SkillCompiler/data/skills-bench/tasks-extra/mhc-layer-impl/environment/skills/modal-gpu/references/getting-started.md

Getting Started with Modal

Basic Structure

Every Modal script follows this pattern:

import modal

# 1. Create an app
app = modal.App("my-training-app")

# 2. Define the container image
image = modal.Image.debian_slim(python_version="3.11").pip_install(
    "torch",
    "einops",
    "numpy",
)

# 3. Define a GPU function
@app.function(gpu="A100", image=image, timeout=3600)
def train():
    import torch
    device = torch.device("cuda")
    print(f"Using GPU: {torch.cuda.get_device_name(0)}")

    # Your training code here
    return results

# 4. Local entrypoint
@app.local_entrypoint()
def main():
    results = train.remote()
    print(results)

Running Scripts

# Set up Modal token (if not already done)
modal token set --token-id <id> --token-secret <secret>

# Run the script
modal run train_modal.py

# Run in background
modal run --detach train_modal.py

Key Concepts

Concept	Description
App	Container for your Modal functions
Image	Container image with dependencies
@app.function	Decorator to mark functions for remote execution
@app.local_entrypoint	Entry point that runs locally
.remote()	Call a function remotely on Modal infrastructure

Image Building

Install dependencies at image build time, not runtime:

# CORRECT - install in image
image = modal.Image.debian_slim(python_version="3.11").pip_install(
    "torch",
    "einops",
)

# WRONG - install at runtime (slow, unreliable)
@app.function(...)
def train():
    import subprocess
    subprocess.run(["pip", "install", "torch"])  # Don't do this

Print Statements

Print statements in remote functions appear in Modal logs in real-time:

@app.function(gpu="A100", image=image, timeout=3600)
def train():
    print("Starting training...")  # Appears in logs
    for step in range(100):
        if step % 10 == 0:
            print(f"Step {step}")  # Progress updates
    return {"done": True}