Llama-recipes v.0.0.4.post1 Release Notes

This release import bug fixes and some doc changes.

What's Changed

• Improve discoverability of 3.2 recipes by @subramen in #684
• fix readme by @wukaixingxp in #679
• fix AutoModel and bump transformers version to 4.45 by @wukaixingxp in #686
• post1 release version bump by @mreso in #687

Llama-recipes v.0.0.4 Release Notes

This release accompanies the release of Llama 3.2 which included new versions of the Llama models in sizes of 1B, 3B, 11B and 90B. To get started with the new models you can find information in the official documentation or the on the HuggingFace hub. Further details can also be found in the model card and the The Llama 3 Herd of Models paper. For this release we updated the documentation and made sure all components work with the new models including multimodal finetuning.

What's Changed

Integrate Llama 3.2

• Upstream merge by @albertodepaola in #677

New and updated recipes

• Adding end-to-end llama chatbot recipe using Retrieval Augmented Fine Tuning (RAFT) by @wukaixingxp in #569
• [WIP] adding chatbot-e2e by @HamidShojanazeri in #462
• [Azure] Update Azure API usage example to 3.1 by @WuhanMonkey in #615
• Corrected wrong order of commands by @BakungaBronson in #602
• Fill in one sentence in the prompt guard tutorial. by @cynikolai in #609
• Llamaguard notebook colab link fix by @tryrobbo in #619
• Updating llama 3 references to 3.1 model by @init27 in #632
• recipes/quickstart/Getting_to_know_Llama.ipynb, typo fix lama -> llama line 127 by @cselip in #635
• Update hello_llama_cloud.ipynb by @MrDlt in #584
• Update hello_llama_cloud.ipynb by @MrDlt in #638
• Add preprocessor to patch PromptGuard scores for inserted characters by @cynikolai in #636
• Eval reproduce recipe using lm-evaluation-harness and our 3.1 evals datasets by @wukaixingxp in #627

Documentation update

• Update readme text to be version-agnostic by @subramen in #614
• Move supported features table to main README by @subramen in #616
• document less obvious training config parameters by @kjslag in #522

Misc fixes

• Enable users to trust remote code in samsum dataset by @mreso in #628
• Use new get_model_state_dict api for save_pretrained peft model by @mreso in #629
• Fix version number in Python example by @wstnmssr in #643
• Fix checkpoint saving by @mreso in #650
• Adding custom dataset file by @goswamig in #659
• Make gradio and langchain optional dependencies by @mreso in #676
• Update get_default_finetune_args.py by @edamamez in #662
• Fix/custom dataset chat template by @mreso in #665
• Create v0.0.4 release by @mreso in #678

New Contributors

• @cynikolai made their first contribution in #609
• @BakungaBronson made their first contribution in #602
• @init27 made their first contribution in #632
• @cselip made their first contribution in #635
• @MrDlt made their first contribution in #584
• @wstnmssr made their first contribution in #643
• @goswamig made their first contribution in #659
• @edamamez made their first contribution in #662

Full Changelog: v0.0.3...v0.0.4

We would like to thank all who contributed to this release and are looking forward to future contributions!

Llama-recipes v.0.0.3 Release Notes

Llama 3.1 Integration

This release accompanies the release of Llama 3.1 which included new versions of the Llama 8B and 70B models as well as the new 405B version. To get started with the new models you can find information in the official documentation or the on the HuggingFace hub. Further details can also be found in the model card and the Llama 3.1 paper. For this release we updated the documentation and made sure all components work with the new models.

• Release update by @albertodepaola @cynikolai @mreso @subramen @tryrobbo @varunfb in #603

New Features

We also added new features like FSDP + QLoRA fine-tuning and H2O algorithm for long context inference.

• Implement H2O for long context inference on summarization tasks by @Kyriection in #411
• Resume the fine-tuning process from the previous PEFT checkpoint folder by @wukaixingxp in #531
• Update hf weight conversion script to llama 3 by @dongwang218 in #551
• Adding support for FSDP+Qlora. by @HamidShojanazeri in #572

Additional Examples

Besides, we added new examples to get you up and running quickly with the Llama models

• Add Groq/Llama3 recipes (cookbook and command line examples) by @dloman118 in #553
• [WIP] Peft Finetuning Quickstart Notebook by @mreso in #558
• 4 notebooks ported from 4 DLAI agent short courses using Llama 3 by @jeffxtang in #560
• [lamini] Add lamini text2sql memory tuning tutorial by @powerjohnnyli in #573
• colab links fixed for dlai agents notebooks by @jeffxtang in #593
• Port of DLAI LlamaIndex Agent short course lessons 2-4 to use Llama 3 by @jeffxtang in #594

Codebase Refactor

We also refactored our codebase to improve discoverability of our documentation and recipes

• New structure and rename for tools, docs and quickstart folder by @pia-papanna in #575
• Add Langchain agent notebooks to 3P_Integrations by @subramen in #576
• Updates to benchmarks code by @subramen in #577
• Add README for quickstart + update to codellama url by @subramen in #578
• Updating the folder name 3p_integrations by @pia-papanna in #581
• Deleting Agents folder and adding llamaindex by @pia-papanna in #582
• Update 3p_integration README.md by @subramen in #586
• Add experimental folder to README by @subramen in #585
  fix typo by @subramen in #588
• Updating chatbot folder names by @pia-papanna in #590
• Move MediaGen notebook to octoai folder by @subramen in #601

Misc improvements and bugfixes

• Fix hsdp_device_mesh=None when enable HSDP and HYBRID_SHARD by @haozhx23 in #402
• bump up version by @mreso in #529
• Fix config file links for FMBench, update business summary chart. by @aarora79 in #532
• fixed alpaca dataset evalset length and make sure len(eval_loader)>0 by @wukaixingxp in #540
• Fix typo in Getting_to_know_Llama.ipynb by @jenyckee in #545
• replace groq llama 2 with replicate by @jeffxtang in #546
• Remove pkg_resources.packaging by @mreso in #547
• Update langgraph tool calling agent, simplify examples and README by @rlancemartin in #549
• Minor update to README by @rlancemartin in #555
• Add ToolMessage import by @rlancemartin in #559
• Make quickstart finetuning notebook ready for T4 by @mreso in #562
• bug fix by @jarvisDang in #570
• changed --pure_bf16 to --fsdp_config.pure_bf16 and corrected "examples/" path by @wukaixingxp in #587
• Update links in README.md by @subramen in #589
• Fix broken image link by @subramen in #597
• Fix relative links to images by @subramen in #596
• Remove max_length from tokenization by @mreso in #604
• Update transformers requirements by @mreso in #605
• Address feedback not possible before launch in LG3 recipe and dataset file by @tryrobbo in #606

New Contributors

• @jenyckee made their first contribution in #545
• @dloman118 made their first contribution in #553
• @Kyriection made their first contribution in #411
• @haozhx23 made their first contribution in #402
• @powerjohnnyli made their first contribution in #573
• @jarvisDang made their first contribution in #570
• @pia-papanna made their first contribution in #575

Full Changelog: v0.0.2...v0.0.3

We would like to thank all who contributed to this release and are looking forward to future contributions!