update speechllm (#8486)

* fix(clustering_diarizer.py): fix typo (#7772)

Signed-off-by: Jean-Louis Queguiner <[email protected]>

* fix(diarization-README): typo (#7771)

Signed-off-by: Jean-Louis Queguiner <[email protected]>

* Fix bug wrt change decoding strategy for bpe models (#7762) (#7764)

* Fix bug wrt change decoding strategy for bpe models



* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: smajumdar <[email protected]>
Co-authored-by: Somshubra Majumdar <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Remove incorrect extra argument for load_from_checkpoint_dir() (#7500)

Signed-off-by: Robin Dong <[email protected]>
Co-authored-by: Eric Harper <[email protected]>

* Add nemo to mcore GPT conversion script  (#7730)

* add conversion script

Signed-off-by: Chen Cui <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove references to 'ckpt'

Signed-off-by: Chen Cui <[email protected]>

* add one more sanity check to make sure there is no unexpected keys in state dict

Signed-off-by: Chen Cui <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* make cpu loading work

Signed-off-by: Chen Cui <[email protected]>

* make script work for llama2 models

Signed-off-by: Chen Cui <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* address code check

Signed-off-by: Chen Cui <[email protected]>

* remove trainer precision (was for old sanity check)

Signed-off-by: Chen Cui <[email protected]>

* fix script for llama2 model

Signed-off-by: Chen Cui <[email protected]>

* remove commented code

Signed-off-by: Chen Cui <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Chen Cui <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>

* Fix bug in ConditionalInput: cat along the feature dim, not the batch dim (#7785)

Signed-off-by: anferico <[email protected]>

* Add some docs and update scripts for ASR (#7790)

* Add some docs and update scripts

Signed-off-by: smajumdar <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: smajumdar <[email protected]>
Signed-off-by: Somshubra Majumdar <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* set context for text memmap to fork (#7784)

* set context for text memmap to fork

Signed-off-by: arendu <[email protected]>

* typo

Signed-off-by: arendu <[email protected]>

---------

Signed-off-by: arendu <[email protected]>

* add training with multiple audios

Signed-off-by: stevehuang52 <[email protected]>

* Support flash decoding (#7744)

* Add flash-decoding

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

---------

Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Yang Zhang <[email protected]>

* Change accelerator to 'auto' in nlp_checkpoint_port.py (#7761)

* Change accelerator to 'auto' in nlp_checkpoint_port.py (#7747)

* Change accelerator to auto

Signed-off-by: Abhishree <[email protected]>

* Pass omegaconf object to trainer in nlp_checkpoint_port.py

Signed-off-by: Abhishree <[email protected]>

* Pass omegaconf object to trainer in export.py

Signed-off-by: Abhishree <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Abhishree <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>
Signed-off-by: Abhishree <[email protected]>

* docs: fix typos (#7758)

Signed-off-by: shuoer86 <[email protected]>
Co-authored-by: Xuesong Yang <[email protected]>
Signed-off-by: Abhishree <[email protected]>

* Snake act (#7736)

Signed-off-by: Abhishree <[email protected]>

* Update gpt_dataset.py (#6963)

Signed-off-by: Xin Yao <[email protected]>
Co-authored-by: Sandeep Subramanian <[email protected]>
Signed-off-by: Abhishree <[email protected]>

---------

Signed-off-by: Abhishree <[email protected]>
Signed-off-by: shuoer86 <[email protected]>
Signed-off-by: Xin Yao <[email protected]>
Co-authored-by: Abhishree Thittenamane <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>
Co-authored-by: shuoer86 <[email protected]>
Co-authored-by: Xuesong Yang <[email protected]>
Co-authored-by: Nithin Rao <[email protected]>
Co-authored-by: Xin Yao <[email protected]>
Co-authored-by: Sandeep Subramanian <[email protected]>

* Add selection criteria for reference audios in the `GlobalStyleToken` submodule (#7788)

* add selection criteria for reference audios

Signed-off-by: anferico <[email protected]>

* Update configuration files

Signed-off-by: anferico <[email protected]>

* add informative comment in config files

Signed-off-by: anferico <[email protected]>

* sample random index for reference audio selection

Signed-off-by: anferico <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: anferico <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* update text server to support compute logprobs (#7733)

* update text server to support compute logprobs

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix typo

---------

Signed-off-by: Zhilin Wang <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* add multi-layer feat extract and fix random question insertion

Signed-off-by: stevehuang52 <[email protected]>

* Configure MCore logger (#7781)

Signed-off-by: Mikołaj Błaż <[email protected]>

* Revert "PEFT eval fix (#7626) (#7638)" (#7693)

This reverts commit f03dd660bd26d88fd569e76c6f74b83a7c203ff9.

* remove TN from ctc_segm tut (#7807)

Signed-off-by: Evelina <[email protected]>

* [TTS] Support audio offsets in TTS data loaders (#7156)

* [TTS] Support audio offsets in TTS data loaders

Signed-off-by: Ryan <[email protected]>

* [TTS] Change docstring mentions of .pt to .npy

Signed-off-by: Ryan <[email protected]>

---------

Signed-off-by: Ryan <[email protected]>

* Update Apex install command in Dockerfile (#7794) (#7804)

* move core install to /workspace (#7706)



* update apex install in dockerfile



* use fetch head



---------

Signed-off-by: Abhinav Khattar <[email protected]>
Signed-off-by: eharper <[email protected]>
Co-authored-by: Eric Harper <[email protected]>
Co-authored-by: Abhinav Khattar <[email protected]>

* fix typo

Signed-off-by: stevehuang52 <[email protected]>

* Nemo to HF converter for LLaMA model (#7770)

* Create config_llama_truncate.yaml

Signed-off-by: Utkarsh <[email protected]>

* Add files via upload

Signed-off-by: Utkarsh <[email protected]>

* Update convert_nemo_llama_to_hf.py

Signed-off-by: Utkarsh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update config_llama_truncate.yaml

Signed-off-by: Utkarsh <[email protected]>

* Update convert_nemo_llama_to_hf.py

Signed-off-by: Utkarsh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update convert_nemo_llama_to_hf.py

Signed-off-by: Utkarsh <[email protected]>

* clean up trainer

* remove dependency on yaml config. load config from nemo file instead.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* enable ckpt saving into other precision formats

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* support 70b + cleanup qkv slice logic

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix bug

* move hf model folder code from comment to function and add instruction to run

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Utkarsh <[email protected]>
Signed-off-by: Chen Cui <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>
Co-authored-by: Chen Cui <[email protected]>

* Save best NeMo model only when necessary (#7836)

Signed-off-by: Ante Jukić <[email protected]>

* add guard if its a distributed checkpoint (#7845)

Signed-off-by: Gerald Shen <[email protected]>

* Fix tn duplex (#7808)

* fix duplex tn infer

Signed-off-by: Evelina <[email protected]>

* fix typo

Signed-off-by: Evelina <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix TN docs

Signed-off-by: Evelina <[email protected]>

---------

Signed-off-by: Evelina <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Update transformers cache on Jenkins (#7854)

* update transformers cache

Signed-off-by: eharper <[email protected]>

* update

Signed-off-by: eharper <[email protected]>

* add cd

Signed-off-by: eharper <[email protected]>

---------

Signed-off-by: eharper <[email protected]>

* Update README.rst for container update (#7844)

Signed-off-by: fayejf <[email protected]>

* Add support for finetuning with huggingface datasets (#7834)

* add finetune with huggingface dataset

Signed-off-by: stevehuang52 <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update yaml

Signed-off-by: stevehuang52 <[email protected]>

* update

Signed-off-by: stevehuang52 <[email protected]>

* update and refactor

Signed-off-by: stevehuang52 <[email protected]>

* add extrac hf text and update

Signed-off-by: stevehuang52 <[email protected]>

* update and refactor

Signed-off-by: stevehuang52 <[email protected]>

* move dataset dependency to common

Signed-off-by: stevehuang52 <[email protected]>

* add docstring

Signed-off-by: stevehuang52 <[email protected]>

* Add to Dics

Signed-off-by: Nithin Rao Koluguri <nithinraok>

* add ci test

Signed-off-by: Nithin Rao Koluguri <nithinraok>

* add max steps in jenkins

Signed-off-by: Nithin Rao Koluguri <nithinraok>

* reduce max steps

Signed-off-by: Nithin Rao Koluguri <nithinraok>

* jenkins test

Signed-off-by: Nithin Rao Koluguri <nithinraok>

* add bs=2

Signed-off-by: Nithin Rao Koluguri <nithinraok>

---------

Signed-off-by: stevehuang52 <[email protected]>
Signed-off-by: Nithin Rao Koluguri <nithinraok>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Nithin Rao Koluguri <nithinraok>
Co-authored-by: Nithin Rao <[email protected]>

* Multimodal merge (#7728)

* ControlNet TRT export

* Final MR before release

* SD2 update

* Fixed export issue

* Fix for instruct p2p and reformat

* Fix SD export issue

* Add nemo clip export for DB

* Fix ins pix2pix

* fix sd2 config

* [Mingyuan Ma] BF16 and SD conversion script

* [Imagen] NHWC Feature

* Fix .nemo loading issue for NeMo CLIP in SD

* NeMo r1.20.0 Multimodal Merge

* fix the inductor issue in inference

* Fix inductor loading .nemo issue

* Add Neva Model Support

* Imagen Optimizations

* Neva inference code

* NeMo TOT 1.21 to Internal/main

* Update neva_inference.yaml

* REBASING  for latest code changes

* Update internal/main to main tot

* Parallel DDIM implementation

* 1. Fixing indentation bug. (#7352)

Signed-off-by: Micha Livne <[email protected]>

* NeMo MCore llama2 support + MCore PEFT adapters (#7299)

* start adding gpt from megatron core path

Signed-off-by: ericharper <[email protected]>

* set model parallel config

Signed-off-by: ericharper <[email protected]>

* use model parallel config object

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update args

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* set vp size to none if it is 1

Signed-off-by: ericharper <[email protected]>

* set vp size to none if it is 1

Signed-off-by: ericharper <[email protected]>

* add TransformerConfig

Signed-off-by: ericharper <[email protected]>

* start updating to TransformerConfig

Signed-off-by: ericharper <[email protected]>

* add todo

Signed-off-by: ericharper <[email protected]>

* revert to model parallel config

Signed-off-by: ericharper <[email protected]>

* add hidden_size to model_parallel_config

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove imports

Signed-off-by: ericharper <[email protected]>

* revert

Signed-off-by: ericharper <[email protected]>

* remove import

Signed-off-by: ericharper <[email protected]>

* small clean up

Signed-off-by: ericharper <[email protected]>

* update hidden size in peft base model, add mcore commit to jenkins

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update module args

Signed-off-by: ericharper <[email protected]>

* add config obj to flash attention tests

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove args

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove sequence parallel arg

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* add config to self

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* add config to test

Signed-off-by: ericharper <[email protected]>

* get hidden_size from config

Signed-off-by: ericharper <[email protected]>

* add try except

Signed-off-by: ericharper <[email protected]>

* use default

Signed-off-by: ericharper <[email protected]>

* update config with hidden size

Signed-off-by: ericharper <[email protected]>

* remove arg

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* comment out jenkins test

Signed-off-by: ericharper <[email protected]>

* revert import

Signed-off-by: ericharper <[email protected]>

* build transformer config

Signed-off-by: ericharper <[email protected]>

* add model to provider func

Signed-off-by: ericharper <[email protected]>

* update forward and float16 wrapper

Signed-off-by: ericharper <[email protected]>

* instantiate model parallel config after init model parallel

Signed-off-by: ericharper <[email protected]>

* set virtual rank

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add GQA config to megatron gpt model (#7096)

* Add GQA config in gpt config file

Signed-off-by: jasonwan <[email protected]>

* Verify mcore is enabled when using GQA

Signed-off-by: jasonwan <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>

* revert

Signed-off-by: ericharper <[email protected]>

* mcore llama2 ckpt conversion & small fix

Signed-off-by: jasonwan <[email protected]>

* Add inference & sft config by Hongbin

Co-authored-by: Hongbin Liu <[email protected]>

Signed-off-by: jasonwan <[email protected]>

* fix config

Signed-off-by: jasonwan <[email protected]>

* add inference param. update TP/PP script to support mcore gpt

Signed-off-by: jasonwan <[email protected]>

* p-tuning

Signed-off-by: jasonwan <[email protected]>

* modify ckpt conversion script (adding model cast)

Signed-off-by: jasonwan <[email protected]>

* ckpt conversion use relative path for config

Signed-off-by: jasonwan <[email protected]>

* start adding gpt from megatron core path

Signed-off-by: ericharper <[email protected]>

* set model parallel config

Signed-off-by: ericharper <[email protected]>

* use model parallel config object

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* set vp size to none if it is 1

Signed-off-by: ericharper <[email protected]>

* set vp size to none if it is 1

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* add TransformerConfig

Signed-off-by: ericharper <[email protected]>

* start updating to TransformerConfig

Signed-off-by: ericharper <[email protected]>

* add todo

Signed-off-by: ericharper <[email protected]>

* revert to model parallel config

Signed-off-by: ericharper <[email protected]>

* add hidden_size to model_parallel_config

Signed-off-by: ericharper <[email protected]>

* remove imports

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove import

Signed-off-by: ericharper <[email protected]>

* small clean up

Signed-off-by: ericharper <[email protected]>

* update hidden size in peft base model, add mcore commit to jenkins

Signed-off-by: ericharper <[email protected]>

* update module args

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* add config obj to flash attention tests

Signed-off-by: ericharper <[email protected]>

* remove args

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove sequence parallel arg

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update args

Signed-off-by: ericharper <[email protected]>

* add config to self

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* add config to test

Signed-off-by: ericharper <[email protected]>

* get hidden_size from config

Signed-off-by: ericharper <[email protected]>

* add try except

Signed-off-by: ericharper <[email protected]>

* use default

Signed-off-by: ericharper <[email protected]>

* update config with hidden size

Signed-off-by: ericharper <[email protected]>

* remove arg

Signed-off-by: ericharper <[email protected]>

* comment out jenkins test

Signed-off-by: ericharper <[email protected]>

* revert import

Signed-off-by: ericharper <[email protected]>

* remove optimizer_idx

Signed-off-by: eharper <[email protected]>

* prefetch num microbatches

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* start adding gpt from megatron core path

Signed-off-by: ericharper <[email protected]>

* set model parallel config

Signed-off-by: ericharper <[email protected]>

* use model parallel config object

Signed-off-by: ericharper <[email protected]>

* update args

Signed-off-by: ericharper <[email protected]>

* fix for p-tuning sequence parallel

Signed-off-by: jasonwan <[email protected]>

* support SFT/distOpt mcore (#7207)

* add inference param. update TP/PP script to support mcore gpt

* p-tuning

Signed-off-by: jasonwan <[email protected]>

* change layer names for SFT

Signed-off-by: Hongbin Liu <[email protected]>

* fix bug in SFT

Signed-off-by: Hongbin Liu <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>
Signed-off-by: Hongbin Liu <[email protected]>
Co-authored-by: Hongbin Liu <[email protected]>
Co-authored-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* start updating to TransformerConfig

Signed-off-by: ericharper <[email protected]>

* revert to model parallel config

Signed-off-by: ericharper <[email protected]>

* add hidden_size to model_parallel_config

Signed-off-by: ericharper <[email protected]>

* remove imports

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update module args

Signed-off-by: ericharper <[email protected]>

* add config to self

Signed-off-by: ericharper <[email protected]>

* build transformer config

Signed-off-by: ericharper <[email protected]>

* add model to provider func

Signed-off-by: ericharper <[email protected]>

* update forward and float16 wrapper

Signed-off-by: ericharper <[email protected]>

* instantiate model parallel config after init model parallel

Signed-off-by: ericharper <[email protected]>

* set virtual rank

Signed-off-by: ericharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add GQA config to megatron gpt model (#7096)

* Add GQA config in gpt config file

Signed-off-by: jasonwan <[email protected]>

* Verify mcore is enabled when using GQA

Signed-off-by: jasonwan <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>

* revert

Signed-off-by: ericharper <[email protected]>

* remove import

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* rollback model cast for p-tuning

Signed-off-by: jasonwan <[email protected]>

* update for dist adam

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* use get_gpt_module_list

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update ckpt conversion script

Signed-off-by: jasonwan <[email protected]>

* ptl2.0 patch for llama config

Signed-off-by: jasonwan <[email protected]>

* add plugins to trainer in scripts

Signed-off-by: jasonwan <[email protected]>

* fix activation checkpointing mcore

Signed-off-by: jasonwan <[email protected]>

* fix variable names

Signed-off-by: jasonwan <[email protected]>

* overwrite normalization type for mcore/te

Signed-off-by: jasonwan <[email protected]>

* Update megatron_llama_sft.yaml

Signed-off-by: Jason Wang <[email protected]>

* add PEFT adapter support for mcore gpt path (#7276)

* implementation for mcore adapter/mxins

Signed-off-by: jasonwan <[email protected]>

* small fix for lora and ptuning

Signed-off-by: jasonwan <[email protected]>

* support layerwise peft

Signed-off-by: jasonwan <[email protected]>

* support multiple target layers

Signed-off-by: jasonwan <[email protected]>

* support lora GQA

Signed-off-by: jasonwan <[email protected]>

* support amp O2

Signed-off-by: jasonwan <[email protected]>

* revert & more O2 fix

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* lora inject to attention

Signed-off-by: jasonwan <[email protected]>

* support lora weight tying

Signed-off-by: jasonwan <[email protected]>

* add copyright header

Signed-off-by: jasonwan <[email protected]>

* rollback ptuning name change. full string match mcore target

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove comment

Signed-off-by: jasonwan <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* clean up config

Signed-off-by: jasonwan <[email protected]>

* Sync llama branch (#7297)

* add inference param. update TP/PP script to support mcore gpt

* p-tuning

Signed-off-by: jasonwan <[email protected]>

* change layer names for SFT

Signed-off-by: Hongbin Liu <[email protected]>

* fix bug in SFT

Signed-off-by: Hongbin Liu <[email protected]>

* fix bug: cpu initialization is not really enabled

Signed-off-by: Hongbin Liu <[email protected]>

* add use_cpu_initialization to TransformerConfig

Signed-off-by: Hongbin Liu <[email protected]>

* fix bug: wrong config path when using relative cjpt path

Signed-off-by: Hongbin Liu <[email protected]>

* revert mcore config change

Signed-off-by: Jason Wang <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>
Signed-off-by: Hongbin Liu <[email protected]>
Signed-off-by: Jason Wang <[email protected]>
Co-authored-by: Hongbin Liu <[email protected]>

* clean up ckpt conversion script

Signed-off-by: jasonwan <[email protected]>

* rollback git merge errors

Signed-off-by: jasonwan <[email protected]>

* update mcore, add check for mcore+te

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* formatting

Signed-off-by: jasonwan <[email protected]>

* make sft test dataset optional. fix indentation in config

Signed-off-by: jasonwan <[email protected]>

* one more fix for optional test set

Signed-off-by: jasonwan <[email protected]>

* support merging lora weights in mcore

Signed-off-by: jasonwan <[email protected]>

* update mcore for cpu init

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update ckpt conversion for code llama

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add seq_len_interpolation_factor support for long-context llama ckpts (#7312)

* add inference param. update TP/PP script to support mcore gpt

* p-tuning

Signed-off-by: jasonwan <[email protected]>

* add seq_len_interpolation_factor

Signed-off-by: Hongbin Liu <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>
Signed-off-by: Hongbin Liu <[email protected]>
Co-authored-by: jasonwan <[email protected]>
Co-authored-by: Hongbin Liu <[email protected]>

* fix old ptuning model, update mcore to support seq_len_interpolation_factor

Signed-off-by: jasonwan <[email protected]>

* support fused layernorm linear, fix ptuning O2

Signed-off-by: jasonwan <[email protected]>

* drop loss mask for mcore for now

Signed-off-by: jasonwan <[email protected]>

* disable dist ckpt in peft

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix loading non dist ckpt

Signed-off-by: jasonwan <[email protected]>

* add ckpt conversion to CI

Signed-off-by: jasonwan <[email protected]>

* update CI

Signed-off-by: jasonwan <[email protected]>

* mcore_mixin docstring

Signed-off-by: jasonwan <[email protected]>

* minor change in mcore peft error message

Signed-off-by: jasonwan <[email protected]>

* fix amp o2 in lora weight tying

Signed-off-by: jasonwan <[email protected]>

* correct mcore fp8 config

Signed-off-by: jasonwan <[email protected]>

* add TE installation

Signed-off-by: jasonwan <[email protected]>

* support mcore adapter tuning

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* comment out new CI test. rollback docker image

Signed-off-by: jasonwan <[email protected]>

* ignore FA tests, try new CI on 23.08

Signed-off-by: jasonwan <[email protected]>

* mark new CI as L2, put to beginning to test

Signed-off-by: jasonwan <[email protected]>

* minor fix for prompt learning

Signed-off-by: jasonwan <[email protected]>

* rollback to 23.06. comment out CI

Signed-off-by: jasonwan <[email protected]>

* minor fix ckpt conversion script

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* minor rollback gpt model change

Signed-off-by: jasonwan <[email protected]>

---------

Signed-off-by: ericharper <[email protected]>
Signed-off-by: jasonwan <[email protected]>
Signed-off-by: eharper <[email protected]>
Signed-off-by: Hongbin Liu <[email protected]>
Signed-off-by: Jason Wang <[email protected]>
Co-authored-by: ericharper <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: eharper <[email protected]>
Co-authored-by: Hongbin Liu <[email protected]>
Co-authored-by: Kelvin Liu <[email protected]>

* Hiddens modules documentation (#7303)

* 1. Changed hiddens transformations module from `transformations` to `hiddens`.

Signed-off-by: Micha Livne <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* 1. Finished doc.

Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

* 1. Debugging. Signed-off-by: Micha Livne <[email protected]>

---------

Signed-off-by: Micha Livne <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>

* Support for flash attention 2.0 (#7063)

* Add flash attn 2

Signed-off-by: MaximumEntropy <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add FA2 feature

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Remove debugging

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: MaximumEntropy <[email protected]>
Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Oleksii Kuchaiev <[email protected]>
Co-authored-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: Cheng-Ping Hsieh <[email protected]>

* lora merge fix for O2 names (#7325)

* wip

Signed-off-by: arendu <[email protected]>

* adjust key names based on O2

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update

Signed-off-by: arendu <[email protected]>

* minor

Signed-off-by: arendu <[email protected]>

---------

Signed-off-by: arendu <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* multiple fields can form a context (#7147)

* list of context fields and flexible prompt template

Signed-off-by: arendu <[email protected]>

* list of fields for context

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix bug

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Fix bug

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Add multiple truncation fields and middle truncation

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Compatible to old ckpt

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix tokenize detokenize issue

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Remove detokenization, add truncation augmentation

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Resolve comments

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Remove unused import

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* revert eos

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Add tokenizer space_sensitive attribute

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix error

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Fix erorr and use re

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix bug

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Change assert logic

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Follow adi suggestion

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Remove merge function

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add example and comment

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Remove context_key and add comment

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* Remove random truncation

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix bug

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix template none

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix bug

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

---------

Signed-off-by: arendu <[email protected]>
Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: Cheng-Ping Hsieh <[email protected]>

* Load buffers in checkpoint (#7357)

Signed-off-by: Jason Wang <[email protected]>

* Add migration guide for lightning 2.0 upgrade (#7360)

* Add lightning 2.0 migration guide in NeMo docs

Signed-off-by: Abhishree <[email protected]>

* Add remaining guide for lightning 2.0 upgrade

Signed-off-by: Abhishree <[email protected]>

* Remove line spill over and continue in next line

Signed-off-by: Abhishree <[email protected]>

* Add missing dataloader_iter in the guide

Signed-off-by: Abhishree <[email protected]>

* Fix minor typo

Signed-off-by: Abhishree <[email protected]>

---------

Signed-off-by: Abhishree <[email protected]>

* adding bias_dropout_add_fusion option for BERT (#7332)

Signed-off-by: Alexander Jipa <[email protected]>
Co-authored-by: Alexander Jipa <[email protected]>

* [TTS] Change audio codec token type to TokenIndex (#7356)

Signed-off-by: Ryan <[email protected]>

* enable selective unfreeze (#7326)

* wip

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* wip

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* avoid PTL method conflicts

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: arendu <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Fix typos (#7361)

* fix typos

Signed-off-by: omahs <[email protected]>

* fix typo

Signed-off-by: omahs <[email protected]>

* fix typos

Signed-off-by: omahs <[email protected]>

* fix typos

Signed-off-by: omahs <[email protected]>

* fix typo

Signed-off-by: omahs <[email protected]>

* fix typos

Signed-off-by: omahs <[email protected]>

* fix typo

Signed-off-by: omahs <[email protected]>

* fix typo

Signed-off-by: omahs <[email protected]>

* fix typo

Signed-off-by: omahs <[email protected]>

---------

Signed-off-by: omahs <[email protected]>

* pin numba=0.57.1 to fix reinstall.sh error (#7366)

Signed-off-by: Xuesong Yang <[email protected]>

* Update new conversion script for converting safetensors.

* Upgrade pytorch container to 23.08 (#7353)

* upgrade pytorch container

Signed-off-by: eharper <[email protected]>

* use mcore

Signed-off-by: eharper <[email protected]>

* revert test change

Signed-off-by: eharper <[email protected]>

* pleasefixme

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* check for ampere

Signed-off-by: eharper <[email protected]>

* comment test temporarily

Signed-off-by: eharper <[email protected]>

---------

Signed-off-by: eharper <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* enable fp32 optimizer for output_layer in mcore (#7355)

Signed-off-by: lhb8125 <[email protected]>

* revert comment (#7368)

Signed-off-by: eharper <[email protected]>

* Update to core 23.08 branch ToT (#7371)

Signed-off-by: Abhinav Khattar <[email protected]>

* upper bounding ptl (#7370)

Signed-off-by: eharper <[email protected]>

* fix pipeline parallel inference (#7367)

* fix pp inference

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: jasonwan <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* fix for peft tied weights (#7372)

Signed-off-by: arendu <[email protected]>

* fixed trainer.strategy=auto from None. (#7369)

Signed-off-by: Xuesong Yang <[email protected]>

* add O2 option in gpt eval (#7358)

* add O2 option in eval

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* add doc for O2 config

Signed-off-by: jasonwan <[email protected]>

* add to llama inference config

Signed-off-by: jasonwan <[email protected]>

---------

Signed-off-by: jasonwan <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>

* Move model precision copy (#7336)

* move cfg precision set to megatron base model

Signed-off-by: Maanu Grover <[email protected]>

* remove copy from other models

Signed-off-by: Maanu Grover <[email protected]>

* modify attribute not arg

Signed-off-by: Maanu Grover <[email protected]>

* fix gpt model test for ptl 2.0

Signed-off-by: Maanu Grover <[email protected]>

* rename function and add docstring

Signed-off-by: Maanu Grover <[email protected]>

* replace precision to dtype conditionals with func call

Signed-off-by: Maanu Grover <[email protected]>

* unnecessary function and cfg reset

Signed-off-by: Maanu Grover <[email protected]>

* set default value

Signed-off-by: Maanu Grover <[email protected]>

* fix precision lookup in a few more places

Signed-off-by: Maanu Grover <[email protected]>

* rename mapping function

Signed-off-by: Maanu Grover <[email protected]>

* ununsed import

Signed-off-by: Maanu Grover <[email protected]>

* save torch datatype to model

Signed-off-by: Maanu Grover <[email protected]>

* set weights precision wrt amp o2

Signed-off-by: Maanu Grover <[email protected]>

* Revert "set weights precision wrt amp o2"

This reverts commit 313a4bfe5eb69d771a6d2433898c0685836aef5c.

Signed-off-by: Maanu Grover <[email protected]>

* revert half precision at inference attempt

Signed-off-by: Maanu Grover <[email protected]>

* move autocast dtype to base model

Signed-off-by: Maanu Grover <[email protected]>

* move params dtype to base model, enable fp16 O2 inf

Signed-off-by: Maanu Grover <[email protected]>

* unused imports

Signed-off-by: Maanu Grover <[email protected]>

---------

Signed-off-by: Maanu Grover <[email protected]>

* Fix PEFT checkpoint loading (#7388)

* Fix PEFT checkpoint loading

Signed-off-by: Jason Wang <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Jason Wang <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Use distributed optimizer support for multiple dtypes (#7359)

* Update distopt wrapper with multiple dtype support

Remove manual handling of separate FP32 optimizer.

Signed-off-by: Tim Moon <[email protected]>

* Use distopt support for contiguous buffers with multiple dtypes

Signed-off-by: Tim Moon <[email protected]>

* Fix typo

Signed-off-by: Tim Moon <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Separate distopt buckets for first GPT layer and non-overlapped params

Signed-off-by: Tim Moon <[email protected]>

* Add distopt logic for int dtypes

Signed-off-by: Tim Moon <[email protected]>

* Update Apex commit

Signed-off-by: Tim Moon <[email protected]>

* Remove unused variables

Signed-off-by: Tim Moon <[email protected]>

* Update Apex commit in README and Jenkensfile

Signed-off-by: Tim Moon <[email protected]>

* Debug Dockerfile and Jenkinsfile

Signed-off-by: Tim Moon <[email protected]>

---------

Signed-off-by: Tim Moon <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Harper <[email protected]>

* minor fix for llama ckpt conversion script (#7387)

* minor fix for llama ckpt conversion script

Signed-off-by: Jason Wang <[email protected]>

* Update Jenkinsfile

Signed-off-by: Jason Wang <[email protected]>

* remove fast_swiglu configuration

Signed-off-by: Jason Wang <[email protected]>

---------

Signed-off-by: Jason Wang <[email protected]>
Co-authored-by: Eric Harper <[email protected]>

* Fix wrong calling of librosa.get_duration() in notebook (#7376)

Signed-off-by: Robin Dong <[email protected]>
Co-authored-by: Somshubra Majumdar <[email protected]>

* [PATCH] PEFT import mcore (#7393)

* [PATCH] PEFT import mcore

Signed-off-by: Jason Wang <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Jason Wang <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* [TTS] Added a callback for logging initial data (#7384)

Signed-off-by: Ante Jukić <[email protected]>

* Update Core Commit (#7402)

* Update Core Commit

Signed-off-by: Abhinav Khattar <[email protected]>

* update commit

Signed-off-by: Abhinav Khattar <[email protected]>

---------

Signed-off-by: Abhinav Khattar <[email protected]>

* Use cfg attribute in bert (#7394)

* use cfg attribute instead of arg

Signed-off-by: Maanu Grover <[email protected]>

* use torch_dtype in place of cfg.precision

Signed-off-by: Maanu Grover <[email protected]>

* move precision copy before super constructor

Signed-off-by: Maanu Grover <[email protected]>

* use trainer arg

Signed-off-by: Maanu Grover <[email protected]>

---------

Signed-off-by: Maanu Grover <[email protected]>

* Add support for bias conversion in Swiglu models (#7386)

* Add support for bias conversion in Swiglu models

Signed-off-by: smajumdar <[email protected]>

* Add support for auto extracting tokenizer model

Signed-off-by: smajumdar <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add support for auto extracting tokenizer model

Signed-off-by: smajumdar <[email protected]>

* Fix issue with missing tokenizer

Signed-off-by: smajumdar <[email protected]>

* Refactor

Signed-off-by: smajumdar <[email protected]>

* Refactor

Signed-off-by: smajumdar <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: smajumdar <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Update save_to and restore_from for dist checkpointing (#7343)

* add dist ckpt to save to, in progress

Signed-off-by: eharper <[email protected]>

* move dist ckpt

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* clean up

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update restore from, need to figure out how to initialize distributed

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* launch distrib if needed when restoring dist ckpt

Signed-off-by: eharper <[email protected]>

* when using mcore we can change tp pp on the fly

Signed-off-by: eharper <[email protected]>

* add load_from_checkpoint support for dist ckpt

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* update llama convert script to save dist .nemo

Signed-off-by: eharper <[email protected]>

* fix load dist ckpt

Signed-off-by: jasonwan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* setup TE TP groups if needed

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* setup te tp groups if needed

Signed-off-by: eharper <[email protected]>

* remove import

Signed-off-by: eharper <[email protected]>

---------

Signed-off-by: eharper <[email protected]>
Signed-off-by: jasonwan <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: jasonwan <[email protected]>

* fix forward for with mcore=false (#7403)

Signed-off-by: Jimmy Zhang <[email protected]>
Co-authored-by: Jimmy Zhang <[email protected]>

* Fix logging to remove 's/it' from progress bar in Megatron models and add train_step_timing (#7374)

* Add CustomProgressBar class to exp_manager and trainer callbacks

Signed-off-by: Abhishree <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix the progress bar to reflect total microbatch cnt

Signed-off-by: Abhishree <[email protected]>

* Modify CustomProgressBar class

1) Modify CustomProgressBar class to update progress bar per global_step instead of per microbatch
2) Add the callback to other megatron training/finetuning files that are not using MegatronTrainerBuilder

Signed-off-by: Abhishree <[email protected]>

* Add CustomProgressBar callback to tuning files

Signed-off-by: Abhishree <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Abhishree <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Set Activation Checkpointing Defaults (#7404)

* Set Activation Checkpointing Defaults

Signed-off-by: Abhinav Khattar <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* check for None

Signed-off-by: Abhinav Khattar <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Abhinav Khattar <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* make loss mask default to false (#7407)

Signed-off-by: eharper <[email protected]>

* Add dummy userbuffer config files (#7408)

Signed-off-by: Sangkug Lym <[email protected]>

* add missing ubconf files (#7412)

Signed-off-by: Abhinav Khattar <[email protected]>

* New tutorial on Speech Data Explorer (#7405)

* Added Google Colab based tutorial on Speech Data Explorer 

Signed-off-by: George Zelenfroynd <[email protected]>

* Update ptl training ckpt conversion script to work with dist ckpt (#7416)

* update ptl convert script

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* don't break legacy

Signed-off-by: eharper <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: eharper <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Allow disabling sanity checking when num_sanity_val_steps=0 (#7413)

* Allow disabling sanity checking when num_sanity_val_steps=0

Signed-off-by: Abhishree <[email protected]>

* Update num_sanity_val_steps to be a multiple of num_microbatches

Signed-off-by: Abhishree Thittenamane <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Abhishree <[email protected]>
Signed-off-by: Abhishree Thittenamane <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add comprehensive error messages (#7261)

Signed-off-by: Anton Peganov <[email protected]>

* check NEMO_PATH (#7418)

Signed-off-by: Nikolay Karpov <[email protected]>

* layer selection for ia3 (#7417)

* layer selection for ia3

Signed-off-by: arendu <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: arendu <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Fix missing pip package 'einops' (#7397)

Signed-off-by: Robin Dong <[email protected]>

* Fix failure of pyaudio in Google Colab (#7396)

Signed-off-by: Robin Dong <[email protected]>

* Update README.md: output_path --> output_manifest_filepath (#7442)

Signed-off-by: Samuele Cornell <[email protected]>

* Updating FlashAttention API to match FlashAttentionV2

* Multiple fixes for mm

* Fix CI inductor issue and update to torch compile

* Remove suppress error

* Fix when conversion config uses fp16 and it complains about precision plugin

* Fixing FAv2 API usage

* Initial release of content filtering model

* Added synthetic dataloader for precached and online mode

* Mingyuanm/dreambooth opt

* Add llama2 support in neva training

* Fix sampler length

* Fix all precision issues in nemo multimodal

* Add rope dynamic linear scaling (#7437)

* Add dynamic linear scaling

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix bug

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

---------

Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Yang Zhang <[email protected]>

* Fix None dataloader issue in PTL2.0 (#7455)

* Fix None dataloader issue in PTL2.0

Signed-off-by: KunalDhawan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* updating values of self._validation_dl and self._test_dl as well

Signed-off-by: KunalDhawan <[email protected]>

* updating values of self._validation_dl and self._test_dl as well

Signed-off-by: KunalDhawan <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: KunalDhawan <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* [ASR] Confidence measure -> method renames (#7434)

* measure -> method

Signed-off-by: Aleksandr Laptev <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Aleksandr Laptev <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add steps for document of getting dataset 'SF Bilingual Speech' (#7378)

* Add steps for document of getting dataset 'SF Bilingual Speech'

Signed-off-by: Robin Dong <[email protected]>

* Update datasets.rst

added a link from a tutorial demonstrating detailed data prep steps.

Signed-off-by: Xuesong Yang <[email protected]>

---------

Signed-off-by: Robin Dong <[email protected]>
Signed-off-by: Xuesong Yang <[email protected]>
Co-authored-by: Xuesong Yang <[email protected]>

* RNN-T confidence and alignment bugfix (#7381)

* new frame_confidence and alignments lists are now always created after the while loop

Signed-off-by: Aleksandr Laptev <[email protected]>

* tests added

Signed-off-by: Aleksandr Laptev <[email protected]>

---------

Signed-off-by: Aleksandr Laptev <[email protected]>

* Fix resume from checkpoint in exp_manager (#7424) (#7426)

Signed-off-by: Abhishree <[email protected]>
Co-authored-by: Abhishree Thittenamane <[email protected]>
Co-authored-by: Eric Harper <[email protected]>

* Fix checking of cuda/cpu device for inputs of Decoder (#7444)

* Fix checking of cuda/cpu device for inputs of Decoder

Signed-off-by: Robin Dong <[email protected]>

* Update tacotron2.py

Signed-off-by: Jason <[email protected]>

---------

Signed-off-by: Robin Dong <[email protected]>
Signed-off-by: Jason <[email protected]>
Co-authored-by: Jason <[email protected]>

* Fix failure of ljspeech's get_data.py (#7430)

* Fix failure of ljspeech's get_data.py

Signed-off-by: Robin Dong <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Robin Dong <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* [TTS] Fix audio codec type checks (#7373)

* [TTS] Fix audio codec type checks

Signed-off-by: Ryan <[email protected]>

* [TTS] Fix audio codec tests

Signed-off-by: Ryan <[email protected]>

---------

Signed-off-by: Ryan <[email protected]>

* [TTS] Add dataset to path of logged artifacts (#7462)

* [TTS] Add dataset to path of logged artifacts

Signed-off-by: Ryan <[email protected]>

* [TTS] Revert axis name back to Audio Frames

Signed-off-by: Ryan <[email protected]>

---------

Signed-off-by: Ryan <[email protected]>

* Fix sft dataset truncation (#7464)

* Add fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix

Signed-off-by: Cheng-Ping Hsieh <[email protected]>

---------

Signed-off-by: Cheng-Ping Hsieh <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Automatic Lip Reading Recognition (ALR) - ASR/CV (Visual ASR) (#7330)

* striding_conv1d_k5 and dw_striding_conv1d_k5 subsampling

Signed-off-by: mburchi <[email protected]>

* transpose conv1d inputs

Signed-off-by: mburchi <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

Signed-off-by: mburchi <[email protected]>

* Update subsampling.py

change striding_conv1d_k5 to striding_conv1d

Signed-off-by: Maxime Burchi <[email protected]>

* cv branch

Signed-off-by: mburchi <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* video manifest

Signed-off-by: mburchi <[email protected]>

* add collection classes

Signed-off-by: mburchi <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* add test_step_outputs

Signed-off-by: mburchi <[email protected]>

* correct manifest bug when having only audio or only videos

Signed-off-by: mburchi <[email protected]>

* correct manifest bug when having only audio or only videos

Signed-off-by: mburchi <[email protected]>

* clean references

Signed-off-by: mburchi <[email protected]>

* freeze unfreeze transcribe cv models

Signed-off-by: mburchi <[email protected]>

* correct manifest get_full_path bug

Signed-off-by: mburchi <[email protected]>

* update for PR

Signed-off-by: mburchi <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* guard torchvision

Signed-off-by: mburchi <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update nemo/collections/cv/data/video_to_text_dataset.py

Co-aut…

examples/multimodal/speechllm/README.md

@@ -1,6 +1,131 @@
-# Scripts for NeMo SpeechLM
+# Modular SpeechLLM
 
-This is the repository of the ICASSP'24 paper [SALM: Speech-augmented Language Model with In-context Learning for Speech Recognition and Translation
-](https://arxiv.org/abs/2310.09424)
+This directory contains example scripts to train and evaluate modular SpeechLLM models [1]. 
 
-We will release the scripts and checkpoints soon.
+## Requirements
+You will need to install this specific branch of NeMo, or use the provided Dockerfile in the root directory of this repository to build a Docker image with all the necessary dependencies. This branch is based on NeMo main branch by 2/14/2024, while diverging from the main branch in the following ways:
+- Migrating to pytorch_lightning==2.2 to fix some bugs with multiple validation dataloader_iter and saving -last.ckpt files.
+- Pinning to megatron-core==0.4.0 to avoid possible unstable behavior of the latest versions or not well supported NeMo components.
+
+
+## Architecture
+
+In general, there're three main components of a modular SpeechLLM model: 
+- An audio encoder that processes the input audio and produces a sequence of audio embeddings.
+- A modality adapter that processes the audio embeddings and produces a sequence of embeddings in the same latent space as the token embeddings of a pretrained large language model (LLM).
+- A pretrained large language model (LLM) that processes embeddings from the modality adapter as well as token embeddings of input prompt, and produces the text output. The audio embeddings and text token embeddings are concatenated in time dimension before going into the LLM.
+
+
+## Usage
+
+### Input Format
+
+You'll need to prepare data in the NeMo manifest format, where each line is a python dictionary with some keys, for example:
+```
+{
+    "audio_filepath": "path/to/audio.wav",
+    "offset": 0.0, # offset of the audio in seconds, this is an optional field
+    "duration": 10.0 , # duration of the audio in seconds, can set to `None` to load the whole audio
+    "question": "what is the transcription of the audio?", # this is an optional field, see below for more details
+    "answer": "the transcription of the audio", # optional for inference
+}
+```
+
+The `question` field in the manifest is optional, and you can put a list of questions in a file then set `++model.data.train_ds.question_file=<path to to question file>` to ask the dataloader to randomly pick a question from the file for each audio sample. This is useful for training with multiple prompts for the same task. If neither `question` field nor `question_file` is provided, the dataloader will use a default question `what does the audio mean?` for all aduios.
+
+
+### Training
+
+There are several configs for training a SpeechLLM:
+- `conf/modular_audio_gpt_config_peft.yaml`: a config for training a SpeechLLM model with PEFT (e.g., LoRA), where you don't want to tune the whole LLM but still want to adapt the LLM to your needs.
+- `conf/modular_audio_gpt_config_sft.yaml`: a config for training a SpeechLLM model without PEFT, where you might want to tune the whole LLM or simply freeze it and use as is.
+- `conf/modular_audio_gpt_multi_enc_config_peft.yaml`: a config for training a SpeechLLM model with multiple audio encoders and PEFT, where you can add speaker embeddings to the audio embeddings. Currently only TitaNet is supported as the speaker encoder.
+
+With any config, you can set the following flags to control which components to train or freeze:
+- `model.freeze_llm` # Generally set to `True` unless you want to fine-tune the whole LLM.
+- `model.freeze_audio_encoder` # Generally set to `False` unless you want to freeze the audio encoder.
+- `model.freeze_modality_adapter` # Generally set to `False` since we want to train the modality adapter.
+
+In addition to the config file, you will also need two prepare the audio encoder and the LLM as `*.nemo` files.
+
+To train a SpeechLLM model, you can run the following script:
+```bash
+MEGATRON_MODEL=/path/to/megatron-model.nemo
+ASR_MODEL=/path/to/audio-encoder.nemo
+
+TRAIN_MANIFESTS="[/data/train_1.json,/data/train_2.json]"
+VAL_MANIFESTS="[/data/dev_1.json,/data/dev_2.json]"
+VAL_NAMES="[dev-1,dev-2]"
+
+NVTE_FLASH_ATTN=0 \
+NVTE_FUSED_ATTN=0 \
+NVTE_MASKED_SOFTMAX_FUSION=0 \
+CUDA_VISIBLE_DEVICES="0,1" python modular_audio_gpt_train.py --config-path="./conf" --config-name "modular_audio_gpt_config_peft" \
+    trainer.devices=-1 \
+    model.freeze_audio_encoder=True \
+    model.freeze_llm=True \
+    model.global_batch_size=4 \  # global_batch_size = micro_batch_size * num_gpus_per_node * num_nodes * gradient_accumulation_steps
+    model.micro_batch_size=2 \  # micro_batch_size = batch_size_per_gpu
+    model.pretrained_audio_model=$ASR_MODEL \
+    model.restore_from_path=$MEGATRON_MODEL \
+    model.data.train_ds.manifest_filepath=$TRAIN_MANIFESTS \
+    model.data.validation_ds.manifest_filepath=$VAL_MANIFESTS \
+    ++model.data.validation_ds.names=$VAL_NAMES \
+```
+
+You can also use tarred datasets for faster training by converting normal NeMo datasets to tarred datasets using this [script](https://github.com/NVIDIA/NeMo/blob/main/scripts/speech_recognition/convert_to_tarred_audio_dataset.py) and follow the same dataset setting as shown in the script.
+
+
+#### Multi-task training
+In order to use a question file, you can set `++model.data.train_ds.question_file=<path to to question file>` in the command line or use multiple question files with `++model.data.train_ds.question_file=[<path to to question file1>,<path to question file2>,...]`. If the number of question files is equal to the number of provided datasets, the dataloader will assigne each question file to a dataset. Otherwise, the dataloader will randomly pick a question file from all provided question files for each audio sample. Using multiple question files is useful for training with multiple tasks, where each task has its own set of prompts. Meanwhile, you can control the weights for different tasks/datasets by using concatentated tarred datasets, where you can assign weights to datasets by:
+```
+++model.data.train_ds.is_tarred=True \
+++model.data.train_ds.is_concat=True \
+++model.data.train_ds.manifest_filepath=[/path/to/data1/tarred_audio_manifest.json,/path/to/data2/tarred_audio_manifest.json] \
+++model.data.train_ds.tarred_audio_filepaths=[/path/to/data1/audio__OP_0..1023_CL_.tar,/path/to/data2/audio__OP_0..1023_CL_.tar] \
+++model.data.train_ds.concat_sampling_technique='random' \
+++model.data.train_ds.concat_sampling_probabilities=[0.4,0.6] \
+```
+
+#### Available Audio Encoders
+Currently all NeMo ASR models are supported, others may also work if they have an `encoder` attribute that returns a sequence of audio embeddings, and a `preprocessor` that takes raw audios and returns a sequence of features for the encoder. The model should also have a `cfg` attribute that returns a `omegaconf.DictConfig` object of model configuration. In addition to a local model, you can also set `pretrained_audio_model` to a model from NGC (e.g., `stt_en_fastconformer_transducer_large`) or Huggingface (e.g., `nvidia/parakeet-rnnt-1.1b`), and the script will download the model and use it for training.
+
+
+### Inference
+
+The config file for inference is `conf/modular_audio_gpt_config_eval.yaml`, where you mainly need to set the `model.data.test_ds` fields. An example of running inference is shown below:
+
+```bash
+ASR_MODEL=/path/to/asr-model.nemo  # required only if you freeze the audio encoder during training
+MEGATRON_CKPT=/path/to/megatron-llm.nemo
+ALM_DIR=/path/to/nemo_experiments/job_name
+ALM_YAML=$ALM_DIR/version_0/hparams.yaml
+ALM_CKPT="$ALM_DIR/checkpoints/AudioGPT--validation_wer\=0.2-step\=100000-epoch\=0-last.ckpt"  # this checkpoint file only contains the trainable params
+
+VAL_MANIFESTS="[/data/libri-test-other.json,/data/MCV_7.1_test.json,/data/wsj-test.json]"
+VAL_NAMES="[ls-test-other,mcv7.1-test,wsj-test]"
+
+NVTE_MASKED_SOFTMAX_FUSION=0 \
+NVTE_FLASH_ATTN=0 \
+NVTE_FUSED_ATTN=0 \
+CUDA_VISIBLE_DEVICES=0 python modular_audio_gpt_eval.py \
+    model.restore_from_path=$MEGATRON_CKPT \
+    model.pretrained_audio_model=$ASR_MODEL \  # required only if you freeze the audio encoder during training
+    model.peft.restore_from_path=$ALM_CKPT \
+    model.peft.restore_from_hparams_path=$ALM_YAML \
+    model.data.test_ds.manifest_filepath=$VAL_MANIFESTS \
+    model.data.test_ds.names=$VAL_NAMES \
+    model.data.test_ds.global_batch_size=8 \
+	model.data.test_ds.micro_batch_size=8 \
+	model.data.test_ds.tokens_to_generate=256 \
+    ++inference.greedy=False \
+    ++inference.top_k=50 \
+    ++inference.top_p=0.95 \
+    ++inference.temperature=0.4 \
+    ++inference.repetition_penalty=1.2 \
+    ++model.data.test_ds.output_dir=${ALM_DIR}
+```
+
+
+## Reference
+[1] Chen, Z.\*, Huang, H.\*, Andrusenko, A., Hrinchuk, O., Puvvada, K.C., Li, J., Ghosh, S., Balam, J. and Ginsburg, B., 2023. SALM: Speech-augmented Language Model with In-context Learning for Speech Recognition and Translation. ICASSP'24.

examples/multimodal/speechllm/conf/speechllm/modularized_speech_gpt_config_eval.yaml → examples/multimodal/speechllm/conf/modular_audio_gpt_config_eval.yaml

@@ -1,13 +1,13 @@
-name: megatron_gpt_peft_${model.peft.peft_scheme}_eval
+name: megatron_audio_gpt_eval
 
 trainer:
   devices: 1
   accelerator: gpu
   num_nodes: 1
-  precision: 16
+  precision: bf16
   logger: False # logger provided by exp_manager
   enable_checkpointing: False
-  replace_sampler_ddp: False
+  use_distributed_sampler: False
   max_epochs: 9999
   max_steps: 20000 # consumed_samples = global_step * micro_batch_size * data_parallel_size * accumulate_grad_batches
   log_every_n_steps: 10 # frequency with which training steps are logged 
@@ -42,9 +42,9 @@ model:
 
   global_batch_size: 1
   micro_batch_size: 1
+  pretrained_audio_model: null  # Path to a .nemo model for audio encoder
   restore_from_path: ??? # Path to an existing .nemo model you wish to add new tasks to or run inference with
   resume_from_checkpoint: null # The path to a checkpoint file to continue the training, restores the whole state including the epoch, step, LR schedulers, apex, etc.
-  save_nemo_on_validation_end: True # Saves an inference ready .nemo file every time a checkpoint is saved during training. 
   sync_batch_comm: False
   megatron_amp_O2: False
 
@@ -68,28 +68,10 @@ model:
   attention_dropout: 0.0
   ffn_dropout: 0.0
 
-  peft:
-    peft_scheme: "adapter"  # can be either adapter,ia3, or ptuning
+  peft: # keep these basic params for reusing in both sft and peft SpeechLMs
     restore_from_path: null
     restore_from_ckpt_name: null
     restore_from_hparams_path: null
-
-    # Used for adapter peft training
-    adapter_tuning:
-      type: 'parallel_adapter' # this should be either 'parallel_adapter' or 'linear_adapter'
-      adapter_dim: 32
-      adapter_dropout: 0.0
-      norm_position: 'pre' # This can be set to 'pre' or 'post', 'pre' is normally what is used.
-      column_init_method: 'xavier' # IGNORED if linear_adapter is used, options: xavier, zero or normal
-      row_init_method: 'zero' # IGNORED if linear_adapter is used, options: xavier, zero or normal
-      norm_type: 'mixedfusedlayernorm' # IGNORED if layer_adapter is used,  options are ['layernorm', 'mixedfusedlayernorm']
-
-    # Used for p-tuning peft training
-    p_tuning:
-      virtual_tokens: 10  # The number of virtual tokens the prompt encoder should add at the start of the sequence
-      bottleneck_dim: 1024  # the size of the prompt encoder mlp bottleneck
-      embedding_dim: 1024  # the size of the prompt encoder embeddings
-      init_std: 0.023
 
   data:
     test_ds:
@@ -103,21 +85,21 @@ model:
       max_seq_length: 2048
       min_seq_length: 1
       drop_last: False
-      end_string: ${data.train_ds.end_string}
-      context_key: ${data.train_ds.context_key}
-      label_key: ${data.train_ds.label_key}
-      add_eos: ${data.train_ds.add_eos}
-      add_sep: ${data.train_ds.add_sep}
-      add_bos: ${data.train_ds.add_bos}
+      end_string: ${data.train_ds.end_string}  # don't change, let hydra resolve from saved config
+      context_key: ${data.train_ds.context_key} # don't change, let hydra resolve from saved config
+      label_key: ${data.train_ds.label_key} # don't change, let hydra resolve from saved config
+      add_eos: ${data.train_ds.add_eos} # don't change, let hydra resolve from saved config
+      add_sep: ${data.train_ds.add_sep} # don't change, let hydra resolve from saved config
+      add_bos: ${data.train_ds.add_bos} # don't change, let hydra resolve from saved config
       separate_prompt_and_response_with_newline: ${data.train_ds.separate_prompt_and_response_with_newline}
       write_predictions_to_file: True
       output_file_path_prefix: "preds" # Prefix of the file to write predictions to.
-      truncation_field: ${data.train_ds.truncation_field} # Options: ['context', 'answer']
+      truncation_field: ${data.train_ds.truncation_field}  # don't change, let hydra resolve from saved config
       index_mapping_dir: null # Path to a directory to write index mapping files.
-      prompt_template: ${data.train_ds.prompt_template}
+      prompt_template: ${data.train_ds.prompt_template} # don't change, let hydra resolve from saved config
       tokens_to_generate: 512
       log_every_n_steps: 1
-      sample_rate: ${data.train_ds.sample_rate}
+      sample_rate: ${data.train_ds.sample_rate} # don't change, let hydra resolve from saved config
       audio_locator: null # set it to allow multiple audios in a sample, e.g. '|audio|', and use it in the context field of manifest to specify the locations of audios (`audio_filepath` is a list of audios).
 
       metric:

examples/multimodal/speechllm/conf/speechllm/modularized_speech_gpt_config.yaml → examples/multimodal/speechllm/conf/modular_audio_gpt_config_peft.yaml

@@ -1,4 +1,4 @@
-# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-name: megatron_audio_gpt_peft_tuning
+name: megatron_audio_gpt_peft
 
 trainer:
   devices: 1
@@ -21,9 +21,9 @@ trainer:
   precision: 16
   logger: False # logger provided by exp_manager
   enable_checkpointing: False
-  replace_sampler_ddp: False
-  max_epochs: 9999
-  max_steps: -1 # consumed_samples = global_step * micro_batch_size * data_parallel_size * accumulate_grad_batches
+  use_distributed_sampler: False
+  max_epochs: 100
+  max_steps: 1000000 # 1M steps
   log_every_n_steps: 10 # frequency with which training steps are logged 
   val_check_interval: 1.0 # If is an int n > 1, will run val every n training steps, if a float 0.0 - 1.0 will run val every epoch fraction, e.g. 0.25 will run val every quarter epoch
   gradient_clip_val: 1.0
@@ -102,25 +102,31 @@ model:
   # override_vocab_size: 1024
 
   peft:
-    peft_scheme: "adapter"  # can be either adapter,ia3, or ptuning
+    peft_scheme: "lora"  # can be either lora, adapter, ia3 or ptuning
     restore_from_path: null
 
     # Used for adapter peft training
     adapter_tuning:
       type: 'parallel_adapter' # this should be either 'parallel_adapter' or 'linear_adapter'
       adapter_dim: 32
       adapter_dropout: 0.0
-      norm_position: 'pre' # This can be set to 'pre' or 'post', 'pre' is normally what is used.
+      norm_position: 'pre' # This can be set to 'pre', 'post' or null, 'pre' is normally what is used.
       column_init_method: 'xavier' # IGNORED if linear_adapter is used, options: xavier, zero or normal
       row_init_method: 'zero' # IGNORED if linear_adapter is used, options: xavier, zero or normal
       norm_type: 'mixedfusedlayernorm' # IGNORED if layer_adapter is used,  options are ['layernorm', 'mixedfusedlayernorm']
-
+      layer_selection: null  # selects in which layers to add adapters, e.g. [1,12] will add adapters to layer 1 (lowest) and 12. null will apply adapters to all layers
+      weight_tying: False
+      position_embedding_strategy: null # used only when weight_tying is True
+
     lora_tuning:
       adapter_dim: 32
       adapter_dropout: 0.0
       column_init_method: 'xavier' # IGNORED if linear_adapter is used, options: xavier, zero or normal
       row_init_method: 'zero' # IGNORED if linear_adapter is used, options: xavier, zero or normal
-
+      layer_selection:  null  # selects in which layers to add lora adapters. e.g. [1,12] will add lora to layer 1 (lowest) and 12. null will apply adapters to all layers
+      weight_tying: False
+      position_embedding_strategy: null # used only when weight_tying is True
+
     # Used for p-tuning peft training
     p_tuning:
       virtual_tokens: 10  # The number of virtual tokens the prompt encoder should add at the start of the sequence
@@ -129,6 +135,14 @@ model:
       init_std: 0.023
 
   perception:
+    use_multi_layer_feat: false  # whether to extract multi-layer features, only supports conformer encoder
+    multi_layer_feat:
+      layer_idx_list: [0,16]  # layer indices to extract features from
+      aggregator:
+        mode: "cat"  # ways to combine features from different layers, choices=['cat','sum','mean', 'max', 'min'], default to concat ('cat')
+        pooling: "avg"  # ways to pool features if they have different temporal lengths and align_mode=min, choices=['mean', 'max', 'min']
+        align_mode: "min"  # if features have different temporal lengths, set `min` to pool to the shortest length or `max` to repeat to the longest.
+
     modality_adapter: 
       _target_: nemo.collections.asr.modules.ConformerEncoder
       feat_in: 1024
@@ -219,8 +233,8 @@ model:
       concat_sampling_probabilities: null # When providing a list of datasets, this arg defines the sampling probabilities from each dataset when strategy='random'
       context_key: 'input'
       label_key: 'output'
-      # add_eos: True
-      add_eos: False
+      add_eos: True
+      # add_eos: False
       end_string: ${model.data.end_string}
       add_sep: False
       add_bos: False
@@ -267,9 +281,9 @@ model:
       # ASR configs
       sample_rate: 16000 #${model.audio_encoder.preprocessor.sample_rate}
 
-      log_every_n_steps: 1
+      log_every_n_steps: 10
       metric:
-        name: "wer" # Name of the evaluation metric to use. Options: ['exact_string_match', 'loss']
+        name: "loss" # Name of the evaluation metric to use. Options: ['exact_string_match', 'loss', 'wer', 'bleu', 'rouge']
         average: null # Average the metric over the dataset. Options: ['macro', 'micro']. Works only for 'F1', 'accuracy' etc. Refer to torchmetrics for metrics where this is supported.
         num_classes: null