🧠 @dxo/core

Developer preview — API unstable (0.0.x).

TypeScript entry for DXO. Loads the platform native addon via optionalDependencies (@dxo/dxo-<platform>).

Install

npm install @dxo/[email protected]

Contract (G3 preview)

| Surface | Behavior | |----------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------| | tensor(data, shape, { requiresGrad? }) | CPU leaf; shape is required (product must match flat length) | | zeros / ones / randn | Same options; only device: 'cpu' in this slice | | backend() | "cpu" / "cuda" / … (DXO product labels; never an internal engine brand) | | TypedBuffer family | TypedBuffer / TensorView / DeviceBuffer / mapped·stream / CodecHandle + domain carriers (ImageBuffer, TokenBuffer, …) | | ImageBuffer / decodeImageBuffer | Pixel carrier + decode bridge stubs; not vision Image | | t.add/mul/matmul/relu/reshape/transpose/sum/mean | Eager ops; record Tape when requiresGrad and grad enabled | | t.backward() | Scalar only (numel === 1, typically shape [1]) | | t.grad | Row-major number[] or undefined | | t.zeroGrad() | Clears this leaf's accumulated grad | | t.detach() | Values only — no tape / requiresGrad | | withoutGrad(fn) | Disables tape for fn, then restores prior flag |

import { tensor, withoutGrad } from '@dxo/core';

const x = tensor([1, 2, 3, 4], [2, 2], { requiresGrad: true });
const y = x.matmul(tensor([0.5, 0, 0, 0.5], [2, 2])).sum();
y.backward();
console.log(x.grad);

withoutGrad(() => {
  // no tape
});

Develop (monorepo)

pnpm build:native
pnpm verify -- g3-contract