v1.0.0

New major features:

1. Trainer Controller and ability to stop training earlier using metrics and thresholds. See https://github.com/foundation-model-stack/fms-hf-tuning?tab=readme-ov-file#trainer-controller-framework
2. Experiment Tracker while tuning for more insights (choice of frameworks)

Dependency upgrades:

1. Transformers to minimum version 4.42.0 needed for tuning of granite with llama models

API /interface changes:

1. Introduction of tokenizer field in ModelArguments https://github.com/foundation-model-stack/fms-hf-tuning/blob/main/tuning/config/configs.py#L51 to allow custom tokenizer while tuning
2. Deprecation of tokenizer field from PromptTuningConfig and moved to ModelArguments

Full List of what's Changed:

• minor refactor to allow modular functions by @Ssukriti in #224
• feat: support some metrics being 'None' without stopping training by @HarikrishnanBalagopal in #169
• Support custom embedding layer resizing to the desired multiple by @kmehant in #227
• ReIntroduce Package for FMS Accel by @fabianlim in #223
• feat: History based loss metric by @seshapad in #156
• docs: instructions for using the trainer controller framework by @HarikrishnanBalagopal in #214
• feat: support custom tokenizer by @kmehant in #229
• tests: tracker unit tests by @dushyantbehl in #172
• feat: Aim runid export by @dushyantbehl in #90
• Move default operations and metrics to variables by @alex-jw-brooks in #249
• Update README.md by @Ssukriti in #240
• deps: Update transformers to latest and skip broken prompt tuning tests by @Abhishek-TAMU in #246
• deps: Update transformers lower bound version by @Abhishek-TAMU in #250
• V100rc1 release by @olson-ibm in #252

New Contributors

• @Abhishek-TAMU made their first contribution in #246
• @olson-ibm made their first contribution in #252

Full Changelog: v0.4.0...v1.0.0