A browser-based, WebGL2 implementation of GPT-2.

• Full GPT-2 small (117M) forward pass in the GPU via WebGL2 shaders
• BPE tokenization using js-tiktoken in the browser (no WASM fetch)
• Simple Python script to download the pretrained weights

• Node.js ≥ 16.x and npm
• Python ≥ 3.8
• A modern browser with WebGL2 support (Chrome, Firefox, Safari, Edge)

We rely on HuggingFace’s transformers to pull down the official GPT-2 weights and emit raw Float32Array blobs:

1. Install Python dependencies:

   pip install torch numpy transformers

2. Run the downloader:

   python download_weights.py

   This will fetch:
   • wte.bin (token embeddings)
   • wpe.bin (positional embeddings)
   • c_attn_q_w_0.bin … c_attn_q_w_11.bin
   • c_attn_k_w_0.bin … etc.
   • lm_head_w.bin, lm_head_b.bin
   • And a generated manifest.json mapping names → URLs

We use Vite to bundle TS, serve ESM modules & handle js-tiktoken:

1. Install JS dependencies:
2. Start the local dev server:
3. Open your browser at

Any changes under src/ will trigger HMR and live-reload.

.
├── public/                  # static assets served at `/`
│   └── weights/             # GPT-2 weight binaries + manifest.json
├── src/
│   ├── gpt2_webgl.ts        # WebGL2 inference + shaders + tokenizer
│   └── main.ts              # bootstrap: loads manifest, sets up UI
├── download_weights.py      # Python script to fetch & dump weights
├── index.html               # (copied by Vite) entrypoint HTML
├── vite.config.ts           # Vite config
├── package.json
└── tsconfig.json

MIT