Pull requests / #1786
#1786 serve: scope watchdog diagnostics to the current engine run
open · draft · @CC-David-CC · 0 comentários · No GitHub
Descrição
Upstream now searches the last 256 KB of the engine log for a watchdog, fixing the usual long-trace case. This follow-up searches the current engine run in reverse, so an older run's watchdog cannot explain a new exit and a trace longer than 256 KB cannot hide the current cause. Preserves server-stop and current-error handling. Adds regressions for prior-run attribution, trace length, reverse-read boundaries, missing trailing newlines, and diagnostic precedence. Documents the reported HIP runtime workaround without claiming it fixes hangs or reboots. No native engine or inference default changes remain in the diff. Validation on current main 61b3fb5d: 8 focused EngineDeath tests passed; the full serving suite ran 696 tests successfully, with 15 skips. git diff --check passed. Related: #1705 and #872. #1931 adds separate opt-in HIP runtime stability controls. A stability-profile run validates its specific runtime configuration; it is not evidence that this diagnostics PR fixes the underlying failure.
No site
Links install, modelos, releases.