v0.1.0 - First release

Summary of Changes

• Supports and validated tuning technique: full fine tuning using single-GPU and multi-GPU
  • Multi-GPU training using HuggingFace accelerate library, focused on FSDP
• Experimental tuning techniques:
  • Single GPU Prompt tuning
  • Single GPU LoRA tuning
• Scripts to allow local inference and evaluation of tuned models
• Build scripts for containerization of library
• Initial trainer controller framework for controlling the trainer loop using user-defined rules and metrics

Pip package: pip install fms-hf-tuning==0.1.0

What's Changed

• Init by @raghukiran1224 in #1
• allows disable flash attn and torch dtype param by @Ssukriti in #2
• First refactor train by @Ssukriti in #3
• fix : the way args are passed by @Ssukriti in #10
• fix full param tuning by @lchu-ibm in #14
• fix import of aim_loader by @anhuong in #13
• fix: set model max length to either passed in or tokenizer value by @anhuong in #17
• fix: do not set model max length when loading model by @anhuong in #21
• add EOS token to dataset by @Ssukriti in #15
• Local inference by @alex-jw-brooks in #27
• feat: add validation dataset to train by @anhuong in #26
• feat: support str in target_modules for LoraConfig by @VassilisVassiliadis in #39
• Add formatting tools by @hickeyma in #31
• Enable code formatting by @hickeyma in #40
• Enable daily dependabot updates by @hickeyma in #41
• Add file logger callback & export train loss json file by @alex-jw-brooks in #22
• Merge models by @alex-jw-brooks in #32
• Local inference merged models by @alex-jw-brooks in #43
• feat: track validation loss in logs file by @anhuong in #51
• Add linting capability by @hickeyma in #52
• Add PR/Issue templates by @tedhtchang in #65
• Add sample unit tests by @tedhtchang in #61
• Initial commit for trainer image by @tharapalanivel in #69
• Adding copyright notices by @tharapalanivel in #77
• Enable pylint in the github workflow by @tedhtchang in #63
• Bump aim from 3.17.5 to 3.18.1 by @dependabot in #42
• Add Contributing file by @jbusche in #58
• docs: lora and getting modules list by @anhuong in #46
• Allow SFT_TRAINER_CONFIG_JSON_ENV_VAR to be encoded json string by @kellyaa in #82
• Document lint by @tedhtchang in #84
• Let Huggingface Properly Initialize Arguments, and Fix FSDP-LORA Checkpoint-Saves and Resumption by @fabianlim in #53
• Unit tests by @tharapalanivel in #83
• Update CONTRIBUTING.md by @Ssukriti in #86
• Update input args to max_seq_length and training_data_path by @anhuong in #94
• feat: move to accelerate launch for distributed training by @kmehant in #92
• Update README.md by @Ssukriti in #95
• Modify copyright notice by @tharapalanivel in #96
• Switches dependencies from txt file to toml file by @jbusche in #68
• fix: use attn_implementation="flash_attention_2" by @kmehant in #101
• fix: not passing PEFT argument should default to full parameter finetuning by @kmehant in #100
• feat: update launch training with accelerate for multi-gpu by @anhuong in #98
• Setting default values in training job config by @tharapalanivel in #104
• add refactored build utils into docker image by @anhuong in #108
• feat: combine train and eval loss into one file by @anhuong in #109
• docs: add note on ephemeral storage by @anhuong in #106
• Move accelerate launch args parsing by @tharapalanivel in #107
• Docs improvements by @Ssukriti in #111
• feat: add env var SET_NUM_PROCESSES_TO_NUM_GPUS by @anhuong in #110
• feat: Trainer controller framework by @seshapad in #45
• Copying logs file by @tharapalanivel in #113
• Fix copying over logs by @tharapalanivel in #114
• Add eval script by @alex-jw-brooks in #102
• Lint tests by @tharapalanivel in #112
• Move sklearn to optional, install optionals for linting by @alex-jw-brooks in #117
• Build Wheel Action by @jbusche in #105
• rstrip eos in evaluation by @alex-jw-brooks in #121
• Fix eos token suffix removal by @alex-jw-brooks in #125
• Make use of instruction field optional by @alex-jw-brooks in #123
• Deprecating the requirements.txt for dependencies management by @tedhtchang in #116
• Add unit tests for various edge cases by @alex-jw-brooks in #97
• fix typo in build gha by @jbusche in #138
• Install whl in Dockerfile by @tedhtchang in #126
• feat: add flash attn to inference and eval scripts by @anhuong in #132
• OS update in dockerfile by @jbusche in #127
• fix: ignore the build output and auto-generated files by @HarikrishnanBalagopal in #140
• Propose ADR for Training Acceleration by @fabianlim in #119
• feat: new format for the controller metrics and operations by @HarikrishnanBalagopal in #130
• adr: Format change to the trainer controller configuration by @seshapad in #128
• Generic tracker API and implementation of Aimstack tracker by @dushyantbehl in #89
• fix: Allow makefile to run test independent of fmt/lint by @dushyantbehl in #145
• feat: Trainer state as a trainer controller metric by @seshapad in #150
• Bump aim from 3.18.1 to 3.19.0 by @dependabot in #93
• fix: launch_training.py arguments with new tracker api by @dushyantbehl in #153
• feat: Exposed the evaluation metrics for rules within trainer controller by @seshapad in #146
• Comment out aim in dockerfile by @jbusche in #155
• fix: replace eval with a safer alternative by @HarikrishnanBalagopal in #147
• doc...

v0.1.0-rc.1

What's Changed

• fix: replace eval with a safer alternative by @HarikrishnanBalagopal in #147
• docs: ADR for moving from eval to simpleeval for evaluating trainer controller rules by @HarikrishnanBalagopal in #151
• Add exception catching / writing to termination log by @kellyaa in #149
• fix: merging of model for multi-gpu by @anhuong in #158
• add .complete file to output dir when done by @kellyaa in #159

Full Changelog: v0.0.2rc2...v0.1.0-rc.1

v0.0.2rc.2

What's Changed

• fix typo in build gha by @jbusche in #138
• Install whl in Dockerfile by @tedhtchang in #126
• feat: add flash attn to inference and eval scripts by @anhuong in #132
• OS update in dockerfile by @jbusche in #127
• fix: ignore the build output and auto-generated files by @HarikrishnanBalagopal in #140
• Propose ADR for Training Acceleration by @fabianlim in #119
• feat: new format for the controller metrics and operations by @HarikrishnanBalagopal in #130
• adr: Format change to the trainer controller configuration by @seshapad in #128
• Generic tracker API and implementation of Aimstack tracker by @dushyantbehl in #89
• fix: Allow makefile to run test independent of fmt/lint by @dushyantbehl in #145
• feat: Trainer state as a trainer controller metric by @seshapad in #150
• Bump aim from 3.18.1 to 3.19.0 by @dependabot in #93
• fix: launch_training.py arguments with new tracker api by @dushyantbehl in #153
• feat: Exposed the evaluation metrics for rules within trainer controller by @seshapad in #146
• Comment out aim in dockerfile by @jbusche in #155

New Contributors

• @HarikrishnanBalagopal made their first contribution in #140
• @dushyantbehl made their first contribution in #89

Full Changelog: v0.0.2rc1...v0.0.2rc2

v0.0.2rc1

What's Changed

• Init by @raghukiran1224 in #1
• allows disable flash attn and torch dtype param by @Ssukriti in #2
• First refactor train by @Ssukriti in #3
• fix : the way args are passed by @Ssukriti in #10
• fix full param tuning by @lchu-ibm in #14
• fix import of aim_loader by @anhuong in #13
• fix: set model max length to either passed in or tokenizer value by @anhuong in #17
• fix: do not set model max length when loading model by @anhuong in #21
• add EOS token to dataset by @Ssukriti in #15
• Local inference by @alex-jw-brooks in #27
• feat: add validation dataset to train by @anhuong in #26
• feat: support str in target_modules for LoraConfig by @VassilisVassiliadis in #39
• Add formatting tools by @hickeyma in #31
• Enable code formatting by @hickeyma in #40
• Enable daily dependabot updates by @hickeyma in #41
• Add file logger callback & export train loss json file by @alex-jw-brooks in #22
• Merge models by @alex-jw-brooks in #32
• Local inference merged models by @alex-jw-brooks in #43
• feat: track validation loss in logs file by @anhuong in #51
• Add linting capability by @hickeyma in #52
• Add PR/Issue templates by @tedhtchang in #65
• Add sample unit tests by @tedhtchang in #61
• Initial commit for trainer image by @tharapalanivel in #69
• Adding copyright notices by @tharapalanivel in #77
• Enable pylint in the github workflow by @tedhtchang in #63
• Bump aim from 3.17.5 to 3.18.1 by @dependabot in #42
• Add Contributing file by @jbusche in #58
• docs: lora and getting modules list by @anhuong in #46
• Allow SFT_TRAINER_CONFIG_JSON_ENV_VAR to be encoded json string by @kellyaa in #82
• Document lint by @tedhtchang in #84
• Let Huggingface Properly Initialize Arguments, and Fix FSDP-LORA Checkpoint-Saves and Resumption by @fabianlim in #53
• Unit tests by @tharapalanivel in #83
• Update CONTRIBUTING.md by @Ssukriti in #86
• Update input args to max_seq_length and training_data_path by @anhuong in #94
• feat: move to accelerate launch for distributed training by @kmehant in #92
• Update README.md by @Ssukriti in #95
• Modify copyright notice by @tharapalanivel in #96
• Switches dependencies from txt file to toml file by @jbusche in #68
• fix: use attn_implementation="flash_attention_2" by @kmehant in #101
• fix: not passing PEFT argument should default to full parameter finetuning by @kmehant in #100
• feat: update launch training with accelerate for multi-gpu by @anhuong in #98
• Setting default values in training job config by @tharapalanivel in #104
• add refactored build utils into docker image by @anhuong in #108
• feat: combine train and eval loss into one file by @anhuong in #109
• docs: add note on ephemeral storage by @anhuong in #106
• Move accelerate launch args parsing by @tharapalanivel in #107
• Docs improvements by @Ssukriti in #111
• feat: add env var SET_NUM_PROCESSES_TO_NUM_GPUS by @anhuong in #110
• feat: Trainer controller framework by @seshapad in #45
• Copying logs file by @tharapalanivel in #113
• Fix copying over logs by @tharapalanivel in #114
• Add eval script by @alex-jw-brooks in #102
• Lint tests by @tharapalanivel in #112
• Move sklearn to optional, install optionals for linting by @alex-jw-brooks in #117
• Build Wheel Action by @jbusche in #105
• rstrip eos in evaluation by @alex-jw-brooks in #121
• Fix eos token suffix removal by @alex-jw-brooks in #125
• Make use of instruction field optional by @alex-jw-brooks in #123
• Deprecating the requirements.txt for dependencies management by @tedhtchang in #116
• Add unit tests for various edge cases by @alex-jw-brooks in #97

New Contributors

• @raghukiran1224 made their first contribution in #1
• @Ssukriti made their first contribution in #2
• @lchu-ibm made their first contribution in #14
• @anhuong made their first contribution in #13
• @alex-jw-brooks made their first contribution in #27
• @VassilisVassiliadis made their first contribution in #39
• @hickeyma made their first contribution in #31
• @tedhtchang made their first contribution in #65
• @tharapalanivel made their first contribution in #69
• @dependabot made their first contribution in #42
• @jbusche made their first contribution in #58
• @kellyaa made their first contribution in #82
• @fabianlim made their first contribution in #53
• @kmehant made their first contribution in #92
• @seshapad made their first contribution in #45

Full Changelog: https://github.com/foundation-model-stack/fms-hf-tuning/commits/v.0.0.2rc1