fix(mlx): strip file:// LocalPrefix before loading filesystem-imported models

MLX backends passed request.Model verbatim to mlx_lm/mlx_vlm load(). For a
model imported from the filesystem, LocalAI hands the backend a file:// URI
(its LocalPrefix), which load() rejects: the scheme is neither a valid HF
repo id nor an existing path (Path(model).exists() fails on the scheme),
producing "Repo id must be in the form 'repo_name' or 'namespace/repo_name'
... Use repo_type argument if needed".

Add a pure, unit-testable resolve_model_path(model, model_file) helper in the
shared python_utils: it prefers the resolved ModelFile, strips a file://
scheme and percent-decodes the path, and leaves plain repo ids and local
paths untouched. Wire it into the mlx, mlx-vlm and mlx-distributed backends
(load, model_key, and the distributed broadcast all use the normalized path).

Fixes #7461.

Assisted-by: claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

backend/python/common/python_utils.py

@@ -5,6 +5,31 @@ imported by any backend that needs to parse LocalAI gRPC options or build a
 chat-template-compatible message list from proto Message objects.
 """
 import json
+from urllib.parse import unquote
+
+
+def resolve_model_path(model, model_file=""):
+    """Resolve a LocalAI model reference to something an HF/MLX loader accepts.
+
+    LocalAI hands backends either a plain HuggingFace repo id
+    (``namespace/name``), an already-local filesystem path, or a
+    ``file://`` URI (its ``LocalPrefix``) for models imported from disk.
+    Loaders such as ``mlx_lm.load`` reject the ``file://`` form because the
+    scheme is neither a valid repo id nor an existing path, so we normalize
+    it here before loading.
+
+    Resolution order:
+      1. Prefer ``model_file`` when set and non-empty - that is the resolved
+         local path LocalAI computed for the model.
+      2. Strip a ``file://`` scheme and percent-decode it to a plain path.
+      3. Leave plain repo ids and already-local paths unchanged.
+    """
+    candidate = model_file if model_file else model
+    if candidate is None:
+        return candidate
+    if candidate.startswith("file://"):
+        return unquote(candidate[len("file://"):])
+    return candidate
 
 
 def parse_options(options_list):

backend/python/mlx-distributed/backend.py

@@ -28,7 +28,7 @@ import grpc
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'common'))
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'common'))
 from grpc_auth import get_auth_interceptors
-from python_utils import messages_to_dicts, parse_options as _shared_parse_options
+from python_utils import messages_to_dicts, parse_options as _shared_parse_options, resolve_model_path
 from mlx_utils import parse_tool_calls, split_reasoning
 
 
@@ -99,7 +99,11 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             from mlx_lm import load
             from mlx_lm.models.cache import make_prompt_cache, can_trim_prompt_cache, trim_prompt_cache
 
-            print(f"[Rank 0] Loading model: {request.Model}", file=sys.stderr)
+            # Normalize the model reference: strip LocalAI's file:// LocalPrefix
+            # and prefer the resolved ModelFile so mlx_lm.load() gets a plain
+            # repo id or filesystem path (it rejects file:// URIs).
+            model_path = resolve_model_path(request.Model, request.ModelFile)
+            print(f"[Rank 0] Loading model: {model_path}", file=sys.stderr)
 
             self.options = parse_options(request.Options)
             print(f"Options: {self.options}", file=sys.stderr)
@@ -128,7 +132,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 )
                 self.coordinator = DistributedCoordinator(self.group)
                 self.coordinator.broadcast_command(CMD_LOAD_MODEL)
-                self.coordinator.broadcast_model_name(request.Model)
+                self.coordinator.broadcast_model_name(model_path)
             else:
                 print("[Rank 0] No hostfile configured, running single-node", file=sys.stderr)
 
@@ -144,9 +148,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
             if tokenizer_config:
                 print(f"Loading with tokenizer_config: {tokenizer_config}", file=sys.stderr)
-                self.model, self.tokenizer = load(request.Model, tokenizer_config=tokenizer_config)
+                self.model, self.tokenizer = load(model_path, tokenizer_config=tokenizer_config)
             else:
-                self.model, self.tokenizer = load(request.Model)
+                self.model, self.tokenizer = load(model_path)
 
             if self.group is not None:
                 from sharding import pipeline_auto_parallel
@@ -157,7 +161,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 from mlx_cache import ThreadSafeLRUPromptCache
                 max_cache_entries = self.options.get("max_cache_entries", 10)
                 self.max_kv_size = self.options.get("max_kv_size", None)
-                self.model_key = request.Model
+                self.model_key = model_path
                 self.lru_cache = ThreadSafeLRUPromptCache(
                     max_size=max_cache_entries,
                     can_trim_fn=can_trim_prompt_cache,

backend/python/mlx-vlm/backend.py

@@ -18,7 +18,7 @@ import grpc
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'common'))
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'common'))
 from grpc_auth import get_auth_interceptors
-from python_utils import messages_to_dicts, parse_options
+from python_utils import messages_to_dicts, parse_options, resolve_model_path
 from mlx_utils import parse_tool_calls, split_reasoning
 
 from mlx_vlm import load, stream_generate
@@ -67,7 +67,11 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             backend_pb2.Result: The load model result.
         """
         try:
-            print(f"Loading MLX-VLM model: {request.Model}", file=sys.stderr)
+            # Normalize the model reference: strip LocalAI's file:// LocalPrefix
+            # and prefer the resolved ModelFile so mlx_vlm.load() gets a plain
+            # repo id or filesystem path (it rejects file:// URIs).
+            model_path = resolve_model_path(request.Model, request.ModelFile)
+            print(f"Loading MLX-VLM model: {model_path}", file=sys.stderr)
             print(f"Request: {request}", file=sys.stderr)
 
             # Parse Options[] key:value strings into a typed dict
@@ -76,10 +80,10 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
             # Load model and processor using MLX-VLM
             # mlx-vlm load function returns (model, processor) instead of (model, tokenizer)
-            self.model, self.processor = load(request.Model)
+            self.model, self.processor = load(model_path)
 
             # Load model config for chat template support
-            self.config = load_config(request.Model)
+            self.config = load_config(model_path)
 
             # Auto-infer the tool parser from the chat template. mlx-vlm has
             # its own _infer_tool_parser that falls back to mlx-lm parsers.

backend/python/mlx/backend.py

@@ -17,7 +17,7 @@ import grpc
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'common'))
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'common'))
 from grpc_auth import get_auth_interceptors
-from python_utils import messages_to_dicts, parse_options
+from python_utils import messages_to_dicts, parse_options, resolve_model_path
 from mlx_utils import parse_tool_calls, split_reasoning
 
 from mlx_lm import load, stream_generate
@@ -63,7 +63,11 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             backend_pb2.Result: The load model result.
         """
         try:
-            print(f"Loading MLX model: {request.Model}", file=sys.stderr)
+            # Normalize the model reference: strip LocalAI's file:// LocalPrefix
+            # and prefer the resolved ModelFile so mlx_lm.load() gets a plain
+            # repo id or filesystem path (it rejects file:// URIs).
+            model_path = resolve_model_path(request.Model, request.ModelFile)
+            print(f"Loading MLX model: {model_path}", file=sys.stderr)
             print(f"Request: {request}", file=sys.stderr)
 
             # Parse Options[] key:value strings into a typed dict (shared helper)
@@ -89,9 +93,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             # Load model and tokenizer using MLX
             if tokenizer_config:
                 print(f"Loading with tokenizer_config: {tokenizer_config}", file=sys.stderr)
-                self.model, self.tokenizer = load(request.Model, tokenizer_config=tokenizer_config)
+                self.model, self.tokenizer = load(model_path, tokenizer_config=tokenizer_config)
             else:
-                self.model, self.tokenizer = load(request.Model)
+                self.model, self.tokenizer = load(model_path)
 
             # mlx_lm.load() returns a TokenizerWrapper that detects tool
             # calling and thinking markers from the chat template / vocab.
@@ -111,7 +115,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             # Initialize thread-safe LRU prompt cache for efficient generation
             max_cache_entries = self.options.get("max_cache_entries", 10)
             self.max_kv_size = self.options.get("max_kv_size", None)
-            self.model_key = request.Model
+            self.model_key = model_path
             self.lru_cache = ThreadSafeLRUPromptCache(
                 max_size=max_cache_entries,
                 can_trim_fn=can_trim_prompt_cache,

backend/python/mlx/test.py

@@ -12,7 +12,7 @@ import backend_pb2_grpc
 # Make the shared helpers importable so we can unit-test them without a
 # running gRPC server.
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'common'))
-from python_utils import messages_to_dicts, parse_options
+from python_utils import messages_to_dicts, parse_options, resolve_model_path
 from mlx_utils import parse_tool_calls, split_reasoning
 
 class TestBackendServicer(unittest.TestCase):
@@ -322,6 +322,42 @@ class TestSharedHelpers(unittest.TestCase):
         self.assertEqual(r, "")
         self.assertEqual(c, "just text")
 
+    def test_resolve_model_path_file_uri(self):
+        # file:// LocalPrefix (LocalAI import) is stripped to a plain path.
+        self.assertEqual(resolve_model_path("file:///a/b"), "/a/b")
+
+    def test_resolve_model_path_file_uri_percent_decoded(self):
+        # Percent-encoded characters (e.g. spaces) are decoded.
+        self.assertEqual(
+            resolve_model_path("file:///Users/me/My%20Models/Qwen3"),
+            "/Users/me/My Models/Qwen3",
+        )
+
+    def test_resolve_model_path_hf_repo_id_unchanged(self):
+        # Plain HuggingFace repo ids must pass through untouched.
+        self.assertEqual(
+            resolve_model_path("mlx-community/Qwen3-Coder-30B"),
+            "mlx-community/Qwen3-Coder-30B",
+        )
+
+    def test_resolve_model_path_local_path_unchanged(self):
+        # An already-local absolute path is left as-is.
+        self.assertEqual(resolve_model_path("/models/Qwen3"), "/models/Qwen3")
+
+    def test_resolve_model_path_prefers_model_file(self):
+        # The resolved ModelFile wins over Model when both are set.
+        self.assertEqual(
+            resolve_model_path("file:///ignored", "/resolved/local/path"),
+            "/resolved/local/path",
+        )
+
+    def test_resolve_model_path_model_file_file_uri(self):
+        # A ModelFile that is itself a file:// URI is also normalized.
+        self.assertEqual(
+            resolve_model_path("ignored", "file:///a/b"),
+            "/a/b",
+        )
+
     def test_parse_tool_calls_with_shim(self):
         tm = types.SimpleNamespace(
             tool_call_start="<tool_call>",

core/http/endpoints/openai/realtime_webrtc.go

@@ -2,8 +2,6 @@ package openai
 
 import (
 	"net/http"
-	"runtime"
-	"strings"
 	"time"
 
 	"github.com/labstack/echo/v4"
@@ -13,52 +11,6 @@ import (
 	"github.com/pion/webrtc/v4"
 )
 
-// opusBackendName is the canonical gallery name/alias of the opus audio codec
-// backend that the realtime WebRTC transport needs.
-const opusBackendName = "opus"
-
-// resolveOpusBackend picks which installed opus-codec backend the realtime
-// WebRTC transport should load. The transport historically hardcoded the
-// literal "opus" backend name, but on darwin/arm64 the only installable opus
-// codec is "metal-opus" (it shares the gallery alias "opus"). When that
-// platform-specific variant is registered under its concrete directory name
-// rather than the "opus" alias key, loading the literal "opus" fails with
-// "opus backend not available" (issue #9813). Given the set of currently
-// loadable backend names, this returns the best opus codec to load for the
-// running platform, falling back to the literal name so the caller surfaces
-// the same error as before when no opus codec is installed at all.
-func resolveOpusBackend(installed []string, goos, goarch string) string {
-	// An exact match wins: this covers the plain "opus" backend as well as the
-	// "opus" alias key registered by gallery alias resolution for a
-	// user-installed platform variant.
-	for _, b := range installed {
-		if b == opusBackendName {
-			return opusBackendName
-		}
-	}
-
-	// No "opus" key is registered (e.g. a system-path metal-opus whose alias
-	// was never collected). Fall back to a platform-appropriate "*opus*" codec
-	// backend; on darwin/arm64 prefer the metal build.
-	var fallback string
-	for _, b := range installed {
-		if !strings.Contains(strings.ToLower(b), opusBackendName) {
-			continue
-		}
-		if goos == "darwin" && goarch == "arm64" && strings.Contains(strings.ToLower(b), "metal") {
-			return b
-		}
-		if fallback == "" {
-			fallback = b
-		}
-	}
-	if fallback != "" {
-		return fallback
-	}
-
-	return opusBackendName
-}
-
 // RealtimeCallRequest is the JSON body for POST /v1/realtime/calls.
 type RealtimeCallRequest struct {
 	SDP   string `json:"sdp"`
@@ -142,25 +94,15 @@ func RealtimeCalls(application *application.Application) echo.HandlerFunc {
 			}
 		}()
 
-		// Load the Opus backend. The opus codec ships under different backend
-		// names per platform (e.g. "metal-opus" on darwin/arm64), so resolve the
-		// platform-appropriate variant from the installed backends instead of
-		// hardcoding the literal "opus" name (issue #9813).
-		ml := application.ModelLoader()
-		installed := make([]string, 0)
-		for name := range ml.GetAllExternalBackends(nil) {
-			installed = append(installed, name)
-		}
-		opusName := resolveOpusBackend(installed, runtime.GOOS, runtime.GOARCH)
-
-		opusBackend, err := ml.Load(
-			model.WithBackendString(opusName),
+		// Load the Opus backend
+		opusBackend, err := application.ModelLoader().Load(
+			model.WithBackendString("opus"),
 			model.WithModelID("__opus_codec__"),
-			model.WithModel(opusName),
+			model.WithModel("opus"),
 		)
 		if err != nil {
 			pc.Close()
-			xlog.Error("failed to load opus backend", "error", err, "backend", opusName)
+			xlog.Error("failed to load opus backend", "error", err)
 			return c.JSON(http.StatusInternalServerError, map[string]string{"error": "opus backend not available"})
 		}
 

core/http/endpoints/openai/realtime_webrtc_opus_test.go

@@ -1,32 +0,0 @@
-package openai
-
-import (
-	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
-)
-
-var _ = Describe("resolveOpusBackend", func() {
-	It("prefers the exact opus backend when it is installed", func() {
-		Expect(resolveOpusBackend([]string{"opus", "metal-opus"}, "linux", "amd64")).To(Equal("opus"))
-	})
-
-	It("resolves to the opus alias key on linux", func() {
-		Expect(resolveOpusBackend([]string{"opus"}, "linux", "amd64")).To(Equal("opus"))
-	})
-
-	It("selects metal-opus on darwin/arm64 when no plain opus is installed", func() {
-		Expect(resolveOpusBackend([]string{"metal-opus"}, "darwin", "arm64")).To(Equal("metal-opus"))
-	})
-
-	It("selects metal-opus on darwin/arm64 even when other backends are present", func() {
-		Expect(resolveOpusBackend([]string{"silero-vad", "metal-opus", "whisper"}, "darwin", "arm64")).To(Equal("metal-opus"))
-	})
-
-	It("falls back to any opus codec backend when there is no exact match (non-darwin)", func() {
-		Expect(resolveOpusBackend([]string{"metal-opus"}, "linux", "amd64")).To(Equal("metal-opus"))
-	})
-
-	It("returns the literal opus name when no opus codec is installed", func() {
-		Expect(resolveOpusBackend([]string{"silero-vad", "whisper"}, "darwin", "arm64")).To(Equal("opus"))
-	})
-})