AMEVA-Forge Documentation

Release 1.0.0 [ Live WebGPU Studio ] [ GitHub Repository ]

API Reference (Release 1.0.0)

Complete documentation for all modules, classes, and methods supported in AMEVA-Forge Release 1.0.

1. Core Tensor (forge.Tensor)

Method / Attribute Description
forge.tensor(data, requires_grad=False, device='cpu') Creates a new multi-dimensional tensor from array or nested lists.
tensor.to(device) / tensor.move_to_(device) Transfers tensor storage between CPU and WebGPU in-place preserving parameter identity.
tensor.backward(gradient=None) Computes vector-Jacobian products (VJP) across the autograd graph.
await tensor.numpy_async() Asynchronously reads back GPU buffer to NumPy array without blocking browser UI.

2. Neural Network Modules (forge.nn)

Module Description
nn.Linear(in_features, out_features, bias=True) Applies an affine linear transformation \(y = xA^T + b\).
nn.LayerNorm(normalized_shape, eps=1e-5) Applies Layer Normalization over the specified normalized shape.
nn.BatchNorm2d(num_features, eps=1e-5, momentum=0.1) Applies 2D Batch Normalization with train/eval state tracking.
nn.PositionalEncoding(d_model, max_len=5000) Sinusoidal positional encoding with LRU dynamic caching.
nn.Dropout(p=0.5) Applies randomized zeroing during training mode; identity during eval.
nn.Sequential(*layers) Sequential container for chaining neural network modules.

3. Functional & Loss Operations (forge.functional)

Function Description
F.scaled_dot_product_attention(q, k, v, is_causal=False) Computes attention \(\text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V\) with optional causal masking.
F.cross_entropy(input, target) Computes Cross-Entropy loss with mathematically exact closed-form GPU derivative.
F.softmax(input, dim=-1) Applies numerically stabilized softmax along the specified dimension.
nn.MSELoss() Measures mean squared error between prediction and ground truth.

4. Optimizers (forge.optim)

Optimizer Description
optim.SGD(params, lr=0.01, momentum=0.0) Stochastic Gradient Descent optimizer supporting asynchronous WebGPU in-place AXPY kernel updates.