ameva-cluster (PyPI)

Symmetric Disaggregated Mobile RAM Pooling & On-Device AI Acceleration Runtime with Compute-Memory Decoupling

Architecture & Overview

Executing large frontier neural networks on standalone edge mobile devices inevitably causes out-of-memory crashes via the Android Low Memory Killer (LMK) or severe thermal throttling.

ameva-cluster resolves this limitation by decoupling compute from memory:

• Compute-Memory Decoupling: Master flagship GPU performs 100% of forward passes; worker nodes act purely as remote LPDDR memory pools with zero GPU workload.
• Protocol Guard Proxy & MasterTunnel: Enforces time-windowed HMAC-SHA256 mutual authentication, blocking unauthorized probes in < 0.001s.
• Android LMK 300MB Guardband: Protects against unexpected kernel process terminations.
• Storage-Assisted Hybrid Failover: Gracefully transitions to local UFS 3.1 flash memory-mapped streaming if an edge worker disconnects.

Installation

pip install ameva-cluster

Quickstart

1) Start Protected Worker Daemon (Remote Android Device)

# Starts worker with automatic WakeLock acquisition and Guard Proxy on port 50052
ameva-cluster worker --port 50052 --guard-band 300

2) Execute Disaggregated Master Inference (Python SDK)

from ameva_cluster import AmevaCluster, ClusterHardwareProbe

# Inspect available physical RAM minus safety buffer
mem = ClusterHardwareProbe.get_memory_info(guard_band_mb=300)
print(f"Usable RAM: {mem['usable_mb']}MB")

# Run distributed inference
AmevaCluster.run_master(
    model_path="Qwen2.5-7B-Instruct-Q4_K_M.gguf",
    workers=["192.0.2.10:50052", "192.0.2.11:50052"],
    tensor_split="40,30,30",
    prompt="Explain disaggregated neural tensor memory architecture."
)

6-Modality Distributed CLI Integration

Interconnects with all six official on-device AI modality runtimes:

# 1. Large Language Models (LLM)
termux-llama run -m Qwen2.5-7B.gguf --rpc 192.0.2.10:50052 -ts auto -p "Hello"

# 2. Image Diffusion
termux-diffusion generate --model z_image_turbo-q4.gguf --rpc 192.0.2.11:50052 -ts 50,50

# 3. Multimodal Vision
termux-vision analyze --model Qwen2.5-VL-7B.gguf --mmproj mmproj.gguf --image test.png --rpc 192.0.2.10:50052

# 4. Speech-to-Text (STT)
termux-stt transcribe audio.wav --model whisper-large-v3-turbo.bin --rpc 192.0.2.10:50052

# 5. Text-to-Speech (TTS)
termux-tts synthesize "Clustering active." --rpc 192.0.2.11:50052

# 6. LoRA Fine-Tuning
termux-train lora --base-model Qwen2.5-3B.gguf --dataset data.jsonl --rpc 192.0.2.10:50052

Official Documentation

• Architecture & API Documentation
• Guard Protocol Specification
• 6-Modality Sharding Guide
• GitHub Repository

License

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