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: list compute devices off the startup path (#2160)

Aktan Azat29bd2c0d6b4b
diff --git a/src-tauri/src/lib.rs b/src-tauri/src/lib.rs
index facc9786..6a3baf45 100644
--- a/src-tauri/src/lib.rs
+++ b/src-tauri/src/lib.rs
@@ -915,6 +915,7 @@ pub fn run(cli_args: CliArgs) {
app_handle.manage(model_manager);
app_handle.manage(transcription_manager);
managers::transcription::init_transcribe_backend();
+ managers::transcription::report_compute_devices();
managers::transcription::apply_accelerator_settings(&app_handle);
let handle = app_handle.clone();
@@ -1023,11 +1024,14 @@ pub fn run(cli_args: CliArgs) {
);
// Pre-warm GPU/accelerator enumeration on a background thread. The first
- // get_available_accelerators call enumerates ORT execution providers and
- // transcribe-cpp compute devices, which can take a moment; without this
+ // device listing opens the GPU, which on macOS loads ggml's Metal library
+ // and compiles it when the system shader cache does not have it yet, so it
+ // stays off the startup path. get_available_accelerators then enumerates
+ // ORT execution providers and transcribe-cpp compute devices; without this
// the cost is paid synchronously when the user first opens Advanced
// settings, freezing the UI. Result is cached in a OnceLock.
std::thread::spawn(|| {
+ crate::managers::transcription::report_compute_devices();
let _ = crate::managers::transcription::get_available_accelerators();
});
diff --git a/src-tauri/src/managers/transcription.rs b/src-tauri/src/managers/transcription.rs
index bed8d92c..a8d0dddd 100644
--- a/src-tauri/src/managers/transcription.rs
+++ b/src-tauri/src/managers/transcription.rs
@@ -1891,21 +1891,31 @@ pub fn init_transcribe_backend() {
disabling transcribe.cpp GPU acceleration and using CPU"
);
}
- let devices = transcribe_compute_devices();
- info!(
- "transcribe-cpp initialized with {} compute device(s): [{}]",
- devices.len(),
- devices
- .iter()
- .map(|d| format!("{} ({})", d.name, d.kind))
- .collect::<Vec<_>>()
- .join(", ")
- );
}
Err(e) => warn!("Failed to initialize transcribe-cpp backends: {}", e),
}
}
+/// Log the compute devices [`init_transcribe_backend`] registered.
+///
+/// Listing devices is what first opens the GPU. On macOS that loads ggml's
+/// Metal library, which is compiled from source whenever the system's shader
+/// cache does not hold it yet (the first launch after an install or update),
+/// so the app calls this from a background thread instead of its startup
+/// path. A model load that comes first waits on the same one-time compile.
+pub fn report_compute_devices() {
+ let devices = transcribe_compute_devices();
+ info!(
+ "transcribe-cpp initialized with {} compute device(s): [{}]",
+ devices.len(),
+ devices
+ .iter()
+ .map(|d| format!("{} ({})", d.name, d.kind))
+ .collect::<Vec<_>>()
+ .join(", ")
+ );
+}
+
/// Human-readable list of the transcribe-cpp compute devices registered at
/// startup, for the `--list-devices` flag. The reported `index` is the
/// value to pass to `--device-index`. Backends must be initialized first