Packages

All packages are on npm under @johnhenry/*. Each one has its own README with the full API and a ## Limitations section, and a CHANGELOG; the links below go to them. Versions are the latest on npm.

Every release is built and published by the repo’s GitHub Actions release workflow through npm trusted publishing (OIDC, no long-lived token), so every current version carries an npm provenance attestation linking it to the commit and workflow run that built it (the first 0.1.0/0.1.1 releases, published from a local machine, do not).

Laya

These packages port code from laya-mlx and Laya, so they are Apache-2.0 and ship a NOTICE file.

Package	Version	What it is
@johnhenry/laya	0.3.0	load() / predict() / createAgent() / shortlist, plus quantized-checkpoint loading and quantizeSafetensors. A port of laya-mlx’s Agent, load and shortlist.py.
@johnhenry/laya-core	0.1.1	The logic that needs no tensors: validation, prompts, tokenization, collation, calibration, result formatting.
@johnhenry/laya-router	0.1.4	Routes by language and task across the three checkpoints (router.py).
@johnhenry/laya-presets	0.1.1	Ready-made question sets and email cleaning (presets.py, email.py).
@johnhenry/laya-cli	0.3.0	The laya command: laya predict, laya bench, laya quantize.
@johnhenry/modernbert	0.3.0	ModernBERT / mmBERT encoder on any backend, from safetensors weights, including quantized Linear weights. Loading and forward are async.
@johnhenry/langdetect-lite	0.1.1	Script and Latin-language detection with no model (lang.py).

Tensor backends

Original code, MIT.

Package	Version	What it is
@johnhenry/tensor-backend	0.3.0	The op contract (async uploads, 22 optional numerics ops, optional quantized ops), plus a conformance suite whose golden cases come from Python MLX.
@johnhenry/backend-cpu	0.3.2	Pure-TypeScript f32 reference. Since 0.3, a re-export of @johnhenry/math-plus-tensor-cpu.
@johnhenry/backend-mlx	0.4.0	Native MLX through mlx-c (koffi on Node, bun:ffi on Bun, Deno.dlopen on Deno).
@johnhenry/backend-mlx-darwin-arm64	0.1.1	The prebuilt MLX 0.32.2 runtime, an optional dependency of backend-mlx. MIT: it contains Apple’s MLX binaries, unmodified.
@johnhenry/backend-webgpu	0.4.0	WGSL kernels with f16, flash attention and quantized Linears, for browsers, Deno, and Node/Bun through Dawn. Exposes its runtime for custom kernels.

Building blocks

Original code, MIT. Both are usable without Laya.

Package	Version	What it is
@johnhenry/hf-cache	0.1.2	Hugging Face Hub resolution plus a cache compatible with huggingface_hub: on disk in Node/Bun, in the Cache API in browsers.
@johnhenry/pyjson	0.1.2	CPython’s json.dumps, repr(float) and round(), byte for byte.

Safetensors reading is not in this repo. It comes from @johnhenry/math-plus-safetensors in Math Plus.

Dependency order

From the leaves up:

1. pyjson, tensor-backend, hf-cache: no dependencies.
2. langdetect-lite and laya-core: depend on pyjson. laya-core also depends on @huggingface/tokenizers.
3. backend-cpu, backend-mlx, backend-webgpu and modernbert: depend on tensor-backend. backend-cpu also depends on @johnhenry/math-plus-tensor-cpu, which it re-exports.
4. laya: depends on laya-core, modernbert, hf-cache, backend-cpu and math-plus-safetensors. backend-mlx and backend-webgpu are optional peer dependencies (^0.2.0 || ^0.3.0 || ^0.4.0).
5. laya-presets and laya-router: depend on laya-core. The router also depends on laya and langdetect-lite.
6. laya-cli: depends on laya and laya-router.

In the other direction, two math-plus packages build on laya-js backends: @johnhenry/math-plus-tensor-webgpu 0.3 is a facade over backend-webgpu, and the experimental @johnhenry/math-plus-tensor-mlx is built on backend-mlx.

Upgrading

• From 0.1: device uploads are async. load() and predict() are unchanged. Await backend.fromHost(...), createAgent, loadDecisionModel, loadModernBert, encoder.forward / embed, model.forwardTensors and model.uploadBatch. If you construct DecisionWeights by hand, add constants. Backend implementers: make fromHost async.
• From 0.2: additive. The quantized ops are optional, and laya 0.3 keeps quantized checkpoints on the device by default; pass quantized: "dequantize" for the 0.2 behaviour.
• Peer ranges: laya and laya-cli 0.3 accept backend-mlx and backend-webgpu 0.2, 0.3 or 0.4. Quantized weights stay on the device only with the 0.4 backends; older ones fall back to dequantizing on load.

JSR

These packages have a jsr.json and are set up for JSR trusted publishing, which publishes the TypeScript sources:

• tensor-backend
• backend-cpu
• backend-mlx
• backend-webgpu
• modernbert
• laya-core
• laya-presets
• laya-router
• pyjson
• langdetect-lite
• hf-cache

None of them is on JSR yet.

These packages stay npm-only:

• laya: it chooses its I/O module per runtime (browser or Node) with a package.json imports split that JSR can’t express. Under Deno, install it from npm.
• laya-cli: a Node/Bun command-line tool.
• backend-mlx-darwin-arm64: the native binaries. Deno users add it with deno add npm:@johnhenry/backend-mlx-darwin-arm64.