flexium 0.2.0a1

Flexium.AI - Flexible Resource Allocation for GPU training with live migration

Flexium

Zero-downtime GPU migration for PyTorch training
Move running training jobs between GPUs without stopping them.

Quick Start • Features • How It Works • Sign Up Free

---

What is Flexium?

Flexium enables live GPU migration for PyTorch training. Move your training between GPUs without stopping, checkpointing, or losing progress.

import flexium.auto

with flexium.auto.run():
    # Your existing training code - unchanged
    for epoch in range(100):
        for batch in dataloader:
            loss = model(batch).sum()
            loss.backward()
            optimizer.step()
    # Training transparently migrates between GPUs when needed

Zero code changes. Zero downtime. Zero memory residue.

Features

• Live Migration - Move training between GPUs without stopping
• Zero Code Changes - Just wrap your training loop
• Zero Memory Residue - Source GPU is completely freed
• Works Everywhere - PyTorch, Lightning, HuggingFace, and more
• Web Dashboard - Visual GPU management and monitoring
• Spot Instance Support - Survive preemptions automatically

Quick Start

1. Install

pip install flexium

2. Get Your API Key

Sign up for free at https://flexium.ai

3. Add Two Lines to Your Training

import flexium.auto

with flexium.auto.run():
    # Your existing training code - no changes needed
    model = MyModel().cuda()
    for epoch in range(100):
        train_one_epoch(model)

That's it! Your training is now migratable.

How It Works

┌─────────────────────────────────────────────────────────────┐
│                    YOUR TRAINING CODE                        │
│                                                              │
│   with flexium.auto.run():                                  │
│       model.train()   ◄─── Training runs normally           │
└─────────────────────────────────────────────────────────────┘
                              │
                              │ Heartbeat
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                   FLEXIUM CLOUD                              │
│                                                              │
│   Monitors your GPUs and triggers migrations when needed    │
│   - Spot preemption? Auto-migrate to another GPU            │
│   - GPU failure? Auto-recover to healthy GPU                │
│   - Rebalancing? Move to optimize utilization               │
└─────────────────────────────────────────────────────────────┘

When migration is needed, Flexium:

1. Captures complete GPU state at driver level
2. Frees source GPU completely (0 MB residue)
3. Restores on target GPU seamlessly

Your training code never knows it moved.

Use Cases

Spot Instance Training

# Survive spot preemptions without losing progress
with flexium.auto.run():
    train_large_model()  # Auto-migrates when spot is reclaimed

GPU Rebalancing

# Free up GPUs for teammates without killing your job
with flexium.auto.run():
    train_model()  # Can be migrated via dashboard anytime

Hardware Failure Recovery

# Auto-recover from GPU errors
with flexium.auto.run():
    train_model()  # Auto-migrates on ECC errors, driver crashes

Requirements

• Python 3.8+
• PyTorch 2.0+
• NVIDIA GPU with driver 580+
• CUDA 12.0+
• Linux x86_64

Documentation

• Getting Started
• API Reference
• Examples
• Troubleshooting

Support

• Docs: docs.flexium.ai
• Discord: Join our community
• Issues: GitHub Issues
• Email: [e-mail]

License

Apache 2.0 - See LICENSE for details.

---

Built with ❤️ by the Flexium team
flexium.ai