@dvai-bridge/ios-llama-core 4.0.0 → 4.0.2

package/ios/Sources/DVAILlamaCore/PluginState.swift

@@ -1,158 +1,158 @@
-// Internal/PluginState.swift
-import Foundation
-#if !COCOAPODS
-import DVAILlamaCoreObjC
-#endif
-#if !COCOAPODS
-import DVAISharedCore
-#endif
-
-/// Owns the running state of the capacitor-llama plugin: the model bridge,
-/// the HTTP server, and the model metadata. All access is serialised through
-/// the actor isolation.
-public actor PluginState {
-    private var server: HttpServer?
-    private var bridge: LlamaCppBridge?
-    private(set) var modelId: String = ""
-    private(set) var isRunning: Bool = false
-    private(set) var baseUrl: String?
-    private(set) var port: Int?
-
-    public init() {}
-
-    /// Start the plugin: load model, bind server, install routes.
-    /// - Returns dictionary suitable for Capacitor's `call.resolve(...)`.
-    public func start(opts: [String: Any]) async throws -> [String: Any] {
-        if isRunning { try await stopInternal() }
-
-        guard let modelPath = opts["modelPath"] as? String, !modelPath.isEmpty else {
-            throw NSError(
-                domain: "DVAIBridgeLlama",
-                code: 400,
-                userInfo: [NSLocalizedDescriptionKey: "modelPath is required for llama backend"]
-            )
-        }
-
-        let mmprojPath = opts["mmprojPath"] as? String
-        let chatTemplate = opts["chatTemplate"] as? String
-        let gpuLayers = opts["gpuLayers"] as? Int ?? 99
-        let contextSize = opts["contextSize"] as? Int ?? 2048
-        let threads = opts["threads"] as? Int ?? 4
-        let embeddingMode = opts["embeddingMode"] as? Bool ?? false
-        let httpBasePort = opts["httpBasePort"] as? Int ?? 38883
-        let httpMaxPortAttempts = opts["httpMaxPortAttempts"] as? Int ?? 16
-        let corsRaw = opts["corsOrigin"]
-        let corsConfig = parseCors(corsRaw)
-
-        // Load model via the ObjC++ bridge (real llama.cpp under the hood).
-        let bridge = LlamaCppBridge()
-        try bridge.loadModel(
-            atPath: modelPath,
-            mmprojPath: mmprojPath,
-            gpuLayers: Int32(gpuLayers),
-            contextSize: Int32(contextSize),
-            threads: Int32(threads),
-            embeddingMode: embeddingMode
-        )
-
-        // Phase 2A Pass 2: load mmproj (if provided) so multimodal handlers
-        // can light up. A failed mmproj load is fatal for this start() call —
-        // the caller asked for a multimodal model and we couldn't deliver.
-        if let mmprojPath = mmprojPath, !mmprojPath.isEmpty {
-            do {
-                try bridge.loadMmproj(atPath: mmprojPath)
-            } catch {
-                bridge.unload()
-                throw error
-            }
-        }
-        let mmprojLoaded = bridge.isMmprojLoaded
-        // Audio encoder support implies mmproj is loaded AND mtmd reports
-        // an audio encoder is present in the projector.
-        let modelHasAudioEncoder = mmprojLoaded && bridge.hasAudioEncoder()
-
-        // Build handlers + context first; Hummingbird requires routes
-        // to be registered at Application construction time, so the
-        // installRoutes → tryBind order is mandatory. Phase 2A Pass 2:
-        // real flags mirrored from the bridge state. embeddingMode
-        // comes straight from the start opts so /v1/embeddings can
-        // short-circuit when off. chatTemplate is an optional
-        // Jinja-compatible override; nil/empty falls through to the
-        // model's bundled `tokenizer.chat_template`.
-        let handlers = LlamaHandlers(
-            bridge: bridge,
-            modelId: modelPath,
-            mmprojLoaded: mmprojLoaded,
-            modelHasAudioEncoder: modelHasAudioEncoder,
-            embeddingMode: embeddingMode,
-            chatTemplate: chatTemplate
-        )
-        let ctx = HandlerContext(modelId: modelPath, backendName: "llama")
-        let server = HttpServer()
-        await server.installRoutes(handlers: handlers, ctx: ctx, corsConfig: corsConfig)
-
-        // Bind server (with port-fallback). If bind fails, release the
-        // bridge so the loaded llama context doesn't leak until next
-        // start().
-        let port: Int
-        do {
-            port = try await server.tryBind(
-                basePort: httpBasePort,
-                maxAttempts: httpMaxPortAttempts,
-                host: "127.0.0.1"
-            )
-        } catch {
-            bridge.unload()
-            throw error
-        }
-
-        self.bridge = bridge
-        self.server = server
-        self.modelId = modelPath
-        self.port = port
-        self.baseUrl = "http://127.0.0.1:\(port)/v1"
-        self.isRunning = true
-
-        return [
-            "baseUrl": self.baseUrl!,
-            "port": port,
-            "backend": "llama",
-            "modelId": modelPath,
-        ]
-    }
-
-    /// Stop the plugin: release model, stop server.
-    public func stop() async throws {
-        try await stopInternal()
-    }
-
-    private func stopInternal() async throws {
-        await server?.stop()
-        bridge?.unload()
-        server = nil
-        bridge = nil
-        modelId = ""
-        baseUrl = nil
-        port = nil
-        isRunning = false
-    }
-
-    /// Snapshot of the current running state, suitable for Capacitor `call.resolve(...)`.
-    public func statusInfo() -> [String: Any] {
-        var dict: [String: Any] = ["running": isRunning]
-        if let baseUrl = baseUrl { dict["baseUrl"] = baseUrl }
-        if isRunning { dict["backend"] = "llama" }
-        return dict
-    }
-
-    /// Parse the CORS option from the start opts dict.
-    private func parseCors(_ raw: Any?) -> CORSConfig {
-        if let s = raw as? String {
-            return s == "*" ? .wildcard : .exact(s)
-        }
-        if let arr = raw as? [String] {
-            return .allowlist(arr)
-        }
-        return .wildcard
-    }
-}
+// Internal/PluginState.swift
+import Foundation
+#if !COCOAPODS
+import DVAILlamaCoreObjC
+#endif
+#if !COCOAPODS
+import DVAISharedCore
+#endif
+
+/// Owns the running state of the capacitor-llama plugin: the model bridge,
+/// the HTTP server, and the model metadata. All access is serialised through
+/// the actor isolation.
+public actor PluginState {
+    private var server: HttpServer?
+    private var bridge: LlamaCppBridge?
+    private(set) var modelId: String = ""
+    private(set) var isRunning: Bool = false
+    private(set) var baseUrl: String?
+    private(set) var port: Int?
+
+    public init() {}
+
+    /// Start the plugin: load model, bind server, install routes.
+    /// - Returns dictionary suitable for Capacitor's `call.resolve(...)`.
+    public func start(opts: [String: Any]) async throws -> [String: Any] {
+        if isRunning { try await stopInternal() }
+
+        guard let modelPath = opts["modelPath"] as? String, !modelPath.isEmpty else {
+            throw NSError(
+                domain: "DVAIBridgeLlama",
+                code: 400,
+                userInfo: [NSLocalizedDescriptionKey: "modelPath is required for llama backend"]
+            )
+        }
+
+        let mmprojPath = opts["mmprojPath"] as? String
+        let chatTemplate = opts["chatTemplate"] as? String
+        let gpuLayers = opts["gpuLayers"] as? Int ?? 99
+        let contextSize = opts["contextSize"] as? Int ?? 2048
+        let threads = opts["threads"] as? Int ?? 4
+        let embeddingMode = opts["embeddingMode"] as? Bool ?? false
+        let httpBasePort = opts["httpBasePort"] as? Int ?? 38883
+        let httpMaxPortAttempts = opts["httpMaxPortAttempts"] as? Int ?? 16
+        let corsRaw = opts["corsOrigin"]
+        let corsConfig = parseCors(corsRaw)
+
+        // Load model via the ObjC++ bridge (real llama.cpp under the hood).
+        let bridge = LlamaCppBridge()
+        try bridge.loadModel(
+            atPath: modelPath,
+            mmprojPath: mmprojPath,
+            gpuLayers: Int32(gpuLayers),
+            contextSize: Int32(contextSize),
+            threads: Int32(threads),
+            embeddingMode: embeddingMode
+        )
+
+        // Phase 2A Pass 2: load mmproj (if provided) so multimodal handlers
+        // can light up. A failed mmproj load is fatal for this start() call —
+        // the caller asked for a multimodal model and we couldn't deliver.
+        if let mmprojPath = mmprojPath, !mmprojPath.isEmpty {
+            do {
+                try bridge.loadMmproj(atPath: mmprojPath)
+            } catch {
+                bridge.unload()
+                throw error
+            }
+        }
+        let mmprojLoaded = bridge.isMmprojLoaded
+        // Audio encoder support implies mmproj is loaded AND mtmd reports
+        // an audio encoder is present in the projector.
+        let modelHasAudioEncoder = mmprojLoaded && bridge.hasAudioEncoder()
+
+        // Build handlers + context first; Hummingbird requires routes
+        // to be registered at Application construction time, so the
+        // installRoutes → tryBind order is mandatory. Phase 2A Pass 2:
+        // real flags mirrored from the bridge state. embeddingMode
+        // comes straight from the start opts so /v1/embeddings can
+        // short-circuit when off. chatTemplate is an optional
+        // Jinja-compatible override; nil/empty falls through to the
+        // model's bundled `tokenizer.chat_template`.
+        let handlers = LlamaHandlers(
+            bridge: bridge,
+            modelId: modelPath,
+            mmprojLoaded: mmprojLoaded,
+            modelHasAudioEncoder: modelHasAudioEncoder,
+            embeddingMode: embeddingMode,
+            chatTemplate: chatTemplate
+        )
+        let ctx = HandlerContext(modelId: modelPath, backendName: "llama")
+        let server = HttpServer()
+        await server.installRoutes(handlers: handlers, ctx: ctx, corsConfig: corsConfig)
+
+        // Bind server (with port-fallback). If bind fails, release the
+        // bridge so the loaded llama context doesn't leak until next
+        // start().
+        let port: Int
+        do {
+            port = try await server.tryBind(
+                basePort: httpBasePort,
+                maxAttempts: httpMaxPortAttempts,
+                host: "127.0.0.1"
+            )
+        } catch {
+            bridge.unload()
+            throw error
+        }
+
+        self.bridge = bridge
+        self.server = server
+        self.modelId = modelPath
+        self.port = port
+        self.baseUrl = "http://127.0.0.1:\(port)/v1"
+        self.isRunning = true
+
+        return [
+            "baseUrl": self.baseUrl!,
+            "port": port,
+            "backend": "llama",
+            "modelId": modelPath,
+        ]
+    }
+
+    /// Stop the plugin: release model, stop server.
+    public func stop() async throws {
+        try await stopInternal()
+    }
+
+    private func stopInternal() async throws {
+        await server?.stop()
+        bridge?.unload()
+        server = nil
+        bridge = nil
+        modelId = ""
+        baseUrl = nil
+        port = nil
+        isRunning = false
+    }
+
+    /// Snapshot of the current running state, suitable for Capacitor `call.resolve(...)`.
+    public func statusInfo() -> [String: Any] {
+        var dict: [String: Any] = ["running": isRunning]
+        if let baseUrl = baseUrl { dict["baseUrl"] = baseUrl }
+        if isRunning { dict["backend"] = "llama" }
+        return dict
+    }
+
+    /// Parse the CORS option from the start opts dict.
+    private func parseCors(_ raw: Any?) -> CORSConfig {
+        if let s = raw as? String {
+            return s == "*" ? .wildcard : .exact(s)
+        }
+        if let arr = raw as? [String] {
+            return .allowlist(arr)
+        }
+        return .wildcard
+    }
+}