New UI and improvments across the app

.github/workflows/cd.yml

@@ -215,7 +215,7 @@ jobs:
         if: env.SIGNED != 'true'
         run: scripts/bundle-app.sh
       - name: Start check
-        run: dist/stemd.app/Contents/MacOS/stemd --version
+        run: dist/Stemd.app/Contents/MacOS/stemd --version
       - name: Remove credentials
         if: always()
         run: |
@@ -271,14 +271,14 @@ jobs:
           Test-Version dist\stemd-windows\stemd-server.exe
           $setup = Get-Item dist\stemd-*-setup.exe
           Start-Process $setup.FullName -ArgumentList '/S' -Wait
-          $installed = Join-Path $env:LOCALAPPDATA 'Programs\stemd'
+          $installed = Join-Path $env:LOCALAPPDATA 'Programs\Stemd'
           Test-Version (Join-Path $installed 'stemd-server.exe')
           Test-Version (Join-Path $installed 'stemd-cli.exe')
       # The runner has no NVIDIA driver.
       - name: CPU separation
         timeout-minutes: 30
         run: |
-          $installed = Join-Path $env:LOCALAPPDATA 'Programs\stemd'
+          $installed = Join-Path $env:LOCALAPPDATA 'Programs\Stemd'
           .github/scripts/cpu-separation.ps1 -Server (Join-Path $installed 'stemd-server.exe') -Cli (Join-Path $installed 'stemd-cli.exe')
       - uses: actions/upload-artifact@v4
         with:

Cargo.toml

@@ -15,9 +15,8 @@ members = [
 [workspace.package]
 version = "0.1.4"
 edition = "2024"
-# Stated so cargo says "requires rustc 1.85" rather than failing to parse the
-# edition. It is the edition's floor, not a tested one.
-rust-version = "1.85"
+# 1.89 for `File::try_lock`, which keeps a second server off a cache in use.
+rust-version = "1.89"
 license = "MIT OR Apache-2.0"
 repository = "https://github.com/nsaintot/stemd"
 readme = "README.md"

README.md

@@ -1,15 +1,15 @@
 <div align="center">
 
-<img src="resources/stemd-icon-mac-256.png" width="120" alt="stemd">
+<img src="resources/stemd-icon-mac-256.png" width="120" alt="Stemd">
 
-# stemd
+# Stemd
 
 **Split any track into stems, on your own machine.**
 
 Drop a file on the window. Get `harmonics`, `vocals` and `drums` in a folder
 beside it.
 
-<img src="resources/screenshot.png" width="440" alt="The stemd window, with a drop zone reading Drop a track">
+<img src="resources/screenshot.png" width="440" alt="The Stemd window: a drop zone reading Drop a track, and the Separated list below it">
 
 [![License](https://img.shields.io/badge/license-MIT%20OR%20Apache--2.0-blue)](#license)
 [![Platforms](https://img.shields.io/badge/platforms-macOS%20%7C%20Windows%20%7C%20Linux-lightgrey)](#install)
@@ -43,7 +43,7 @@ track bit for bit.
 
 **macOS.** Open the `macos-arm64` disk image from the
 [latest release](https://github.com/nsaintot/stemd/releases/latest) and drag
-stemd to Applications. Signed and notarized, and the ticket is stapled to the
+Stemd to Applications. Signed and notarized, and the ticket is stapled to the
 app as well as to the image, so it opens without argument and the first launch
 does not need a network.
 
@@ -54,7 +54,7 @@ installer.
 
 On a machine with an NVIDIA card, tick the box at the end of the installer or run
 `install-cuda.cmd` once afterwards: about 1.2 GB of NVIDIA's own runtime, fetched
-once and pinned by digest. No toolkit and no repository. Without it stemd runs on
+once and pinned by digest. No toolkit and no repository. Without it Stemd runs on
 the CPU, correctly and far slower.
 
 **Linux.** Take the package for your distribution and architecture
@@ -81,7 +81,7 @@ Drop a track on the window, or click to choose one. The stems appear in
 
 ### From another machine
 
-stemd is also a server. It announces itself over mDNS, so a client does not need
+Stemd is also a server. It announces itself over mDNS, so a client does not need
 to be told an address, and it speaks plain HTTP:
 
 ```bash
@@ -126,7 +126,7 @@ by the audio and by everything else that changes the output.
 
 ## Build from source
 
-Needs `cmake`, Rust 1.85 or later, and a recursive clone. Per-platform
+Needs `cmake`, Rust 1.89 or later, and a recursive clone. Per-platform
 toolchains, packaging and every flag are in
 [docs/building.md](docs/building.md).
 

crates/stemd-cli/src/discovery.rs

@@ -30,7 +30,7 @@ pub fn discover(timeout: Duration) -> Result<String> {
     let _ = daemon.shutdown();
     bail!(
         "no stemd server found on the network after {}s — is it running? \
-         start it with `open dist/stemd.app`, or pass --host host:port",
+         start it with `open dist/Stemd.app`, or pass --host host:port",
         timeout.as_secs()
     )
 }

crates/stemd-core/src/stemfmt.rs

@@ -131,9 +131,9 @@ impl StemFormat {
         match self {
             Self::Flac => "FLAC",
             Self::Wav => "WAV",
-            Self::Mp3 => "MP3 320",
-            Self::Pcm16 => "PCM 16",
-            Self::Pcm32 => "PCM 32",
+            Self::Mp3 => "MP3",
+            Self::Pcm16 => "PCM 16-bit",
+            Self::Pcm32 => "PCM 32-bit float",
         }
     }
 }

crates/stemd-mlx/src/device.rs

@@ -74,7 +74,7 @@ impl Accelerator {
         if !gpu {
             return (Self::Cpu, false);
         }
-        if !gpu_computes() {
+        if !gpu_computes_on_kept_thread() {
             return (Self::demote(), true);
         }
         let which = if cfg!(target_vendor = "apple") {
@@ -118,6 +118,24 @@ impl std::fmt::Display for Accelerator {
     }
 }
 
+/// [`gpu_computes`] on a thread that never exits.
+///
+/// MLX keeps CUDA state per thread, and on Windows its destructor throws when the
+/// thread exits. Windows reports that as a crash and freezes the whole process for
+/// about 3 s while it writes the report.
+fn gpu_computes_on_kept_thread() -> bool {
+    let (tx, rx) = std::sync::mpsc::channel();
+    let spawned = std::thread::Builder::new()
+        .name("stemd-gpu-probe".into())
+        .spawn(move || {
+            let _ = tx.send(gpu_computes());
+            loop {
+                std::thread::park();
+            }
+        });
+    spawned.is_ok() && rx.recv().unwrap_or(false)
+}
+
 /// Whether the GPU MLX offered returns right answers.
 ///
 /// One operation of each kind the models use, on inputs small enough that the

crates/stemd-server/Cargo.toml

@@ -22,7 +22,8 @@ clap = { version = "4.6.6", features = ["derive"] }
 tokio-util = { version = "0.7.19", features = ["io"] }
 parking_lot = "0.12.5"
 thiserror.workspace = true
-eframe = { version = "0.35.0", default-features = false, features = ["glow", "default_fonts"] }
+# `persistence` saves the window's position and size between launches.
+eframe = { version = "0.35.0", default-features = false, features = ["glow", "default_fonts", "persistence"] }
 if-addrs = "0.15.0"
 # First-run model fetch: HTTP plus a pinned digest.
 ureq = "3.4.0"
@@ -51,6 +52,12 @@ winit = { version = "0.30.13", default-features = false, features = ["x11"] }
 # download only there. See `cuda`.
 [target.'cfg(windows)'.dependencies]
 zip = { version = "2", default-features = false, features = ["deflate"] }
+# Free disk space, see `disk.rs`.
+windows-sys = { version = "0.61", features = ["Win32_Storage_FileSystem"] }
+
+[target.'cfg(unix)'.dependencies]
+# statvfs and sysctl, see `disk.rs`.
+libc = "0.2"
 
 # `default-features = false` above is what keeps wgpu and accesskit out, but it
 # also drops `x11` and `wayland`, and those are not extras — winit has no Linux
@@ -62,6 +69,7 @@ zip = { version = "2", default-features = false, features = ["deflate"] }
 eframe = { version = "0.35.0", default-features = false, features = [
     "glow",
     "default_fonts",
+    "persistence",
     "x11",
     "wayland",
 ] }