Remove benchmark model weights after their last use

kleb commented

2026-10-31 12:48:21 +01:00

Owner

Benchmark runs accumulated model downloads until the kEmbed disks filled. Matrix lanes now finish one model across its assigned engines and datasets, then remove only downloads owned by that lane after their last use. Existing caches, supplied models, datasets and evidence stay intact; cleanup failures stop dispatch. Standalone runs retain caller-owned caches.

Proven with 331 passing Python tests, including cleanup lifecycle and ownership tests, plus Ruff checks and formatting. No inference or remote rollout was performed. Disk must still fit the largest single model.

Benchmark runs accumulated model downloads until the kEmbed disks filled. Matrix lanes now finish one model across its assigned engines and datasets, then remove only downloads owned by that lane after their last use. Existing caches, supplied models, datasets and evidence stay intact; cleanup failures stop dispatch. Standalone runs retain caller-owned caches. Proven with 331 passing Python tests, including cleanup lifecycle and ownership tests, plus Ruff checks and formatting. No inference or remote rollout was performed. Disk must still fit the largest single model.

kleb added 1 commit

2026-10-31 12:48:21 +01:00

Remove benchmark model weights after their last use 091de537f4

kAgent commented

2026-10-31 12:48:22 +01:00

First-time contributor

kReview review

Verdict: no findings

No verified defects were found in the reviewed changes.

kReview checked only the PR diff, metadata, and supplied repository context at 80e55f3a31. It reports only issues supported by this material.

Estimated cost: ~$0.14 total (401.3k in / 1.9k out); this run ~$0.08 (241.4k in / 817 out) / gpt-6.1-terra.

Run time: 1:01 total; this run 0:20.

## kReview review **Verdict:** no findings No verified defects were found in the reviewed changes. _kReview checked only the PR diff, metadata, and supplied repository context at `80e55f3a31`. It reports only issues supported by this material._ _Estimated cost: ~$0.14 total (401.3k in / 1.9k out); this run ~$0.08 (241.4k in / 817 out) / gpt-6.1-terra._ _Run time: 1:01 total; this run 0:20._

kleb added 1 commit

2026-10-31 12:49:26 +01:00

Reuse existing OpenVINO downloads during matrix runs 80e55f3a31

kleb merged commit 0547b45237 into main

2026-10-31 12:50:07 +01:00

kleb referenced this pull request from a commit

2026-10-31 12:50:08 +01:00

Merge pull request 'Remove benchmark model weights after their last use' (#5) from benchmark/model-weight-cleanup into main

Rows
Columns

Remove benchmark model weights after their last use #5

kReview review