Complete CPU and Iris Xe decision benchmarks #8
Loading…
Reference in a new issue
No description provided.
Delete branch "feat/decision-cpu-igpu-results"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
The decision catalogue was missing CPU and Iris Xe measurements for the expanded self-hosted model set. This publishes 161 new result/evidence pairs covering 48,000 answered requests: 133 CPU results and 28 Iris Xe results. Existing measurements remain byte-for-byte unchanged.
CPU bindings now verify CPU residency, CPU dependency profiles are locked, and H2O selects its measured Q5_K_M artifact on Iris Xe. Matrix runs reserve disk space and can clean fresh owned caches after verified child shutdown, including failed downloads. Open-Jev SciFact uses two disjoint 150-request shards with separate dataset digests, cohorts and per-host metrics rather than a fabricated full-run result.
Proven with 435 Python tests, 47 Bun tests, type checks, schema/evidence/privacy/device audits and desktop/mobile browser checks. Unsupported hardware/runtime variants remain excluded; no private corpus, raw request text or credentials are included.
kReview review
Verdict: no findings
No findings to address across the reviewed diff.
This review did not check generated or vendored files:
packages/runners/decision/requirements-author-cpu.lock,packages/runners/decision/requirements-label-gguf-cpu.lock,packages/runners/decision/requirements-laya-cpu.lock,packages/runners/decision/requirements-mapika-gguf-cpu.lock,packages/runners/decision/requirements-zefan-cpu.lock.kReview checked only the PR diff, metadata, and supplied repository context at
1d0b1ea750. It reports only issues supported by this material.Estimated cost: ~$1.64 (1.7M in / 1.3k out / gpt-6.1-terra).
Run time: 2:28 total; this run 2:28.