GitHub - zolotukhin/zinc: Zig INferenCe Engine — Local LLM inference on AMD GPUs and Apple Silicon
Zig INferenCe Engine — Local LLM inference on AMD GPUs and Apple Silicon - zolotukhin/zinc
Deep Learning Framework
Summary
tinygrad is an end-to-end deep learning stack designed to be intentionally tiny, hackable, and efficient. It combines elements from PyTorch, JAX, and TVM to provide a tensor library, IR-based compiler, and JIT execution.
Highlights
- Provides a complete stack including a tensor library, autograd, IR/compiler, and JIT execution
- Designed to be lightweight and hackable compared to larger frameworks like PyTorch
- Utilizes a lazy execution model to automatically fuse operations into single kernels
- Supports a wide array of accelerators including CUDA, Metal, OpenCL, and WebGPU
auto-generated
Context
Audience
Machine Learning Engineers and Deep Learning Researchers
DomainMachine Learning
Formatopen source repository
Accessopen source
Source Overview
tinygrad is an open-source deep learning stack providing a tensor library with autograd, an IR-based compiler for kernel fusion, and built-in nn/optim/datasets modules for training.
It supports hardware backends including CPU, CUDA, OpenCL, METAL, AMD, NV, QCOM, and WebGPU, with new backends requiring approximately 25 low-level ops. The project prioritizes minimal codebase size and readability, using a PyTorch-like tensor API with IR-based autodiff inspired by JAX.
Topics
Related
PyTorchJAXTVMAutogradNeural NetworksCUDA
