- Category
- AI Tools
- Rank
- No. 1538Tools index
Previous survey · No. 1543 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- badlogic
- GitHub
- 45 stars
- Date
About
Pure TypeScript implementation of GPT-2 — runs the classic transformer entirely in the browser/Node without native deps.
What it does
It turns original GPT-2 checkpoints into a compact tensor layout, then performs text generation one token at a time. The runtime handles byte-pair tokenization, attention, normalization, feed-forward layers, cached keys and values, and either greedy or top-k sampling. Output can stream as tokens arrive.
Why it's ranked here
Its value comes from a small, readable implementation of the complete inference path. Model dimensions come from converted configuration rather than fixed constants, and the command interface exposes useful sampling controls. The narrow scope is also the catch: setup still depends on Python, TensorFlow, downloaded checkpoints, and substantial local storage.
What's good
The code makes inference mechanics unusually concrete. Typed floating-point buffers hold tensors and reusable workspace memory. A key-value cache avoids recomputing earlier attention states. Tensor loading validates magic bytes, format version, dimensions, byte length, and shape consistency. Strict TypeScript checks strengthen the compact implementation.
Tradeoffs
Model preparation is heavier than the runtime. It requires Python, uv, TensorFlow, checkpoint downloads, and a conversion pass before generation works. Weights remain floating point, and the largest documented model needs many gigabytes of memory and disk. The runtime requires Node support for direct TypeScript execution and uses synchronous file loading.
How to use it well
Use it to study GPT-2 inference, inspect transformer mechanics, or run controlled local text-generation experiments from TypeScript. Start with the smallest documented checkpoint, verify types before running, and tune greedy or top-k sampling through command options. It does not train models, provide a hosted service, or eliminate checkpoint preparation.
Technical notes+
gpt-2.ts exports model loading, token encoding, KV-cache and workspace creation, single-token forwarding, sampling, and generation, while also providing the CLI entry point. tensor.ts implements validated loading of the 64-byte little-endian GPT2TNS\0 format plus scalar-loop matrix-vector operations, layer normalization, GELU, and row access. tokenizer.ts implements GPT-2 byte-to-Unicode mapping, BPE merges, caching, encoding, and decoding. convert/download_model.py downloads original checkpoint assets, and convert/convert.py uses TensorFlow and NumPy to write float32 tensor files and a SHA-256 manifest. tsconfig.json enables strict checking with unchecked-index and exact-optional-property safeguards.
Observed
- License
- MIT
- Primary language
- TypeScript runtime with Python conversion tooling
- Packaging
- ES module npm project with TypeScript and Node typings as development dependencies
- Interfaces
- Command-line interface plus exported TypeScript library functions
- Runtime platform
- Node.js with built-in TypeScript type stripping
- Model preparation
- Python tooling managed through uv downloads and converts original TensorFlow checkpoints
- Tensor storage
- Custom little-endian float32 files with fixed 64-byte headers
- Quality checks
- TypeScript no-emit checking is configured; no test directory appears in the supplied repository tree
Read from README.md, package.json, gpt-2.ts, tensor.ts, tokenizer.ts, convert/convert.py, convert/download_model.py, tsconfig.json, convert/uv.lock, convert/README.md, convert/pyproject.toml.
What it can do
Generate text from prompts
Text prompt or seed text → AI-generated text continuation
Complete partial sentences or paragraphs
Incomplete text → Completed text with natural language flow
Run GPT-2 inference in browser
Text input and model parameters → Generated text response
Run GPT-2 inference in Node.js
Text input and model parameters → Generated text response
Process text without external dependencies
Natural language text → AI-processed text output
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.