tinygrad on macOS — GPU Guide
Run deep learning models locally on your Mac with full Metal GPU acceleration. Built by comma.ai - the same team behind openpilot.
examples/openpilot/ folder in tinygrad directly powers the neural networks in openpilot. If you are a comma user, tinygrad is part of your car's AI stack.tinygrad is an end-to-end deep learning stack built by tiny corp, comma.ai's research arm. It is intentionally minimal and hackable - the entire compiler, IR, and runtime are readable and modifiable. Think PyTorch for people who want to understand what is actually happening.
Tensor library with autograd (automatic differentiation)
IR and compiler that fuses and lowers kernels (lazy evaluation)
JIT + graph execution @TinyJit decorator makes inference fast
nn / optim / datasets for real training
| Framework | Similarity | Key Difference |
|---|---|---|
| PyTorch | Same eager Tensor API, autograd, optim | In tinygrad, the entire compiler is visible and hackable |
| JAX | IR-based autodiff, function-level JIT | Simpler - far easier to read and modify |
| TVM | Multiple lowering passes, BEAM search over kernels | tinygrad ships the full front-end too (tensors, nn, optim) |
.realize() or .numpy(). This allows the compiler to fuse multiple operations into a single GPU kernel, which is why performance is surprisingly good despite the small codebase.tinygrad has a first-class Metal backend in tinygrad/runtime/ops_metal.py that uses Apple's Metal API directly.
Full GPU acceleration on all Apple Silicon Macs (M1, M2, M3, M4)
Full GPU acceleration on Intel Macs with AMD or Intel graphics via Metal
Automatic detection - on macOS, Metal is the default backend
Metal 3.1 support on macOS Sonoma (14+), Metal 3.0 on Ventura (13+), Metal 2.0 on Monterey (12+)