flush_cache_on_commit is fixed at serve() and applied to every reload — all-or-nothing. A publisher can't request a prompt-cache flush on one specific version's reload.
Why it's non-trivial: the reconciler converges to the latest pointer, so a single /update_weights_from_disk can coalesce several published versions. A per-version flush flag then has no unambiguous version to attach to (which coalesced version's flag wins?).
Options if we later want per-version flush:
- accumulate — flush if any coalesced version asked (safe; an extra flush is harmless, a missed one isn't).
- carry the directive on
wake and consume it at the next reload (transient, nothing stored).
Low priority: the static policy covers the common case (one flush policy per run). Filing to track.
flush_cache_on_commitis fixed atserve()and applied to every reload — all-or-nothing. A publisher can't request a prompt-cache flush on one specific version's reload.Why it's non-trivial: the reconciler converges to the latest pointer, so a single
/update_weights_from_diskcan coalesce several published versions. A per-version flush flag then has no unambiguous version to attach to (which coalesced version's flag wins?).Options if we later want per-version flush:
wakeand consume it at the next reload (transient, nothing stored).Low priority: the static policy covers the common case (one flush policy per run). Filing to track.