Virtual Adaptive Offloading System (vAOS-Engine)

The vAOS-Engine is a lightweight, PyTorch-native runtime controller designed to break the GPU Memory Wall on consumer-grade hardware (e.g., 6GB VRAM laptops).

By virtualizing host CPU RAM as an OS-bypass mmap pool, dynamically scheduling asynchronous PCIe transfers, and utilizing late-stage register-level INT4 dequantization, vAOS allows you to train Massive Transformer models that natively trigger Out-Of-Memory (OOM) crashes.

🚀 Installation

pip install vaos-engine 

🛠️ Complete Setup Guide & Live Demo

Want to see vAOS in action on your own hardware? You can run a live diagnostic test. The engine will auto-probe your hardware, intentionally crash a heavy model to prove your physical memory limits, and then use the vAOS runtime to train it successfully.

Step 1: Create the Test Script

Create a new file named demo.py anywhere on your computer and paste the following code:

import torch
import torch.nn as nn

# Import the vAOS API and our packaged Heavy Transformer
from src.api import evaluate_system
from src.baseline.model import DummyHeavyTransformer

# 1. Setup a heavy model that normally requires >9GB VRAM
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = DummyHeavyTransformer(
    d_model=2048, 
    n_layers=4, 
    n_heads=16, 
    d_ff=8192, 
    vocab_size=32000, 
    max_seq_len=512
).to(device)

# 2. Generate dummy training data
dummy_input = torch.randint(0, 32000, (16, 512), device=device)
dummy_target = torch.randint(0, 32000, (16, 512), device=device)
criterion = nn.CrossEntropyLoss()

# 3. Trigger the vAOS Auto-Analyzer
print("Starting vAOS Hardware Evaluation...")
evaluate_system(model, dummy_input, dummy_target, criterion)

Step 2: Run the Script

Open your terminal (PowerShell, CMD, or Linux terminal) and execute the file:

python demo.py

Step 3: View your Personalized Report

The engine will execute a ruthless native control trace (until the GPU runs out of memory), followed immediately by the vAOS optimized trace. Once finished, it will automatically generate a comprehensive, personalized Markdown file named vAOS_Evaluation_Report.md in your current directory. Open it to see your exact mathematical metrics, including:

VRAM effectively saved.

PCIe bandwidth reduction (via INT4 dequantization).

Network synchronization payload optimization (via 90th-percentile Adaptive Sparsification).