[feat] added integration with OpenVINO Model Server

README.md

@@ -319,6 +319,45 @@ docker buildx build \
 
 The vLLM wheels are sourced from the official vLLM GitHub Releases at `https://github.com/vllm-project/vllm/releases`, which provides prebuilt wheels for each release version.
 
+### OVMS integration
+
+Docker Model Runner can also run an OVMS backend.
+
+- Default OVMS binary: resolved from `PATH` (looks up `ovms`)
+- Override binary path with: `OVMS_SERVER_PATH`
+
+OVMS can be installed based on this [guide](https://docs.openvino.ai/2026/model-server/ovms_docs_deploying_server_baremetal.html). Minimal version is 2026.2.
+
+When the runner starts, it will try to initialize OVMS as an available backend.
+
+```sh
+MODEL_RUNNER_PORT=13434 ./model-runner
+```
+
+Create a new model
+Use models from HugginFace Hub using repository with OpenVINO format. 
+```sh
+curl http://localhost:13434/models/create -X POST -d '{"from": "hf.co/OpenVINO/Qwen3-0.6B-int4-ov"}'
+```
+
+You can target OVMS explicitly through the backend-prefixed OpenAI-compatible routes:
+```sh
+# List models exposed via OVMS backend routing
+curl http://localhost:13434/engines/ovms/v1/models
+
+# Example chat/completions call through OVMS backend routing
+curl http://localhost:13434/engines/ovms/v1/chat/completions -X POST -d '{
+  "model": "hf.co/OpenVINO/Qwen3-0.6B-int4-ov",
+  "messages": [
+    {"role": "user", "content": "Hello from OVMS"}
+  ]
+}'
+```
+Delete model
+```sh
+curl http://localhost:13434/models/hf.co/OpenVINO/Qwen3-0.6B-int4-ov -X DELETE
+```
+
 ## API Examples
 
 The Model Runner exposes a REST API that can be accessed via TCP port. You can interact with it using curl commands.
@@ -335,7 +374,7 @@ curl http://localhost:8080/models
 curl http://localhost:8080/models/create -X POST -d '{"from": "ai/smollm2"}'
 
 # Get information about a specific model
-curl http://localhost:8080/models/ai/smollm2
+curl http://localhost:13434/models/hf.co/OpenVINO/Qwen3-0.6B-int4-ov
 
 # Chat with a model
 curl http://localhost:8080/engines/llama.cpp/v1/chat/completions -X POST -d '{
@@ -347,12 +386,11 @@ curl http://localhost:8080/engines/llama.cpp/v1/chat/completions -X POST -d '{
 }'
 
 # Delete a model
-curl http://localhost:8080/models/ai/smollm2 -X DELETE
+curl http://localhost:13434/models/hf.co/OpenVINO/Qwen3-0.6B-int4-ov -X DELETE
 
 # Get metrics
 curl http://localhost:8080/metrics
 ```
-
 The response will contain the model's reply:
 
 ```json

pkg/distribution/builder/from_directory.go

@@ -34,6 +34,10 @@ type DirectoryOptions struct {
 	// This is useful for producing deterministic OCI digests.
 	Created *time.Time
 
+	// AllowNoWeightFiles allows packaging a directory even when it contains no
+	// GGUF/SafeTensors/DDUF weight files.
+	AllowNoWeightFiles bool
+
 	// Format is the output artifact format. Defaults to BuildFormatDocker.
 	Format BuildFormat
 }
@@ -66,6 +70,15 @@ func WithCreatedTime(t time.Time) DirectoryOption {
 	}
 }
 
+// WithAllowNoWeightFiles allows FromDirectory to succeed even when no standard
+// model weight files are present. This is used for formats such as OpenVINO IR
+// where model files are represented differently (for example .xml + .bin pairs).
+func WithAllowNoWeightFiles() DirectoryOption {
+	return func(opts *DirectoryOptions) {
+		opts.AllowNoWeightFiles = true
+	}
+}
+
 // WithOutputFormat sets the output artifact format for the directory builder.
 // Defaults to BuildFormatDocker if not specified.
 // This is the DirectoryOption equivalent of WithFormat (BuildOption).
@@ -208,7 +221,7 @@ func FromDirectory(dirPath string, opts ...DirectoryOption) (*Builder, error) {
 		return nil, fmt.Errorf("no files found in directory: %s", dirPath)
 	}
 
-	if len(weightFiles) == 0 {
+	if len(weightFiles) == 0 && !options.AllowNoWeightFiles {
 		return nil, fmt.Errorf("no weight files (safetensors, GGUF, or DDUF) found in directory: %s", dirPath)
 	}
 

pkg/distribution/builder/from_directory_test.go

@@ -3,6 +3,7 @@ package builder
 import (
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
 	"time"
 )
@@ -170,6 +171,44 @@ func TestFromDirectoryWithExclusions(t *testing.T) {
 	}
 }
 
+func TestFromDirectoryNoStandardWeights(t *testing.T) {
+	tmpDir := t.TempDir()
+	createTestFile(t, tmpDir, "openvino/model.xml", "<net></net>")
+	createTestFile(t, tmpDir, "openvino/model.bin", "weights")
+	createTestFile(t, tmpDir, "openvino/config.json", "{}")
+
+	_, err := FromDirectory(tmpDir)
+	if err == nil {
+		t.Fatal("expected error when directory has no GGUF/SafeTensors/DDUF weights")
+	}
+
+	if got := err.Error(); got == "" || !strings.Contains(got, "no weight files") {
+		t.Fatalf("expected no weight files error, got %q", got)
+	}
+}
+
+func TestFromDirectoryAllowNoWeightFiles(t *testing.T) {
+	tmpDir := t.TempDir()
+	createTestFile(t, tmpDir, "openvino/model.xml", "<net></net>")
+	createTestFile(t, tmpDir, "openvino/model.bin", "weights")
+	createTestFile(t, tmpDir, "openvino/config.json", "{}")
+
+	b, err := FromDirectory(tmpDir, WithAllowNoWeightFiles())
+	if err != nil {
+		t.Fatalf("FromDirectory with WithAllowNoWeightFiles failed: %v", err)
+	}
+
+	mdl := b.Model()
+	layers, err := mdl.Layers()
+	if err != nil {
+		t.Fatalf("Failed to get layers: %v", err)
+	}
+
+	if len(layers) != 3 {
+		t.Errorf("Expected 3 layers, got %d", len(layers))
+	}
+}
+
 func TestShouldExclude(t *testing.T) {
 	tests := []struct {
 		name       string

pkg/distribution/huggingface/model.go

@@ -31,8 +31,9 @@ func BuildModel(ctx context.Context, client *Client, repo, revision, tag string,
 
 	// Filter to model files (weights + configs)
 	weightFiles, configFiles := FilterModelFiles(files)
+	isOpenVINORepo := IsOpenVINOModel(files)
 
-	if len(weightFiles) == 0 {
+	if len(weightFiles) == 0 && !isOpenVINORepo {
 		return nil, fmt.Errorf("no model weight files (GGUF or SafeTensors) found in repository %s", repo)
 	}
 
@@ -54,10 +55,20 @@ func BuildModel(ctx context.Context, client *Client, repo, revision, tag string,
 		}
 	}
 
-	// Combine all files to download
-	allFiles := append(weightFiles, configFiles...)
-	if mmprojFile != nil {
-		allFiles = append(allFiles, *mmprojFile)
+	// Combine all files to download.
+	// For OpenVINO repositories, pull all repository files so the full IR layout is preserved.
+	var allFiles []RepoFile
+	if isOpenVINORepo {
+		for _, f := range files {
+			if f.Type == "file" {
+				allFiles = append(allFiles, f)
+			}
+		}
+	} else {
+		allFiles = append(weightFiles, configFiles...)
+		if mmprojFile != nil {
+			allFiles = append(allFiles, *mmprojFile)
+		}
 	}
 
 	if progressWriter != nil {
@@ -91,7 +102,7 @@ func BuildModel(ctx context.Context, client *Client, repo, revision, tag string,
 	}
 
 	model, err := buildModelFromFiles(
-		result.LocalPaths, weightFiles, configFiles, mmprojFile, tempDir, createdTime,
+		result.LocalPaths, weightFiles, configFiles, mmprojFile, tempDir, createdTime, isOpenVINORepo,
 	)
 	if err != nil {
 		return nil, fmt.Errorf("build model: %w", err)
@@ -111,26 +122,30 @@ func buildModelFromFiles(
 	mmprojFile *RepoFile,
 	tempDir string,
 	createdTime *time.Time,
+	allowNoStandardWeights bool,
 ) (types.ModelArtifact, error) {
-	// Check if this is a safetensors model - use V0.2 packaging
-	if isSafetensorsModel(weightFiles) {
-		return buildSafetensorsModelV02(tempDir, createdTime)
+	// Safetensors and OpenVINO repos are packaged with V0.2 layer-per-file packaging.
+	if isSafetensorsModel(weightFiles) || allowNoStandardWeights {
+		return buildDirectoryModelV02(tempDir, createdTime, allowNoStandardWeights)
 	}
 
 	// For GGUF models, use V0.1 packaging (backward compatible)
 	return buildGGUFModelV01(localPaths, weightFiles, configFiles, mmprojFile, createdTime)
 }
 
-// buildSafetensorsModelV02 builds a safetensors model using V0.2 layer-per-file packaging.
+// buildDirectoryModelV02 builds a model using V0.2 layer-per-file packaging.
 // It uses builder.FromDirectory which recursively scans the tempDir and creates one layer
 // per file, preserving nested directory structure with filepath annotations.
 // If createdTime is non-nil, it is used as the creation timestamp for the OCI config
 // to produce deterministic digests. Otherwise time.Now() is used.
-func buildSafetensorsModelV02(tempDir string, createdTime *time.Time) (types.ModelArtifact, error) {
+func buildDirectoryModelV02(tempDir string, createdTime *time.Time, allowNoStandardWeights bool) (types.ModelArtifact, error) {
 	var dirOpts []builder.DirectoryOption
 	if createdTime != nil {
 		dirOpts = append(dirOpts, builder.WithCreatedTime(*createdTime))
 	}
+	if allowNoStandardWeights {
+		dirOpts = append(dirOpts, builder.WithAllowNoWeightFiles())
+	}
 
 	b, err := builder.FromDirectory(tempDir, dirOpts...)
 	if err != nil {

pkg/distribution/huggingface/repository.go

@@ -62,6 +62,40 @@ func FilterModelFiles(repoFiles []RepoFile) (weights []RepoFile, configs []RepoF
 	return weights, configs
 }
 
+// IsOpenVINOModel returns true when the repository contains at least one OpenVINO
+// IR weight pair (.xml + .bin with the same stem), including nested paths.
+func IsOpenVINOModel(repoFiles []RepoFile) bool {
+	xmlFiles := make(map[string]struct{})
+	binFiles := make(map[string]struct{})
+
+	for _, f := range repoFiles {
+		if f.Type != "file" {
+			continue
+		}
+
+		ext := strings.ToLower(path.Ext(f.Path))
+		if ext != ".xml" && ext != ".bin" {
+			continue
+		}
+
+		stem := strings.TrimSuffix(f.Path, path.Ext(f.Path))
+		switch ext {
+		case ".xml":
+			xmlFiles[stem] = struct{}{}
+		case ".bin":
+			binFiles[stem] = struct{}{}
+		}
+	}
+
+	for stem := range xmlFiles {
+		if _, ok := binFiles[stem]; ok {
+			return true
+		}
+	}
+
+	return false
+}
+
 // TotalSize calculates the total size of files
 func TotalSize(repoFiles []RepoFile) int64 {
 	var total int64

pkg/distribution/huggingface/repository_test.go

@@ -102,3 +102,53 @@ func TestIsSafetensorsModel(t *testing.T) {
 		})
 	}
 }
+
+func TestIsOpenVINOModel(t *testing.T) {
+	tests := []struct {
+		name  string
+		files []RepoFile
+		want  bool
+	}{
+		{
+			name: "matching xml/bin pair at root",
+			files: []RepoFile{
+				{Type: "file", Path: "openvino_model.xml"},
+				{Type: "file", Path: "openvino_model.bin"},
+			},
+			want: true,
+		},
+		{
+			name: "matching xml/bin pair in subdirectory",
+			files: []RepoFile{
+				{Type: "file", Path: "int4/openvino_model.xml"},
+				{Type: "file", Path: "int4/openvino_model.bin"},
+				{Type: "file", Path: "int4/config.json"},
+			},
+			want: true,
+		},
+		{
+			name: "xml without matching bin",
+			files: []RepoFile{
+				{Type: "file", Path: "openvino_model.xml"},
+				{Type: "file", Path: "other_model.bin"},
+			},
+			want: false,
+		},
+		{
+			name: "no openvino files",
+			files: []RepoFile{
+				{Type: "file", Path: "model.safetensors"},
+				{Type: "file", Path: "config.json"},
+			},
+			want: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := IsOpenVINOModel(tt.files); got != tt.want {
+				t.Errorf("IsOpenVINOModel() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}

pkg/envconfig/envconfig.go

@@ -152,6 +152,12 @@ func VLLMMetalServerPath() string {
 	return Var("VLLM_METAL_SERVER_PATH")
 }
 
+// OVMSServerPath returns the optional path to the OVMS server binary.
+// Configured via OVMS_SERVER_PATH.
+func OVMSServerPath() string {
+	return Var("OVMS_SERVER_PATH")
+}
+
 // LogDir returns the directory containing DMR log files.
 // Configured via MODEL_RUNNER_LOG_DIR. When empty, the server
 // auto-creates a default log directory so that the /logs API