GPU deployment Tutorial - SciPy MN 2026

This repository contains materials for the tutorial: Deploying and debugging GPU accelerated Python workloads

Running the tutorial

Brev

For the in-person version of this tutorial we will use NVIDIA Brev

Content

Module	Link
0 - Introduction Slides	introduction-to-gpu-stack
1 - Setup Brev	1-setup
2 - Setting Up Your GPU Python Environment	2-environments
3 - Verifying Your Environment with the RAPIDS CLI	3-rapids-cli-verify-env
4 - Easy GPU Acceleration Wins with RAPIDS Accelerators	4-acceleration-examples
5 - Monitoring and Debugging GPU Python Workloads	5-monitoring-and-debugging

1-Setup Brev

• Covers how to get access to a cloud GPU using NVIDIA Brev.
• Walks through launching a VM (choosing GPU type, provider, and deploying).
• Explains how to connect to the VM via SSH using the brev CLI on macOS, Linux, and Windows.

2-Setting Up Your GPU Python Environment

• Guides users through installing Python packages for GPU workloads.
• Supports three package managers: uv/pip, conda, and pixi via config files (pyproject.toml, environment.yaml, pixi.toml) to manage dependencies— pick one and follow that path.
• Covers checking for core CUDA libraries and creating isolated, reproducible environments.

3-Verifying Your Environment with the RAPIDS CLI

• Introduces two commands to validate a GPU environment after setup: rapids doctor and rapids debug.
• rapids doctor runs health checks on the driver, CUDA toolkit, GPU compute capability, and NVLink.
• Failed checks return actionable messages explaining how to fix the issue.
• Showcases how libraries can register their own checks, using cuml as an example.
• rapids debug produces a full snapshot of driver versions, installed packages, and build tools.
• rapids debug --json outputs a machine-readable version, useful for bug reports and sharing environment state.

4-Easy GPU Acceleration Wins with RAPIDS Accelerators

• Shows how to get GPU speedups on existing Python code with zero rewrites.
• Shows how to use the accelerators via command line and in notebooks.
• Discusses built-in --profile and --line-profile flags show which operations ran on the GPU vs. CPU.
• Introduces JupyterLab NVDashboard GPU Accelerators toggle that enables the same zero-code acceleration interactively in notebooks.

5-Monitoring and Debugging GPU Python Workloads

• A hands-on diagnostic guide for when GPU code runs but doesn't deliver the expected speedup.
• Covers a progression of tools, each answering a different question about what the GPU is doing.
• nvidia-smi and nvtop confirm driver-level visibility and live GPU utilization.
• cProfile + SnakeViz identify CPU-side bottlenecks before any GPU work begins.
• Nsight Systems (nsys) produces a full CPU/GPU timeline showing transfers, kernel launches, and idle gaps.
• Uses a real climate data workload (xarray EOF analysis) to walk through two common pitfalls:
  • Moving data back to the CPU before the expensive operation.
  • Naively porting a Python loop to the GPU, causing thousands of tiny kernel launches.
• Shows how to identify and fix both issues using profiler output and CUDA timeline traces.