Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

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

Daniel Hiltgend950ff12c09c
diff --git a/server/sched.go b/server/sched.go
index 3fc54e55a..612e47020 100644
--- a/server/sched.go
+++ b/server/sched.go
@@ -387,6 +387,17 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
s.loadedMu.Unlock()
runner.refMu.Unlock()
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 {
slog.Debug("starting background wait for VRAM recovery", "runner", runner)
finished := runner.waitForVRAMRecovery()