@ameva/forge

High-Performance Browser-Native Tensor Computation Engine & Reverse-Mode Autograd Framework Powered by WebGPU.

An industrial-grade, zero-server-cost deep learning runtime engineered to execute high-throughput tensor operations, automatic differentiation, and end-to-end neural model training natively inside client WebGPU and WebAssembly runtimes.

Key Features

1. Deterministic Autograd & Topological Execution: Reverse-mode automatic differentiation graph with cycle detection, multi-output tuple bindings, in-place version tracking, and scalar-tensor memory optimization.
2. WebGPU Hardware Acceleration: Direct-to-silicon WGSL compute shaders featuring 8-dimensional non-contiguous stride dispatching, 2D workgroup partitioning ($65,535 \times 65,535$), and explicit buffer lifecycle recycling.
3. PyTorch 1:1 API Parity: Seamless drop-in compatibility across neural layers (nn.Module, nn.MultiheadAttention, nn.Conv2d), mathematical primitives (linalg, fft, special), and probabilistic models (distributions).
4. Zero-Server Infrastructure: Execute full model fine-tuning and inference directly inside the browser using Pyodide and WebGPU with zero cloud compute cost and total data privacy.

Installation

npm install @ameva/forge

Quick Start

import { Tensor, nn, optim } from '@ameva/forge';

// Initialize WebGPU Tensor Context
const x = new Tensor([1.0, 2.0, 3.0, 4.0], { shape: [2, 2], requiresGrad: true });
const w = new Tensor([0.5, -0.5, 1.0, 2.0], { shape: [2, 2], requiresGrad: true });

// Forward Pass
const y = x.matmul(w).relu().sum();

// Reverse-Mode Automatic Differentiation
y.backward();

console.log('Output Value:', y.item());
console.log('Gradients of x:', x.grad.toArray());

Architecture

+-----------------------------------------------------------------------------------+
|                            AMEVA-Forge User Space                                 |
|   forge.nn  |  forge.optim  |  forge.linalg  |  forge.fft  |  forge.distributions |
+-----------------------------------------------------------------------------------+
|                      Reverse-Mode Autograd DAG Engine                             |
|       Vector-Jacobian Products (VJP)  *  In-Place Mutation Version Locks          |
+-----------------------------------------------------------------------------------+
|                         Hardware Abstraction Layer                                |
|   CPU Backend (Vectorized C/NumPy)  <--->  WebGPU Backend (Async WGSL Kernels)    |
|   Staging Buffer Recycling Pool     <--->  Zero-Leak Allocation Token Ring        |
+-----------------------------------------------------------------------------------+

License

MIT License. Copyright (c) 2026 uno-km (AMEVA Foundation).