| { | |
| "id": "qk-fc2-cache", | |
| "change_class": "exact", | |
| "hypothesis": "A 512 MiB free-allocator cache can remove the small physical-footprint regression of the fused path while retaining its speed. It cannot reduce live model memory.", | |
| "candidate_config": { | |
| "kernel": "metal-exact", | |
| "tune_matmul": true, | |
| "allocator_cache_gb": 0.5 | |
| }, | |
| "control_allocator_cache_gb": 1, | |
| "reject_if": "Full declared quality/timing/memory contract fails; do not change gates or pool samples from other campaigns.", | |
| "screen_before_full": "Real compiled block at both cache limits; no GPU concurrency.", | |
| "bounded_search": "One additional full-image finalist at most, within original 120-minute and 5GiB global budget. No inference-code edits until current benchmark completes.", | |
| "result": { | |
| "decision": "rejected-at-block-screen", | |
| "evidence": "reports/optimization-r3/cache-screen.json", | |
| "slowdown_pct": 1.6453551990850412, | |
| "cache_1_peak_physical_bytes": 2387051720, | |
| "cache_05_peak_physical_bytes": 2581791944, | |
| "accuracy": "all BF16 block outputs identical", | |
| "scope": "Isolated real first block; does not establish whole-image behavior. No allocator-limit change shipped." | |
| } | |
| } | |