fix(podman): discover CDI GPUs from server info

Signed-off-by: Evan Lezar <elezar@nvidia.com>

Author: elezar

crates/openshell-driver-podman/README.md

@@ -149,3 +149,21 @@ authenticated health signal.
 
 Gateway OTLP configuration continues to export compute-driver spans under the
 `openshell-driver-podman` service, preserving gateway trace context.
+
+## CDI GPU discovery
+
+The driver reads `host.cdiSpecDirs` and `host.discoveredDevices` from
+`GET /v5.0.0/libpod/info` during initialization and before each default GPU
+preflight or create selection. Only devices with `source: "cdi"` enter the CDI
+inventory. The shared GPU selector retains its NVIDIA class and count policy;
+collecting other CDI IDs does not add GPU workload support for other vendors.
+
+Presence of either CDI field means server discovery is authoritative. Podman
+can omit `discoveredDevices` when its inventory is empty while still reporting
+`cdiSpecDirs`; this yields an empty inventory rather than probing local devices.
+When both fields are absent, legacy discovery scans local `/dev/nvidia<N>`
+paths and retains the `/dev/dxg` all-device fallback. Spec directories are
+server paths for diagnostics, not directories to scan on the gateway host.
+Refresh API or decoding failures fail the default GPU request without updating
+the selector or selecting from stale inventory. Explicit CDI selectors continue
+to be validated and passed to Podman without inventory lookup.

crates/openshell-driver-podman/src/client.rs

@@ -311,6 +311,11 @@ pub struct SystemInfo {
 #[derive(Debug, Clone, serde::Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct HostInfo {
+    /// Presence marks servers that expose CDI discovery, including empty inventories.
+    #[serde(default)]
+    pub cdi_spec_dirs: Option<Vec<String>>,
+    #[serde(default)]
+    pub discovered_devices: Option<Vec<DiscoveredDevice>>,
     #[serde(default)]
     pub cgroup_version: String,
     #[serde(default)]
@@ -321,6 +326,13 @@ pub struct HostInfo {
     pub security: SecurityInfo,
 }
 
+/// A device discovered by the Podman server, independent of GPU selection policy.
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct DiscoveredDevice {
+    pub source: String,
+    pub id: String,
+}
+
 /// Security-related fields from the Podman system info response.
 ///
 /// Podman returns `host.security.rootless: true` when the daemon is
@@ -1158,6 +1170,41 @@ mod tests {
         assert!(validate_name(&exact_name).is_ok());
     }
 
+    #[tokio::test]
+    async fn system_info_retains_cdi_inventory_and_spec_directories() {
+        let (socket, requests, handle) = spawn_podman_stub(
+            "info-cdi",
+            vec![
+                StubResponse::new(
+                    StatusCode::OK,
+                    r#"{"host":{"cdiSpecDirs":["/etc/cdi","/var/run/cdi"],"discoveredDevices":[{"source":"cdi","id":"example.com/gpu=test0"}]}}"#,
+                ),
+                StubResponse::new(StatusCode::OK, r#"{"host":{"cdiSpecDirs":["/etc/cdi"]}}"#),
+                StubResponse::new(StatusCode::OK, r#"{"host":{}}"#),
+            ],
+        );
+        let client = PodmanClient::new(socket);
+        let host = client.system_info().await.unwrap().host;
+        assert_eq!(
+            host.cdi_spec_dirs.unwrap(),
+            vec!["/etc/cdi", "/var/run/cdi"]
+        );
+        let devices = host.discovered_devices.unwrap();
+        assert_eq!(devices[0].source, "cdi");
+        assert_eq!(devices[0].id, "example.com/gpu=test0");
+        let empty = client.system_info().await.unwrap().host;
+        assert!(empty.cdi_spec_dirs.is_some());
+        assert!(empty.discovered_devices.is_none());
+        let legacy = client.system_info().await.unwrap().host;
+        assert!(legacy.cdi_spec_dirs.is_none());
+        assert!(legacy.discovered_devices.is_none());
+        handle.await.unwrap();
+        assert_eq!(
+            *requests.lock().unwrap(),
+            vec!["GET /v5.0.0/libpod/info"; 3]
+        );
+    }
+
     #[test]
     fn system_info_parses_rootless_network_helper() {
         let info: SystemInfo = serde_json::from_str(