API Reference

Context

• init(options?)
• gpu.array(data, shape?)
• gpu.zeros(shape) / gpu.ones(shape) / gpu.full(shape, value)
• gpu.arange(start, stop, step?) / gpu.linspace(start, stop, num)
• gpu.random(shape) / gpu.randn(shape)
• gpu.fromArrow(column, options?)
• gpu.fromBuffer(buffer, options?)
• gpu.scoped(fn) / gpu.tidy(fn)

Math Ops

• add, sub, mul, div, sum, max, min, mean

Linear Algebra

• matmul, dot, transpose, inv, det, solve, qr, svd

ML & Convolution

• softmax, layerNorm, batchNorm, attentionScores
• conv2d, maxPool2d, avgPool2d

Signal Processing

• fft, ifft, fftMagnitude, spectrogram

Training

• gradients, sgdStep

Shape Notes

• matmul(gpu, A, B) infers dimensions from array shape metadata.
• layerNorm(gpu, input, gamma, beta, rows?, cols?) supports explicit row/column shape override.

TypeScript Shape Safety

Use tuple shapes to get stronger compile-time checks in IDEs:

const A = gpu.array<[2, 3]>(new Float32Array(6), [2, 3]);
const B = gpu.array<[3, 4]>(new Float32Array(12), [3, 4]);
const C = await matmul(gpu, A, B); // GPUArray<[2, 4]>

const v1 = gpu.array<[3]>(new Float32Array([1, 2, 3]), [3]);
const v2 = gpu.array<[3]>(new Float32Array([4, 5, 6]), [3]);
const d = await dot(gpu, v1, v2);

Live API Playground