torchscript

PyTorch/libtorch compiled to WebAssembly for TypeScript.

This package provides TypeScript bindings for PyTorch tensor operations compiled to WebAssembly using Emscripten. It works in both Node.js and browser environments.

Features

• Tensor creation (ones, zeros, from data)
• Basic tensor operations (add, mul, matmul)
• TorchScript model loading and inference
• Works in Node.js and browsers via WebAssembly

Installation

npm install torchscript
# or
pnpm add torchscript

Usage

import { loadModule, ones, zeros, tensor, DType } from 'torchscript';

// Initialize the WASM module
await loadModule();

// Create tensors
const a = await ones([2, 3]);
const b = await zeros([2, 3]);
const c = await tensor([1, 2, 3, 4, 5, 6], [2, 3]);

// Tensor operations
const sum = a.add(b);
const product = a.mul(c);

// Matrix multiplication
const x = await ones([2, 3]);
const y = await ones([3, 4]);
const result = x.matmul(y);

// Get tensor data
console.log(result.shape);  // [2, 4]
console.log(result.toArray());

// Free tensors when done
a.free();
b.free();
c.free();
sum.free();
product.free();
x.free();
y.free();
result.free();

Loading TorchScript Models

import { loadModule, load, tensor } from 'torchscript';

await loadModule();

// Load a TorchScript model
const model = await load('model.pt');

// Create input tensor
const input = await tensor([/* your input data */], [1, 3, 224, 224]);

// Run inference
const output = model.forward(input);

console.log(output.shape);
console.log(output.toArray());

// Clean up
input.free();
output.free();
model.free();

Building from Source

Prerequisites

• Docker
• Node.js 18+
• pnpm

Build Steps

1. Build the Docker image with Emscripten and PyTorch:

   pnpm wasm:docker-build

2. Build the WASM module:

   pnpm wasm:build

3. Build the TypeScript wrapper:

   pnpm build:js

Or run all steps at once:

pnpm build

Note: The initial Docker build is slow as it compiles PyTorch from source.

API Reference

Functions

• loadModule(): Promise<void> - Initialize the WASM module (must be called before other operations)
• version(): Promise<string> - Get PyTorch version
• ones(shape: number[], dtype?: DType): Promise<Tensor> - Create tensor filled with ones
• zeros(shape: number[], dtype?: DType): Promise<Tensor> - Create tensor filled with zeros
• tensor(data: number[] | Float32Array, shape: number[]): Promise<Tensor> - Create tensor from data
• load(filename: string): Promise<Module> - Load a TorchScript model

Tensor Class

• shape: number[] - Tensor dimensions
• numel: number - Number of elements
• toArray(): Float32Array - Get tensor data
• add(other: Tensor): Tensor - Element-wise addition
• mul(other: Tensor): Tensor - Element-wise multiplication
• matmul(other: Tensor): Tensor - Matrix multiplication
• free(): void - Free tensor memory

Module Class

• forward(input: Tensor): Tensor - Run model inference
• free(): void - Free module memory

DType Enum

• DType.Float32 - 32-bit floating point
• DType.Float64 - 64-bit floating point
• DType.Int32 - 32-bit integer
• DType.Int64 - 64-bit integer

License

MIT

Credits

Based on:

• libtorch-wasm - PyTorch WASM compilation
• libpg-query-node - Package structure patterns