root
/
ollama
mirror of https://github.com/ollama/ollama.git
1
0
Fork 0
Code Issues Actions Packages Projects Releases Wiki Activity

sched: fix runner leak during reloading unload (#10819) 

When the same model is being reloaded rapidly with client connections
being canceled before the model finishes loading, the queued unload
event could cause a leak of runners by deleting a different runner from
the loaded list.
This commit is contained in:
Daniel Hiltgen 2025-05-22 14:31:36 -07:00 committed by GitHub
parent adff143bcd
commit d950ff12c0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
server/sched.go
View File

@ -387,6 +387,17 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
s.loadedMu.Unlock() s.loadedMu.Unlock()
runner.refMu.Unlock() runner.refMu.Unlock()
slog.Debug("duplicate expired event, ignoring", "runner", runner) slog.Debug("duplicate expired event, ignoring", "runner", runner)
} else if runner.pid != runnerToUnload.pid {
// If the pids do not match, we likely had multiple load
// failures for the same model in quick succession due to
// request context canceled and are draining the queue of
// events. Ensure the orphaned runner is properly shut down, but
// do not delete the mismatched loaded runner, or wait for VRAM
// convergence.
slog.Debug("orphaned runner shutting down", "orphan", runner, "loaded", runnerToUnload)
runner.unload()
s.loadedMu.Unlock()
runner.refMu.Unlock()
} else { } else {
slog.Debug("starting background wait for VRAM recovery", "runner", runner) slog.Debug("starting background wait for VRAM recovery", "runner", runner)
finished := runner.waitForVRAMRecovery() finished := runner.waitForVRAMRecovery()