Reproducible Setup & Compute Fallbacks

Pins, preflight checks, and a route through the course when a GPU is unavailable

What the version record means

Model software changes quickly. Every installation lesson now records an immutable package version or Git commit in data/tools.yml, and every capstone target uses the structure, chain, and hotspots in data/targets.yml.

The pins were checked against the upstream projects on August 30, 2026. A pin makes the instructions repeatable; it does not prove that every GPU, driver, cluster module, or model download has been tested. In the tool manifest, a blank last_full_smoke_test deliberately means that a maintainer has not recorded a complete end-to-end rerun. Do not interpret “pinned” as “guaranteed on this cluster.”

Run a one-design smoke test first

Before requesting a long allocation, run the smallest test in the lesson and save its command, output log, environment export, GPU name, driver, and date. That record is the evidence that your stack works.

1. Run the course preflight

From the course repository:

python scripts/preflight.py --output preflight-report.txt

Select a tool to add its GPU and container expectations:

python scripts/preflight.py --tool rfdiffusion-original --check-network

PASS means the dependency was detected. WARN means resolve it or take the documented fallback. INFO records useful context but does not block the course. The script is read-only except when you explicitly request an output report.

2. Keep environments isolated and record the solve

Use one environment per model. After a successful smoke test, save both a human-readable history and a fully resolved record:

mkdir -p run-record
conda env export --from-history > run-record/environment-history.yml
conda list --explicit > run-record/environment-explicit.txt
python -m pip freeze > run-record/pip-freeze.txt
nvidia-smi > run-record/nvidia-smi.txt
git rev-parse HEAD > run-record/tool-commit.txt

Also save the exact command or configuration, random seed, input identifiers, and model-weight filenames. If the upstream project publishes checksums, record those too.

3. Choose the compute or analysis route

Compute access is not a learning objective. When a queue, account, model license, download, or incompatible GPU blocks execution, continue with the same scientific checkpoint using existing outputs.

Lesson type	Run-it route	No-compute route	Equivalent evidence
Structure prediction	Generate the prediction locally or through an approved service	Use the prediction and confidence examples supplied in Tuesday’s lessons	Annotated pLDDT/PAE interpretation and a method-selection memo
Sequence design	Run LigandMPNN on a selected backbone	Start from the candidate sequences in a published/tool example and apply the lesson’s ranking questions	Candidate table and selection rationale
Backbone generation	Run RFdiffusion and retain successes and failures	Use the official, revision-pinned RFdiffusion output set in Wednesday’s lesson	Backbone gallery, contact/topology observations, and keep/reject rationale
Complex validation	Run Chai-1, Boltz, or an approved prediction service	Analyze the supplied confidence figures or an archived prediction from your project group	Interface-confidence interpretation with limitations
Benchmarking	Run the CPU/GPU notebook on both devices	Use the provided benchmark results in Thursday’s activity	Plot interpretation and a prediction of the crossover point
Capstone	Produce a new end-to-end campaign	Build a staged evidence portfolio from canonical targets and known-good outputs, clearly labeled as re-analysis	The same rubric-scored portfolio; no claim of a novel generated binder

The no-compute route changes the provenance of the result, not the standard of reasoning. State which files you received, where they came from, and which steps you did not run.

4. Verify immutable pins

For a Git-based tool, the lesson performs a detached checkout. Confirm it before running:

git rev-parse HEAD
git status --short

The first line must match the full commit in the lesson. A detached HEAD is expected for this use. For Python packages, confirm the resolved version:

python -m pip show PACKAGE_NAME

If you intentionally use a newer version, create a new environment and document the deviation. Do not silently replace the course pin with the default branch or pip install -U.

5. Validate course records

Course maintainers can detect stale target data, missing lesson metadata, typographic quotes in commands, stale generated browser data, and a stale Python-refresher archive with one command:

python scripts/validate_course_data.py

After intentionally editing a manifest, regenerate its checked-in derivatives and then validate again:

python scripts/validate_course_data.py --generate
python scripts/validate_course_data.py

Report reproducibility problems with the tool ID, course pin, preflight report, failing command, and the first complete error traceback. Never include access tokens, private keys, or credentials in a report.

← Back to Monday