tinygrad/tinygrad

Health v2 · maintenance only

Adoption not part of the score

33509 stars · 4286 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

tinygrad is a small, hackable deep learning framework providing a PyTorch-like tensor library with autograd, plus a visible IR and compiler that fuses and lowers kernels across hardware backends. It includes nn, optim, and dataset modules for real training, with JIT and graph execution support.

Use cases

• train neural networks with a lightweight pytorch-like api
• learn how a deep learning framework works internally
• compile and fuse gpu kernels with a hackable ir
• run deep learning on diverse hardware backends
• prototype autograd and tensor operations like micrograd
• experiment with kernel scheduling and codegen

When to choose

• you want a minimal, readable alternative to pytorch for training models
• you need to understand or modify the compiler and kernel generation
• you want deep learning support across many accelerators from one codebase
• you are learning how autodiff and jit compilation work

When to avoid

• you need the full ecosystem, vmap/pmap transforms, or production maturity of pytorch or jax
• you rely on extensive pretrained model hubs and third-party integrations
• you need guaranteed stability for critical production workloads

Facets

library · maturity active

machine-learning deep-learning compiler llm-training gpu-computing deep-learning machine-learning compilers gpu-computing python cross-platform tensor-library autograd deep-learning-framework kernel-fusion jit pytorch-alternative gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="tinygrad/tinygrad")