Skip to content

Releases: MaybeIcanShow/Relax

Task 22 evidence v9

Choose a tag to compare

@MaybeIcanShow MaybeIcanShow released this 11 Aug 16:43

Full Task 22 hybrid async text-generation performance evidence.

  • Tested a frozen dirty workspace snapshot based on commit 4603e03ffde5769813346b313ce7284c8f726640; the complete tested snapshot and provenance metadata are attached.
  • Optimized vs zero-KL response throughput: +12.90% (paired trials: +13.20%, +12.60%, +12.91%).
  • Formal benchmark completion: 9/9 jobs successful, each with 20 steps and 640 samples.
  • Optimized vs baseline response throughput: +13.51%; end-to-end time vs zero-KL: -8.26%.
  • One initial excluded attempt failed during Ray runtime-environment packaging. It was an infrastructure failure and is not included in the formal measurements.
  • raw-evidence.tar.gz SHA-256: e76f53da3ff064da4045b9b539c9aebddeff5a959884ebb884092c1b39106d81.

See report.md, source-provenance.txt, and raw-evidence-index.csv for methodology and reproducibility details.

Task 22 evidence v8

Choose a tag to compare

@MaybeIcanShow MaybeIcanShow released this 06 Aug 09:24

Task 22 Hybrid-async paired benchmark rerun on commit d8893619bf2ce6a4393c77a49a0af297df51b972.

  • 3 variants x 3 paired trials, all 9 jobs succeeded with 20 steps and 640 samples each.
  • optimized throughput: +13.54% versus zero_kl, +12.74% versus baseline.
  • zero_kl versus baseline: -0.71%, consistent with an equivalent-configuration noise control.
  • No runtime errors, OOMs, unexpected NaN/Inf values, or missing samples.
  • Physical GPUs 2 and 3: NVIDIA RTX PRO 6000 Blackwell Server Edition.
  • Redacted raw evidence SHA-256: f44e0bfd65d59a9fada980000d5596f4b6dbab957e2ff6fd9b2e6e7235cf3b88.

Includes the report, CSV/SVG summaries, redacted raw evidence, checksum index, and the benchmark source archive carried forward from v7.

Task 22 evidence v7

Choose a tag to compare

@MaybeIcanShow MaybeIcanShow released this 05 Aug 08:48

Re-run of Task 22 Hybrid-async paired experiments on commit 6a70677. Includes report, CSV/SVG summaries, and redacted raw evidence. The benchmark source archive is carried forward from v6.

Task 22 evidence v6

Choose a tag to compare

@MaybeIcanShow MaybeIcanShow released this 04 Aug 16:46

Task 22 review-fix rerun. Training evidence was produced from clean commit 946a0dc on physical GPUs 0 and 1; 9/9 final component runs completed 20 steps. Commit 6d2d71b only isolates benchmark GPU telemetry and renders GPU IDs from manifests. raw-evidence.tar.gz SHA-256: 740f39d14c993d9240d646ddb6600ceff9c17c12810689ba765134f6c8efe2cf.

Task 22 hybrid-async evidence v4

Choose a tag to compare

@MaybeIcanShow MaybeIcanShow released this 03 Aug 20:04

Rerun evidence for redai-infra/Relax PR redai-infra#211 at commit a0b1676. Sequential 3-variant x 3-trial experiment under unchanged Task 22 conditions. The report, CSV tables, SVG curves, raw evidence, checksum, and file index are published as 9 individual release assets. raw-evidence.tar.gz SHA-256: da497ee37d05f97e43c5c6277b21dd7793ee0c4236587d4fc60e054f78674ff6.