Fix: dedupe model/quantization types, buffer downloads, GGUF offsets and device scanner isolation

ModelQuantizer.xcodeproj/project.pbxproj

@@ -23,6 +23,7 @@
 		AA00001B /* HuggingFaceAPI.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA00001A /* HuggingFaceAPI.swift */; };
 		AA00001D /* GGUFBuilder.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA00001C /* GGUFBuilder.swift */; };
 		AA00001F /* ModelTypes.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA00001E /* ModelTypes.swift */; };
+		AA000100 /* KeychainTokenStore.swift in Sources */ = {isa = PBXBuildFile; fileRef = AA000101 /* KeychainTokenStore.swift */; };
 		AA000021 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = AA000020 /* Assets.xcassets */; };
 		AA000023 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = AA000022 /* Preview Assets.xcassets */; };
 /* End PBXBuildFile section */
@@ -44,6 +45,7 @@
 		AA00001A /* HuggingFaceAPI.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = HuggingFaceAPI.swift; sourceTree = "<group>"; };
 		AA00001C /* GGUFBuilder.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = GGUFBuilder.swift; sourceTree = "<group>"; };
 		AA00001E /* ModelTypes.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ModelTypes.swift; sourceTree = "<group>"; };
+		AA000101 /* KeychainTokenStore.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = KeychainTokenStore.swift; sourceTree = "<group>"; };
 		AA000020 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
 		AA000022 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
 		AA000024 /* ModelQuantizer.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = ModelQuantizer.app; sourceTree = BUILT_PRODUCTS_DIR; };
@@ -120,6 +122,7 @@
 				AA000008 /* SettingsSuggester.swift */,
 				AA00001A /* HuggingFaceAPI.swift */,
 				AA00001C /* GGUFBuilder.swift */,
+				AA000101 /* KeychainTokenStore.swift */,
 			);
 			path = Services;
 			sourceTree = "<group>";
@@ -235,6 +238,7 @@
 				AA00001B /* HuggingFaceAPI.swift in Sources */,
 				AA00001D /* GGUFBuilder.swift in Sources */,
 				AA00001F /* ModelTypes.swift in Sources */,
+				AA000100 /* KeychainTokenStore.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};

ModelQuantizer/ContentView.swift

@@ -48,7 +48,7 @@ struct ContentView: View {
 
                 DeviceInfoView()
                     .tabItem {
-                        Image(systemName: "iphone.gen3")
+                        Image(systemName: "iphone")
                         Text("Device")
                     }
                     .tag(3)

ModelQuantizer/Models/ModelTypes.swift

@@ -67,7 +67,7 @@ enum ModelArchitecture: String, Codable, CaseIterable {
     var supportedQuantizations: [QuantizationType] {
         switch self {
         case .llama, .mistral, .qwen2, .gemma, .phi:
-            return [.q4_0, .q4_1, .q5_0, .q5_1, .q8_0, .fp16, .fp32]
+            return [.q4_0, .q4_1, .q8_0, .fp16, .fp32]
         case .falcon, .gpt2:
             return [.q4_0, .q4_1, .q8_0, .fp16]
         case .bert:
@@ -206,6 +206,8 @@ struct QuantizationJob: Codable, Identifiable {
     let startTime: Date
     let endTime: Date
     let contextLength: Int
+    let estimatedTokensPerSecond: Double?
+    let validationScore: Double?
 
     var duration: TimeInterval {
         return endTime.timeIntervalSince(startTime)
@@ -294,6 +296,7 @@ struct HFModelConfig: Codable {
 
 struct HFSibling: Codable {
     let rfilename: String
+    let size: Int64?
 }
 
 // MARK: - Performance Estimate
@@ -324,3 +327,113 @@ struct InferenceSettings {
     let maxTokens: Int
     let quantizationType: QuantizationType
 }
+
+enum QuantizationError: Error, LocalizedError {
+    case noDownloadURL
+    case noModelFiles
+    case downloadFailed
+    case invalidModelFormat
+    case unsupportedVersion
+    case quantizationFailed
+    case invalidOutput
+    case insufficientMemory
+    case cancelled
+
+    var errorDescription: String? {
+        switch self {
+        case .noDownloadURL: return "No download URL provided for model"
+        case .noModelFiles: return "No model files found in repository"
+        case .downloadFailed: return "Failed to download model files"
+        case .invalidModelFormat: return "Invalid or unsupported model format"
+        case .unsupportedVersion: return "Unsupported GGUF version"
+        case .quantizationFailed: return "Quantization process failed"
+        case .invalidOutput: return "Generated model file is invalid"
+        case .insufficientMemory: return "Insufficient memory for quantization"
+        case .cancelled: return "Quantization was cancelled"
+        }
+    }
+}
+
+
+enum ModelCatalog {
+    static let curatedModels: [HFModel] = [
+        HFModel(
+            modelId: "microsoft/Phi-3-mini-4k-instruct",
+            name: "Phi-3 Mini 4K",
+            description: "Microsoft's efficient 3.8B parameter model with excellent performance",
+            parameters: "3.8B",
+            architecture: .phi,
+            downloadURL: URL(string: "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/model.safetensors"),
+            sizeBytes: 7_600_000_000,
+            recommendedContextLength: 4096,
+            tags: ["instruct", "chat", "efficient"],
+            downloads: 2_500_000,
+            likes: 8500
+        ),
+        HFModel(
+            modelId: "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            name: "Llama 3.1 8B Instruct",
+            description: "Meta's latest 8B parameter instruction-tuned model",
+            parameters: "8B",
+            architecture: .llama,
+            downloadURL: URL(string: "https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct/resolve/main/model.safetensors"),
+            sizeBytes: 16_000_000_000,
+            recommendedContextLength: 8192,
+            tags: ["instruct", "chat", "meta"],
+            downloads: 5_000_000,
+            likes: 15000
+        ),
+        HFModel(
+            modelId: "mistralai/Mistral-7B-Instruct-v0.3",
+            name: "Mistral 7B Instruct v0.3",
+            description: "Mistral's powerful 7B instruction model",
+            parameters: "7B",
+            architecture: .mistral,
+            downloadURL: URL(string: "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/resolve/main/model.safetensors"),
+            sizeBytes: 14_000_000_000,
+            recommendedContextLength: 32768,
+            tags: ["instruct", "chat", "long-context"],
+            downloads: 8_000_000,
+            likes: 22000
+        ),
+        HFModel(
+            modelId: "google/gemma-2-2b-it",
+            name: "Gemma 2 2B IT",
+            description: "Google's lightweight 2B instruction model",
+            parameters: "2B",
+            architecture: .gemma,
+            downloadURL: URL(string: "https://huggingface.co/google/gemma-2-2b-it/resolve/main/model.safetensors"),
+            sizeBytes: 4_000_000_000,
+            recommendedContextLength: 8192,
+            tags: ["instruct", "chat", "lightweight"],
+            downloads: 1_200_000,
+            likes: 5600
+        ),
+        HFModel(
+            modelId: "Qwen/Qwen2.5-7B-Instruct",
+            name: "Qwen2.5 7B Instruct",
+            description: "Alibaba's Qwen2.5 with improved reasoning",
+            parameters: "7B",
+            architecture: .qwen2,
+            downloadURL: URL(string: "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/resolve/main/model.safetensors"),
+            sizeBytes: 15_000_000_000,
+            recommendedContextLength: 32768,
+            tags: ["instruct", "chat", "multilingual"],
+            downloads: 3_000_000,
+            likes: 9800
+        ),
+        HFModel(
+            modelId: "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+            name: "SmolLM2 1.7B Instruct",
+            description: "Hugging Face's tiny but capable model",
+            parameters: "1.7B",
+            architecture: .llama,
+            downloadURL: URL(string: "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct/resolve/main/model.safetensors"),
+            sizeBytes: 3_400_000_000,
+            recommendedContextLength: 8192,
+            tags: ["instruct", "chat", "tiny"],
+            downloads: 800_000,
+            likes: 4200
+        )
+    ]
+}

ModelQuantizer/Services/DeviceScanner.swift

@@ -20,6 +20,7 @@ import AppKit
 /// Represents the device capability profile for model quantization
 struct DeviceCapabilityProfile: Codable, Equatable {
     let deviceModel: String
+    let deviceIdentifier: String
     let operatingSystem: String
     let operatingSystemVersion: String
     let deviceClass: DeviceClass
@@ -98,14 +99,15 @@ struct DeviceCapabilityProfile: Codable, Equatable {
 }
 
 /// Comprehensive device scanner for ML model optimization
-class DeviceScanner: ObservableObject, @unchecked Sendable {
+@MainActor
+final class DeviceScanner: ObservableObject {
     static let shared = DeviceScanner()
 
     @Published var currentProfile: DeviceCapabilityProfile?
     @Published var isScanning = false
     @Published var lastScanDate: Date?
 
-    private var timer: Timer?
+    private var monitoringTask: Task<Void, Never>?
     private let metalDevice: MTLDevice?
 
     private init() {
@@ -114,9 +116,7 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
         startMonitoring()
     }
 
-    deinit {
-        timer?.invalidate()
-    }
+    deinit { monitoringTask?.cancel() }
 
     // MARK: - Public Methods
 
@@ -125,11 +125,9 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
 
         Task {
             let profile = await createProfile()
-            await MainActor.run {
-                self.currentProfile = profile
-                self.lastScanDate = Date()
-                self.isScanning = false
-            }
+            self.currentProfile = profile
+            self.lastScanDate = Date()
+            self.isScanning = false
         }
     }
 
@@ -145,15 +143,19 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
     // MARK: - Private Methods
 
     private func startMonitoring() {
-        timer = Timer.scheduledTimer(withTimeInterval: 30.0, repeats: true) { [weak self] _ in
-            self?.performScan()
+        monitoringTask?.cancel()
+        monitoringTask = Task { [weak self] in
+            while !Task.isCancelled {
+                try? await Task.sleep(nanoseconds: 30_000_000_000)
+                self?.performScan()
+            }
         }
     }
 
     private func createProfile() async -> DeviceCapabilityProfile {
-        let deviceModel = getDeviceModel()
+        let device = getDeviceModel()
         let osInfo = getOperatingSystemInfo()
-        let deviceClass = classifyDevice(deviceModel)
+        let deviceClass = classifyDevice(device.name)
         let ram = getRAMInfo()
         let cpu = getCPUInfo()
         let gpu = getGPUInfo()
@@ -162,7 +164,8 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
         let storage = getStorageInfo()
 
         return DeviceCapabilityProfile(
-            deviceModel: deviceModel,
+            deviceModel: device.name,
+            deviceIdentifier: device.identifier,
             operatingSystem: osInfo.name,
             operatingSystemVersion: osInfo.version,
             deviceClass: deviceClass,
@@ -188,16 +191,20 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
 
     // MARK: - Device Information Gathering
 
-    private func getDeviceModel() -> String {
-        var systemInfo = utsname()
-        uname(&systemInfo)
-        let machineMirror = Mirror(reflecting: systemInfo.machine)
-        let identifier = machineMirror.children.reduce("") { identifier, element in
-            guard let value = element.value as? Int8, value != 0 else { return identifier }
-            return identifier + String(UnicodeScalar(UInt8(value)))
-        }
-
-        return mapToMarketingName(identifier)
+    private func getDeviceModel() -> (identifier: String, name: String) {
+        #if targetEnvironment(simulator)
+        let simId = ProcessInfo.processInfo.environment["SIMULATOR_MODEL_IDENTIFIER"] ?? "Simulator"
+        return (simId, mapToMarketingName(simId))
+        #else
+        var sysInfo = utsname()
+        uname(&sysInfo)
+        let mirror = Mirror(reflecting: sysInfo.machine)
+        let identifier = mirror.children.compactMap { element -> Character? in
+            guard let value = element.value as? Int8, value != 0 else { return nil }
+            return Character(UnicodeScalar(UInt8(value)))
+        }.reduce("") { $0 + String($1) }
+        return (identifier, mapToMarketingName(identifier))
+        #endif
     }
 
     private func getOperatingSystemInfo() -> (name: String, version: String) {
@@ -332,7 +339,7 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
         let name = device.name
 
         // Estimate GPU cores based on device class
-        let model = getDeviceModel()
+        let model = getDeviceModel().name
         var cores = 4 // Default
 
         if model.contains("Pro") || model.contains("Max") {
@@ -372,7 +379,7 @@ class DeviceScanner: ObservableObject, @unchecked Sendable {
 
     private func getNeuralEngineInfo() -> (cores: Int, tops: Double) {
         // Estimate Neural Engine cores based on device
-        let model = getDeviceModel()
+        let model = getDeviceModel().name
         var cores = 8
         var tops = 15.8
 

ModelQuantizer/Services/GGUFBuilder.swift

@@ -61,36 +61,29 @@ public struct GGUFBuilder {
             try appendMetadataValue(value, to: &data)
         }
 
-        // Write tensor info
+        // Write tensor info into a temporary buffer so offsets are stable
+        var tensorInfoData = Data()
         var tensorDataOffset = data.count + calculateTensorInfoSize()
-        // Align to 32 bytes
         tensorDataOffset = ((tensorDataOffset + 31) / 32) * 32
-        
+
         for tensor in tensors {
-            // Tensor name
-            data.append(UInt64(tensor.name.utf8.count).littleEndianData)
-            data.append(Data(tensor.name.utf8))
-
-            // Number of dimensions
-            data.append(UInt32(tensor.shape.count).littleEndianData)
-
-            // Shape dimensions
+            tensorInfoData.append(UInt64(tensor.name.utf8.count).littleEndianData)
+            tensorInfoData.append(Data(tensor.name.utf8))
+            tensorInfoData.append(UInt32(tensor.shape.count).littleEndianData)
+
             for dim in tensor.shape {
-                data.append(UInt64(dim).littleEndianData)
+                tensorInfoData.append(UInt64(dim).littleEndianData)
             }
-
-            // Data type
-            data.append(tensor.type.rawValue.littleEndianData)
-
-            // Offset to tensor data
-            data.append(UInt64(tensorDataOffset).littleEndianData)
-
+
+            tensorInfoData.append(tensor.type.rawValue.littleEndianData)
+            tensorInfoData.append(UInt64(tensorDataOffset).littleEndianData)
+
             tensorDataOffset += tensor.data.count
-            // Align each tensor to 32 bytes
             tensorDataOffset = ((tensorDataOffset + 31) / 32) * 32
         }
-
-        // Pad to alignment
+
+        data.append(tensorInfoData)
+
         while data.count % 32 != 0 {
             data.append(0)
         }
@@ -240,17 +233,3 @@ extension FixedWidthInteger {
         return withUnsafeBytes(of: &value) { Data($0) }
     }
 }
-
-extension UInt32 {
-    var littleEndianData: Data {
-        var value = self.littleEndian
-        return withUnsafeBytes(of: &value) { Data($0) }
-    }
-}
-
-extension UInt64 {
-    var littleEndianData: Data {
-        var value = self.littleEndian
-        return withUnsafeBytes(of: &value) { Data($0) }
-    }
-}