# termux-train (AMEVA-Termux)
> Native On-Device Deep Learning & LoRA Training Framework for Android Termux

## Core Overview
`termux-train` is a lightweight, zero-dependency deep learning and automatic differentiation (Autograd) training framework engineered specifically for Android Termux and edge devices. It enables on-device backpropagation, RoPE Transformers, and LoRA fine-tuning directly on smartphone hardware without requiring heavy PyTorch binaries or PRoot Linux virtualization.

## Distribution & Installation
- **PyPI Package**: `pip install termux-train`
- **Optional C Acceleration**: `pip install termux-train[accelerated]` (NumPy)
- **Termux Native**: `pkg update && pkg install python python-numpy && pip install termux-train`

## Key Architecture
1. **Core Autograd (`termux_train.Tensor`)**: Pure-Python dynamic computation graph with reverse-mode DAG autograd.
2. **Pluggable Backends**: `PythonBackend` (zero dependencies) and `NumPyBackend` (C-level OpenBLAS vectorization).
3. **Neural Network Layers (`termux_train.nn`)**: `Linear`, `LoRALinear`, `Embedding`, `LayerNorm`, `MultiHeadAttention`, `TransformerBlock`, `TinyTransformerLM`.
4. **Optimizers (`termux_train.optim`)**: `SGD` (Momentum, Nesterov), `Adam`, `AdamW`.
5. **Modern Checkpointing (`termux_train.checkpoint`)**: HuggingFace-compatible `.safetensors` binary zero-copy, lightweight LoRA adapter serialization (<100KB), atomic JSON.
6. **Data Streaming (`termux_train.data`)**: `MMapTokenDataset` for streaming multi-GB token corpora directly from disk without filling RAM.

## Canonical 10-Line Quickstart
```python
from termux_train import Tensor, nn, optim, checkpoint

# Define model
model = nn.Sequential(nn.Linear(16, 32), nn.ReLU(), nn.Linear(32, 2))
optimizer = optim.AdamW(model.parameters(), lr=0.01)
criterion = nn.CrossEntropyLoss()

# Forward, backward, and step
x = Tensor([[0.5] * 16])
target = Tensor([1], dtype="int64")

optimizer.zero_grad()
loss = criterion(model(x), target)
loss.backward()
optimizer.step()
```