CSCS Summer School 2026 event branch by brycelelbach · Pull Request #215 · NVIDIA/accelerated-computing-hub

brycelelbach · 2026-07-21T13:08:04Z

Summary

merge the July 2026 CSCS Summer School event branch into main
add the PyHPC tutorial and CSCS rootless Slurm Container Engine support
build and publish the PyHPC image for AMD64 and ARM64 with GitHub Actions

Review

This is a draft so the complete event-branch diff can be reviewed before merge.

Co-Authored-By: Claude <noreply@anthropic.com>

Port exercises 01, 03, 05, 06, 40, 41, 60 from the accelerated-python tutorial into tutorials/pyhpc/notebooks (flat, no subdirectories) and renumber the full set 00-13 in the order: 00 numpy_intro__ndarray_basics (AP 01) 01 numpy_to_cupy__ndarray_basics (AP 03) 02 memory_spaces__power_iteration (AP 05) 03 mpi4py__collectives (AP 60) 04 intro (SWE 00) 05 swe_core__reference_solver (SWE 01) 06 jax__lax_scan (SWE 02) 07 pyomp__parallel_for (SWE 03) 08 nanobind__cpp_kernel (SWE 04) 09 cppjit__gpu_thrust (SWE 05) 10 synthesis (SWE 06) 11 asynchrony__power_iteration (AP 06) 12 kernel_authoring__copy (AP 40) 13 kernel_authoring__book_histogram (AP 41) Give the SWE notebooks consistent NN__<tech>__<example> names and update their titles and cross-references for the new numbering. Normalize all notebooks to canonical format. Co-Authored-By: Claude <noreply@anthropic.com>

- requirements.txt: add cupy-cuda13x, numba-cuda[cu13], cuda-cccl[test-cu13], nvtx, nsightful, mpi4py, and pytest alongside the existing jax/pyomp/ nanobind/CppJIT stack. All resolve together under Python 3.12. - dockerfile: rename ACH_TUTORIAL to pyhpc, add OpenMPI (mpirun for notebook 03) and the OMPI-run-as-root / CuPy cache env vars. - docker-compose.yml: rename tutorial, dockerfile path, image tag, working-dir. - test_notebooks.py: expand the ordered ladder to all 14 notebooks (00-13), preferring solutions and falling back to exercises, with per-cell timing. - test.bash + test_packages.py: new, mirroring accelerated-python, so the tutorial runs through the standard ACH entrypoint test machinery. - Track the Shallow_water_waves.gif via Git LFS (was committed as a raw blob). Co-Authored-By: Claude <noreply@anthropic.com>

- README: cover the full arc (NumPy/CuPy fundamentals, mpi4py, the SWE programming-models ladder, and CUDA kernel authoring), grouped into Fundamentals / Distributed / Programming models / Kernels with Colab badges pointing at the new pyhpc paths and 00-13 numbering. Add a Syllabi section and refresh the run instructions. - Add notebooks/syllabi/pyhpc__numpy_cupy_jax_mpi4py_kernels__8_hours.ipynb in the accelerated-python syllabus style (Notebooks + Materials + Introduction). Co-Authored-By: Claude <noreply@anthropic.com>

Notebooks 11-13 %%writefile helper .py scripts and emit Nsight reports, a downloaded text corpus, and a rendered plot. Ignore those (keeping the tracked swe_core.py) so a test run leaves the working tree clean. Co-Authored-By: Claude <noreply@anthropic.com>

The SWE solution notebooks (06-09) import swe_core, but swe_core.py and swe_step.cpp live in notebooks/ while the solutions execute from notebooks/solutions/. Under nbclient/JupyterLab only the notebook's own directory is on sys.path, so the import failed. Symlink both helpers into solutions/ (swe_core.py resolves __file__ to notebooks/, so timings.json and swe_step.cpp still resolve to the single shared copy). This was a latent bug inherited from the original tutorial, whose solutions were never executed in CI (it shipped no test.bash). Verified: full suite now passes on an L4 -- test_packages (9) and all 14 notebooks (00-13) green. Co-Authored-By: Claude <noreply@anthropic.com>

Adapt the build workflow for this fork: - Trigger on event/** branches (in addition to main and pull-request/N). - Add a tutorials_filter input (default 'pyhpc') and filter the build matrix accordingly, so the fork builds only the pyhpc image rather than every tutorial. An empty filter restores the build-everything behavior. The runner labels (linux-amd64-cpu4 build, linux-amd64-gpu-l4-latest-1 test) are satisfied by a self-hosted L4 runner registered on the fork. Co-Authored-By: Claude <noreply@anthropic.com>

The build bake uses context ../../.. (repo root), which reads outside the bake directory and requires an fs.read entitlement. The hardcoded --allow=fs.read=/home/runner only covers NVIDIA's self-hosted runner home; on other runners the checkout path differs, so buildx denied it with 'additional privileges requested'. Use --allow=fs.read=* so the build works regardless of the runner's workspace layout. Co-Authored-By: Claude <noreply@anthropic.com>

The test job pulls the image from the generated docker-compose, whose image: anchor is committed as ghcr.io/nvidia/... The build job, however, pushes to ghcr.io/<repo_owner>/... (derived from github.repository_owner). On a fork these differ, so the test failed pre-pull with 'not found'. Add an --image-owner option to generate-tagged-docker-composes.py that rewrites the GHCR owner of the &image anchor, and pass github.repository_owner from the build workflow. Empty owner is a no-op, so upstream and sibling tutorials are unaffected. Co-Authored-By: Claude <noreply@anthropic.com>

…cheme Rename all PyHPC exercise and solution notebooks to the new <NN>[__<EXAMPLE>][__<TECH>][__<CONCEPT>] naming scheme and merge the standalone intro into the reference-solver notebook. - Merge 04__intro into 05__swe_core__reference_solver as 04__swe__intro (framing + NumPy baseline in one notebook); the SWE ladder and the kernel/async notebooks shift down by one (05-13 -> 04-12). - Renumber in-notebook titles, "Next:" pointers, and cross-references. - Update README tables/prose, the 8-hour syllabus, test_notebooks.py ladder, and .gitignore comments to the new names and numbers. Co-Authored-By: Claude <noreply@anthropic.com>

- Rename the 8-hour syllabus to drop "numpy" and add "cppinterop" (pyhpc__cupy_jax_mpi4py_cppinterop_kernels__8_hours.ipynb); update its title, internal generated-compose link, and the README Syllabi entry. - Strip the leading notebook number from the title cells of notebooks 04-09 and their solutions (titles no longer carry "NN." / "NN ·"). - Add brev/docker-compose.deploy.yml: publishes JupyterLab on 0.0.0.0:8888 and opens Lab on the renamed syllabus. Co-Authored-By: Claude <noreply@anthropic.com>

Set each notebook's H1 (and solution H1) to the readable form of its new <NN>[__EXAMPLE][__TECH][__CONCEPT] filename, in the same component order (e.g. "Memory Spaces - Power Iteration" -> "Power Iteration - CuPy - Memory Spaces"; "Kernel Authoring - Copy" -> "Copy - Kernel Authoring"; SWE ladder titled "SWE - <tool>"). Update the matching titles in the README Notebook column and the syllabus Topic column. Co-Authored-By: Claude <noreply@anthropic.com>

Merge aaronj0:pyhpc-cppjit-and-interop (PR #2) into the event branch. The PR branched before our rename/restructure, so on conflicts we keep our structure and conform the incoming content: - New interop notebook 14__python_cpp_interop -> 13__cpp_interop (+ solution), title set to our scheme ("C++ Interop"); added to README and the syllabus. - README + test_notebooks ladder conflicts resolved to our numbering/names; kept Aaron's CppInterOp beta-release note. - Preserved Aaron's non-conflicting work: swe_thrust_solver.cpp extraction, cppjit notebook edits (now in 08__swe__cppjit__thrust), dockerfile python3-config + include path, cffi/memory_profiler deps and package tests, swe_core SWE_THRUST_CPP constant. - docker-compose.deploy.yml: add ?file-browser-path so Lab opens the syllabus but the file browser lands in the notebooks dir, not syllabi/. Co-Authored-By: Aaron Jomy <aaronj0@users.noreply.github.com> Co-Authored-By: Claude <noreply@anthropic.com>

@01

Reorder the syllabus filename tokens to match the tutorial's actual progression (CuPy @01 -> mpi4py @03 -> JAX @05): pyhpc__cupy_jax_mpi4py_cppinterop_kernels__8_hours -> pyhpc__cupy_mpi4py_jax_cppinterop_kernels__8_hours Updates the syllabus H1 title, its internal generated-compose link, the README Syllabi entry, and the local deploy compose default URL to match. Co-Authored-By: Claude <noreply@anthropic.com>

Executing the solution notebooks (e.g. in CI) writes the same generated artifacts the notebooks/ patterns already ignore — %%writefile scripts, Nsight reports (.nsys-rep/.sqlite/.ncu-rep), the downloaded corpus, and generated C/C++ — but into notebooks/solutions/, which the existing patterns didn't cover. Mirror the patterns there, re-including the checked-in tutorial source (swe_step.cpp, swe_thrust_solver.cpp). Co-Authored-By: Claude <noreply@anthropic.com>

copy-pr-bot · 2026-07-21T13:08:08Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

robobryce · 2026-07-26T08:16:58Z

Addressed all of this feedback in follow-up PR #242: #242

Copilot