Appearance
延伸資源
原始論文
| 資源 | 連結 |
|---|---|
| Attention Is All You Need (2017) | arXiv |
| Language Models are Unsupervised Multitask Learners (GPT-2, 2019) | OpenAI |
| Improving Language Understanding by Generative Pre-Training (GPT, 2018) | OpenAI |
| Root Mean Square Layer Normalization (RMSNorm, 2019) | arXiv |
| Adam: A Method for Stochastic Optimization (2014) | arXiv |
Karpathy 相關資源
| 資源 | 說明 | 連結 |
|---|---|---|
| Let's build GPT (YouTube) | Karpathy 從零建立 GPT 的影片教學 | YouTube |
| nanoGPT | microgpt 的 production 版本(使用 PyTorch) | GitHub |
| minGPT | 更早期的精簡 GPT 實作 | GitHub |
| micrograd | microgpt 的 Autograd 引擎的前身 | GitHub |
| makemore | 資料集來源(names.txt) | GitHub |
其他教學資源
| 資源 | 說明 |
|---|---|
| The Illustrated Transformer (Jay Alammar) | 最知名的 Transformer 圖文解說 |
| The Annotated Transformer (Harvard NLP) | Transformer 論文搭配程式碼逐行解讀 |
| 3Blue1Brown — Attention 系列 | 直覺化的注意力機制動畫解說 |
| Let's build GPT from Scratch (Andrej Karpathy) | 1 小時 53 分的完整教學影片 |
實作參考
- nanoGPT 的
model.py:比較 microgpt 與 production PyTorch 實作的差異 - Hugging Face Transformers 的 GPT2Model:業界標準實作
- llama.c (karpathy):在 C 語言中實作 LLM 推論,與 microgpt 是同一脈絡