Awesome Llama Resources

Llama2 is a part open source commercial model released from Meta, including 7B/13B/70B and chat models with 4096 context window.

Models

• [Original Model] 202307 Meta Released Llama2
  • Github
  • Meta's llama-recipes: provide examples for finetuning at SingleGPU/Multiple GPU and the recipe to convert model to HuggingFace transformers's LLama2 model definition
  • Paper: Llama 2: Open Foundation and Fine-Tuned Chat Models
  • Download Applications
• [Togehter AI] 202307 TogetherAI released Llama2-7B context window with 32k context window based on Meta's research Extending Context Window of Large Language Models via Positional Interpolation
  • Together.ai's blog about Preparing for the era of 32K context: Early learnings and explorations
• Codellama: Meta finetuned Llama2 for code generation usage. Support C++/ Java/ PHP/ Type Script/ C#/ Bash/ Python generation. Include models 7B/13B/34B，and 3 kind of variation (Generatl/python/instruction). Extend maximum context window from 4,096 tokens to 100k(like claude2).

Demo

• Llama2 70B Chatbot at HuggingFace
• A16z's Llama2-chatbot: provide a streamlit chatbot app for LLaMA2

Finetune Method/ Scripts

• Finetune with PEFT
• Finetune together.ai 32k context window model: script to finetune on booksum/mqa dataset
  • Llama-2-7B-32K-Instruct — and fine-tuning for Llama-2 models with Together API: Together AI show their 32k context instruct 7b model.
• Finetune with QLora at 13b model: a colab about finetuning llama2
• HuggingFace SFT training script
• Pytorch-lightening's script to finetune Llama2 on custom dataset
• Instuction-tune Llama2: HuggingFace's Tech Lead Philschmid introduced how to instruct finetune Llama2
• Finetune LLaMA2 7-70B on Amazon SageMaker: Philschmid introduce preparing datasets/using QLoRA/Deploy model on Amazon SageMaker
• Finetune LLaMa2 with QLoRA at colab
• Fine-tune Llama 2 with DPO by huggingface
• Fine-tune Llama2 on specific usage like SQL Gen/Functional Representation: Anyscale's member used their lib ray to demo finetune Llama2 70B.Their scripts

Porting

• Karpathy's Llama2.c: Karpathy's weekend project to build a LLama2 at C

• web-llm: Bringing large-language models and chat to web browsers

• HuggingFace release Swift Transformers to help run LLM on Apple Device: Provide Swift based Swift Transformers Lib, a swift chat app and a exporters for exporting model to coreml.

• pyllama: LLaMA: Open and Efficient Foundation Language Models

Tutorial

• Meta's started guide to use Llama
• Llama2.c for dummies: a description about Karpathy's LLama2 line by line
• NeurIPS 2023 LLM Efficiency Challenge Quickstart Guide: A competition focused on training 1 LLM for 24 hours on 1 GPU – the team with the best LLM gets to present their results at NeurIPS 2023.
• Huggingface share how to train and deploy an open source LLM?

Prompt

• LLama2 prompt template

For specific usage Model/ Finetuned model

• Huggingface trend about llama2
• Chinese-Llama-2-7b: finetune on a chinese and english instruction dataset with 10 millions size
• Chinese-LLaMA-Alpaca
• Finetuned on code with qLoRA
• ToolLLaMA: An open source project to train LLaMa on ToolBench, to make LLaMa support function call
• Llama2-Code-Interpreter: make Llama2 use Code Execution, Debug, Save Code, Reuse it, Access to Internet
• Llama2-Medical-Chatbot: A medical bot built using Llama2 and Sentence Transformers
• Finetune LLaMA 7B with Traditional Chinese instruction datasets
• Taiwan-LLaMa: NTU's MiuLab finetune 13B Llama2 with 5B traditional chinese tokens and 490k instruction dataset.
• Finetuning LLaMa + Text-to-SQL: LlamaIndex show how to fine-tune LLaMa 2 7B on a Text-to-SQL dataset

Multimodal LLM

• LLaSM: Large Language and Speech Model: Support chinese/english voice chat model based on whisper features
• LLaVA : Large Language-and-Vision Assistant
• Chinese-LLaVA: support vision input and chinese text input/output

Toolkits

• [TogetherAI] OpenChatKit: Together.ai's open toolkit for LLM finetune/moderation
• LLaMA2-Accessory:An Open-source Toolkit for LLM Development
  • LLaMA-Adapter: Fine-tuning LLaMA to follow Instructions within 1 Hour and 1.2M Parameters
• text-generation-webui:A gradio web UI for running Large Language Models like LLaMA, llama.cpp, GPT-J, OPT, and GALACTICA.
• text-generation-inference: Huggingface's Large Language Model Text Generation Inference.
• FlexFlow Serve: Low-Latency, High-Performance LLM Serving: An open-source compiler and distributed system for low latency, high performance LLM serving.
• LLM-As-Chatbot:Use lots of open sourced instruction-following fine-tuned LLM models as a Chatbot service.

Optimiztion(Latency/Size)

• Optimizing LLM latency: A great blog about exploration of inference tools for open source LLMs
• Series Quantized LLama2 Model from The Bloke with GPTQ/GGML
  • TheBloke/llama-2-7B-Guanaco-QLoRA-GPTQ
  • OpenAssistant-Llama2-13B-Orca-8K-3319-GPTQ
• Quantization
  • GPTQ: Accurate Post Training Quantization for generative pre-trained transformers
    • AutoGPTQ: An easy-to-use LLMs quantization package with user-friendly apis, based on GPTQ algorithm.
• Together AI's Medusa to accelerate decoding
• NVIDIA TensorRT-LLM Supercharges Large Language Model Inference on NVIDIA H100 GPUs:TensorRT-LLM is an open-source library that accelerates and optimizes inference performance on the latest LLMs on NVIDIA Tensor Core GPUs.
• 20231130 Pytorch Team use pytorch tool to accelerate

Optimization(Reasoning)

• LLM Reasoners: LLM Reasoners is a library to enable LLMs to conduct complex reasoning, with advanced reasoning algorithms.
• Deepminds LLM as Optimizers

Use

• Run Llama 2 on your own Mac using LLM and Homebrew
• Deploy Llama2 7B/13B/70B model on AWS SageMaker: Based on Hugging Face LLM DLC(Deep Learning Container) which is powered by huggingface's text generation inference. HuggingFace's text generation inference is a Rust, Python and gRPC server for text generation inference. Used in production at HuggingFace to power Hugging Chat, the Inference API and Inference Endpoint.

Other Resources

• Llama 2 の情報まとめ

• LLaMA2-Every Resource you need

• LLaMA-efficient-tuning: Easy-to-use fine-tuning framework using PEFT (PT+SFT+RLHF with QLoRA) (LLaMA-2, BLOOM, Falcon, Baichuan)

• awesome-llm and aigc

• Finetune Falcon-7B on Your GPU with TRL and QLoRA: A blog about tuning falcon-7b on your consumer GPU

• A Definitive Guide to QLoRA: Fine-tuning Falcon-7b with PEFT

• Amazon sagemaker generativeai: Fine-tune Falcon-40B with QLoRA

• Llama with FlashAttention2: Reduces VRAM usage, especially during training.Full finetune Llama 2 7b:51.3->40.3GiB

• Anti-hype LLM reading list: A reading list about LLM.

Move on to production

• Patterns for Building LLM-based Systems & Products: Amazon's LLM Engineer Eugene Yan wrote a blog about patterns of LLM based system
• Finetuning an LLM: RLHF and alternatives
• Github:A developer’s guide to prompt engineering and LLMs: Github engineer shares their experiences to to prompt engineering for their copilot product.
• The Rise and Potential of Large Language Model Based Agents: A Survey: A survey from Fudan NLP Group about LLM based Agents. Their github repo https://github.com/WooooDyy/LLM-Agent-Paper-List

Evaluation

• 🤗Open LLM Leaderboard: A huggingface space which track, rank and evaluate LLMs and chatbots as they are released.

Calculation

• How is LLaMa.cpp possible: The post showed why Llama is limited by memory bound with some calculations of the transformers parameters.
  • Transformer Math101
  • Transformer inference Arithmetic
  • The Illustrated Transformer

Some theory

• Why we should train smaller LLMs on more tokens

  • harms law on hugging face for calculating the model size/dataset size's compute overhead

• LLMSurvey: A Survey of Large Language Models

• Open challenges in LLM research: Chip Huyen's post about LLM's challenge

• Stanford CS324 - Large Language Models: The fundamentals about the modeling, theory, ethics, and systems aspects of large language models.

  • CS221:Artificial Intelligence: Principles and Techniques

• Why you(Propbably) Don't Need to Fine-tune an LLM: Finetuning maynot reduce hallucinating. You could use few-shot prompting/ Retrieval Augmented Generation(RAG)

• Challenges and Applications of Large Language Models

Some basics

• Some Intuition on Attention and the Transformer: A post introduces the big deal about attention/what are query,key and value
• Intro to transformers