Skip to content

fix(nemotron/test): persist aiperf results via --artifact-dir (space-safe) - #57

Closed
dmvevents wants to merge 1 commit into
aws-samples:mainfrom
dmvevents:fix/test-aiperf-artifact-dir
Closed

dmvevents wants to merge 1 commit into
aws-samples:mainfrom
dmvevents:fix/test-aiperf-artifact-dir

Conversation

@dmvevents

Copy link
Copy Markdown
Contributor

Problem

test-aiperf.sh runs aiperf profile with no --artifact-dir, so results live only in the ephemeral do-aiperf pod and vanish when the script deletes it at the end. Adding the flag by hand is a footgun: a missing space between the artifact-dir value and the next flag makes aiperf glue them and reject the URL as an unattached positional —

Unused Tokens: ['http://nemotron3-ultra-550b-agg-frontend...:8000']
command terminated with exit code 1

— even though the server is healthy (chat/completions work).

Fix

  • Add --artifact-dir "${ARTIFACT_DIR}" to the built command, space-delimited, before --url.
  • Default ARTIFACT_DIR=${MODEL_PATH}/aiperf in the test .env (.env-overridable).
  • The do-aiperf pod already mounts fsx-pvc at /shared, so results now persist on the shared filesystem across pod restarts (also sidesteps the empty-CSV-on-pod-delete gotcha).

Verification

  • envsubst render shows --artifact-dir "/shared/.../aiperf" --url "http://...:8000" with correct spacing.
  • bash -n clean.
  • Context: verified against a live agg-LWS run where chat + completions served correctly and only the aiperf harness failed on this quoting issue.

… persist to /shared

test-aiperf.sh ran aiperf profile with no --artifact-dir, so results lived only in the
ephemeral do-aiperf pod and were lost on pod delete. Adding --artifact-dir by hand is
error-prone: a missing space between the artifact-dir value and the next flag makes aiperf
glue them and reject the URL as an unattached positional ('Unused Tokens: [http://...:8000]',
exit 1) even though serving is healthy.

Add --artifact-dir ${ARTIFACT_DIR} to the built command (space-delimited, before --url) and
default ARTIFACT_DIR=${MODEL_PATH}/aiperf in the test .env. The do-aiperf pod already mounts
fsx-pvc at /shared, so results now persist on the shared filesystem across pod restarts
(also sidesteps the empty-CSV-on-pod-delete gotcha). .env-overridable.

Verified: envsubst render shows '--artifact-dir "..." --url "..."' with correct spacing;
bash -n clean.

Signed-off-by: Anton Alexander <dmvevents@gmail.com>
@iankouls-aws

Copy link
Copy Markdown
Contributor

suggested change is already in place, closing PR

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants