Normalize measurement protocol contracts

#4
by nv-simonec - opened
NVIDIA org

Summary

  • replace campaign-specific measurement protocol IDs with forward-pass-measurement-v1 and forward-pass-record-v1
  • describe wave and bucket summaries with evidence_format_id instead of treating them as measurement protocols
  • classify the Qwen wave summaries as supporting evidence and keep FPM artifacts on the existing aic_fpm_forward_perf v6 schema
  • document that availability status is consumer-derived and that unknown truth protocols are invalid
  • regenerate manifests, catalogs, hashes, and audit reports

Validation

  • dataset validator: 17 current configurations, 7 historical snapshots, 24 measurement bundles
  • JSON Schema: 24 configuration manifests, 24 measurement manifests, and catalog index
  • Ruff format/check and git diff --check
  • parsed by the updated gym consumer: 12 ready, 4 no measurements, 1 supporting-evidence-only

Follow-up to merged discussion #3.

NVIDIA org

Consumer implementation: https://gitlab-master.nvidia.com/dl/ai-dynamo/aisim-fpm-gym/-/merge_requests/19 (stacked on the existing HF dashboard MR !16).

nv-simonec changed pull request status to merged

Sign up or log in to comment