AMEVA-Forge (ameva-forge)

High-Performance Client-Side Tensor Computation Engine & Reverse-Mode Autograd Framework Powered by WebGPU.

Developed and maintained by the AMEVA Foundation (아메바 재단), AMEVA-Forge is an industrial-grade, zero-server-cost deep learning library engineered to execute high-throughput tensor operations, automated differentiation, and end-to-end neural model training natively within client runtimes (WebGPU, WASM/Pyodide, and native Python environments).

---

Architectural Pillars

+-----------------------------------------------------------------------------------+
|                            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        |
+-----------------------------------------------------------------------------------+

1. Deterministic Autograd & Topological Execution
   Strict reverse-mode automatic differentiation graph with cycle detection, multi-output tuple bindings, in-place version invalidation, 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 tracking.
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 graphical models (distributions).
4. Zero-Server Infrastructure (Edge & Browser)
   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

Install the official package from PyPI:

pip install ameva-forge

Or install from source with development dependencies:

git clone https://github.com/uno-km/ameva-forge.git
cd ameva-forge/packages/forge-py
pip install -e .

---

Quick Start

1. Basic Tensor & Automated Differentiation

import forge as fg

# Initialize tensors with gradient tracking
x = fg.tensor([[1.0, 2.0], [3.0, 4.0]], requires_grad=True)
w = fg.tensor([[0.5, -0.5], [1.0, 2.0]], requires_grad=True)
b = fg.tensor([0.1, -0.1], requires_grad=True)

# Forward pass: Linear projection + GELU activation
y = fg.matmul(x, w) + b
loss = fg.sum(fg.nn.functional.gelu(y))

# Compute Vector-Jacobian Products (Autograd backward)
loss.backward()

print("Loss Value :", loss.numpy())
print("Gradient dL/dw :\n", w.grad.numpy())

---

2. Character-Level Transformer (NanoGPT)

Train a complete causal autoregressive transformer directly on your local device:

import forge as fg
import forge.nn as nn
from forge.models.nanogpt import GPT, GPTConfig

# Define model configuration
config = GPTConfig(
    block_size=32,
    vocab_size=64,
    n_layer=4,
    n_head=4,
    n_embd=64,
    bias=False
)
model = GPT(config)
optimizer = fg.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-2)
criterion = nn.CrossEntropyLoss()

# Training step (Batch Size: 8, Sequence Length: 32)
input_tokens = fg.tensor([[1, 5, 12, 3]], dtype="int32")
target_tokens = fg.tensor([[5, 12, 3, 18]], dtype="int32")

optimizer.zero_grad()
logits = model(input_tokens)
loss = criterion(logits, target_tokens)
loss.backward()
optimizer.step()

---

3. Speech-to-Text & Acoustic Signal Processing (forge.fft + forge.nn)

Compute real Fourier Mel-spectrograms from raw acoustic waveforms:

import forge as fg
import forge.nn as nn

# 16kHz PCM audio waveform (Batch: 4, Samples: 8000)
raw_audio = fg.tensor(audio_data, dtype="float32")

# Fast Fourier Transform (Complex Spectrum)
fft_complex = fg.fft.rfft(raw_audio, n=1024, dim=-1)

# Power Spectrogram Energy
power_spec = (fft_complex.real.pow(2.0) + fft_complex.imag.pow(2.0) + 1e-6).log()

# 1D Convolutional Audio Feature Extractor
conv = nn.Conv1d(in_channels=513, out_channels=64, kernel_size=3, padding=1)
audio_features = conv(power_spec)

---

Comprehensive Module Directory

Module	Core Functionality	Key Operators / Classes
forge	Core Tensor Engine & Factories	tensor, zeros, ones, randn, matmul, einsum, reshape, permute, where
forge.nn	Deep Learning Layers & Containers	Linear, Conv1d, Conv2d, MultiheadAttention, LayerNorm, RMSNorm, BatchNorm2d, Embedding, CrossEntropyLoss, MSELoss
forge.optim	Optimizers & Rate Schedulers	SGD, Adam, AdamW, RMSprop, CosineAnnealingLR, StepLR
forge.linalg	Linear Algebra Decomposition	norm, svd, qr, cholesky, inv, pinv, det, matrix_rank, solve, eigh
forge.fft	Discrete Fourier Transforms	rfft, irfft, fft, ifft, fft2, ifft2, rfft2, irfft2, fftfreq, fftshift
forge.special	Transcendental & Error Functions	erf, erfc, erfinv, gammaln, digamma, expm1, log1p, expit, logit, sinc, i0, xlogy
forge.distributions	Probability Distributions & KL	Normal(rsample), Uniform, Bernoulli, Categorical, kl_divergence
forge.models	Pre-architected Reference Models	GPT, GPTConfig, LLaMA

---

In-Browser Zero-Install Execution (WebGPU + Pyodide)

AMEVA-Forge packages a single bundled JavaScript distribution (forge-py-bundle.js) that mounts into browser-native Pyodide runtimes:

<script src="https://cdn.jsdelivr.net/pyodide/v0.26.2/full/pyodide.js"></script>
<script src="https://uno-km.github.io/ameva-forge/dist/forge-py-bundle.js"></script>
<script>
  async function runClientDeepLearning() {
    let pyodide = await loadPyodide();
    await window.loadAmevaForgeBundle(pyodide);
    
    await pyodide.runPythonAsync(`
      import forge as fg
      x = fg.randn((1024, 1024), device="gpu")
      y = fg.matmul(x, x)
      print("Computed 1024x1024 on WebGPU Hardware:", y.shape)
    `);
  }
  runClientDeepLearning();
</script>

---

🌍 The AMEVA Foundation (아메바 재단)

AMEVA-Forge is an open-source deep learning initiative created and governed by the AMEVA Foundation (아메바 재단).

🎯 Our Mission & Zero-Server Paradigm

The AMEVA Foundation is dedicated to the democratisation of client-side artificial intelligence and edge computing. We envision a decentralized web where deep learning training, fine-tuning, and neural inference occur directly on user devices (browsers, laptops, tablets, and smartphones)—eliminating massive cloud server costs, safeguarding user data sovereignty with zero network leakage, and providing zero-latency neural intelligence everywhere.

🌟 Core Foundation Initiatives

• 🌐 AMEVA-Forge Tensor Engine: High-performance WebGPU tensor computation & reverse-mode autograd engine.
• 🎙️ On-Device Whisper STT: Real-time acoustic FFT signal processing & speech recognition running locally in browser.
• 🤖 Decentralized LLM Transformers: Zero-install character and subword transformer models with zero cloud dependency.
• 🔒 Privacy-First AI Sovereignty: Sovereign on-device machine learning where private data never leaves client silicon.

• Foundation Portal: https://uno-km.github.io/ameva-forge/
• Official Repository: https://github.com/uno-km/ameva-forge
• Community & Governance: https://github.com/uno-km/ameva-forge/issues
• Sponsorship & Research Partnerships: https://github.com/uno-km

---

Quality Assurance & Verification

Every release of AMEVA-Forge undergoes rigorous multi-tier verification:

• 292 Unit & Stress Tests: 100% automated pass rate across CPU, GPU fallback, mathematical accuracy, and memory quota managers.
• Finite-Difference Gradcheck: Numerical gradient validation against analytical Vector-Jacobian backward formulations.
• Memory Lifecycle Audit: Zero-leak allocation token reclamation and buffer recycling verification across 10,000+ continuous execution cycles.

---

License & Citation

AMEVA-Forge is licensed under the MIT License.

@software{ameva_forge_2026,
  author = {AMEVA Foundation},
  title = {AMEVA-Forge: High-Performance WebGPU-Accelerated Tensor Computation Engine},
  year = {2026},
  publisher = {GitHub},
  url = {https://github.com/uno-km/ameva-forge}
}