Skip to content

armgit / Handy

Public Imported from GitHub
0

A free, open source, and extensible speech-to-text application that works completely offline.

fix: auto-unload model after idle timeout to reduce memory (#1051)

Viren Mohindrad1da935479b5
diff --git a/src-tauri/src/managers/transcription.rs b/src-tauri/src/managers/transcription.rs
index 7f6e7ac9..73e4e5b5 100644
--- a/src-tauri/src/managers/transcription.rs
+++ b/src-tauri/src/managers/transcription.rs
@@ -97,6 +97,7 @@ impl TranscriptionManager {
let manager_cloned = manager.clone();
let shutdown_signal = manager.shutdown_signal.clone();
let handle = thread::spawn(move || {
+ debug!("Idle watcher thread started");
while !shutdown_signal.load(Ordering::Relaxed) {
thread::sleep(Duration::from_secs(10)); // Check every 10 seconds
@@ -114,28 +115,29 @@ impl TranscriptionManager {
.duration_since(SystemTime::UNIX_EPOCH)
.unwrap()
.as_millis() as u64;
+ let idle_ms = now_ms.saturating_sub(last);
+ let limit_ms = limit_seconds * 1000;
- if now_ms.saturating_sub(last) > limit_seconds * 1000 {
+ if idle_ms > limit_ms {
// idle -> unload
if manager_cloned.is_model_loaded() {
let unload_start = std::time::Instant::now();
- debug!("Starting to unload model due to inactivity");
-
- if let Ok(()) = manager_cloned.unload_model() {
- let _ = app_handle_cloned.emit(
- "model-state-changed",
- ModelStateEvent {
- event_type: "unloaded".to_string(),
- model_id: None,
- model_name: None,
- error: None,
- },
- );
- let unload_duration = unload_start.elapsed();
- debug!(
- "Model unloaded due to inactivity (took {}ms)",
- unload_duration.as_millis()
- );
+ info!(
+ "Model idle for {}s (limit: {}s), unloading",
+ idle_ms / 1000,
+ limit_seconds
+ );
+ match manager_cloned.unload_model() {
+ Ok(()) => {
+ let unload_duration = unload_start.elapsed();
+ info!(
+ "Model unloaded due to inactivity (took {}ms)",
+ unload_duration.as_millis()
+ );
+ }
+ Err(e) => {
+ error!("Failed to unload idle model: {}", e);
+ }
}
}
}
@@ -399,6 +401,15 @@ impl TranscriptionManager {
*current_model = Some(model_id.to_string());
}
+ // Reset idle timer so the watcher doesn't immediately unload a just-loaded model
+ self.last_activity.store(
+ SystemTime::now()
+ .duration_since(SystemTime::UNIX_EPOCH)
+ .unwrap()
+ .as_millis() as u64,
+ Ordering::Relaxed,
+ );
+
// Emit loading completed event
let _ = self.app_handle.emit(
"model-state-changed",
@@ -771,7 +782,15 @@ pub fn get_available_accelerators() -> AvailableAccelerators {
impl Drop for TranscriptionManager {
fn drop(&mut self) {
- debug!("Shutting down TranscriptionManager");
+ // Skip shutdown unless this is the very last clone. TranscriptionManager
+ // is cloned by initiate_model_load() and the watcher thread — those
+ // clones dropping must not kill the watcher. The watcher thread holds
+ // its own clone, so engine's strong_count is always >= 2 while the
+ // watcher is alive. When it reaches 1, only this instance remains
+ // and we can safely shut down.
+ if Arc::strong_count(&self.engine) > 1 {
+ return;
+ }
// Signal the watcher thread to shutdown
self.shutdown_signal.store(true, Ordering::Relaxed);
diff --git a/src-tauri/src/settings.rs b/src-tauri/src/settings.rs
index 95321689..a713dae4 100644
--- a/src-tauri/src/settings.rs
+++ b/src-tauri/src/settings.rs
@@ -180,7 +180,7 @@ impl Default for KeyboardImplementation {
impl Default for ModelUnloadTimeout {
fn default() -> Self {
- ModelUnloadTimeout::Never
+ ModelUnloadTimeout::Min5
}
}
@@ -735,7 +735,7 @@ pub fn get_default_settings() -> AppSettings {
debug_mode: false,
log_level: default_log_level(),
custom_words: Vec::new(),
- model_unload_timeout: ModelUnloadTimeout::Never,
+ model_unload_timeout: ModelUnloadTimeout::default(),
word_correction_threshold: default_word_correction_threshold(),
history_limit: default_history_limit(),
recording_retention_period: default_recording_retention_period(),