Skip to content

YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context

fix(audio): stop building throwaway ORT sessions in boot-time model pre-download (#4991)

Ezra Ellette766f06d9f2b3
diff --git a/apps/screenpipe-app-tauri/src-tauri/src/main.rs b/apps/screenpipe-app-tauri/src-tauri/src/main.rs
index d0cb5c452..7f2536d3c 100644
--- a/apps/screenpipe-app-tauri/src-tauri/src/main.rs
+++ b/apps/screenpipe-app-tauri/src-tauri/src/main.rs
@@ -1388,7 +1388,13 @@ async fn main() {
}
},
async {
- match screenpipe_audio::speaker::models::get_or_download_model(
+ // File-only fetch — NOT get_or_download_model, which also builds
+ // an ORT session. That session would be immediately discarded here
+ // and would compete for CPU with the real session build that
+ // SegmentationManager does moments later for the same file (root
+ // cause of the "ort session init: timed out after 30s" boot warning
+ // observed on the macos-15-intel CI runner).
+ match screenpipe_audio::speaker::models::ensure_model_file(
screenpipe_audio::speaker::models::PyannoteModel::Segmentation
).await {
Ok(p) => info!("segmentation model pre-download complete: {:?}", p),
@@ -1396,7 +1402,7 @@ async fn main() {
}
},
async {
- match screenpipe_audio::speaker::models::get_or_download_model(
+ match screenpipe_audio::speaker::models::ensure_model_file(
screenpipe_audio::speaker::models::PyannoteModel::Embedding
).await {
Ok(p) => info!("embedding model pre-download complete: {:?}", p),
diff --git a/crates/screenpipe-audio/src/speaker/models.rs b/crates/screenpipe-audio/src/speaker/models.rs
index e932f96e5..f4a4aa533 100644
--- a/crates/screenpipe-audio/src/speaker/models.rs
+++ b/crates/screenpipe-audio/src/speaker/models.rs
@@ -36,14 +36,39 @@ pub async fn get_or_download_model(model_type: PyannoteModel) -> Result<LoadedMo
get_or_download_model_with_retries(model_type, MAX_RECOVERY_RETRIES).await
}
+/// Ensure the model file is on disk without building an ORT session for it.
+///
+/// Callers that only want the file warmed in cache (e.g. main.rs's opportunistic
+/// boot-time pre-download) must use this instead of `get_or_download_model` —
+/// that one unconditionally calls `create_session`, which is a real CPU-bound
+/// ONNX Runtime graph load/optimize, not just a disk check. A pre-download call
+/// that builds (and immediately discards) a session competes for CPU with the
+/// *real* session build that `SegmentationManager`/`EmbeddingExtractor` does
+/// moments later on the same file, and on a CPU-constrained host that
+/// self-inflicted contention is enough to trip `ORT_INIT_TIMEOUT` on both sides
+/// (observed on GitHub's `macos-15-intel` runner: concurrent pre-download +
+/// on-demand session builds for the same model pushed session build time from
+/// the normal sub-second range past the 30s watchdog).
+pub async fn ensure_model_file(model_type: PyannoteModel) -> Result<PathBuf> {
+ let (url, filename, model_path_lock, downloading_flag) = model_state(model_type);
+ let cache_dir = get_cache_dir()?;
+ let downloader = ModelDownloader::new(
+ url.to_string(),
+ filename.to_string(),
+ cache_dir,
+ downloading_flag,
+ model_path_lock,
+ );
+ downloader.ensure_model_available().await
+}
+
async fn get_or_download_model_with_retries(
model_type: PyannoteModel,
max_retries: u8,
) -> Result<LoadedModel> {
let mut retry_count = 0;
loop {
- let (url, filename, model_path_lock, downloading_flag) = model_state(model_type);
- let cache_dir = get_cache_dir()?;
+ let (_, filename, model_path_lock, _) = model_state(model_type);
{
let mut cached = model_path_lock.lock().await;
@@ -58,14 +83,7 @@ async fn get_or_download_model_with_retries(
}
}
- let downloader = ModelDownloader::new(
- url.to_string(),
- filename.to_string(),
- cache_dir,
- downloading_flag,
- model_path_lock,
- );
- let path = downloader.ensure_model_available().await?;
+ let path = ensure_model_file(model_type).await?;
match create_session(&path) {
Ok(session) => return Ok(LoadedModel { path, session }),