Material for the SC26 tutorial Accelerating and Scaling Python for HPC, Monday, 16 November 2026, 8:30am to 5:00pm CST, room W185a.
Python is powering breakthrough exascale scientific discoveries - come learn how to program the world's largest supercomputers with it! In this interactive tutorial you'll learn how to write, debug, profile, and optimize high-performance, multi-node GPU applications in Python. You'll learn and master: CuPy for drop-in GPU acceleration of NumPy workflows, Numba for writing custom kernels that match the performance of C++ and Fortran, mpi4py for scaling across thousands of nodes, DragonHPC for workflow acceleration with cuPhoton for astronomical image processing, and CppJIT for on-demand Python/C++ interoperability and automatic binding generation, including compiler-based automatic differentiation of C++ with Clad. Everything is hands-on: short, interactive lectures by expert instructors are paired with guided Jupyter notebooks that introduce each concept and have you apply it immediately in bite-sized exercises, among them a mini-app applying several programming models (JAX, cuda.compute, Numba CUDA) to simulated cosmic-web data and a CARTopiaX CAR T-cell therapy example. We'll use a web-based environment that requires no installs or setup - just a laptop and a browser. Whether you're a domain scientist seeking faster turnaround or a software engineer evaluating portable acceleration strategies, you'll leave with a roadmap, skills, and code for bringing scalable Python practices back to your own HPC facility.
In this tutorial we will cover:
- NumPy fundamentals and the
ndarraymemory model. - Accelerating array workflows on the GPU with CuPy, including memory spaces and asynchrony.
- Distributed computing with mpi4py for multi-node parallelism.
- Programming models in Python for GPU applications: array expressions, fusion, library iterators, kernel authoring, and interoperability with CUDA C++ kernels.
- Reading an application's profile with Nsight Systems and Nsight Compute.
- Case studies of horizontal Python and C++ integration in domain sciences such as cancer research and high energy physics.
- DragonHPC fundamentals enabling
multiprocessingto scale across HPC multi-node resources. - cuPhoton astronomical image processing workflows orchestrated using DragonHPC.
Attendees are expected to have general knowledge of HPC and Python programming. Some familiarity with NumPy and C/C++ is helpful but not required. The outline and the notebook list are in notebooks/syllabi/sc26__8_hours.ipynb.
| Path | Content |
|---|---|
notebooks/fundamentals/ |
NumPy and CuPy |
notebooks/distributed/ |
mpi4py, DragonHPC and cuPhoton |
notebooks/gpu-mini-app/ |
Analysis of cosmic-web density fields on GPU |
notebooks/science-integration/ |
Python/C++ integration examples for domain science |
notebooks/syllabi/ |
Tutorial outline |
slides/ |
Slides and material |
brev/ |
Docker image and compose stack. Build scripts and checks |
test/ |
Package and notebook tests run during image build |
The tutorial runs in the container image ghcr.io/pyhpc/sc2026-tutorial. brev/docker-compose.yml
starts JupyterLab on port 8888, opening on the outline, with Nsight Systems on port 8080 and Nsight Compute on
port 8081 as launcher tiles. Brev starts this file; docs/brev.md is the attendees' guide. Locally,
with Docker and the NVIDIA Container Toolkit installed:
docker compose -f brev/docker-compose.yml up -d
docker compose -f brev/docker-compose.yml down -vDevelopment
docker compose -f brev/docker-compose.yml -f brev/docker-compose.dev.yml up -d # this checkout mounted in the containers
docker compose -f brev/docker-compose.yml run --rm --no-deps jupyter test # the test suite
docker compose -f brev/docker-compose.yml run --rm --no-deps jupyter shell # a shell in the container
brev/build.bash [--chain] [--push] # build and test the image
python3 brev/check-notebooks.py [--fix] # notebook format
python3 brev/check-links.py # relative linksThe two checks also run as pre-commit hooks (pre-commit install) and in the Checks workflow. The Image
workflow builds the image on a self-hosted GPU runner set up with brev/runner-setup.bash, runs every notebook
in it, starts the Compose stack from it and publishes it: a pull request from a branch of this repository gets
the tag pr-<number>, a push to main the tags git-<sha> and latest.
Written material is licensed under CC BY-NC-SA 4.0 (LICENSE) and code under Apache-2.0 (LICENSE-CODE).