Welcome to AMEVA-Forge (Release 1.0.0)
AMEVA-Forge is a lightweight deep learning framework engineered for educational autograd, small model experimentation, and client-side privacy-preserving computation across both native Python environments and modern web browsers via WebGPU.
1-Line Quick Installation
Install via pip for native/Pyodide usage, or launch the zero-install WebGPU Live Studio:
🐍 Python Edition (PyPI):
pip install ameva-forgeRelease 1.0 Scope: 2-Layer MLP & Causal Attention
Our Release 1.0 focus is laser-targeted on one goal: Enable users to build, train, and understand neural networks (2-Layer MLP, Causal Attention, In-Place SGD) entirely in the browser using Python without server compute costs.
Core Capabilities
- 100% PyTorch-like API: Drop-in compatibility with PyTorch syntax (
torch.Tensor,nn.Linear,nn.LayerNorm,optim.SGD,loss.backward()). - Browser WebGPU Execution: Empowers deep learning operations to run client-side using Pyodide (WASM) and native WGSL compute shaders.
- Autograd Engine: Implements reverse-mode automatic differentiation with mathematically closed-form GPU gradients.
- Memory Quota Manager: Zero-leak staging buffer recycling and deterministic garbage collection finalizers.
Supported Operations & Modules in Release 1.0
| Category | Supported Operations & Features |
|---|---|
| Tensor Operations | add, sub, mul, div, matmul (2D), batched_matmul (3D BMM), transpose, permute, scatter, gather, 8D broadcasting |
| Neural Network (nn) | nn.Linear, nn.LayerNorm, nn.BatchNorm2d (train/eval), nn.PositionalEncoding (LRU cache), nn.Dropout, nn.Sequential |
| Functional & Loss | F.scaled_dot_product_attention (Causal), F.softmax, F.log_softmax, F.cross_entropy, nn.MSELoss, F.relu |
| Optimization (optim) | optim.SGD (In-place parameter updates with zero allocation overhead) |
| GPU Shader Kernels | elementwise.wgsl, matmul.wgsl (16x16 workgroup tiled), softmax.wgsl, layernorm.wgsl, cross_entropy.wgsl |