// the take
A from-scratch ML framework with a TypeScript front end and a Rust backend offering CPU, CUDA, and WebGPU execution, built explicitly as a learning exercise in how frameworks like PyTorch work under the hood. Good for someone who wants to see autograd and hand-written GPU kernels implemented end to end in Rust; not yet something to train real models on.
Three real backends (pure Rust CPU, CUDA via cudarc, WebGPU via wgpu) share one autograd tape, with custom kernels for flash attention, KV-cache quantization, and mixed precision, not just a CPU toy with a GPU label. Ships prebuilt native binaries for every target triple (darwin-arm64, linux-x64-gnu-cuda, win32-x64-msvc-webgpu, etc.) through npm so consumers don't need a Rust or CUDA toolchain installed. Decent op and module coverage for a from-scratch framework (conv1d/conv2d, pooling, layernorm, embedding, cross-entropy, dropout) backed by an actual test suite covering autograd, the native bridge, nn modules, and tensor ops rather than just examples.
No commits since 2026-04-20, over six months stale, which is risky for a project with hand-written CUDA kernels and WebGPU shaders that need to track driver and toolkit changes. No model serialization or checkpointing anywhere in the API, so there's no way to save or load trained weights, which is basic functionality for using a framework beyond a single session. Only two optimizers (SGD, Adam) and a single XOR example in the docs, with no benchmarks showing the CUDA/WebGPU backends are actually competitive, so the 'GPU speed' claim is unverified. Single-device only, no distributed or multi-GPU training path, capping it to small experiments regardless of backend.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →