Add support to docker runtime for OCI access to NVIDIA GPUs (#20959)

* Add support for --device=nvidia.com/gpu=all

* Add "nvidia.com" to valid options for gpus flag in TestValidateGPUs.

Author: aaron-nall

cmd/minikube/cmd/start.go

@@ -1479,8 +1479,8 @@ func validateGPUs(value, drvName, rtime string) error {
 	if err := validateGPUsArch(); err != nil {
 		return err
 	}
-	if value != "nvidia" && value != "all" && value != "amd" {
-		return errors.Errorf(`The gpus flag must be passed a value of "nvidia", "amd" or "all"`)
+	if value != "nvidia" && value != "all" && value != "amd" && value != "nvidia.com" {
+		return errors.Errorf(`The gpus flag must be passed a value of "nvidia", "nvidia.com", "amd" or "all"`)
 	}
 	if drvName == constants.Docker && (rtime == constants.Docker || rtime == constants.DefaultContainerRuntime) {
 		return nil

cmd/minikube/cmd/start_test.go

@@ -814,7 +814,7 @@ func TestValidateGPUs(t *testing.T) {
 		{"nvidia", "docker", "", ""},
 		{"all", "kvm", "docker", "The gpus flag can only be used with the docker driver and docker container-runtime"},
 		{"nvidia", "docker", "containerd", "The gpus flag can only be used with the docker driver and docker container-runtime"},
-		{"cat", "docker", "docker", `The gpus flag must be passed a value of "nvidia", "amd" or "all"`},
+		{"cat", "docker", "docker", `The gpus flag must be passed a value of "nvidia", "nvidia.com", "amd" or "all"`},
 		{"amd", "docker", "docker", ""},
 		{"amd", "docker", "", ""},
 		{"amd", "docker", "containerd", "The gpus flag can only be used with the docker driver and docker container-runtime"},

pkg/drivers/kic/oci/oci.go

@@ -194,6 +194,8 @@ func CreateContainerNode(p CreateParams) error { //nolint to suppress cyclomatic
 	switch p.GPUs {
 	case "all", "nvidia":
 		runArgs = append(runArgs, "--gpus", "all", "--env", "NVIDIA_DRIVER_CAPABILITIES=all")
+	case "nvidia.com":
+		runArgs = append(runArgs, "--device", "nvidia.com/gpu=all")
 	case "amd":
 		/* https://rocm.docs.amd.com/projects/install-on-linux/en/latest/how-to/docker.html
 		 * "--security-opt seccomp=unconfined" is also required but included above.

pkg/minikube/cruntime/docker.go

@@ -595,7 +595,7 @@ func (r *Docker) configureDocker(driver string) error {
 	}
 
 	switch r.GPUs {
-	case "all", "nvidia":
+	case "all", "nvidia", "nvidia.com":
 		assets.Addons["nvidia-device-plugin"].EnableByDefault()
 		daemonConfig.DefaultRuntime = "nvidia"
 		runtimes := &dockerDaemonRuntimes{}

site/content/en/docs/commands/start.md

@@ -57,7 +57,7 @@ minikube start [flags]
       --feature-gates string              A set of key=value pairs that describe feature gates for alpha/experimental features.
       --force                             Force minikube to perform possibly dangerous operations
       --force-systemd                     If set, force the container runtime to use systemd as cgroup manager. Defaults to false.
-  -g, --gpus string                       Allow pods to use your GPUs. Options include: [all,nvidia,amd] (Docker driver with Docker container-runtime only)
+  -g, --gpus string                       Allow pods to use your GPUs. Options include: [all,nvidia,amd,nvidia.com] (Docker driver with Docker container-runtime only)
       --ha                                Create Highly Available Multi-Control Plane Cluster with a minimum of three control-plane nodes that will also be marked for work.
       --host-dns-resolver                 Enable host resolver for NAT DNS requests (virtualbox driver only) (default true)
       --host-only-cidr string             The CIDR to be used for the minikube VM (virtualbox driver only) (default "192.168.59.1/24")

site/content/en/docs/tutorials/nvidia.md

@@ -30,7 +30,9 @@ date: 2018-01-02
   sudo sysctl -p
   ```
 
-- Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) on your host machine
+- Install NVIDIA support using one of:
+  - Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) on your host machine
+  - Enable [NVIDIA CDI resources](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html) on your host machine.
 
 - Configure Docker:
   ```shell
@@ -45,10 +47,15 @@ date: 2018-01-02
   ```
   This will make sure minikube does any required setup or addon installs now that the nvidia runtime is available.
 
-- Start minikube:
-  ```shell
-  minikube start --driver docker --container-runtime docker --gpus all
-  ```
+- Start minikube with one of:
+  - The NVIDIA Container Toolkit
+    ```shell
+    minikube start --driver docker --container-runtime docker --gpus all
+    ```
+  - NVIDIA CDI resources
+    ```shell
+    minikube start --driver docker --container-runtime docker --gpus nvidia.com
+    ```
 
 {{% /tab %}}
 {{% tab none %}}