== fpoff-1 ==
12:46:21 INFO    compile-warmup ledger: 12.5s total, 12 dynamo frame compile(s) — decode-compile 3.2s/4f · prefill-multirow 4.9s/5f · prefill-single-row 2.6s/2f · tkv-cute-prewarm 0.3s/0f · decode-recurrent 1.4s/1f · verify-hidden 0.1s/0f (step seconds/frames; a frame is a Dynamo trace, which the Inductor caches do NOT skip — they cover the backend only)
12:46:44 INFO    Boot: KV cache ready at 4.08 GiB (17% of 23.52 GiB) — 926 pages × 256 tok/page = 237056 servable tokens, free 0.14 GiB (1% of 23.52 GiB)
  NOTE: the address_space.released_pool_va row (896M; cuMem tracked-but-unmapped 896M) is VA whose PHYSICAL IS ALREADY BACK with the driver — a pool this boot destroyed (release_empty_pool -> cuMemUnmap/cuMemRelease) whose torch segment records outlive the release. It reads reserved AND fully-free and is neither: those bytes were already grown into KV, and empty_cache returning 0 for them is the correct result, not a missed reclaim.
12:46:45 INFO    Boot: ready in 56.7s — serving now. GPU memory: KV cache 4.08 GiB (17% of 23.52 GiB), free 0.13 GiB (1% of 23.52 GiB) (memory locked — won't grow while serving).
