modal-gpu

1# Modal GPU Training
2 
3## Overview
4 
5Modal is a serverless platform for running Python code on cloud GPUs. It provides:
6 
7- **Serverless GPUs**: On-demand access to T4, A10G, A100 GPUs
8- **Container Images**: Define dependencies declaratively with pip
9- **Remote Execution**: Run functions on cloud infrastructure
10- **Result Handling**: Return Python objects from remote functions
11 
12Two patterns:
13- **Single Function**: Simple script with `@app.function` decorator
14- **Multi-Function**: Complex workflows with multiple remote calls
15 
16## Quick Reference
17 
18| Topic | Reference |
19|-------|-----------|
20| Basic Structure | [Getting Started](references/getting-started.md) |
21| GPU Options | [GPU Selection](references/gpu-selection.md) |
22| Data Handling | [Data Download](references/data-download.md) |
23| Results & Outputs | [Results](references/results.md) |
24| Troubleshooting | [Common Issues](references/common-issues.md) |
25 
26## Installation
27 
28```bash
29pip install modal
30modal token set --token-id <id> --token-secret <secret>
31```
32 
33## Minimal Example
34 
35```python
36import modal
37 
38app = modal.App("my-training-app")
39 
40image = modal.Image.debian_slim(python_version="3.11").pip_install(
41    "torch",
42    "einops",
43    "numpy",
44)
45 
46@app.function(gpu="A100", image=image, timeout=3600)
47def train():
48    import torch
49    device = torch.device("cuda")
50    print(f"Using GPU: {torch.cuda.get_device_name(0)}")
51 
52    # Training code here
53    return {"loss": 0.5}
54 
55@app.local_entrypoint()
56def main():
57…