CSCS Summer School 2026 event branch

[brycelelbach]

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.

Related: #213, #214

[copy-pr-bot]

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.

[Copilot]

Pull request overview

Copilot reviewed 73 out of 77 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (5)

tutorials/pyhpc/test/test_notebooks.py:56

• If NOTEBOOKS_DIR is missing/empty in an image build, ladder becomes empty and pytest will report a passing test suite (0 parametrized cases). That can mask broken packaging where no notebooks were included.
  tutorials/pyhpc/notebooks/swe_thrust_solver.cpp:36
• gpu_swe_solve dereferences managed pointers immediately after cudaMallocManaged, but never checks the return codes. If an allocation fails, this can segfault or corrupt outputs; likewise cudaDeviceSynchronize/cudaFree errors are silently ignored (there’s even a TODO comment).
  brev/entrypoint-base-user.bash:26
• When TURN_USERNAME/TURN_PASSWORD are provided via environment (and the shared credentials file doesn’t exist), the script currently does not write /accelerated-computing-hub/.turn-credentials. Sibling services (nsys/ncu/jupyter) rely on that file unless they also receive TURN_* via their own environment, so they can start without TURN credentials.

if [ -z "${TURN_USERNAME:-}" ] && [ ! -f "${TURN_CREDENTIALS_FILE}" ]; then
  TURN_USERNAME="turn_$(openssl rand -base64 24 | tr -dc 'a-zA-Z0-9' | head -c 16)"
  TURN_PASSWORD="$(openssl rand -base64 48 | tr -dc 'a-zA-Z0-9' | head -c 32)"

  echo "TURN_USERNAME=${TURN_USERNAME}" > "${TURN_CREDENTIALS_FILE}"
  echo "TURN_PASSWORD=${TURN_PASSWORD}" >> "${TURN_CREDENTIALS_FILE}"
fi

brev/generate-cscs-edf.bash:91

• This helper is documented as usable on CSCS systems where Docker isn’t available, but it currently shells out to docker compose ... config to extract x-config fields. On Daint/login nodes without Docker, EDF generation will fail even though this script is part of the CSCS workflow.

if [ -z "${IMAGE}" ] || [ -z "${WORKDIR}" ]; then
    COMPOSE_CONFIG=$(docker compose -f "${COMPOSE_FILE}" config --format json)
fi

docs/cscs.md:44

• This step implies brev/generate-cscs-edf.bash can be run on CSCS, but the generator currently requires docker compose to read x-config from docker-compose.yml. On CSCS where Docker isn’t available, the guidance should clarify where to run the generator (or the script needs a Docker-free implementation).

For a reproducible run, generate a no-mount EDF pinned to the CI commit tag:

[Copilot]

Pull request overview

Copilot reviewed 74 out of 78 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (3)

brev/generate-cscs-edf.bash:91

• brev/generate-cscs-edf.bash shells out to docker compose ... | jq to read x-config.image and x-config.working-dir. The CSCS docs explicitly instruct running this script on Daint (which typically has no Docker daemon), so this will fail even though the EDF generator is meant for non-Compose environments. Add a fallback that parses these two values directly from the compose YAML when Docker/jq are unavailable (or require --image/--workdir in that case).

if [ -z "${IMAGE}" ] || [ -z "${WORKDIR}" ]; then
    COMPOSE_CONFIG=$(docker compose -f "${COMPOSE_FILE}" config --format json)
fi

tutorials/pyhpc/notebooks/swe_thrust_solver.cpp:36

• cudaMallocManaged() return values are not checked. If an allocation fails, the subsequent host writes (h[i] = ...) will dereference null/invalid pointers and likely crash, and any prior allocations may leak. Capture and validate the cudaError_t from each allocation and free already-allocated buffers on failure.
  tutorials/pyhpc/notebooks/swe_thrust_solver.cpp:67
• cudaDeviceSynchronize() return value is ignored. If any kernel launch in the loop fails, the code will still copy back potentially-uninitialized data and then free without surfacing the failure. Check the synchronize result and bail out (after freeing managed allocations) on error.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

[brycelelbach]

Feedback on the Nsightful cell magic notebook refactorings. Please address all of this in a separate PR targeting this branch:

Notebook 03

"First, we need to ensure the Nsight Systems profiler (nsys), Nsightful, and NVTX are installed and available." => Change to "First, we need to ensure NVIDIA's developer tools are installed and available and do all of our imports."

In the estimate_device_baseline cell, this is wrong:

estimate_device_baseline(
  A_device,
  cfg=PowerIterationConfig(max_steps=1, check_frequency=1, progress=False),
)

The testing should be in a separate code cell, with a "Now let's make sure it works:" text cell right before it. The code cell should look like this:

lam_est_baseline = estimate_device_baseline(A_device)

assert isinstance(lam_est_baseline, (np.ndarray, np.generic)), "Must return a NumPy array or NumPy scalar"
np.testing.assert_allclose(lam_est_baseline, 1, atol=1e-4)

print()
print("Dominant Eigenvalue:", lam_est_baseline)

The testing cell for estimate_device_async current has:

lam_est_async = estimate_device_async(A_device)
np.testing.assert_allclose(lam_est_async, 1, atol=1e-4)

It should have:

lam_est_async = estimate_device_async(A_device)

assert isinstance(lam_est_async, (np.ndarray, np.generic)), "Must return a NumPy array or NumPy scalar"
np.testing.assert_allclose(lam_est_async, 1, atol=1e-4)

print()
print("Dominant Eigenvalue:", lam_est_baseline)

"The timeline is displayed below the profiled cell. Explore what's going on in the program.

EXTRA CREDIT: Download the Nsight Systems GUI and open the report in it to see even more information."
^ The first sentence is inaccurate and was not in the original notebook. What is displayed below the profiled cell is a button to open the results in Perfetto. The second sentence is unnecessary because nsys is available in the container (and it shouldn't be extra credit). Rewrite this cell entirely; it should tell users to explore the profile, either in Perfetto (by clicking the button), by clicking the "+" in list of JupyterLab tabs and then opening Nsight Systems, or by installing the Nsight Systems GUI and downloading the report onto their local system.

"# The native report is saved as power_iteration__baseline.nsys-rep." <= Delete this entire cell.

"This notebook uses the Python 3 (Nsight Systems) kernel so we can profile the baseline in place." <= This is an implementation detail, the user doesn't need to know this. What they need to know is what the %%nsys cell magic does.

"# The native report is saved as power_iteration__async.nsys-rep." <= Delete this entire cell.

Notebook 04

"First, we need to make sure the Nsight Compute profiler, Nsightful, Numba CUDA, and CuPy are available in our notebook:" <= Change to "First, we need to ensure NVIDIA's developer tools are installed and available and do all of our imports."

"The profiling sections require the Python 3 (Nsight Compute) custom kernel provided by the ACH environment. Google Colab cannot select this kernel." <= Delete this text entirely

launch_blocked and launch_optimized are unnecessary, remove them and inline all uses.

There are multiple unnecessary newlines at the end of code cells, clean that up. In some places within a code block, there are blank lines in a row; collapse those to just one blank line.

"The profiling report is displayed directly below the cell above. The first tab summarizes Nsight recommendations and advisories; subsequent tabs contain detailed sections." <= Add a sentence explaining that the Nsight Compute GUI has richer information, charts, and diagrams. Explain that they can access it by clicking the "+" in list of JupyterLab tabs and then opening Nsight Compute, or by installing the Nsight Compute GUI and downloading the report onto their local system.

"EXTRA CREDIT: Download the Nsight Compute GUI and open the report in it to see even more information." <= Delete this sentence.

The further optimization section (7) should have a code cell with some comments providing guidance for the extra credit.

Make sure to reset the copy output before launching the copy kernel when doing correctness checks or profiling, but do NOT do this when benchmarking.

The NCU profiling needs to regex filter for the kernel we want. This was done in the prior version but you dropped support for it.

Notebook 05

There are multiple unnecessary newlines at the end of code cells, clean that up.

Use the same # of items per thread for all histogram variants. It was a mistake that it wasn't this way before. That number should be 8.

This:

bins = 256
values = cp.fromfile("books__15m.txt", dtype=cp.uint8)
histogram = cp.zeros(bins, dtype=cp.int32)
threads_per_block = 512
items_per_thread = 8
items_per_block = threads_per_block * items_per_thread
blocks = len(values) // items_per_block
assert values.size % items_per_block == 0

Should be:

bins = 256

values = cp.fromfile("books__15m.txt", dtype=cp.uint8)
histogram = cp.zeros(bins, dtype=cp.int32)

threads_per_block = 512
items_per_thread = 8
items_per_block = threads_per_block * items_per_thread
blocks = len(values) // items_per_block
assert values.size % items_per_block == 0

Remove launch_global and inline its definition. Inline the histogram[:] = 0, but NOT when benchmarking (fine when profiling).

Remove the cp.cuda.runtime.deviceSynchronize() calls.

Combine the cells that launches the histogram with the cell that prints out the most frequently occurring characters. This occurs for both local and global histograms.

"assert cp.sum(histogram) < len(values), "The intentionally racy kernel should lose updates."" => Replace this with "assert cp.sum(histogram) < len(values)".

The NCU profiling needs to regex filter for the kernel we want. This was done in the prior version but you dropped support for it.

This notebook erroneously refers to "cooperative groups". This is a misunderstanding from a prior contributor. I believe it has been fixed on main. Fix it here too.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

[robobryce]

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

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.