Skip to content

feat(perf): add flag-gated composite metric for serving KEEP decisions - #1316

Open
hmahmad26 wants to merge 2 commits into
AMD-AGI:mainfrom
hmahmad26:feat/composite-perf-metric
Open

feat(perf): add flag-gated composite metric for serving KEEP decisions#1316
hmahmad26 wants to merge 2 commits into
AMD-AGI:mainfrom
hmahmad26:feat/composite-perf-metric

Conversation

@hmahmad26

Copy link
Copy Markdown

Summary

  • Adds an opt-in composite KEEP metric for serving runs: S = 0.55Δ_in + 0.30Δ_intv + 0.15Δ_out, with Δ_x = max(0, (x_cand - x_B) / x_B).
  • Off by default. Enable with HYPERLOOM_PERF_METRIC=composite_v1. Scriptable frameworks stay on output-throughput grading.
  • Noise floors default to 0 (raw Δ). Override with HYPERLOOM_PERF_NOISE_PCT. Intvty p90 is a hard gate: a regression fails KEEP.
  • Mapping / grid / writeback now carry the input-tput, intvty p90, and p90 TPOT triple so explore can grade on it.

Test plan

  • test_perf_metric.py (score, noise=0, intvty gate)
  • test_map_aiperf.py (input tput / p90 tpot / intvty mapping)
  • test_grid_runner.py and related grid-runner unit files
  • test_benchmark_result.py, test_explore_executor.py
  • CI on this PR

Made with Cursor

Co-authored-by: Cursor <cursoragent@cursor.com>
@hmahmad26
hmahmad26 requested a review from a team as a code owner August 27, 2026 23:58
When the composite flag is on with a full triple, session totals, conc-sweep
speedup, and the 10% roofline refresh step use S instead of output tok/s.

Co-authored-by: Cursor <cursoragent@cursor.com>
hmahmad26 pushed a commit to hmahmad26/Hyperloom that referenced this pull request Sep 1, 2026
Local integration of PR AMD-AGI#1315 plus-sign JSON with PR AMD-AGI#1316 composite S.
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