zigTensor

zigTensor is a fast, flexible machine learning library written entirely in Zig; the design is heavily inspired by the Flashlight library.

Minimum Requirements

Requirement	Notes
Zig version	main
ArrayFire	latest via homebrew
OS	OSx & Linux (verified on Ubuntu)

Install with zigup:

zigup master

OSx Setup

Install ArrayFire with homebrew:

brew install arrayfire

Linux Setup

See Github Action Workflow for example setup on Ubuntu.

Status

The project is in the incredibly early stages of development and only the core Tensor/Tensor Ops functionality is currently functional. Autograd is a WIP.

Currently, development is solely focused on OSx/Linux and leveraging ArrayFire for the backend.

Basic Tensor Usage

const zt = @import("zigTensor");
const allocator = std.heap.c_allocator; // or your preferred allocator

const DType = zt.tensor.DType;
const deinit = zt.tensor.deinit;
const Tensor = zt.tensor.Tensor;

defer deinit(); // deinit global singletons (e.g. ArrayFire Backend/DeviceManager)

var a = try zt.tensor.rand(allocator, &.{5, 5}, DType.f32);
defer a.deinit();

var b = try zt.tensor.rand(allocator, &.{5, 5}, DType.f32);
defer b.deinit();

var c = try zt.tensor.add(allocator, Tensor, a, Tensor, b); // operator overloading pending (likely utilize comath library)
defer c.deinit();

Roadmap

• bindings
  • ArrayFire // TODO: optimize
  • oneDNN (WIP)
• autograd (WIP)
  • Functions // TODO: debug oneDNN (TensorExtension) fns
  • Utils
  • Variable // TODO: optimize
• common (WIP)
• contrib
• dataset
• distributed
• meter
• nn
• optim
• runtime (WIP)
  • CUDADevice
  • CUDAStream
  • CUDAUtils
  • Device
  • DeviceManager
  • DeviceType
  • Stream
  • SynchronousStream
• tensor
  • backend
    • ArrayFire (current focus - WIP)
      • mem
        • CachingMemoryManager
        • DefaultMemoryManager
        • MemoryManagerAdapter
        • MemoryManagerAdapterDeviceInterface
        • MemoryManagerInstaller
      • AdvancedIndex (CUDA specific API)
      • ArrayFireBLAS
      • ArrayFireBackend
      • ArrayFireBinaryOps
      • ArrayFireCPUStream
      • ArrayFireReductions
      • ArrayFireShapeAndIndex
      • ArrayFireTensor
      • ArrayFireUnaryOps
      • Utils
    • JIT
    • oneDNN
    • Stub
  • CUDAProfile
  • Compute
  • DefaultTensorType
  • Index
  • Init
  • Profile
  • Random
  • Shape
  • TensorAdapter
  • TensorBackend
  • TensorBase
  • TensorExtension
  • Types

Backend Roadmap

• ArrayFire (current focus - WIP)
  • CPU
  • OpenCL
  • CUDA
• JIT
• oneDNN (WIP - first implementing for use with Autograd ops)
• Stub

Other TODOs:

• Language APIs
  • C compatible API
  • JS Bindings
    • FFI
    • NAPI
    • WASM