Improvements to Swift LLM app

examples/llm_inference/ios/InferenceExample.xcodeproj/project.pbxproj

@@ -7,7 +7,8 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
-		8D60EEED2B9A8A180019075E /* gemma-2b-it-cpu-int4.bin in Resources */ = {isa = PBXBuildFile; fileRef = 8D60EEEC2B9A8A180019075E /* gemma-2b-it-cpu-int4.bin */; };
+		0687D6A32BEFC65D00167F37 /* gemma-1.1-2b-it-gpu-int4.bin in Resources */ = {isa = PBXBuildFile; fileRef = 0687D6A22BEFC65D00167F37 /* gemma-1.1-2b-it-gpu-int4.bin */; };
+		0B6641881CDD8AB5E5F6C3FE /* Pods_InferenceExample.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = F02369998C59A0B7AAC3A1D5 /* Pods_InferenceExample.framework */; };
 		8DCF4C452B99289E00427D77 /* InferenceExampleApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8DCF4C442B99289E00427D77 /* InferenceExampleApp.swift */; };
 		8DCF4C472B99289E00427D77 /* ConversationScreen.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8DCF4C462B99289E00427D77 /* ConversationScreen.swift */; };
 		8DCF4C492B99289E00427D77 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8DCF4C482B99289E00427D77 /* Assets.xcassets */; };
@@ -17,7 +18,8 @@
 /* End PBXBuildFile section */
 
 /* Begin PBXFileReference section */
-		8D60EEEC2B9A8A180019075E /* gemma-2b-it-cpu-int4.bin */ = {isa = PBXFileReference; lastKnownFileType = archive.macbinary; path = "gemma-2b-it-cpu-int4.bin"; sourceTree = "<group>"; };
+		0687D6A22BEFC65D00167F37 /* gemma-1.1-2b-it-gpu-int4.bin */ = {isa = PBXFileReference; lastKnownFileType = archive.macbinary; name = "gemma-1.1-2b-it-gpu-int4.bin"; path = "../../../../../../Downloads/gemma-1.1-2b-it-gpu-int4.bin"; sourceTree = "<group>"; };
+		84EC0509CA2D0A791FE6A50D /* Pods-InferenceExample.release.xcconfig */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = text.xcconfig; name = "Pods-InferenceExample.release.xcconfig"; path = "Target Support Files/Pods-InferenceExample/Pods-InferenceExample.release.xcconfig"; sourceTree = "<group>"; };
 		8DCF4C412B99289E00427D77 /* InferenceExample.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = InferenceExample.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		8DCF4C442B99289E00427D77 /* InferenceExampleApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = InferenceExampleApp.swift; sourceTree = "<group>"; };
 		8DCF4C462B99289E00427D77 /* ConversationScreen.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConversationScreen.swift; sourceTree = "<group>"; };
@@ -26,33 +28,47 @@
 		8DCF4C4C2B99289E00427D77 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
 		8DCF4C572B992B9C00427D77 /* ConversationViewModel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConversationViewModel.swift; sourceTree = "<group>"; };
 		8DCF4C5B2B9939D700427D77 /* OnDeviceModel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = OnDeviceModel.swift; sourceTree = "<group>"; };
+		E0D7FB7ACC6233403D36FDF8 /* Pods-InferenceExample.debug.xcconfig */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = text.xcconfig; name = "Pods-InferenceExample.debug.xcconfig"; path = "Target Support Files/Pods-InferenceExample/Pods-InferenceExample.debug.xcconfig"; sourceTree = "<group>"; };
+		F02369998C59A0B7AAC3A1D5 /* Pods_InferenceExample.framework */ = {isa = PBXFileReference; explicitFileType = wrapper.framework; includeInIndex = 0; path = Pods_InferenceExample.framework; sourceTree = BUILT_PRODUCTS_DIR; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
 		8DCF4C3E2B99289E00427D77 /* Frameworks */ = {
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				0B6641881CDD8AB5E5F6C3FE /* Pods_InferenceExample.framework in Frameworks */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
 /* End PBXFrameworksBuildPhase section */
 
 /* Begin PBXGroup section */
+		32B29128DD33338DDF219030 /* Frameworks */ = {
+			isa = PBXGroup;
+			children = (
+				F02369998C59A0B7AAC3A1D5 /* Pods_InferenceExample.framework */,
+			);
+			name = Frameworks;
+			sourceTree = "<group>";
+		};
 		540B7F154909C4C9EF376B57 /* Pods */ = {
 			isa = PBXGroup;
 			children = (
+				E0D7FB7ACC6233403D36FDF8 /* Pods-InferenceExample.debug.xcconfig */,
+				84EC0509CA2D0A791FE6A50D /* Pods-InferenceExample.release.xcconfig */,
 			);
 			path = Pods;
 			sourceTree = "<group>";
 		};
 		8DCF4C382B99289D00427D77 = {
 			isa = PBXGroup;
 			children = (
+				0687D6A22BEFC65D00167F37 /* gemma-1.1-2b-it-gpu-int4.bin */,
 				8DCF4C432B99289E00427D77 /* InferenceExample */,
 				8DCF4C422B99289E00427D77 /* Products */,
 				540B7F154909C4C9EF376B57 /* Pods */,
-				8D60EEEC2B9A8A180019075E /* gemma-2b-it-cpu-int4.bin */,
+				32B29128DD33338DDF219030 /* Frameworks */,
 			);
 			sourceTree = "<group>";
 		};
@@ -93,6 +109,7 @@
 			isa = PBXNativeTarget;
 			buildConfigurationList = 8DCF4C502B99289E00427D77 /* Build configuration list for PBXNativeTarget "InferenceExample" */;
 			buildPhases = (
+				3C9FFF9E17DFA2F74C306852 /* [CP] Check Pods Manifest.lock */,
 				8DCF4C3D2B99289E00427D77 /* Sources */,
 				8DCF4C3E2B99289E00427D77 /* Frameworks */,
 				8DCF4C3F2B99289E00427D77 /* Resources */,
@@ -146,12 +163,37 @@
 			files = (
 				8DCF4C4D2B99289E00427D77 /* Preview Assets.xcassets in Resources */,
 				8DCF4C492B99289E00427D77 /* Assets.xcassets in Resources */,
-				8D60EEED2B9A8A180019075E /* gemma-2b-it-cpu-int4.bin in Resources */,
+				0687D6A32BEFC65D00167F37 /* gemma-1.1-2b-it-gpu-int4.bin in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
 /* End PBXResourcesBuildPhase section */
 
+/* Begin PBXShellScriptBuildPhase section */
+		3C9FFF9E17DFA2F74C306852 /* [CP] Check Pods Manifest.lock */ = {
+			isa = PBXShellScriptBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			inputFileListPaths = (
+			);
+			inputPaths = (
+				"${PODS_PODFILE_DIR_PATH}/Podfile.lock",
+				"${PODS_ROOT}/Manifest.lock",
+			);
+			name = "[CP] Check Pods Manifest.lock";
+			outputFileListPaths = (
+			);
+			outputPaths = (
+				"$(DERIVED_FILE_DIR)/Pods-InferenceExample-checkManifestLockResult.txt",
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+			shellPath = /bin/sh;
+			shellScript = "diff \"${PODS_PODFILE_DIR_PATH}/Podfile.lock\" \"${PODS_ROOT}/Manifest.lock\" > /dev/null\nif [ $? != 0 ] ; then\n    # print error to STDERR\n    echo \"error: The sandbox is not in sync with the Podfile.lock. Run 'pod install' or update your CocoaPods installation.\" >&2\n    exit 1\nfi\n# This output is used by Xcode 'outputs' to avoid re-running this script phase.\necho \"SUCCESS\" > \"${SCRIPT_OUTPUT_FILE_0}\"\n";
+			showEnvVarsInLog = 0;
+		};
+/* End PBXShellScriptBuildPhase section */
+
 /* Begin PBXSourcesBuildPhase section */
 		8DCF4C3D2B99289E00427D77 /* Sources */ = {
 			isa = PBXSourcesBuildPhase;
@@ -283,13 +325,15 @@
 		};
 		8DCF4C512B99289E00427D77 /* Debug */ = {
 			isa = XCBuildConfiguration;
+			baseConfigurationReference = E0D7FB7ACC6233403D36FDF8 /* Pods-InferenceExample.debug.xcconfig */;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CODE_SIGN_ENTITLEMENTS = InferenceExample/InferenceExample.entitlements;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
 				DEVELOPMENT_ASSET_PATHS = "\"InferenceExample/Preview Content\"";
+				DEVELOPMENT_TEAM = M3D535GFVK;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				"INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=iphoneos*]" = YES;
@@ -307,7 +351,7 @@
 				"LD_RUNPATH_SEARCH_PATHS[sdk=macosx*]" = "@executable_path/../Frameworks";
 				MACOSX_DEPLOYMENT_TARGET = 14.2;
 				MARKETING_VERSION = 1.0;
-				PRODUCT_BUNDLE_IDENTIFIER = com.mediapipe.InferenceExample;
+				PRODUCT_BUNDLE_IDENTIFIER = com.mediapipe.InferenceExample.foo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;
 				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
@@ -319,13 +363,15 @@
 		};
 		8DCF4C522B99289E00427D77 /* Release */ = {
 			isa = XCBuildConfiguration;
+			baseConfigurationReference = 84EC0509CA2D0A791FE6A50D /* Pods-InferenceExample.release.xcconfig */;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
 				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CODE_SIGN_ENTITLEMENTS = InferenceExample/InferenceExample.entitlements;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
 				DEVELOPMENT_ASSET_PATHS = "\"InferenceExample/Preview Content\"";
+				DEVELOPMENT_TEAM = M3D535GFVK;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				"INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=iphoneos*]" = YES;
@@ -343,7 +389,7 @@
 				"LD_RUNPATH_SEARCH_PATHS[sdk=macosx*]" = "@executable_path/../Frameworks";
 				MACOSX_DEPLOYMENT_TARGET = 14.2;
 				MARKETING_VERSION = 1.0;
-				PRODUCT_BUNDLE_IDENTIFIER = com.mediapipe.InferenceExample;
+				PRODUCT_BUNDLE_IDENTIFIER = com.mediapipe.InferenceExample.foo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;
 				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";

examples/llm_inference/ios/InferenceExample/ConversationViewModel.swift

@@ -89,7 +89,12 @@ class ConversationViewModel: ObservableObject {
       messages.append(systemMessage)
 
       do {
-        let response = try await chat.sendMessage(text)
+        let response = try await chat.sendMessage(text, progress : { [weak self] partialResult in
+          guard let self = self else { return }
+          DispatchQueue.main.async {
+            self.messages[self.messages.count - 1].message = partialResult
+          }
+        })
 
         // replace pending message with model response
         messages[messages.count - 1].message = response

examples/llm_inference/ios/InferenceExample/OnDeviceModel.swift

@@ -17,14 +17,27 @@ import MediaPipeTasksGenAI
 
 final class OnDeviceModel {
 
-  private var inference: LlmInference! = {
-    let path = Bundle.main.path(forResource: "gemma-2b-it-cpu-int4", ofType: "bin")!
-    let llmOptions = LlmInference.Options(modelPath: path)
-    return LlmInference(options: llmOptions)
-  }()
 
-  func generateResponse(prompt: String) async throws -> String {
+  private var cachedInference: LlmInference?
+
+  private var inference: LlmInference
+  {
+    get throws {
+      if let cached = cachedInference {
+         return cached
+      } else {
+        let path = Bundle.main.path(forResource: "gemma-1.1-2b-it-gpu-int4", ofType: "bin")!
+        let llmOptions = LlmInference.Options(modelPath: path)
+        cachedInference = try LlmInference(options: llmOptions)
+        return cachedInference!
+      }
+    }
+  }
+
+  func generateResponse(prompt: String, progress: @escaping (String) -> Void) async throws -> String {
     var partialResult = ""
+
+    let inference = try inference
     return try await withCheckedThrowingContinuation { continuation in
       do {
         try inference.generateResponseAsync(inputText: prompt) { partialResponse, error in
@@ -34,6 +47,7 @@ final class OnDeviceModel {
           }
           if let partial = partialResponse {
             partialResult += partial
+            progress(partialResult.trimmingCharacters(in: .whitespacesAndNewlines))
           }
         } completion: {
           let aggregate = partialResult.trimmingCharacters(in: .whitespacesAndNewlines)
@@ -62,16 +76,24 @@ final class Chat {
     self.model = model
   }
 
+  private func composeUserTurn(_ newMessage: String) -> String {
+    return "<start_of_turn>user\n\(newMessage)<end_of_turn>\n"
+  }
+
+  private func composeModelTurn(_ newMessage: String) -> String {
+    return "<start_of_turn>model\n\(newMessage)<end_of_turn>\n"
+  }
+
   private func compositePrompt(newMessage: String) -> String {
     return history.joined(separator: "\n") + "\n" + newMessage
   }
 
-  func sendMessage(_ text: String) async throws -> String {
-    let prompt = compositePrompt(newMessage: text)
-    let reply = try await model.generateResponse(prompt: prompt)
 
-    history.append(text)
-    history.append(reply)
+  func sendMessage(_ text: String, progress: @escaping (String) -> Void) async throws -> String {
+    let prompt = compositePrompt(newMessage: composeUserTurn(text))
+    let reply = try await model.generateResponse(prompt: prompt, progress: progress)
+
+    history = [prompt, composeModelTurn(reply)]
 
     print("Prompt: \(prompt)")
     print("Reply: \(reply)")

examples/llm_inference/ios/Podfile.lock

@@ -1,7 +1,7 @@
 PODS:
-  - MediaPipeTasksGenAI (0.10.11):
-    - MediaPipeTasksGenAIC (= 0.10.11)
-  - MediaPipeTasksGenAIC (0.10.11)
+  - MediaPipeTasksGenAI (0.10.14):
+    - MediaPipeTasksGenAIC (= 0.10.14)
+  - MediaPipeTasksGenAIC (0.10.14)
 
 DEPENDENCIES:
   - MediaPipeTasksGenAI
@@ -12,8 +12,8 @@ SPEC REPOS:
     - MediaPipeTasksGenAIC
 
 SPEC CHECKSUMS:
-  MediaPipeTasksGenAI: 9fb3fc0e9f9329d0b3f89d741dbdb4cb4429b87a
-  MediaPipeTasksGenAIC: 9bb1f037b742d7d642c8b8fbec8b4626f73c18c5
+  MediaPipeTasksGenAI: 8cd77fa32ea21f7a6319b025aa28cfc3e20ab73b
+  MediaPipeTasksGenAIC: 270ec81f85e96fac283945702e34112ebbfd5e77
 
 PODFILE CHECKSUM: b561fe84c5e19b81e1111ba0f8f21564f7006b85