transformers

• A Mathematical Framework for Transformer Circuits
• Fact Finding: Attempting to Reverse-Engineer Factual Recall on the Neuron Level (Post 1) — AI Alignment Forum
• An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2 — AI Alignment Forum
• Toy Models of Superposition
• Towards Monosemanticity: Decomposing Language Models With Dictionary Learning
• Softmax Linear Units
• Concrete Steps to Get Started in Transformer Mechanistic Interpretability — Neel Nanda
• Amazing series from 3Blue1Brown:
  • How large language models work, a visual intro to transformers | Chapter 5, Deep Learning - YouTube
  • [Attention in transformers, visually explained | Chapter 6, Deep Learning - YouTube](https://www.youtube.com/watch?v=9-Jl0dxWQs8)
  • How might LLMs store facts | Chapter 7, Deep Learning - YouTube
• Neel Nanda Transformer Series
  • What is a Transformer? (Transformer Walkthrough Part 1/2) - YouTube
  • Implementing GPT-2 From Scratch (Transformer Walkthrough Part 2/2) - YouTube
  • A Walkthrough of A Mathematical Framework for Transformer Circuits - YouTube

Encoder-Decoder

T5

Decoder

GPT

Why decoder only beats encoder-decoder (Stanford CS25: V4 I Hyung Won Chung of OpenAI - YouTube):

	Encoder-decoder	Decoder-only
Additional cross attention	Separate cross attention	Self-attention serving both roles
Parameter sharing	Separate parameters for input and target	Shared parameters
Target-to-input attention pattern	Only attends to the last layer of encoder’s output	Within-layer (i.e. layer 1 attends to layer 1)
Input attention	Bidirectional	Unidirectional*

1. In generative applications causal attention allows us to cache previous steps since tokens only attend up to their step (KV Cache)
   1. With encoder models need to recompute full sequence at each step to update previous tokens

from Stanford CS 25 - Google Slides

Decoder Improvements

1. pre norm (norm before attention)
2. RMS norm - cheaper
3. RoPE position embeddings
4. Grouped Query Attention - Smaller KV cache
5. Mixture of Experts - More Parameters with less Flops
6. SwiGLU / other GLU variants

Encoder

BERT ViT (MAE)

---

Transformers and LLMs (Large Language Models)

llm

Components

Tokenization

• GitHub - SumanthRH/tokenization: A comprehensive deep dive into the world of tokens
• GitHub - openai/tiktoken: tiktoken is a fast BPE tokeniser for use with OpenAI’s models.

Positional Encoding

RoPE

• Rotary Embeddings: A Relative Revolution | EleutherAI Blog
• [2410.06205] Round and Round We Go! What makes Rotary Positional Encodings useful?

Embedding Tables

Attention

Flash Attention

Grouped Query Attention

Flex Attention

Normalization

RMS Norm

Why not BatchNorm

Can’t use batchnorm in causal models because it leads to information leakage (since it uses batch level statistics)

Residual

Feed Forward

SwiGLU

Activation Functions

Classifier

MoE - Mixture of Experts

Tricks

Speculative Decoding

KV Cache

Context Length

Long Context Transformers

Transformer Variants

nGPT - Normalized Transformer

TokenFormer

2024-11-03 - TokenFormer - RETHINKING TRANSFORMER SCAL-ING WITH TOKENIZED MODEL PARAMETERS

Sigmoid Attention

---

Other

• LLM Worksheet - Google Sheets

• GitHub - lm-sys/FastChat: The release repo for “Vicuna: An Open Chatbot Impressing GPT-4”

• GitHub - kernelmachine/cbtm: Code repository for the c-BTM paper

Chat

• GitHub - imoneoi/openchat: OpenChat: Advancing Open-source Language Models with Imperfect Data
• LangChain

Alignment

Instruction Tuning

• GitHub - huggingface/alignment-handbook: Robust recipes for to align language models with human and AI preferences
• GitHub - allenai/open-instruct

RLHF

• Fine-tuning 20B LLMs with RLHF on a 24GB consumer GPU

• Illustrating Reinforcement Learning from Human Feedback (RLHF)

• [29 Nov 2023] RLHF Lecture @ Stanford - Google Slides

DPO

• [2305.18290] Direct Preference Optimization: Your Language Model is Secretly a Reward Model

• RLHF progress: Scaling DPO to 70B, DPO vs PPO update, Tülu 2, Zephyr-β, meaningful evaluation, data contamination

• Fine-tune a Mistral-7b model with DPO.ipynb

• GitHub - eric-mitchell/direct-preference-optimization: Reference implementation for DPO (Direct Preference Optimization)

• Fine-tune Llama 2 with DPO

• RLHF and DPO compared: user feedback methods for LLM optimization | by Automata | 𝐀𝐈 𝐦𝐨𝐧𝐤𝐬.𝐢𝐨 | Oct, 2023 | Medium

Training

• The Novice’s LLM Training Guide

• [2310.18313] FP8-LM: Training FP8 Large Language Models

Fine Tuning

• GitHub - OpenAccess-AI-Collective/axolotl: Go ahead and axolotl questions
• GitHub - unslothai/unsloth: Finetune Llama 3.1, Mistral, Phi & Gemma LLMs 2-5x faster with 80% less memory

Distillation / Pruning / Compression

• [2407.14679] Compact Language Models via Pruning and Knowledge Distillation

Inference

GGML / LLAMA.CPP

Introduction to ggml

Open Models

• Mistral NeMo | Mistral AI | Frontier AI in your hands

Serving

• Hamel’s Blog - Optimizing latency

Quantization

• GitHub - PanQiWei/AutoGPTQ: An easy-to-use LLMs quantization package with user-friendly apis, based on GPTQ algorithm.
• GitHub - IST-DASLab/gptq: Code for the ICLR 2023 paper “GPTQ: Accurate Post-training Quantization of Generative Pretrained Transformers”.

Tutorials

• GitHub - RahulSChand/llama2.c-for-dummies: Step by step explanation/tutorial of llama2.c

Visualization

• LLM Visualization

Decoding

• X

Inference Benchmarks

• Performance of llama.cpp on Apple Silicon · ggerganov/llama.cpp · Discussion #4167 · GitHub

Evaluation

Evaluate LLMs and RAG a practical example using Langchain and Hugging Face

Applications

Text to SQL

• GitHub - defog-ai/sqlcoder: SoTA LLM for converting natural language questions to SQL queries

Links

Articles

• Transformer Math 101 | EleutherAI Blog

• The Annotated Transformer (2022 - Rush)

• The Illustrated Transformer (Alammar)

• Transformers From Scratch (2019 - Bloem)

• Transformer Walkthrough: A walkthrough of transformer architecture code (2022 - Riedl)

• https://github.com/cmhungsteve/Awesome-Transformer-Attention

• Transformer models: an introduction and catalog — 2023 Edition - AI, software, tech, and people, not in that order… by X

Lectures

• Stanford CS25: Transformers United (2021)
• Lecture 20 - Transformers and Attention - YouTube

Code

• xformers: Hackable and optimized Transformers building blocks, supporting a composable construction.
• GitHub - lucidrains/x-transformers: A simple but complete full-attention transformer with a set of promising experimental features from various papers
• GitHub - explosion/curated-transformers: 🤖 A PyTorch library of curated Transformer models and their composable components