Run on the CPU without a CUDA device

.github/scripts/cpu-separation.ps1

@@ -0,0 +1,39 @@
+# Separate a short generated mix on a machine without a GPU: the server starts,
+# runs on the CPU and returns stems.
+#
+#   cpu-separation.ps1 -Server stemd-server.exe -Cli stemd-cli.exe
+param(
+    [Parameter(Mandatory)] [string]$Server,
+    [Parameter(Mandatory)] [string]$Cli
+)
+$ErrorActionPreference = 'Stop'
+
+$work = Join-Path $env:RUNNER_TEMP 'cpu-separation'
+New-Item -ItemType Directory -Force -Path $work | Out-Null
+python (Join-Path $PSScriptRoot 'test-mix.py') "$work\mix.wav"
+if ($LASTEXITCODE -ne 0) { throw 'test-mix.py failed' }
+
+$out = "$work\server.out"
+$err = "$work\server.err"
+$proc = Start-Process $Server -PassThru -NoNewWindow `
+    -ArgumentList '--headless', '--no-mdns', '--bind', '127.0.0.1:7979', '--cache-dir', "$work\cache" `
+    -RedirectStandardOutput $out -RedirectStandardError $err
+function Get-Log { (Get-Content $out, $err -ErrorAction SilentlyContinue) -join "`n" }
+function Stop-WithLog([string]$why) {
+    Get-Log | Write-Host
+    if (-not $proc.HasExited) { Stop-Process $proc -Force }
+    throw $why
+}
+
+$deadline = (Get-Date).AddMinutes(10)
+while ((Get-Log) -notmatch 'listening') {
+    if ($proc.HasExited) { Stop-WithLog "the server exited with 0x$('{0:X8}' -f $proc.ExitCode)" }
+    if ((Get-Date) -gt $deadline) { Stop-WithLog 'the server is not listening after 10 minutes' }
+    Start-Sleep -Seconds 1
+}
+if ((Get-Log) -notmatch 'via mlx on cpu') { Stop-WithLog 'the server is not on the CPU' }
+
+& $Cli --host 127.0.0.1:7979 --out "$work\stems" "$work\mix.wav"
+if ($LASTEXITCODE -ne 0) { Stop-WithLog 'the separation failed' }
+Stop-Process $proc -Force
+Get-ChildItem "$work\stems"

.github/scripts/cpu-separation.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+# Separate a short generated mix on a machine without a GPU: the server starts,
+# runs on the CPU and returns stems.
+#
+#   cpu-separation.sh SERVER CLI
+set -euo pipefail
+server="$1"
+cli="$2"
+work="$(mktemp -d)"
+python3 "$(dirname "$0")/test-mix.py" "$work/mix.wav"
+
+"$server" --headless --no-mdns --bind 127.0.0.1:7979 --cache-dir "$work/cache" >"$work/server.log" 2>&1 &
+pid=$!
+trap 'kill "$pid" 2>/dev/null || true' EXIT
+fail() {
+    cat "$work/server.log"
+    echo "ERROR: $1" >&2
+    exit 1
+}
+for _ in $(seq 1 600); do
+    grep -q listening "$work/server.log" && break
+    kill -0 "$pid" 2>/dev/null || fail "the server exited"
+    sleep 1
+done
+grep -q listening "$work/server.log" || fail "the server is not listening after 10 minutes"
+grep -q "via mlx on cpu" "$work/server.log" || fail "the server is not on the CPU"
+"$cli" --host 127.0.0.1:7979 --out "$work/stems" "$work/mix.wav" || fail "the separation failed"
+ls -l "$work/stems"

.github/scripts/test-mix.py

@@ -0,0 +1,30 @@
+"""Write a 10 s stereo test mix at 44.1 kHz: a chord, a kick and some noise.
+
+Usage: test-mix.py OUT.wav
+"""
+
+import math
+import random
+import struct
+import sys
+import wave
+
+RATE = 44100
+SECONDS = 10
+
+random.seed(1)
+frames = bytearray()
+for i in range(RATE * SECONDS):
+    t = i / RATE
+    chord = 0.2 * sum(math.sin(2 * math.pi * f * t) for f in (220.0, 277.2, 329.6)) / 3
+    beat = t % 0.5
+    kick = 0.5 * math.sin(2 * math.pi * 60 * beat) * math.exp(-20 * beat)
+    left = chord + kick + 0.02 * random.uniform(-1, 1)
+    right = chord + kick + 0.02 * random.uniform(-1, 1)
+    frames += struct.pack("<hh", *(int(max(-1.0, min(1.0, v)) * 32767) for v in (left, right)))
+
+with wave.open(sys.argv[1], "wb") as out:
+    out.setnchannels(2)
+    out.setsampwidth(2)
+    out.setframerate(RATE)
+    out.writeframes(bytes(frames))

.github/workflows/cd.yml

@@ -145,6 +145,12 @@ jobs:
           appimage=$(ls dist/*.AppImage)
           (cd "$RUNNER_TEMP" && "$GITHUB_WORKSPACE/$appimage" --appimage-extract >/dev/null)
           LD_LIBRARY_PATH="$PWD/dist/stub" "$RUNNER_TEMP/squashfs-root/AppRun" --version
+      # The runner has no GPU, and the stub libcuda finds no device.
+      - name: CPU separation
+        timeout-minutes: 30
+        run: |
+          root="$RUNNER_TEMP/squashfs-root"
+          LD_LIBRARY_PATH="$PWD/dist/stub" .github/scripts/cpu-separation.sh "$root/AppRun" "$root/opt/stemd/bin/stemd-cli"
       - uses: actions/upload-artifact@v4
         with:
           name: linux-${{ matrix.arch }}
@@ -268,6 +274,12 @@ jobs:
           $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'
+          .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:
           name: windows-x86_64

crates/stemd-server/src/startup.rs

@@ -50,11 +50,17 @@ pub fn prepare(args: &Args, logs: LogBuffer, addr: SocketAddr) -> Result<Server>
     // loaded at CPU speed. A card whose runtime is missing looks exactly like
     // no card in every later line, and only one of the two has a fix.
     if stemd_core::Accelerator::gpu_refused() {
-        tracing::warn!(
-            "a GPU is installed but its runtime is not usable from here, so this \
-             runs on the CPU. Run with --install-cuda to fetch it (about 1.2 GB, \
-             once), or put the libraries beside the executable yourself."
-        );
+        if cfg!(windows) {
+            tracing::warn!(
+                "a GPU is installed but its runtime is not usable from here, so this \
+                 runs on the CPU. Run with --install-cuda to fetch it (about 1.2 GB, \
+                 once), or put the libraries beside the executable yourself."
+            );
+        } else {
+            tracing::warn!(
+                "a GPU is installed but failed a test computation, so this runs on the CPU."
+            );
+        }
     }
     let settings = Arc::new(open_settings(args)?);
     let (stems, cache_summary) = open_stem_cache(args)?;

docs/building.md

@@ -190,6 +190,10 @@ prebuilt copy instead.
 On Windows the prebuilt also carries OpenBLAS (`libopenblas.lib` and `.dll`)
 and `dl.lib`.
 
+`mlx-sys` patches MLX's CUDA allocator so a CUDA build also runs on the CPU of a
+machine without a CUDA device. CD checks this with a CPU separation on the Linux
+and Windows runners, which have no GPU.
+
 The manifest records the `mlx-sys` version, the MLX tag, the target, the C++
 compiler (only `msvc` on Windows, whose toolsets link older ones), nvcc,
 cuDNN, the GPU architectures and the RMSNorm patch, and `mlx-sys` refuses a

packaging/mlx-prebuilt.env

@@ -5,10 +5,10 @@
 # Changing an architecture list, the toolchain or the MLX pin means a new tag
 # and a new build. 90a rather than 90: MLX builds its Hopper-only kernels for
 # 90a alone. The SHA-256 lines are filled in from that workflow's summary.
-MLX_PREBUILT_TAG=mlx-cuda-2
+MLX_PREBUILT_TAG=mlx-cuda-3
 CUDA_ARCHS_linux_x86_64="80;86;89;90a;120"
 CUDA_ARCHS_linux_aarch64="90a;100;121"
 CUDA_ARCHS_windows_x86_64="80;86;89;90a;120"
-SHA256_linux_x86_64=2837bfdeecfa441c0c7d3bf9bd4974a3b04269e79bfbebecdb53806278d2722d
-SHA256_linux_aarch64=703eacbef4b6d54532417dcb455a743a2199b901c67b3630c3deeeb249b1068e
-SHA256_windows_x86_64=51cb5388c5abf5c128717a294d12fdf8b0530ac4cec4d19bbbce233642fae57f
+SHA256_linux_x86_64=0970ff131deb42cf9b18c9600f6de7c8a47513fe9db37eca0f4a88edf0208eb8
+SHA256_linux_aarch64=ead8f1a706cb1b0944f87289040bfbc6b9da1d8014bf5c5326d880d702fdf375
+SHA256_windows_x86_64=d7664ce18c250ae74efeed3b6a77405884e452fa608ae271d36827bd29161f76