Merge pull request #247 from kerthcet/feat/fungibility

Support hostpath models

api/core/v1alpha1/model_types.go

@@ -82,8 +82,10 @@ type ModelSource struct {
 	// ModelHub represents the model registry for model downloads.
 	// +optional
 	ModelHub *ModelHub `json:"modelHub,omitempty"`
-	// URI represents a various kinds of model sources following the uri protocol, e.g.:
-	// - OSS: oss://<bucket>.<endpoint>/<path-to-your-model>
+	// URI represents a various kinds of model sources following the uri protocol, protocol://<address>, e.g.
+	// - oss://<bucket>.<endpoint>/<path-to-your-model>
+	// - ollama://llama3.3
+	// - host://<path-to-your-model>
 	//
 	// +optional
 	URI *URIProtocol `json:"uri,omitempty"`

config/crd/bases/llmaz.io_openmodels.yaml

@@ -155,8 +155,10 @@ spec:
                     type: object
                   uri:
                     description: |-
-                      URI represents a various kinds of model sources following the uri protocol, e.g.:
-                      - OSS: oss://<bucket>.<endpoint>/<path-to-your-model>
+                      URI represents a various kinds of model sources following the uri protocol, protocol://<address>, e.g.
+                      - oss://<bucket>.<endpoint>/<path-to-your-model>
+                      - ollama://llama3.3
+                      - host://<path-to-your-model>
                     type: string
                 type: object
             required:

docs/examples/README.md

@@ -13,6 +13,7 @@ We provide a set of examples to help you serve large language models, by default
 - [Deploy models via ollama](#ollama)
 - [Speculative Decoding with vLLM](#speculative-decoding-with-vllm)
 - [Deploy multi-host inference](#multi-host-inference)
+- [Deploy host models](#deploy-host-models)
 
 ### Deploy models from Huggingface
 
@@ -59,3 +60,7 @@ By default, we use [vLLM](/~https://github.com/vllm-project/vllm) as the inference
 ### Multi-Host Inference
 
 Model size is growing bigger and bigger, Llama 3.1 405B FP16 LLM requires more than 750 GB GPU for weights only, leaving kv cache unconsidered, even with 8 x H100 Nvidia GPUs, 80 GB size of HBM each, can not fit in a single host, requires a multi-host deployment, see [example](./multi-nodes/) here.
+
+### Deploy Host Models
+
+Models could be loaded in prior to the hosts, especially those extremely big models, see [example](./hostpath/) to serve local models.

docs/examples/hostpath/model.yaml

@@ -0,0 +1,13 @@
+apiVersion: llmaz.io/v1alpha1
+kind: OpenModel
+metadata:
+  name: qwen2-0--5b-instruct
+spec:
+  familyName: qwen2
+  source:
+    uri: host:///workspace/Qwen2-0.5B-Instruct
+  inferenceConfig:
+    flavors:
+      - name: t4 # GPU type
+        requests:
+          nvidia.com/gpu: 1

docs/examples/hostpath/playground.yaml

@@ -0,0 +1,8 @@
+apiVersion: inference.llmaz.io/v1alpha1
+kind: Playground
+metadata:
+  name: qwen2-0--5b-instruct
+spec:
+  replicas: 1
+  modelClaim:
+    modelName: qwen2-0--5b-instruct

pkg/controller_helper/model_source/modelsource.go

@@ -71,13 +71,16 @@ func NewModelSourceProvider(model *coreapi.OpenModel) ModelSourceProvider {
 
 	if model.Spec.Source.URI != nil {
 		// We'll validate the format in the webhook, so generally no error should happen here.
-		protocol, address, _ := util.ParseURI(string(*model.Spec.Source.URI))
-		provider := &URIProvider{modelName: model.Name, protocol: protocol, modelAddress: address}
+		protocol, value, _ := util.ParseURI(string(*model.Spec.Source.URI))
+		provider := &URIProvider{modelName: model.Name, protocol: protocol}
 
 		switch protocol {
 		case OSS:
-			provider.endpoint, provider.bucket, provider.modelPath, _ = util.ParseOSS(address)
-		case OLLAMA:
+			provider.endpoint, provider.bucket, provider.modelPath, _ = util.ParseOSS(value)
+		case HostPath:
+			provider.modelPath = value
+		case Ollama:
+			provider.modelPath = value
 		default:
 			// This should be validated at webhooks.
 			panic("protocol not supported")

pkg/controller_helper/model_source/uri.go

@@ -26,22 +26,24 @@ import (
 var _ ModelSourceProvider = &URIProvider{}
 
 const (
-	OSS    = "OSS"
-	OLLAMA = "OLLAMA"
+	OSS      = "OSS"
+	Ollama   = "OLLAMA"
+	HostPath = "HOST"
 )
 
 type URIProvider struct {
-	modelName    string
-	protocol     string
-	bucket       string
-	endpoint     string
-	modelPath    string
-	modelAddress string
+	modelName string
+	protocol  string
+	bucket    string
+	endpoint  string
+	modelPath string
 }
 
 func (p *URIProvider) ModelName() string {
-	if p.protocol == OLLAMA {
-		return p.modelAddress
+	if p.protocol == Ollama {
+		// model path stores the ollama model name,
+		// the model name is the name of model CRD.
+		return p.modelPath
 	}
 	return p.modelName
 }
@@ -54,18 +56,51 @@ func (p *URIProvider) ModelName() string {
 //   - uri: bucket.endpoint/modelPath/model.gguf
 //     modelPath: /workspace/models/model.gguf
 func (p *URIProvider) ModelPath() string {
+	if p.protocol == HostPath {
+		return p.modelPath
+	}
+
+	// protocol is oss.
+
 	splits := strings.Split(p.modelPath, "/")
 
-	if strings.Contains(p.modelPath, ".") {
+	if strings.Contains(p.modelPath, ".gguf") {
 		return CONTAINER_MODEL_PATH + splits[len(splits)-1]
 	}
 	return CONTAINER_MODEL_PATH + "models--" + splits[len(splits)-1]
 }
 
 func (p *URIProvider) InjectModelLoader(template *corev1.PodTemplateSpec, index int) {
-	if p.protocol == OLLAMA {
+	// We don't have additional operations for Ollama, just load in runtime.
+	if p.protocol == Ollama {
 		return
 	}
+
+	if p.protocol == HostPath {
+		template.Spec.Volumes = append(template.Spec.Volumes, corev1.Volume{
+			Name: MODEL_VOLUME_NAME,
+			VolumeSource: corev1.VolumeSource{
+				HostPath: &corev1.HostPathVolumeSource{
+					Path: p.modelPath,
+				},
+			},
+		})
+
+		for i, container := range template.Spec.Containers {
+			// We only consider this container.
+			if container.Name == MODEL_RUNNER_CONTAINER_NAME {
+				template.Spec.Containers[i].VolumeMounts = append(template.Spec.Containers[i].VolumeMounts, corev1.VolumeMount{
+					Name:      MODEL_VOLUME_NAME,
+					MountPath: p.modelPath,
+					ReadOnly:  true,
+				})
+			}
+		}
+		return
+	}
+
+	// Other protocols.
+
 	initContainerName := MODEL_LOADER_CONTAINER_NAME
 	if index != 0 {
 		initContainerName += "-" + strconv.Itoa(index)

pkg/webhook/openmodel_webhook.go

@@ -47,8 +47,9 @@ func SetupOpenModelWebhook(mgr ctrl.Manager) error {
 var _ webhook.CustomDefaulter = &OpenModelWebhook{}
 
 var SUPPORTED_OBJ_STORES = map[string]struct{}{
-	modelSource.OSS:    {},
-	modelSource.OLLAMA: {},
+	modelSource.OSS:      {},
+	modelSource.Ollama:   {},
+	modelSource.HostPath: {},
 }
 
 // Default implements webhook.Defaulter so a webhook will be registered for the type

test/integration/webhook/model_test.go

@@ -105,6 +105,12 @@ var _ = ginkgo.Describe("model default and validation", func() {
 			},
 			failed: false,
 		}),
+		ginkgo.Entry("model creation with host protocol", &testValidatingCase{
+			model: func() *coreapi.OpenModel {
+				return wrapper.MakeModel("llama3-8b").FamilyName("llama3").ModelSourceWithURI("host:///models/meta-llama-3-8B").Obj()
+			},
+			failed: false,
+		}),
 		ginkgo.Entry("model creation with protocol unknown URI", &testValidatingCase{
 			model: func() *coreapi.OpenModel {
 				return wrapper.MakeModel("llama3-8b").FamilyName("llama3").ModelSourceWithURI("unknown://bucket.endpoint/models/meta-llama-3-8B").Obj()

test/util/wrapper/playground.go

@@ -160,7 +160,7 @@ func (w *PlaygroundWrapper) BackendRuntimeLimit(r, v string) *PlaygroundWrapper
 	return w
 }
 
-func (w *PlaygroundWrapper) ElasticConfig(maxReplicas, minReplicas int32) *PlaygroundWrapper {
+func (w *PlaygroundWrapper) ElasticConfig(minReplicas, maxReplicas int32) *PlaygroundWrapper {
 	w.Spec.ElasticConfig = &inferenceapi.ElasticConfig{
 		MaxReplicas: ptr.To[int32](maxReplicas),
 		MinReplicas: ptr.To[int32](minReplicas),