fix(distributed): order derived usecases deterministically

syncKnownUsecasesFromString rebuilds KnownUsecaseStrings by ranging
GetAllModelConfigUsecases, which is a map. Go randomizes that order per
call, and the field is part of the serialized config, so one unchanged
YAML hashed to a different config revision on every load.

A model that derives a single usecase hid the problem. One that derives
several, such as a chat model with an mmproj, alternated between as many
revisions as there are orderings. The router treats a revision it did
not establish as a config change, so requests failed with "stale model
config revision" until the stored value happened to match again.

Sorting the list makes the revision a function of the file alone.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-opus-5 [golangci-lint]

core/config/model_config.go

@@ -1369,6 +1369,12 @@ func (c *ModelConfig) syncKnownUsecasesFromString() {
 			c.KnownUsecaseStrings = append(c.KnownUsecaseStrings, k)
 		}
 	}
+	// GetAllModelConfigUsecases returns a map, and ranging one yields a random
+	// order per call. KnownUsecaseStrings is part of the serialized config, so
+	// an unsorted list gives the same file a different config revision on every
+	// load. In distributed mode that reads as a config change and the router
+	// rejects the request with ErrStaleModelConfigRevision.
+	slices.Sort(c.KnownUsecaseStrings)
 }
 
 func (c *ModelConfig) UnmarshalYAML(value *yaml.Node) error {

core/config/model_config_revision_stability_test.go

@@ -0,0 +1,90 @@
+package config_test
+
+import (
+	"os"
+	"path/filepath"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/system"
+)
+
+// The distributed controller pins a model's replicas to its config revision and
+// rejects any request carrying a different one. A revision that is not stable
+// for one unchanged file on disk therefore wedges the model.
+var _ = Describe("Model config revision stability", func() {
+	// A chat model with an mmproj derives two usecase flags, FLAG_CHAT and
+	// FLAG_VISION. syncKnownUsecasesFromString builds that list by ranging a
+	// map, so an unstable order shows up with two or more flags and stays
+	// hidden with one.
+	const multiUsecaseModel = `backend: llama-cpp
+context_size: 50000
+known_usecases:
+    - chat
+mmproj: llama-cpp/mmproj/example/mmproj.gguf
+name: example
+options:
+    - use_jinja:true
+    - parallel:2
+parameters:
+    model: llama-cpp/models/example/example.gguf
+template:
+    use_tokenizer_template: true
+`
+
+	var (
+		dir       string
+		appConfig *config.ApplicationConfig
+	)
+
+	BeforeEach(func() {
+		dir = GinkgoT().TempDir()
+		Expect(os.WriteFile(filepath.Join(dir, "example.yaml"), []byte(multiUsecaseModel), 0o600)).To(Succeed())
+		appConfig = config.NewApplicationConfig()
+		appConfig.SystemState = &system.SystemState{Model: system.Model{ModelsPath: dir}}
+	})
+
+	loadRevision := func() string {
+		loader := config.NewModelConfigLoader(dir)
+		Expect(loader.LoadModelConfigsFromPath(dir, appConfig.ToConfigLoaderOptions()...)).To(Succeed())
+		cfg, ok := loader.GetModelConfig("example")
+		Expect(ok).To(BeTrue())
+		revision, err := config.ModelConfigRevision(&cfg)
+		Expect(err).ToNot(HaveOccurred())
+		return revision
+	}
+
+	It("does not change when the same file is loaded repeatedly", func() {
+		baseline := loadRevision()
+		for i := 0; i < 20; i++ {
+			Expect(loadRevision()).To(Equal(baseline), "revision changed between two loads of one unchanged file")
+		}
+	})
+
+	It("orders the derived usecases deterministically", func() {
+		loader := config.NewModelConfigLoader(dir)
+		Expect(loader.LoadModelConfigsFromPath(dir, appConfig.ToConfigLoaderOptions()...)).To(Succeed())
+		cfg, ok := loader.GetModelConfig("example")
+		Expect(ok).To(BeTrue())
+		Expect(len(cfg.KnownUsecaseStrings)).To(BeNumerically(">=", 2), "fixture must derive several usecases to expose ordering")
+		Expect(cfg.KnownUsecaseStrings).To(Equal([]string{"FLAG_CHAT", "FLAG_VISION"}))
+	})
+
+	// The request pipeline reloads the config through LoadModelConfigFileByName,
+	// which applies SetDefaults a second time. That must not move the revision
+	// away from the one model administration publishes from the loader map.
+	It("survives the extra SetDefaults the request path applies", func() {
+		loader := config.NewModelConfigLoader(dir)
+		Expect(loader.LoadModelConfigsFromPath(dir, appConfig.ToConfigLoaderOptions()...)).To(Succeed())
+		stored, ok := loader.GetModelConfig("example")
+		Expect(ok).To(BeTrue())
+		adminRevision, err := config.ModelConfigRevision(&stored)
+		Expect(err).ToNot(HaveOccurred())
+
+		requestCfg, err := loader.LoadModelConfigFileByNameDefaultOptions("example", appConfig)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(requestCfg.PersistedConfigRevision()).To(Equal(adminRevision))
+	})
+})

core/http/middleware/request_config_revision_test.go

@@ -50,7 +50,11 @@ var _ = Describe("Model config revision seen by inference requests", func() {
 
 		Expect(os.WriteFile(
 			filepath.Join(modelDir, "test-model.yaml"),
-			[]byte("name: test-model\nbackend: llama-cpp\ncontext_size: 4096\n"),
+			// The mmproj makes this derive several usecase flags. A single-flag
+			// model hides any instability in how that derived list is ordered.
+			[]byte("name: test-model\nbackend: llama-cpp\ncontext_size: 4096\n"+
+				"mmproj: llama-cpp/mmproj/test-model/mmproj.gguf\n"+
+				"known_usecases:\n    - chat\n"),
 			0o600,
 		)).To(Succeed())