termux-train (v1.1.0)

Production-Grade Native On-Device Deep Learning & LoRA Training Framework for Android Termux & ARM64.
Dual-Engine Architecture: Native Python Autograd DAG + Node.js / TypeScript SDK.

📌 Key Architectural Highlights

• ⚡ Production Dual-Engine: Seamless CLI & SDK parity across Python (termux_train) and Node.js/TypeScript (termux-train).
• 🛡️ HuggingFace SafeTensors Hardening: 100MB Header Bomb defense and zero-copy binary checkpointing with optimizer momentum.
• 💾 2M-Sample Bounded MMap Dataset: High-throughput binary .bin token stream loader operating within constant <50MB RAM.
• 🧬 LoRA & RoPE Attention: Rank-decomposition parameter-efficient fine-tuning with transactional snapshot rollback & Rotary Position Embeddings.
• 🌐 Multilingual ByteTokenizer: Native UTF-8 tokenizer with full Korean Hangul, CJK ideographs, Emoji, and UTF-8 BOM support.
• 🚀 One-Touch Universal Installer: install.sh for one-click setup across Android Termux, Linux, and macOS.

🚀 Installation & Quickstart

1. Universal One-Touch Installation (Recommended)

curl -fsSL https://raw.githubusercontent.com/uno-km/termux-train/main/install.sh | bash

2. Python Package (PyPI)

# Standard pure Python + NumPy Autograd engine
pip install termux-train

# With Vulkan GPU acceleration
pip install "termux-train[vulkan]"

3. Node.js Global CLI (NPM)

npm install -g termux-train

🛠️ CLI Usage Guide

# 1. Hardware & Vulkan GPU Diagnostics
termux-train doctor
# or via Python:
python3 -m termux_train.cli doctor

# 2. On-Device GEMM & Autograd Latency Benchmark
termux-train benchmark --dim 256

# 3. Train MLP / LoRA / Transformer with Checkpoints
termux-train train --model lora --dim 64 --rank 8 --epochs 5 --checkpoint ./adapter.safetensors

# 4. Stream 2,000,000+ Sample MMap Dataset
termux-train train --data ./corpus.bin --batch-size 32 --epochs 3

📖 Node.js & TypeScript SDK Usage

import { TermuxTrainer, runDoctor, runBenchmark } from 'termux-train';

// 1. Diagnostics
const doc = runDoctor();
console.log(`Hardware Tier: ${doc.hardware.tier} | RAM: ${doc.hardware.totalRamMb}MB`);

// 2. Training Session
const trainer = new TermuxTrainer();
const result = await trainer.train({
  modelType: 'lora',
  dim: 64,
  loraRank: 8,
  epochs: 5,
  lr: 0.001,
  checkpointPath: './adapter.safetensors'
});

console.log(`Training complete! Final Loss: ${result.finalLoss}`);

📖 Official Documentation & Portal

• Official Architecture & API Reference
• Ecosystem Metrics & Registry Stats
• AMEVA Open-Source Foundation

📄 License

Licensed under the Apache-2.0 License. Copyright (c) 2026 Eunho Kim (@uno-km).