fix(foundry): keep local ASR warm and stream transcripts (#1052)

* fix(foundry): keep local ASR warm and stream transcripts * fix(foundry): cancel stale release schedules
2026-09-13 12:28:41 +08:00
1 parent c4e8e88782
commit c8bd7aa295
@@ -63,6 +63,7 @@
    "foundry_local_asr_release",
    "foundry_local_asr_reveal_model_dir",
    "foundry_local_asr_set_language_hint",
+    "foundry_local_asr_set_keep_loaded_secs",
    "foundry_local_asr_set_model",
    "foundry_local_asr_set_runtime_source",
    "foundry_local_asr_status",
@@ -319,6 +319,7 @@ pub use settings::*;
 pub use shared_types::{
    CapsulePayload, CapsuleState, CapsuleStyle, CredentialsStatus, HotkeyMode, HotkeyStatus,
    PendingCorrection, PlatformCapabilities, SelectionPolishOutputMode, UserPreferences,
+    LOCAL_ASR_KEEP_LOADED_FOREVER_SECS,
 };
 pub use shortcut_types::{
    binding_from_legacy_trigger, binding_requires_side_aware_hook, bindings_overlap,
@@ -178,6 +178,8 @@ pub trait ModelRuntimeAdapter: Send + Sync {
        self.test_model(target, model_dir)
    }

+    fn invalidate_scheduled_release(&self, _runtime: LocalAsrRuntime) {}
+
    fn invalidate_route(&self, _runtime: LocalAsrRuntime) {}
 }

@@ -1151,6 +1153,12 @@ impl LocalAsrApi for LocalAsrService {
            LocalAsrRuntime::Foundry => preferences.foundry_local_asr_keep_loaded_secs = seconds,
            LocalAsrRuntime::SherpaOnnx => preferences.sherpa_onnx_keep_loaded_secs = seconds,
        });
+        if result.is_ok()
+            && runtime == LocalAsrRuntime::Foundry
+            && seconds == crate::LOCAL_ASR_KEEP_LOADED_FOREVER_SECS
+        {
+            self.runtime.invalidate_scheduled_release(runtime);
+        }
        let preferences = Arc::clone(&self.preferences);
        let adapter = Arc::clone(&self.runtime);
        let model_store = Arc::clone(&self.model_store);
@@ -19,6 +19,9 @@ pub use crate::android_types::{

 pub use crate::types::{HistorySource, PolishMode};

+/// 本地 ASR 保持加载设置的兼容值：不自动释放，仅由显式操作或进程退出卸载。
+pub const LOCAL_ASR_KEEP_LOADED_FOREVER_SECS: u32 = 86_400;
+
 /// 识别管线模式（issue #902）：`traditional` = 两段式 ASR + LLM 润色；
 /// `multimodal` = 单个多模态模型一步完成「音频 + 提示词 → 最终文本」。
 /// 两套配置在凭据库中完全隔离，运行时只读当前模式，切换不删除另一套配置。
@@ -530,7 +533,7 @@ pub struct UserPreferences {
    #[serde(default = "default_local_asr_mirror")]
    pub local_asr_mirror: String,
    /// 本地 ASR 引擎在内存中的保留时长（秒）。0 = 说完话即释放；
-    /// 较大值 = 上次使用后驻留 N 秒再释放；86400 = 一天 ≈ 永不释放。
+    /// 较大值 = 上次使用后驻留 N 秒再释放；86400 = 永不自动释放。
    /// 默认 300（5 分钟）：兼顾连续听写不重加载、长时间不用释放 1.2GB+ RAM。
    #[serde(default = "default_local_asr_keep_loaded_secs")]
    pub local_asr_keep_loaded_secs: u32,
@@ -69,6 +69,7 @@ fn public_local_asr_preferences_keep_legacy_normalization_semantics() {
 #[derive(Default)]
 struct RecordingLocalAsrRuntime {
    invalidated: Mutex<Vec<LocalAsrRuntime>>,
+    invalidated_release_schedules: Mutex<Vec<LocalAsrRuntime>>,
    fail_release: std::sync::atomic::AtomicBool,
    fail_prepare: std::sync::atomic::AtomicBool,
    fail_preload: std::sync::atomic::AtomicBool,
@@ -359,6 +360,13 @@ impl ModelRuntimeAdapter for RecordingLocalAsrRuntime {
    fn invalidate_route(&self, runtime: LocalAsrRuntime) {
        self.invalidated.lock().unwrap().push(runtime);
    }
+
+    fn invalidate_scheduled_release(&self, runtime: LocalAsrRuntime) {
+        self.invalidated_release_schedules
+            .lock()
+            .unwrap()
+            .push(runtime);
+    }
 }

 #[derive(Default)]
@@ -1541,6 +1549,8 @@ async fn backend_local_asr_service_owns_preferences_and_change_events() {
    assert_eq!(preferences.sherpa_onnx_language_hint, "zh-hans");
    assert_eq!(preferences.foundry_local_runtime_source, "ort-nightly");
    assert_eq!(preferences.foundry_local_asr_keep_loaded_secs, 42);
+    assert_eq!(preferences.local_asr_keep_loaded_secs, 300);
+    assert_eq!(preferences.sherpa_onnx_keep_loaded_secs, 300);
    assert_eq!(
        runtime.invalidated.lock().unwrap().as_slice(),
        [LocalAsrRuntime::Foundry, LocalAsrRuntime::Foundry]
@@ -1554,6 +1564,53 @@ async fn backend_local_asr_service_owns_preferences_and_change_events() {
    let _ = std::fs::remove_dir_all(data_dir);
 }

+#[tokio::test]
+async fn foundry_never_release_invalidates_only_the_pending_finite_schedule() {
+    let (data_dir, runtime, backend) = local_asr_backend();
+
+    for seconds in [0, 60, 300, 1_800] {
+        backend
+            .services()
+            .local_asr
+            .set_keep_loaded_secs(LocalAsrRuntime::Foundry, seconds)
+            .await
+            .unwrap();
+    }
+    assert!(runtime
+        .invalidated_release_schedules
+        .lock()
+        .unwrap()
+        .is_empty());
+
+    backend
+        .services()
+        .local_asr
+        .set_keep_loaded_secs(
+            LocalAsrRuntime::Foundry,
+            openless_core::LOCAL_ASR_KEEP_LOADED_FOREVER_SECS,
+        )
+        .await
+        .unwrap();
+
+    let preferences = backend.get_preferences();
+    assert_eq!(
+        preferences.foundry_local_asr_keep_loaded_secs,
+        openless_core::LOCAL_ASR_KEEP_LOADED_FOREVER_SECS
+    );
+    assert_eq!(preferences.local_asr_keep_loaded_secs, 300);
+    assert_eq!(preferences.sherpa_onnx_keep_loaded_secs, 300);
+    assert_eq!(
+        runtime
+            .invalidated_release_schedules
+            .lock()
+            .unwrap()
+            .as_slice(),
+        [LocalAsrRuntime::Foundry]
+    );
+
+    let _ = std::fs::remove_dir_all(data_dir);
+}
+
 #[tokio::test]
 async fn backend_local_asr_storage_change_commits_only_after_runtime_quiesces() {
    let (data_dir, runtime, backend) = local_asr_backend();
@@ -81,6 +81,22 @@ for (const contract of [
  }
 }

+for (const contract of [
+  'setFoundryLocalAsrKeepLoadedSecs',
+  'foundryStatus?.keepLoadedSecs ?? 300',
+]) {
+  if (!source.includes(contract)) {
+    throw new Error(`Foundry keep-loaded UI contract is missing: ${contract}`);
+  }
+}
+
+const keepLoadedOptionUses = source.match(/options=\{keepLoadedOptions\}/g) ?? [];
+if (keepLoadedOptionUses.length !== 2) {
+  throw new Error(
+    `Generic and Foundry keep-loaded selectors must share the options, found ${keepLoadedOptionUses.length}`,
+  );
+}
+
 console.log(
  'LocalAsr keeps one refresh poller, pauses it for the download dialog, and preserves stable JSX component types',
 );
@@ -130,9 +130,12 @@ mod imp {
    }

    use anyhow::{Context, Result};
-    use foundry_local_sdk::{DeviceType, FoundryLocalConfig, FoundryLocalManager, Model};
+    use foundry_local_sdk::{
+        AudioTranscriptionResponse, DeviceType, FoundryLocalConfig, FoundryLocalManager, Model,
+    };
+    use futures_util::{Stream, StreamExt};
    use parking_lot::Mutex;
-    use tokio::sync::Mutex as AsyncMutex;
+    use tokio::sync::{Mutex as AsyncMutex, OnceCell};

    use super::{
        FoundryCpuFallbackTerminalError, FoundryFallbackNotice, FoundryFallbackNoticeCallback,
@@ -493,6 +496,19 @@ mod imp {
            .ok_or_else(|| anyhow::anyhow!("Foundry Local Whisper total timeout exhausted"))
    }

+    async fn collect_foundry_transcription_text<S, E>(
+        mut stream: S,
+    ) -> std::result::Result<String, E>
+    where
+        S: Stream<Item = std::result::Result<AudioTranscriptionResponse, E>> + Unpin,
+    {
+        let mut text = String::new();
+        while let Some(chunk) = stream.next().await {
+            text.push_str(&chunk?.text);
+        }
+        Ok(text)
+    }
+
    struct FoundrySdkExecution<'a> {
        runtime: &'a FoundryLocalRuntime,
        manager: &'static FoundryLocalManager,
@@ -546,16 +562,19 @@ mod imp {
                client = client.language(language_hint);
            }
            let model_id = self.loaded.model_id.clone();
-            let result = tokio::time::timeout(timeout, client.transcribe(audio_path))
-                .await
-                .with_context(|| {
-                    format!(
-                        "transcribe audio with Foundry model {model_id} timed out after {} seconds",
-                        timeout.as_secs()
-                    )
-                })?
-                .with_context(|| format!("transcribe audio with Foundry model {model_id}"))?;
-            Ok(result.text)
+            let result = tokio::time::timeout(timeout, async {
+                let stream = client.transcribe_streaming(audio_path).await?;
+                collect_foundry_transcription_text(stream).await
+            })
+            .await
+            .with_context(|| {
+                format!(
+                    "transcribe audio with Foundry model {model_id} timed out after {} seconds",
+                    timeout.as_secs()
+                )
+            })?
+            .with_context(|| format!("transcribe audio with Foundry model {model_id}"))?;
+            Ok(result)
        }

        async fn switch_to_cpu(
@@ -692,6 +711,8 @@ mod imp {
        /// 仍可中断（`cancel_prepare` + `check_prepare_cancelled`）。若未来要缩小粒度，
        /// 可让下载阶段不持锁、下载完成后重新校验 route epoch 再持锁加载/推理。
        lifecycle: AsyncMutex<()>,
+        /// EP 注册会使 SDK 的模型目录缓存失效；成功后本进程不再重复注册。
+        execution_providers_ready: OnceCell<()>,
        cancel_prepare: Arc<AtomicBool>,
        temporary_cpu_fallback_sequence: AtomicU64,
        route_epoch: AtomicU64,
@@ -708,6 +729,7 @@ mod imp {
        pub fn new() -> Self {
            Self {
                lifecycle: AsyncMutex::new(()),
+                execution_providers_ready: OnceCell::new(),
                cancel_prepare: Arc::new(AtomicBool::new(false)),
                temporary_cpu_fallback_sequence: AtomicU64::new(0),
                route_epoch: AtomicU64::new(0),
@@ -1084,24 +1106,33 @@ mod imp {
            ));
            let runtime_progress = Arc::clone(&progress);
            let runtime_alias = alias.to_string();
-            manager
-                .download_and_register_eps_with_progress(
-                    None,
-                    move |ep_name: &str, percent: f64| {
-                        let label = if ep_name.trim().is_empty() {
-                            "Foundry Local runtime components".to_string()
-                        } else {
-                            format!("Foundry Local runtime component: {ep_name}")
-                        };
-                        runtime_progress.as_ref()(FoundryPrepareProgressPayload::runtime(
-                            runtime_alias.clone(),
-                            label,
-                            percent,
-                        ));
-                    },
-                )
-                .await
-                .context("download/register Foundry execution providers")?;
+            let cancel_prepare = Arc::clone(&self.cancel_prepare);
+            self.execution_providers_ready
+                .get_or_try_init(|| async move {
+                    manager
+                        .download_and_register_eps_with_progress(
+                            None,
+                            move |ep_name: &str, percent: f64| {
+                                let label = if ep_name.trim().is_empty() {
+                                    "Foundry Local runtime components".to_string()
+                                } else {
+                                    format!("Foundry Local runtime component: {ep_name}")
+                                };
+                                runtime_progress.as_ref()(FoundryPrepareProgressPayload::runtime(
+                                    runtime_alias.clone(),
+                                    label,
+                                    percent,
+                                ));
+                            },
+                        )
+                        .await
+                        .context("download/register Foundry execution providers")?;
+                    if cancel_prepare.load(Ordering::SeqCst) {
+                        anyhow::bail!("Foundry Local Whisper prepare cancelled");
+                    }
+                    Ok::<(), anyhow::Error>(())
+                })
+                .await?;
            progress.as_ref()(FoundryPrepareProgressPayload::runtime(
                alias,
                "Foundry Local runtime components",
@@ -1662,15 +1693,16 @@ mod imp {
        }

        use super::{
-            cpu_load_completion, foundry_native_dir_candidates, is_cuda_cudnn_failure,
-            is_cuda_fallback_candidate, may_reuse_loaded_model, normalized_language_hint,
-            select_cpu_variant_id, select_foundry_native_dir,
+            collect_foundry_transcription_text, cpu_load_completion, foundry_native_dir_candidates,
+            is_cuda_cudnn_failure, is_cuda_fallback_candidate, may_reuse_loaded_model,
+            normalized_language_hint, select_cpu_variant_id, select_foundry_native_dir,
            should_release_temporary_cpu_fallback, transcribe_recording_with_adapter,
            FoundryCpuLoadCompletion, FoundryCpuSwitch, FoundryExecutionAdapter,
            FoundryExecutionDevice, FoundryFallbackNotice, FoundryFallbackNoticeCallback,
            FoundryLocalRuntime, FoundryVariantDescriptor,
        };
        use anyhow::Result;
+        use foundry_local_sdk::AudioTranscriptionResponse;
        use std::{
            collections::VecDeque,
            fs,
@@ -1823,6 +1855,45 @@ mod imp {
            (callback, received)
        }

+        fn transcription_response(text: &str) -> AudioTranscriptionResponse {
+            AudioTranscriptionResponse {
+                text: text.to_string(),
+                language: None,
+                duration: None,
+                segments: None,
+                words: None,
+            }
+        }
+
+        #[tokio::test]
+        async fn foundry_streaming_transcription_concatenates_ordered_responses() {
+            let stream = futures_util::stream::iter([
+                Ok::<_, &'static str>(transcription_response("中文")),
+                Ok(transcription_response("")),
+                Ok(transcription_response("转写完成")),
+            ]);
+
+            assert_eq!(
+                collect_foundry_transcription_text(stream).await.unwrap(),
+                "中文转写完成"
+            );
+        }
+
+        #[tokio::test]
+        async fn foundry_streaming_transcription_propagates_chunk_errors() {
+            let stream = futures_util::stream::iter([
+                Ok(transcription_response("partial")),
+                Err("stream failed"),
+            ]);
+
+            assert_eq!(
+                collect_foundry_transcription_text(stream)
+                    .await
+                    .unwrap_err(),
+                "stream failed"
+            );
+        }
+
        #[tokio::test]
        async fn cuda_failure_retries_the_failed_chunk_once_on_cpu_and_keeps_cpu_for_later_chunks()
        {
@@ -38,6 +38,7 @@ pub struct FoundryStatusWire {
    pub runtime_source: String,
    pub active_model: String,
    pub loaded_model_id: Option<String>,
+    pub keep_loaded_secs: u32,
    pub endpoint: Option<String>,
    pub error: Option<String>,
 }
@@ -51,6 +52,7 @@ impl From<openless_core::LocalAsrRuntimeStatus> for FoundryStatusWire {
            runtime_source: status.runtime_source.unwrap_or_default().as_str().into(),
            active_model: status.active_model,
            loaded_model_id: status.model_id,
+            keep_loaded_secs: status.keep_loaded_secs,
            endpoint: status.endpoint,
            error: status.error,
        }
@@ -140,6 +142,19 @@ pub async fn foundry_local_asr_set_runtime_source(
        .map_err(core_error)
 }

+#[tauri::command]
+pub async fn foundry_local_asr_set_keep_loaded_secs(
+    backend: CoreState<'_>,
+    seconds: u32,
+) -> Result<(), String> {
+    backend
+        .services()
+        .local_asr
+        .set_keep_loaded_secs(LocalAsrRuntime::Foundry, seconds)
+        .await
+        .map_err(core_error)
+}
+
 #[tauri::command]
 pub async fn foundry_local_asr_prepare(
    backend: CoreState<'_>,
@@ -248,6 +263,7 @@ mod wire_contract_tests {
                "runtimeSource": "ort-nightly",
                "activeModel": "whisper-small",
                "loadedModelId": null,
+                "keepLoadedSecs": 300,
                "endpoint": null,
                "error": null,
            })
@@ -307,16 +307,6 @@ impl TauriLocalAsrRuntimeAdapter {
            foundry_rebind_pending: Arc::new(AtomicBool::new(false)),
        }
    }
-
-    #[cfg(target_os = "windows")]
-    fn invalidate_release(&self, runtime: openless_core::LocalAsrRuntime) {
-        match runtime {
-            openless_core::LocalAsrRuntime::Foundry => &self.native.foundry_generation,
-            openless_core::LocalAsrRuntime::SherpaOnnx => &self.native.sherpa_generation,
-            openless_core::LocalAsrRuntime::Generic => return,
-        }
-        .fetch_add(1, Ordering::AcqRel);
-    }
 }

 impl openless_core::ModelRuntimeAdapter for TauriLocalAsrRuntimeAdapter {
@@ -550,7 +540,7 @@ impl openless_core::ModelRuntimeAdapter for TauriLocalAsrRuntimeAdapter {
        progress: openless_core::ModelPrepareProgressSink,
    ) -> BoxFuture<'static, Result<String, BackendError>> {
        #[cfg(target_os = "windows")]
-        self.invalidate_release(target.runtime);
+        self.invalidate_scheduled_release(target.runtime);
        let foundry = Arc::clone(&self.native.foundry);
        let sherpa = Arc::clone(&self.native.sherpa);
        let foundry_rebind_pending = Arc::clone(&self.foundry_rebind_pending);
@@ -692,7 +682,7 @@ impl openless_core::ModelRuntimeAdapter for TauriLocalAsrRuntimeAdapter {
        runtime: openless_core::LocalAsrRuntime,
    ) -> BoxFuture<'static, Result<(), BackendError>> {
        #[cfg(target_os = "windows")]
-        self.invalidate_release(runtime);
+        self.invalidate_scheduled_release(runtime);
        let foundry = Arc::clone(&self.native.foundry);
        let sherpa = Arc::clone(&self.native.sherpa);
        #[cfg(any(target_os = "macos", target_os = "linux"))]
@@ -813,6 +803,16 @@ impl openless_core::ModelRuntimeAdapter for TauriLocalAsrRuntimeAdapter {
            self.native.foundry.invalidate_route();
        }
    }
+
+    #[cfg(target_os = "windows")]
+    fn invalidate_scheduled_release(&self, runtime: openless_core::LocalAsrRuntime) {
+        match runtime {
+            openless_core::LocalAsrRuntime::Foundry => &self.native.foundry_generation,
+            openless_core::LocalAsrRuntime::SherpaOnnx => &self.native.sherpa_generation,
+            openless_core::LocalAsrRuntime::Generic => return,
+        }
+        .fetch_add(1, Ordering::AcqRel);
+    }
 }

 async fn test_model_with_provider(
@@ -2275,6 +2275,12 @@ fn pcm_duration_ms(bytes: &[u8]) -> u64 {
    (bytes.len() as u64 / 2).saturating_mul(1_000) / 16_000
 }

+#[cfg(any(target_os = "windows", target_os = "macos", target_os = "linux", test))]
+fn local_asr_release_delay(keep_loaded_secs: u32) -> Option<std::time::Duration> {
+    (keep_loaded_secs != openless_core::LOCAL_ASR_KEEP_LOADED_FOREVER_SECS)
+        .then(|| std::time::Duration::from_secs(keep_loaded_secs as u64))
+}
+
 #[cfg(target_os = "windows")]
 fn schedule_foundry_release(
    runtime: Arc<crate::asr::local::FoundryLocalRuntime>,
@@ -2298,8 +2304,11 @@ fn schedule_foundry_release(
                }
            }
        }
-        if keep_loaded_secs > 0 {
-            tokio::time::sleep(std::time::Duration::from_secs(keep_loaded_secs as u64)).await;
+        let Some(delay) = local_asr_release_delay(keep_loaded_secs) else {
+            return;
+        };
+        if !delay.is_zero() {
+            tokio::time::sleep(delay).await;
        }
        if current_generation.load(Ordering::Acquire) != generation {
            return;
@@ -2327,9 +2336,12 @@ fn schedule_sherpa_release(
    generation: u64,
    current_generation: Arc<AtomicU64>,
 ) {
+    let Some(delay) = local_asr_release_delay(keep_loaded_secs) else {
+        return;
+    };
    tauri::async_runtime::spawn(async move {
-        if keep_loaded_secs > 0 {
-            tokio::time::sleep(std::time::Duration::from_secs(keep_loaded_secs as u64)).await;
+        if !delay.is_zero() {
+            tokio::time::sleep(delay).await;
        }
        if current_generation.load(Ordering::Acquire) == generation {
            if let Err(error) = runtime
@@ -2348,8 +2360,10 @@ fn schedule_qwen_release(
    engine: std::sync::Weak<crate::asr::local::LocalQwenEngine>,
    keep_loaded_secs: u32,
 ) {
+    let Some(threshold) = local_asr_release_delay(keep_loaded_secs) else {
+        return;
+    };
    tauri::async_runtime::spawn(async move {
-        let threshold = std::time::Duration::from_secs(keep_loaded_secs as u64);
        if !threshold.is_zero() {
            tokio::time::sleep(threshold).await;
        }
@@ -2363,8 +2377,10 @@ fn schedule_whisper_release(
    engine: std::sync::Weak<crate::asr::local::WhisperEngine>,
    keep_loaded_secs: u32,
 ) {
+    let Some(threshold) = local_asr_release_delay(keep_loaded_secs) else {
+        return;
+    };
    tauri::async_runtime::spawn(async move {
-        let threshold = std::time::Duration::from_secs(keep_loaded_secs as u64);
        if !threshold.is_zero() {
            tokio::time::sleep(threshold).await;
        }
@@ -3373,6 +3389,22 @@ mod tests {

    struct IgnoreTextStreamSink;

+    #[test]
+    fn local_asr_keep_loaded_delay_distinguishes_immediate_finite_and_forever() {
+        assert_eq!(
+            super::local_asr_release_delay(0),
+            Some(std::time::Duration::ZERO)
+        );
+        assert_eq!(
+            super::local_asr_release_delay(300),
+            Some(std::time::Duration::from_secs(300))
+        );
+        assert_eq!(
+            super::local_asr_release_delay(openless_core::LOCAL_ASR_KEEP_LOADED_FOREVER_SECS),
+            None
+        );
+    }
+
    #[cfg(target_os = "windows")]
    #[tokio::test]
    async fn windows_preload_requires_the_requested_model_to_be_prepared() {
@@ -3395,10 +3427,34 @@ mod tests {
                .unwrap_err();
            assert_eq!(error.code, BackendErrorCode::InvalidState);
        }
-        adapter.invalidate_release(LocalAsrRuntime::Foundry);
+    }
+
+    #[cfg(target_os = "windows")]
+    #[tokio::test]
+    async fn windows_scheduled_release_invalidation_is_runtime_scoped() {
+        use openless_core::{LocalAsrRuntime, ModelRuntimeAdapter};
+
+        let adapter = TauriLocalAsrRuntimeAdapter::new(
+            TauriNativeAsrDependencies::new(
+                Arc::new(crate::asr::local::FoundryLocalRuntime::new()),
+                Arc::new(crate::asr::local::SherpaOnnxRuntime::new()),
+            ),
+            Arc::new(openless_core::PreferencesStore::in_memory()),
+        );
+        let stale_foundry_generation = adapter.native.foundry_generation.load(Ordering::Acquire);
+        adapter.invalidate_scheduled_release(LocalAsrRuntime::Foundry);
        assert_eq!(adapter.native.foundry_generation.load(Ordering::Acquire), 1);
        assert_eq!(adapter.native.sherpa_generation.load(Ordering::Acquire), 0);
-        adapter.invalidate_release(LocalAsrRuntime::SherpaOnnx);
+        assert!(!adapter
+            .native
+            .foundry
+            .release_if_generation(
+                adapter.native.foundry_generation.as_ref(),
+                stale_foundry_generation,
+            )
+            .await
+            .unwrap());
+        adapter.invalidate_scheduled_release(LocalAsrRuntime::SherpaOnnx);
        assert_eq!(adapter.native.foundry_generation.load(Ordering::Acquire), 1);
        assert_eq!(adapter.native.sherpa_generation.load(Ordering::Acquire), 1);
    }
@@ -368,6 +368,7 @@ macro_rules! app_invoke_handler_desktop {
            commands::foundry_local_asr_set_model,
            commands::foundry_local_asr_set_language_hint,
            commands::foundry_local_asr_set_runtime_source,
+            commands::foundry_local_asr_set_keep_loaded_secs,
            commands::foundry_local_asr_prepare,
            commands::foundry_local_asr_cancel_prepare,
            commands::foundry_local_asr_release,
@@ -2219,7 +2219,7 @@ export const de: typeof zhCN = {
    releaseNow: 'Jetzt entladen',
    keepLoadedLabel: 'Geladen halten für',
    keepLoadedDesc:
-      'Wie lange Qwen3-ASR nach der letzten Verwendung im Arbeitsspeicher bleibt, bevor es entladen wird.',
+      'Wie lange die aktuelle lokale ASR-Engine nach der nächsten Transkription geladen bleibt; „Nie entladen“ gilt bis zum manuellen Entladen oder Beenden.',
    keepImmediate: 'Sofort entladen',
    keep1min: '1 Minute nach letzter Verwendung',
    keep5min: '5 Minuten nach letzter Verwendung (Standard)',
@@ -2151,7 +2151,8 @@ export const en: typeof zhCN = {
    loadNow: 'Load now',
    releaseNow: 'Release now',
    keepLoadedLabel: 'Keep loaded for',
-    keepLoadedDesc: 'How long Qwen3-ASR stays in memory after the last use, before being freed.',
+    keepLoadedDesc:
+      'How long the current local ASR engine stays loaded after the next transcription; Never release keeps it loaded until manual release or exit.',
    keepImmediate: 'Release immediately',
    keep1min: '1 minute after last use',
    keep5min: '5 minutes after last use (default)',
@@ -2200,7 +2200,7 @@ export const es: typeof zhCN = {
    releaseNow: 'Liberar ahora',
    keepLoadedLabel: 'Mantener cargado durante',
    keepLoadedDesc:
-      'Tiempo que Qwen3-ASR permanece en memoria después del último uso antes de liberarse.',
+      'Tiempo que el motor ASR local actual permanece cargado tras la siguiente transcripción; «No liberar nunca» se mantiene hasta liberarlo manualmente o salir.',
    keepImmediate: 'Liberar inmediatamente',
    keep1min: '1 minuto tras el último uso',
    keep5min: '5 minutos tras el último uso (predeterminado)',
@@ -2229,7 +2229,7 @@ export const fr: typeof zhCN = {
    releaseNow: 'Libérer maintenant',
    keepLoadedLabel: 'Conserver en mémoire pendant',
    keepLoadedDesc:
-      'Durée pendant laquelle Qwen3-ASR reste en mémoire après sa dernière utilisation avant d’être libéré.',
+      'Durée de chargement du moteur ASR local actuel après la prochaine transcription ; « Ne jamais libérer » le conserve jusqu’à une libération manuelle ou la fermeture.',
    keepImmediate: 'Libérer immédiatement',
    keep1min: '1 minute après la dernière utilisation',
    keep5min: '5 minutes après la dernière utilisation (par défaut)',
@@ -2120,7 +2120,7 @@ export const ja: typeof zhCN = {
    releaseNow: '今すぐ解放',
    keepLoadedLabel: 'ロード保持時間',
    keepLoadedDesc:
-      'ローカル ASR を使用後、何分でメモリから解放するかを決定。1+ GB の RAM 占有を回避。',
+      '現在のローカル ASR を次回の文字起こし後に保持する時間を指定します。「解放しない」は手動解放または終了まで保持します。',
    keepImmediate: '使用直後に解放',
    keep1min: '最終使用から 1 分',
    keep5min: '最終使用から 5 分（既定）',
@@ -2102,7 +2102,7 @@ export const ko: typeof zhCN = {
    releaseNow: '지금 해제',
    keepLoadedLabel: '로드 유지 시간',
    keepLoadedDesc:
-      '로컬 ASR 사용 후 메모리에서 해제되기까지의 시간을 결정. 1+ GB RAM 장기 점유 회피.',
+      '현재 로컬 ASR 엔진을 다음 전사 후 얼마나 유지할지 정합니다. 해제하지 않음은 수동 해제 또는 종료까지 유지합니다.',
    keepImmediate: '말하기 직후 해제',
    keep1min: '마지막 사용 후 1분',
    keep5min: '마지막 사용 후 5분(기본)',
@@ -2030,7 +2030,8 @@ export const zhCN = {
    loadNow: '立即加载',
    releaseNow: '立即释放',
    keepLoadedLabel: '保持加载多久',
-    keepLoadedDesc: '决定 Qwen3-ASR 用完后多久从内存释放，避免长期占用内存。',
+    keepLoadedDesc:
+      '决定当前本地 ASR 引擎在下次转写后保持加载多久；“不释放”会持续驻留至手动释放或退出。',
    keepImmediate: '说完话立即释放',
    keep1min: '上次使用后 1 分钟',
    keep5min: '上次使用后 5 分钟（默认）',
@@ -2016,7 +2016,8 @@ export const zhTW: typeof zhCN = {
    loadNow: '立即加載',
    releaseNow: '立即釋放',
    keepLoadedLabel: '保持加載多久',
-    keepLoadedDesc: '決定 Qwen3-ASR 用完後多久從內存釋放，避免長期佔用內存。',
+    keepLoadedDesc:
+      '決定目前本地 ASR 引擎在下次轉寫後保持載入多久；「不釋放」會持續駐留至手動釋放或退出。',
    keepImmediate: '說完話立即釋放',
    keep1min: '上次使用後 1 分鐘',
    keep5min: '上次使用後 5 分鐘（默認）',
@@ -1,4 +1,4 @@
-import { isLocalAsrModelSupportedOnOs } from './localAsr';
+import { LOCAL_ASR_KEEP_LOADED_OPTIONS, isLocalAsrModelSupportedOnOs } from './localAsr';

 function assertEqual(actual: boolean, expected: boolean, name: string) {
  if (actual !== expected) {
@@ -30,3 +30,8 @@ assertEqual(
  true,
  'Whisper is available on macOS',
 );
+
+const keepLoadedSeconds = LOCAL_ASR_KEEP_LOADED_OPTIONS.map((option) => option.seconds);
+if (JSON.stringify(keepLoadedSeconds) !== JSON.stringify([0, 60, 300, 1800, 86400])) {
+  throw new Error(`unexpected keep-loaded options: ${keepLoadedSeconds.join(', ')}`);
+}
@@ -104,10 +104,19 @@ export interface FoundryLocalAsrStatus {
  runtimeSource: FoundryRuntimeSource;
  activeModel: string;
  loadedModelId: string | null;
+  keepLoadedSecs: number;
  endpoint: string | null;
  error: string | null;
 }

+export const LOCAL_ASR_KEEP_LOADED_OPTIONS = [
+  { seconds: 0, labelKey: 'localAsr.keepImmediate' },
+  { seconds: 60, labelKey: 'localAsr.keep1min' },
+  { seconds: 300, labelKey: 'localAsr.keep5min' },
+  { seconds: 1800, labelKey: 'localAsr.keep30min' },
+  { seconds: 86400, labelKey: 'localAsr.keepForever' },
+] as const;
+
 export const FOUNDRY_LOCAL_ASR_MODEL_ALIASES = [
  'whisper-small',
  'whisper-medium',
@@ -403,6 +412,7 @@ export function getFoundryLocalAsrStatus(): Promise<FoundryLocalAsrStatus> {
    runtimeSource: 'auto',
    activeModel: 'whisper-small',
    loadedModelId: null,
+    keepLoadedSecs: 300,
    endpoint: null,
    error: null,
  }));
@@ -424,6 +434,10 @@ export function setFoundryLocalRuntimeSource(source: string): Promise<void> {
  return invokeOrMock('foundry_local_asr_set_runtime_source', { source }, () => undefined);
 }

+export function setFoundryLocalAsrKeepLoadedSecs(seconds: number): Promise<void> {
+  return invokeOrMock('foundry_local_asr_set_keep_loaded_secs', { seconds }, () => undefined);
+}
+
 export function prepareFoundryLocalAsr(modelAlias: string): Promise<string> {
  return invokeOrMock('foundry_local_asr_prepare', { modelAlias }, () => `mock-${modelAlias}`);
 }
@@ -445,7 +445,7 @@ export interface UserPreferences {
  /** 本地模型下载源镜像（'huggingface' / 'hf-mirror'）。 */
  localAsrMirror: string;
  /** 本地 ASR 引擎在内存中的保留时长（秒）。0 = 说完话即释放；
-   *  300 = 默认 5 分钟；86400 ≈ 不释放（保持加载）。 */
+   *  300 = 默认 5 分钟；86400 = 不自动释放（保持加载）。 */
  localAsrKeepLoadedSecs: number;
  /** Windows Foundry Local Whisper 当前激活的模型 alias。 */
  foundryLocalAsrModel: string;
@@ -15,6 +15,7 @@ import { emitSaved } from '../../lib/savedEvent';
 import { useLayoutStack } from '../../lib/useMobileLayout';
 import {
  FOUNDRY_LOCAL_ASR_MODELS,
+  LOCAL_ASR_KEEP_LOADED_OPTIONS,
  SHERPA_ONNX_ASR_MODELS,
  activateLocalAsr,
  cancelFoundryLocalAsrPrepare,
@@ -49,6 +50,7 @@ import {
  revealSherpaOnnxAsrModelDir,
  setLocalAsrModelsBaseDir,
  setFoundryLocalAsrLanguageHint,
+  setFoundryLocalAsrKeepLoadedSecs,
  setFoundryLocalRuntimeSource,
  setLocalAsrKeepLoadedSecs,
  setLocalAsrMirror,
@@ -161,6 +163,10 @@ type RefreshGuard = () => boolean;

 export function LocalAsr({ embedded = false }: LocalAsrProps = {}) {
  const { t } = useTranslation();
+  const keepLoadedOptions = LOCAL_ASR_KEEP_LOADED_OPTIONS.map(({ seconds, labelKey }) => ({
+    value: String(seconds),
+    label: t(labelKey),
+  }));
  const stackLayout = useLayoutStack(1000);
  const { prefs, updatePrefs } = useHotkeySettings();
  const [settings, setSettings] = useState<LocalAsrSettings | null>(null);
@@ -351,6 +357,7 @@ export function LocalAsr({ embedded = false }: LocalAsrProps = {}) {
        runtimeSource: selectedFoundryRuntimeSource,
        activeModel: selectedFoundryAlias,
        loadedModelId: null,
+        keepLoadedSecs: 300,
        endpoint: null,
        error: message,
      });
@@ -937,6 +944,18 @@ export function LocalAsr({ embedded = false }: LocalAsrProps = {}) {
    }
  };

+  const handleFoundryKeepLoadedChange = async (seconds: number, restoreScroll?: () => void) => {
+    try {
+      setError(null);
+      await setFoundryLocalAsrKeepLoadedSecs(seconds);
+      await refreshFoundryStatus();
+    } catch (e) {
+      setError(e instanceof Error ? e.message : String(e));
+    } finally {
+      restoreScroll?.();
+    }
+  };
+
  const handleEnableFoundry = async (aliasOverride?: FoundryLocalAsrModelAlias) => {
    if (!foundryAvailable) return;
    const alias = aliasOverride ?? selectedFoundryAlias;
@@ -2167,13 +2186,7 @@ export function LocalAsr({ embedded = false }: LocalAsrProps = {}) {
                        value={String(engineStatus?.keepLoadedSecs ?? 300)}
                        onChange={(v) => void handleKeepLoadedChange(Number(v))}
                        ariaLabel={t('localAsr.keepLoadedLabel')}
-                        options={[
-                          { value: '0', label: t('localAsr.keepImmediate') },
-                          { value: '60', label: t('localAsr.keep1min') },
-                          { value: '300', label: t('localAsr.keep5min') },
-                          { value: '1800', label: t('localAsr.keep30min') },
-                          { value: '86400', label: t('localAsr.keepForever') },
-                        ]}
+                        options={keepLoadedOptions}
                        style={{ fontSize: 13, height: 31, minWidth: 200 }}
                      />
                    </div>
@@ -2527,6 +2540,33 @@ export function LocalAsr({ embedded = false }: LocalAsrProps = {}) {
                    }}
                  />
                </label>
+                <label
+                  style={{
+                    display: 'flex',
+                    flexDirection: 'column',
+                    gap: 4,
+                    fontSize: 11,
+                    color: 'var(--ol-ink-4)',
+                  }}
+                >
+                  {t('localAsr.keepLoadedLabel')}
+
+                  <SelectLite
+                    value={String(foundryStatus?.keepLoadedSecs ?? 300)}
+                    onChange={(v) => {
+                      const restoreScroll = preserveEmbeddedScroll(foundryControlsRef.current);
+                      void handleFoundryKeepLoadedChange(Number(v), restoreScroll);
+                    }}
+                    disabled={foundryBusy !== null}
+                    ariaLabel={t('localAsr.keepLoadedLabel')}
+                    options={keepLoadedOptions}
+                    style={{
+                      fontSize: 13,
+                      height: 31,
+                      minWidth: 200,
+                    }}
+                  />
+                </label>
              </div>
            </div>

@@ -2586,6 +2626,7 @@ export function LocalAsr({ embedded = false }: LocalAsrProps = {}) {
                </span>
                {foundryStatus?.loadedModelId ?? t('localAsr.foundryNotLoaded')}
              </div>
+              <div>{t('localAsr.keepLoadedDesc')}</div>
              {foundryStatus?.error && (
                <div style={{ color: '#9b2c2c' }}>
                  <span>{t('localAsr.foundryError')}: </span>