Termux-Train

Unified Multimodal On-Device Deep Learning & LoRA Training Framework for Android Termux with 6-Modality Adapters (LLM, Diffusion, VLM, STT, TTS, BitNet), GPU Slicing, and 44GB Disaggregated Cluster Virtual RAM Pooling

1-Line Quick Installation

Install the official package directly into your runtime:

pip install termux-train
# or via npm:
npm install -g termux-train
# or standalone bootstrap:
bash <(curl -sSL https://raw.githubusercontent.com/uno-km/termux-train/main/install.sh)

The Engineering Challenge

Standard deep learning training systems (PyTorch, DeepSpeed, Megatron) require multi-gigabyte build toolchains unavailable on Android Bionic, lack native mobile GPU autograd dispatch, and rapidly trigger the Android Low Memory Killer (LMK) during backward activation accumulation.

The Architectural Breakthrough

Termux-Train v2.0.1 delivers a dependency-free C-vectorized DAG Autograd engine, 6-Modality target-aware LoRA/DoRA training (LLM, Image Diffusion, Vision VLM, Whisper STT, TTS, BitNet 1.58-bit), Mali/Adreno GPU Slicing, 300MB Guard-Band memory protection, and 44GB disaggregated AMEVA Cluster virtual RAM pooling across heterogeneous mobile fleets.

Key Capabilities & Built-in Hardening

6-Modality Target-Aware Adapter Hub

Directly trains and exports native SafeTensors and _config.json sidecars for LLM (llama.cpp/GGUF), Image Diffusion (ComfyUI/Diffusers), Vision VLM (termux-vision), Whisper STT (termux-stt), TTS (termux-tts), and BitNet 1.58-bit ternary models without format conversion steps.

Single-Device GPU Slicing (Mali & Adreno)

Prevents mobile unified memory allocation spikes via Vocab Slicing (--vocab-slice), Chunked Layer Dispatch (--chunk-layers), and Layer Streaming (--stream-layers), preventing GPU watchdog resets.

44GB Disaggregated Cluster Virtual RAM Pooling

Aggregates physical RAM across heterogeneous mobile devices (S25, S21, S20, A53, A35) into a unified 44GB virtual memory pool with 300MB Guard-Band memory protection against Android LMK.

On-Device Reinforcement Learning (GRPO / DPO / PPO)

Executes client-side DeepSeek-R1 style Group Relative Policy Optimization (GRPO), Direct Preference Optimization (DPO), and PPO directly on mobile hardware.

Persistent Latent & Feature Caching

Avoids redundant VAE decoding and Mel filterbank recalculation across training epochs through deterministic SHA-256 keyed .safetensors on-device disk caching.

100% Dual-Engine Parity (Python & Node.js)

Guarantees identical subcommands, parameter names, and exit codes across Python API, Python CLI, Node.js SDK, and Node.js Global CLI.

Supported Compute Kernels & Operations

Subsystem Category	Operations & Kernels	Status
LLM / Text PEFT	LoRA, DoRA, TinyTransformerLM, RoPE Rotary Embeddings, GGUF PEFT Export	Production Verified
Image Diffusion	DDPMScheduler, Sinusoidal Timestep MLP, CrossAttentionLoRA, ComfyUI SafeTensors	Production Verified
Vision VLM	MultimodalProjectorLoRA, Visual Cross-Attention, LLaVA/Qwen2-VL QA Dataset	Production Verified
Audio STT	Whisper Cross-Attention LoRA, Mel Spectrogram Caching, Acoustic Fine-Tuning	Production Verified
Audio TTS	Speaker Style Adaptation LoRA, Phonetic-to-Mel Alignment, Voice Synthesis	Production Verified
BitNet 1.58-bit	Ternary Quantization-Aware Training, Dynamic Activation Scaling, bitnet.cpp Export	Production Verified
Cluster Virtual RAM	44GB Virtual RAM Pooling, Proportional Sharding, Distributed Pipeline Session	Production Verified

Canonical Usage Example

import termux_train as tt

# 1. Image Diffusion LoRA Training (ComfyUI / Diffusers compatible)
tt.diffusion.train_diffusion_lora(
    image_dir="./training_images",
    output_path="./adapter_diffusion.safetensors",
    prompt="high quality technical schematic",
    resolution=512,
    epochs=5,
    lr=0.0001,
    rank=8,
    backend="vulkan"
)

# 2. Vision Multimodal VLM LoRA Training (LLaVA / Qwen2-VL format)
tt.vision.train_vision_vlm_lora(
    data_source="./vlm_dataset.jsonl",
    output_path="./adapter_vision.safetensors",
    epochs=3,
    lr=0.0002,
    backend="vulkan"
)

# 3. Whisper STT LoRA Acoustic Adaptation
tt.stt.train_stt_lora(
    data_source="./audio_transcripts.jsonl",
    output_path="./adapter_stt.safetensors",
    epochs=4,
    backend="vulkan"
)

import { trainDiffusion, trainVision, trainSTT, TermuxTrainer } from 'termux-train';

// 1. Asynchronous Image Diffusion LoRA Training
await trainDiffusion({
  imageDir: './training_images',
  output: './adapter_diffusion.safetensors',
  prompt: 'high quality technical schematic',
  resolution: 512,
  epochs: 5,
  lr: 0.0001,
  rank: 8,
  backend: 'vulkan'
});

// 2. High-Performance LLM LoRA Fine-Tuning
const trainer = new TermuxTrainer({
  modelType: 'lora',
  loraRank: 8,
  backend: 'vulkan',
  vocabSlice: 4096,
  chunkLayers: 2
});
await trainer.fit({ epochs: 3, lr: 0.0002 });