-T, TDT, CTC)

  #10989 commented on Nov 14, 2024 • 0 new comments

• EMMeTT support in SpeechLLM + tutorial for Lhotse Multimodal Dataloading

  #10927 commented on Nov 16, 2024 • 0 new comments

• fix: properly use enhanced_count_thres and max_num_speakers in diariz…

  #11049 commented on Nov 17, 2024 • 0 new comments

• Update T5 to support new Mcore T5 attention mask shape

  #11059 commented on Nov 16, 2024 • 0 new comments

• Breakout get_checkpoint from AutoResume

  #11060 commented on Nov 13, 2024 • 0 new comments

• update attention

  #11235 commented on Nov 14, 2024 • 0 new comments

• Disable CUDA graphs in DDP (ASR). Improve toggle messages

  #11087 commented on Nov 14, 2024 • 0 new comments

• Akoumparouli/destructor singleton in megatronstrategy dtor

  #11093 commented on Nov 14, 2024 • 0 new comments

• [Bugfix] fix qwen tokenizer config when converting to nemo format

  #11098 commented on Nov 13, 2024 • 0 new comments

• DoRA

  #11104 commented on Nov 15, 2024 • 0 new comments

• Add scripts for importing a ckpt and running a forward step on it for nemo.collections.llm

  #11108 commented on Nov 15, 2024 • 0 new comments

• config hierarchy

  #11145 commented on Nov 13, 2024 • 0 new comments

• Nemo2 batcheval

  #11158 commented on Nov 15, 2024 • 0 new comments

• add nemotron5 conversion

  #11171 commented on Nov 12, 2024 • 0 new comments

• Use Mcore ModelParallelConfig in strategy parallelism property

  #11232 commented on Nov 13, 2024 • 0 new comments

• NeMo 2.0 In-framework deployment support

  #11233 commented on Nov 11, 2024 • 0 new comments

• How can I get stt_en_fastconformer_ctc_small pretrain model??

  #11204 commented on Nov 11, 2024 • 0 new comments

• Eval_beamsearch_ngram_ctc throws got an unexpected keyword argument 'logprobs'

  #10175 commented on Nov 13, 2024 • 0 new comments

• Sortformer Integration Release Inquiry

  #10491 commented on Nov 13, 2024 • 0 new comments

• NameError: name' flash_attn_with_kvcache 'is not defined

  #11200 commented on Nov 13, 2024 • 0 new comments

• Converting trained llama 2 checkpoint to hf gives "invalid key" error

  #10884 commented on Nov 14, 2024 • 0 new comments

• Add Hydrarunner to oomptimizer

  #10882 commented on Nov 14, 2024 • 0 new comments

• SFT stage use context parallel with flash attention error

  #10876 commented on Nov 14, 2024 • 0 new comments

• fastconformer hybrid recipe reports strange val_WER with `nemo:24.07` and `nemo:dev`

  #10299 commented on Nov 14, 2024 • 0 new comments

• Allow OOMtimizer tokenizer point towards just parent directory

  #10870 commented on Nov 14, 2024 • 0 new comments

• `IPython` should be included in the requirements

  #10772 commented on Nov 14, 2024 • 0 new comments

• global batch size at different sequence length

  #10905 commented on Nov 16, 2024 • 0 new comments

• Link Not Found at Mamba Tutorial

  #10899 commented on Nov 16, 2024 • 0 new comments

• ASR - WER not decreasing after certain point (Finetuning hybrid_cache_aware_streaming model)

  #10578 commented on Nov 16, 2024 • 0 new comments

• Modules fail for Dreambooth example

  #10888 commented on Nov 17, 2024 • 0 new comments

• Efficient streaming decoding for RNN-T and TDT (support partial hypotheses)

  #9106 commented on Nov 12, 2024 • 0 new comments

• [TTS] Voicebox for Speech Editing and Zero-Shot TTS

  #10312 commented on Nov 15, 2024 • 0 new comments

• Fix trascribe speech parralel with tarred datasets

  #10372 commented on Nov 14, 2024 • 0 new comments

• Use NCCL bootsrap backend for TP communication overlaps

  #10622 commented on Nov 15, 2024 • 0 new comments

• Add support for limit_train_batches to megatron sampler classes

  #10648 commented on Nov 17, 2024 • 0 new comments

• Context Parallel SFT Support for dataset in THD format

  #10688 commented on Nov 14, 2024 • 0 new comments

• multiturn training support for SALM

  #10759 commented on Nov 16, 2024 • 0 new comments

• Fix checkpoint loading when lm_head is on separate pipeline stage

  #10769 commented on Nov 15, 2024 • 0 new comments

• replace `SIGKILL` with `SIGTERM`

  #10777 commented on Nov 12, 2024 • 0 new comments

• [WIP] Migrate SpeechLM to NeMo 2.0

  #10808 commented on Nov 14, 2024 • 0 new comments

• Fix: When training ASR models, it saves .nemo 2 times in a row when save_last is True

  #10814 commented on Nov 13, 2024 • 0 new comments

• Added averaging script for torch dist

  #10834 commented on Nov 12, 2024 • 0 new comments

• Replace usage of np.sctypes with np.issubdtype

  #10839 commented on Nov 12, 2024 • 0 new comments