Kind gpu support in controller (#211)

I verified it's working by using the following server manifest:
```
apiVersion: substratus.ai/v1
kind: Server
metadata:
  name: falcon-7b-instruct
spec:
  build:
    git:
      url: https://github.com/substratusai/images
      path: model-server-basaran
      branch: basaran-4bit-mode
        #image: substratusai/model-server-basaran
  model:
    name: falcon-7b-instruct
  resources:
    gpu:
      count: 1
```

Note how I leave out the GPU type, which is desired because everyone
might be running different GPUs locally

Makefile

@@ -361,9 +361,8 @@ $(PROTOC): $(LOCALBIN)
 skaffold: $(SKAFFOLD)
 $(SKAFFOLD): $(LOCALBIN)
 	@ test -s $(LOCALBIN)/skaffold || \
-	curl -Lo skaffold https://storage.googleapis.com/skaffold/releases/latest/skaffold-$(SKAFFOLD_PLATFORM) && \
-	chmod +x skaffold && \
-	mv skaffold $(LOCALBIN)/skaffold
+	curl -Lo $(LOCALBIN)/skaffold https://storage.googleapis.com/skaffold/releases/latest/skaffold-$(SKAFFOLD_PLATFORM) && \
+	chmod +x $(LOCALBIN)/skaffold
 
 .PHONY: envsubst
 envsubst:

config/sci/deployment.yaml

@@ -18,7 +18,6 @@ spec:
       containers:
         - name: sci
           image: sci
-          imagePullPolicy: Always
           envFrom:
             - configMapRef:
                 name: system

internal/resources/gpu_info.go

@@ -11,6 +11,17 @@ type GPUInfo struct {
 	NodeSelector map[string]string
 }
 
+func GetGPUInfo(cloudName string, gpuType apiv1.GPUType) (*GPUInfo, bool) {
+	if cloudName == cloud.KindName {
+		return &GPUInfo{
+			ResourceName: corev1.ResourceName("nvidia.com/gpu"),
+			NodeSelector: map[string]string{},
+		}, true
+	}
+	gpuInfo, ok := cloudGPUs[cloudName][gpuType]
+	return gpuInfo, ok
+}
+
 var cloudGPUs = map[string]map[apiv1.GPUType]*GPUInfo{
 	cloud.GCPName: {
 		// https://cloud.google.com/compute/docs/gpus#nvidia_t4_gpus

internal/resources/resources.go

@@ -35,7 +35,7 @@ func Apply(podMetadata *metav1.ObjectMeta, podSpec *corev1.PodSpec, containerNam
 	resources.Requests[corev1.ResourceMemory] = *resource.NewQuantity(res.Memory*gigabyte, resource.BinarySI)
 
 	if res.GPU != nil {
-		gpuInfo, ok := cloudGPUs[cloudName][res.GPU.Type]
+		gpuInfo, ok := GetGPUInfo(cloudName, res.GPU.Type)
 		if !ok {
 			return fmt.Errorf("GPU %s is not supported on cloud %s", res.GPU.Type, cloudName)
 		}
@@ -47,9 +47,7 @@ func Apply(podMetadata *metav1.ObjectMeta, podSpec *corev1.PodSpec, containerNam
 			podSpec.NodeSelector = map[string]string{}
 		}
 
-		// TODO: Make spot configurable.
 		// TODO: Move this GCP code into cloud-specific configuration.
-		podSpec.NodeSelector["cloud.google.com/gke-spot"] = "true"
 		// Toleration is needed to trigger NAP
 		// https://cloud.google.com/kubernetes-engine/docs/how-to/node-auto-provisioning#support_for_spot_vms
 		podSpec.Tolerations = append(podSpec.Tolerations, corev1.Toleration{

skaffold.kind.yaml

@@ -32,3 +32,6 @@ build:
         dockerfile: Dockerfile.sci-kind
   local:
     push: false
+deploy:
+  kubectl:
+    defaultNamespace: substratus