fix(gpu): add Tegra/Jetson GPU support

crates/openshell-bootstrap/src/docker.rs

@@ -559,6 +559,19 @@ pub async fn ensure_container(
         }]);
     }
 
+    // On Tegra platforms the nvidia runtime or CDI spec generation reads
+    // host-file injection config from
+    // /etc/nvidia-container-runtime/host-files-for-container.d on the host.
+    // Bind-mount that directory (read-only) into the gateway so the same
+    // nvidia runtime or CDI spec generation running inside k3s (for sandbox
+    // pods) can apply the same config.
+    const HOST_FILES_DIR: &str = "/etc/nvidia-container-runtime/host-files-for-container.d";
+    if std::path::Path::new(HOST_FILES_DIR).is_dir() {
+        let mut binds = host_config.binds.take().unwrap_or_default();
+        binds.push(format!("{HOST_FILES_DIR}:{HOST_FILES_DIR}:ro"));
+        host_config.binds = Some(binds);
+    }
+
     let mut cmd = vec![
         "server".to_string(),
         "--disable=traefik".to_string(),

crates/openshell-sandbox/src/process.rs

@@ -414,7 +414,26 @@ pub fn drop_privileges(policy: &SandboxPolicy) -> Result<()> {
             target_os = "redox"
         )))]
         {
+            // Snapshot the container-level supplemental GIDs (e.g. injected by
+            // CDI for GPU device access) before initgroups replaces them.
+            // Exclude GID 0 (root) to avoid inadvertent privilege retention.
+            let root_gid = nix::unistd::Gid::from_raw(0);
+            let container_gids: Vec<nix::unistd::Gid> = nix::unistd::getgroups()
+                .unwrap_or_default()
+                .into_iter()
+                .filter(|&g| g != root_gid)
+                .collect();
             nix::unistd::initgroups(user_cstr.as_c_str(), group.gid).into_diagnostic()?;
+            // Merge back any CDI-injected GIDs that initgroups dropped so that
+            // exec'd processes retain access to GPU devices (e.g. /dev/nvmap on
+            // Tegra requires the video GID).
+            let mut merged: Vec<nix::unistd::Gid> = nix::unistd::getgroups().unwrap_or_default();
+            for gid in container_gids {
+                if !merged.contains(&gid) {
+                    merged.push(gid);
+                }
+            }
+            nix::unistd::setgroups(&merged).into_diagnostic()?;
         }
     }
 

deploy/kube/gpu-manifests/nvidia-device-plugin-helmchart.yaml

@@ -27,6 +27,9 @@ spec:
   targetNamespace: nvidia-device-plugin
   createNamespace: true
   valuesContent: |-
+    image:
+      repository: ghcr.io/nvidia/k8s-device-plugin
+      tag: "2ab68c16"
     runtimeClassName: nvidia
     gfd:
       enabled: false

e2e/python/test_sandbox_gpu.py

@@ -20,11 +20,13 @@ def test_gpu_sandbox_reports_available_gpu(
     sandbox: Callable[..., Sandbox],
     gpu_sandbox_spec: datamodel_pb2.SandboxSpec,
 ) -> None:
+    nvidia_smi_args = ["--query-gpu=name", "--format=csv,noheader"]
     with sandbox(spec=gpu_sandbox_spec, delete_on_exit=True) as sb:
-        result = sb.exec(
-            ["nvidia-smi", "--query-gpu=name", "--format=csv,noheader"],
-            timeout_seconds=30,
-        )
+        result = sb.exec(["nvidia-smi", *nvidia_smi_args], timeout_seconds=30)
+        if result.exit_code != 0:
+            # On some platforms (e.g. Tegra/Jetson) nvidia-smi lives in
+            # /usr/sbin rather than /usr/bin and may not be on PATH.
+            result = sb.exec(["/usr/sbin/nvidia-smi", *nvidia_smi_args], timeout_seconds=30)
 
         assert result.exit_code == 0, result.stderr
         assert result.stdout.strip()