server: fix pos_min gate + improve ring buffer log messages
- Fix inverted pos_min gate in create_checkpoint() that caused checkpoint search to skip on warm slots. The condition `it->pos_min < pos_min_thold` was backwards -- should be `>=` to find checkpoints that cover the LCP position. This was silently forcing full re-prefill on warm slots. - Improve ring buffer log message readability in create_checkpoint() and deferred_create_final_checkpoint(). Updated 4 log messages to be more concise and follow the same format as other log messages: - "kv ring buffer, evicted oldest (size=X MiB)" - "kv ring buffer, pushed mid-prompt to slot N/M (pos_min=X size=Y MiB)" - "kv ring buffer, recycled slot N/M (dropped pos_min=X size=Y MiB)" - "kv ring buffer, recycled/pushed final checkpoint to slot N/M (pos_min=X size=Y MiB)" Removed redundant pos_max and n_tokens fields, shortened the message prefix from "[ring buffer] appended new slot at back (buffer=9/16)" style to "kv ring buffer, ..." style.
F
fewtarius committed
6af265fa1f125ae54658be6a2bc2bacaf2fe769f
Parent: da9a6da