aorta

Probe handout templates

Generic mode: probe recipes for customer-facing handouts. Each template ships with a redaction: block so aorta bundle can produce a shareable tarball without manual scrubbing.

When to use which template

Template Customer launch pattern Example trailing argv
probe-template-torchrun.yaml PyTorch distributed / torchrun torchrun --nproc_per_node=8 train.py --config cfg.yaml
probe-template-buck2.yaml Buck2 monorepo targets buck2 run //models:train -- --steps 1000
probe-template-bash.yaml Shell script / Makefile wrapper bash launch.sh --profile production

All three templates use a 2×2 mitigation × diagnostic matrix (none / tf32_off × none / xnack), three trials, and a 30-minute timeout_per_trial (1800 seconds).

Workflow

  1. Copy or symlink the template; set ticket: to the customer’s ticket id.
  2. Run the sweep (customer prepends aorta sweep run, keeps their argv after --):

    aorta sweep run --recipe recipes/probe/probe-template-bash.yaml \
        --output ./probe-out --ticket TICKET-1234 -- \
        bash launch.sh
    

    (aorta probe still works as a deprecated alias.)

  3. Dry-run first when validating the recipe on your side:

    aorta sweep run --recipe recipes/probe/probe-template-bash.yaml --dry-run -- echo hi
    
  4. After the matrix completes, bundle the ticket leaf:

    aorta bundle ./probe-out/TICKET-1234/ --review
    

    Redaction is loaded automatically from ./probe-out/TICKET-1234/recipe.resolved.yaml when that file contains a redaction: block (the runner writes it during aorta probe).

  5. Share the resulting <ticket>-<timestamp>.tar.gz with AMD support.

Customising axes

Replace mitigation_axis / diagnostic_axis entries with any names registered in aorta.registry (run aorta mitigations list on your host). The template defaults use only mitigations present in the public registry snapshot.

See redaction.md for scrubber semantics and usage.md for env-passthrough and artifact layout.