whisper.swiftui : model list & bench methods

examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift

@@ -55,11 +55,85 @@ actor WhisperContext {
         return transcription
     }
 
+    func system_info() -> String {
+        var info = ""
+        if (ggml_cpu_has_neon() != 0) { info += "NEON " }
+        if (ggml_cpu_has_metal() != 0) { info += "METAL " }
+        if (ggml_cpu_has_blas() != 0) { info += "BLAS " }
+        return String(info.dropLast())
+    }
+
+    func bench_full(modelName: String) async -> String {
+        let n_threads = Int32(min(4, cpuCount())) // Default in whisper.cpp
+
+        let n_mels = whisper_model_n_mels(context)
+        if (whisper_set_mel(context, nil, 0, n_mels) != 0) {
+            return "error: failed to set mel"
+        }
+
+        // heat encoder
+        if (whisper_encode(context, 0, n_threads) != 0) {
+            return "error: failed to encode"
+        }
+
+        var tokens = [whisper_token](repeating: 0, count: 512)
+
+        // prompt heat
+        if (whisper_decode(context, &tokens, 256, 0, n_threads) != 0) {
+            return "error: failed to decode"
+        }
+
+        // text-generation heat
+        if (whisper_decode(context, &tokens, 1, 256, n_threads) != 0) {
+            return "error: failed to decode"
+        }
+
+        whisper_reset_timings(context)
+
+        // actual run
+        if (whisper_encode(context, 0, n_threads) != 0) {
+            return "error: failed to encode"
+        }
+
+        // text-generation
+        for i in 0..<256 {
+            if (whisper_decode(context, &tokens, 1, Int32(i), n_threads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        // batched decoding
+        for i in 0..<64 {
+            if (whisper_decode(context, &tokens, 5, 0, n_threads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        // prompt processing
+        for i in 0..<16 {
+            if (whisper_decode(context, &tokens, 256, 0, n_threads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        whisper_print_timings(context)
+
+        let system_info = self.system_info()
+        let timings: whisper_timings = whisper_get_timings(context)
+        let encode_ms = String(format: "%.2f", timings.encode_ms)
+        let decode_ms = String(format: "%.2f", timings.decode_ms)
+        let batchd_ms = String(format: "%.2f", timings.batchd_ms)
+        let prompt_ms = String(format: "%.2f", timings.prompt_ms)
+        return "| <todo> | iOS | \(system_info) | \(modelName) | \(n_threads) | 1 | \(encode_ms) | \(decode_ms) | \(batchd_ms) | \(prompt_ms) | <todo> |"
+    }
+
     static func createContext(path: String) throws -> WhisperContext {
         var params = whisper_context_default_params()
 #if targetEnvironment(simulator)
         params.use_gpu = false
         print("Running on the simulator, using CPU")
+#else
+        params.flash_attn = true // Enabled by default for Metal
 #endif
         let context = whisper_init_from_file_with_params(path, params)
         if let context {

examples/whisper.swiftui/whisper.swiftui.demo/Models/Model.swift

@@ -0,0 +1,17 @@
+import Foundation
+
+struct Model: Identifiable {
+    var id = UUID()
+    var name: String
+    var info: String
+    var url: String
+
+    var filename: String
+    var fileURL: URL {
+        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
+    }
+
+    func fileExists() -> Bool {
+        FileManager.default.fileExists(atPath: fileURL.path)
+    }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift

@@ -2,6 +2,7 @@ import Foundation
 import SwiftUI
 import AVFoundation
 
+
 @MainActor
 class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     @Published var isModelLoaded = false
@@ -14,7 +15,7 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     private var recordedFile: URL? = nil
     private var audioPlayer: AVAudioPlayer?
 
-    private var modelUrl: URL? {
+    private var builtInModelUrl: URL? {
         Bundle.main.url(forResource: "ggml-base.en", withExtension: "bin", subdirectory: "models")
     }
 
@@ -28,23 +29,51 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
 
     override init() {
         super.init()
+        loadModel()
+    }
+
+    func loadModel(path: URL? = nil, log: Bool = true) {
         do {
-            try loadModel()
+            whisperContext = nil
+            if (log) { messageLog += "Loading model...\n" }
+            let modelUrl = path ?? builtInModelUrl
+            if let modelUrl {
+                whisperContext = try WhisperContext.createContext(path: modelUrl.path())
+                if (log) { messageLog += "Loaded model \(modelUrl.lastPathComponent)\n" }
+            } else {
+                if (log) { messageLog += "Could not locate model\n" }
+            }
             canTranscribe = true
         } catch {
             print(error.localizedDescription)
-            messageLog += "\(error.localizedDescription)\n"
+            if (log) { messageLog += "\(error.localizedDescription)\n" }
         }
     }
-
-    private func loadModel() throws {
-        messageLog += "Loading model...\n"
-        if let modelUrl {
-            whisperContext = try WhisperContext.createContext(path: modelUrl.path())
-            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
-        } else {
-            messageLog += "Could not locate model\n"
+
+    func benchCurrentModel() async {
+        if whisperContext == nil {
+            messageLog += "Cannot bench without loaded model\n"
+            return
+        }
+        messageLog += "Benchmarking current model\n"
+        let result = await whisperContext?.bench_full(modelName: "<current>")
+        if (result != nil) { messageLog += result! + "\n" }
+    }
+
+    func bench(models: [Model]) async {
+        messageLog += "Benchmarking models\n"
+        messageLog += "| CPU | OS | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |\n"
+        messageLog += "| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n"
+        for model in models {
+            loadModel(path: model.fileURL, log: false)
+            if whisperContext == nil {
+                messageLog += "Cannot bench without loaded model\n"
+                break
+            }
+            let result = await whisperContext?.bench_full(modelName: model.name)
+            if (result != nil) { messageLog += result! + "\n" }
         }
+        messageLog += "Benchmarking completed\n"
     }
 
     func transcribeSample() async {

examples/whisper.swiftui/whisper.swiftui.demo/UI/ContentView.swift

@@ -1,5 +1,6 @@
 import SwiftUI
 import AVFoundation
+import Foundation
 
 struct ContentView: View {
     @StateObject var whisperState = WhisperState()
@@ -29,15 +30,125 @@ struct ContentView: View {
                     Text(verbatim: whisperState.messageLog)
                         .frame(maxWidth: .infinity, alignment: .leading)
                 }
+                .font(.footnote)
+                .padding()
+                .background(Color.gray.opacity(0.1))
+                .cornerRadius(10)
+
+                HStack {
+                    Button("Clear Logs", action: {
+                        whisperState.messageLog = ""
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+
+                    Button("Copy Logs", action: {
+                        UIPasteboard.general.string = whisperState.messageLog
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+
+                    Button("Bench", action: {
+                        Task {
+                            await whisperState.benchCurrentModel()
+                        }
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+
+                    Button("Bench All", action: {
+                        Task {
+                            await whisperState.bench(models: ModelSettingsView.getDownloadedModels())
+                        }
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+                }
+
+                NavigationLink(destination: ModelSettingsView(whisperState: whisperState)) {
+                    Text("View Models")
+                }
+                .font(.footnote)
+                .padding()
             }
             .navigationTitle("Whisper SwiftUI Demo")
             .padding()
         }
     }
-}
 
-struct ContentView_Previews: PreviewProvider {
-    static var previews: some View {
-        ContentView()
+    struct ModelSettingsView: View {
+        @ObservedObject var whisperState: WhisperState
+        @Environment(\.dismiss) var dismiss
+
+        private static let models: [Model] = [
+            Model(name: "tiny", info: "(F16, 75 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin", filename: "tiny.bin"),
+            Model(name: "tiny-q5_1", info: "(31 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q5_1.bin", filename: "tiny-q5_1.bin"),
+            Model(name: "tiny-q8_0", info: "(42 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q8_0.bin", filename: "tiny-q8_0.bin"),
+            Model(name: "tiny.en", info: "(F16, 75 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin", filename: "tiny.en.bin"),
+            Model(name: "tiny.en-q5_1", info: "(31 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q5_1.bin", filename: "tiny.en-q5_1.bin"),
+            Model(name: "tiny.en-q8_0", info: "(42 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q8_0.bin", filename: "tiny.en-q8_0.bin"),
+            Model(name: "base", info: "(F16, 142 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin", filename: "base.bin"),
+            Model(name: "base-q5_1", info: "(57 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base-q5_1.bin", filename: "base-q5_1.bin"),
+            Model(name: "base-q8_0", info: "(78 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base-q8_0.bin", filename: "base-q8_0.bin"),
+            Model(name: "base.en", info: "(F16, 142 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin", filename: "base.en.bin"),
+            Model(name: "base.en-q5_1", info: "(57 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q5_1.bin", filename: "base.en-q5_1.bin"),
+            Model(name: "base.en-q8_0", info: "(78 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q8_0.bin", filename: "base.en-q8_0.bin"),
+            Model(name: "small", info: "(F16, 466 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin", filename: "small.bin"),
+            Model(name: "small-q5_1", info: "(181 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small-q5_1.bin", filename: "small-q5_1.bin"),
+            Model(name: "small-q8_0", info: "(252 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small-q8_0.bin", filename: "small-q8_0.bin"),
+            Model(name: "small.en", info: "(F16, 466 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin", filename: "small.en.bin"),
+            Model(name: "small.en-q5_1", info: "(181 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q5_1.bin", filename: "small.en-q5_1.bin"),
+            Model(name: "small.en-q8_0", info: "(252 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q8_0.bin", filename: "small.en-q8_0.bin"),
+            Model(name: "medium", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.bin", filename: "medium.bin"),
+            Model(name: "medium-q5_0", info: "(514 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium-q5_0.bin", filename: "medium-q5_0.bin"),
+            Model(name: "medium-q8_0", info: "(785 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium-q8_0.bin", filename: "medium-q8_0.bin"),
+            Model(name: "medium.en", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en.bin", filename: "medium.en.bin"),
+            Model(name: "medium.en-q5_0", info: "(514 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q5_0.bin", filename: "medium.en-q5_0.bin"),
+            Model(name: "medium.en-q8_0", info: "(785 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q8_0.bin", filename: "medium.en-q8_0.bin"),
+            Model(name: "large-v1", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large.bin", filename: "large.bin"),
+            Model(name: "large-v2", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2.bin", filename: "large-v2.bin"),
+            Model(name: "large-v2-q5_0", info: "(1.1 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2-q5_0.bin", filename: "large-v2-q5_0.bin"),
+            Model(name: "large-v2-q8_0", info: "(1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2-q8_0.bin", filename: "large-v2-q8_0.bin"),
+            Model(name: "large-v3", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin", filename: "large-v3.bin"),
+            Model(name: "large-v3-q5_0", info: "(1.1 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-q5_0.bin", filename: "large-v3-q5_0.bin"),
+            Model(name: "large-v3-turbo", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin", filename: "large-v3-turbo.bin"),
+            Model(name: "large-v3-turbo-q5_0", info: "(547 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin", filename: "large-v3-turbo-q5_0.bin"),
+            Model(name: "large-v3-turbo-q8_0", info: "(834 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q8_0.bin", filename: "large-v3-turbo-q8_0.bin"),
+        ]
+
+        static func getDownloadedModels() -> [Model] {
+            // Filter models that have been downloaded
+            return models.filter {
+                FileManager.default.fileExists(atPath: $0.fileURL.path())
+            }
+        }
+
+        func loadModel(model: Model) {
+            Task {
+                dismiss()
+                whisperState.loadModel(path: model.fileURL)
+            }
+        }
+
+        var body: some View {
+            List {
+                Section(header: Text("Models")) {
+                    ForEach(ModelSettingsView.models) { model in
+                        DownloadButton(model: model)
+                            .onLoad(perform: loadModel)
+                    }
+                }
+            }
+            .listStyle(GroupedListStyle())
+            .navigationBarTitle("Model Settings", displayMode: .inline).toolbar {}
+        }
     }
 }
+
+//struct ContentView_Previews: PreviewProvider {
+//    static var previews: some View {
+//        ContentView()
+//    }
+//}