Search the user's own runtime directory before the system library

The reference desktop's only system ONNX Runtime is Arch's
onnxruntime-opt-cuda: 1.29, built without TensorRT and against cuDNN 8
on a cuDNN 9 machine. The probe rejects both providers correctly and
the app runs on the CPU provider, which is right and not what anyone
wants. runtime/ beside the models is now searched ahead of /usr/lib,
tools/fetch-desktop-runtime.sh fills it with the four libraries from
the current onnxruntime-gpu wheel (cuDNN 9, TensorRT 10), and the
About caption lists every rung that lost and why, not only the first.
Verified: the app selects TensorRT from that directory with no
environment variable set.

apps/darkroom-desktop/src/main.rs

@@ -83,10 +83,14 @@ fn main() -> anyhow::Result<()> {
 /// `DARKROOM_ORT_DIR` is for a developer pointing at a runtime that is not
 /// installed — the wheel's `capi` directory, say. Then beside the executable
 /// and in the package's private library directory, for a package that
-/// bundles its own; then the Flatpak prefix; then the system library
-/// directory, for a distribution that ships ONNX Runtime as a package of its
-/// own. A system copy whose GPU providers do not load is not a problem: the
-/// probe builds a real session before believing a provider.
+/// bundles its own; then the user's own `runtime/` beside the models, where
+/// `tools/fetch-desktop-runtime.sh` puts one; then the Flatpak prefix; then
+/// the system library directory, for a distribution that ships ONNX Runtime
+/// as a package of its own. The user's copy outranks the system's because
+/// the system's is the one most likely to be built without the GPU
+/// providers, or against the wrong cuDNN — and a system copy whose providers
+/// do not load is not a problem, only a slower app: the probe builds a real
+/// session before believing a provider.
 fn runtime_dirs() -> Vec<PathBuf> {
     let mut dirs = Vec::new();
     if let Some(dir) = std::env::var_os("DARKROOM_ORT_DIR") {
@@ -98,6 +102,7 @@ fn runtime_dirs() -> Vec<PathBuf> {
             dirs.push(bin.join("../lib/darkroom"));
         }
     }
+    dirs.push(dr_ui::inference::user_runtime_dir());
     #[cfg(target_os = "linux")]
     dirs.extend([
         PathBuf::from("/app/lib/darkroom"),

core/dr-inference-engine/src/lib.rs

@@ -155,6 +155,8 @@ pub struct Status {
     /// Engines compiled and engines wanted, for a compiling rung; `(0, 0)`
     /// otherwise.
     pub engines: (usize, usize),
+    /// Every rung above the selected one that was tried, and why it lost.
+    pub failed: Vec<(Rung, String)>,
 }
 
 impl Status {
@@ -399,6 +401,7 @@ pub fn status() -> Status {
         runtime: api::runtime(),
         rung,
         reason: s.cache.reason.clone(),
+        failed: s.cache.failed.clone(),
         probing: s.probing,
         engines: if rung.compiles() {
             (s.cache.compiled.len(), s.wanted)

tools/fetch-desktop-runtime.sh

@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# Put a GPU-capable ONNX Runtime where the desktop app looks for one
+# (docs/inference.md §3): `runtime/` beside the models in the user data
+# directory, ahead of the system library.
+#
+#   ./tools/fetch-desktop-runtime.sh [DEST]
+#
+# The source is the `onnxruntime-gpu` wheel: the one build that carries the
+# CUDA *and* TensorRT providers against the cuDNN and TensorRT majors current
+# on this machine. Distribution packages tend to have neither — Arch's
+# `onnxruntime-opt-cuda` is built without TensorRT and against cuDNN 8 — and
+# the probe rejects them correctly and leaves the app on the CPU provider,
+# which is what this script exists to fix. Nothing NVIDIA is bundled here:
+# the providers load CUDA, cuDNN and TensorRT from the system, and if those
+# are missing the probe says so and the app stays on the CPU.
+set -euo pipefail
+DEST="${1:-${XDG_DATA_HOME:-${HOME}/.local/share}/darkroom/runtime}"
+WORK="$(mktemp -d -p /var/tmp fetch-desktop-runtime.XXXXXX)"
+trap 'rm -rf "${WORK}"' EXIT
+
+echo "==> downloading the onnxruntime-gpu wheel"
+uv venv --python 3.12 "${WORK}/venv" >/dev/null
+VIRTUAL_ENV="${WORK}/venv" uv pip install --quiet onnxruntime-gpu
+CAPI="$(find "${WORK}/venv" -type d -path '*/onnxruntime/capi' | head -1)"
+[[ -n "${CAPI}" ]] || { echo "error: no capi directory in the wheel" >&2; exit 1; }
+
+mkdir -p "${DEST}"
+# The runtime and its provider libraries; not the Python binding.
+cp "${CAPI}"/libonnxruntime.so* "${CAPI}"/libonnxruntime_providers_*.so "${DEST}/"
+echo "==> runtime in ${DEST}:"
+ls -1 "${DEST}" | sed 's/^/    /'
+echo "    (the app finds it on its next launch; Settings › About › Inference says what it chose)"

ui/dr-ui/src/inference.rs

@@ -60,6 +60,18 @@ pub fn init(runtime_dirs: Vec<PathBuf>) {
     crate::memory::evict_at(crate::memory::Tier::Gpu, dr_inference_engine::release_all);
 }
 
+/// Where a person can put a runtime by hand: `runtime/` beside the models,
+/// searched before any system library. The system copy on the reference
+/// desktop is built without TensorRT and against the wrong cuDNN, and a
+/// working one is four files from the `onnxruntime-gpu` wheel; this is
+/// where they go, and `tools/fetch-desktop-runtime.sh` puts them there.
+pub fn user_runtime_dir() -> PathBuf {
+    crate::library::shared_face_models_dir()
+        .parent()
+        .map(|p| p.join("runtime"))
+        .unwrap_or_else(|| PathBuf::from("runtime"))
+}
+
 /// Which form the current backend loads `detector` in, given the files on
 /// this device — the fact `faces.model_id` has to carry (§7).
 ///
@@ -94,8 +106,18 @@ pub fn about_lines() -> (String, String) {
             status.engines.0,
             status.engines.1
         )
-    } else {
+    } else if status.failed.is_empty() {
         status.reason
+    } else {
+        // Every rung that was tried and why it lost, not only the first:
+        // "TensorRT: not enabled in this build" says nothing about why CUDA
+        // was not taken instead.
+        status
+            .failed
+            .iter()
+            .map(|(rung, why)| format!("{}: {why}", rung.label()))
+            .collect::<Vec<_>>()
+            .join(" · ")
     };
     (line, detail)
 }