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.
| 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).
ticket: to the customer’s ticket id.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.)
Dry-run first when validating the recipe on your side:
aorta sweep run --recipe recipes/probe/probe-template-bash.yaml --dry-run -- echo hi
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).
<ticket>-<timestamp>.tar.gz with AMD support.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.