Merge 8d11c13b44 into 5952e33570

iOS LLM Chat App Bugfix and Local Model Mode

apps/iOS/MNNLLMChat/MNNLLMiOS.xcodeproj/project.pbxproj

@@ -77,11 +77,6 @@
 			path = MNNLLMiOS;
 			sourceTree = "<group>";
 		};
-		3EAD33932D2BF38800021CF7 /* LocalModel */ = {
-			isa = PBXFileSystemSynchronizedRootGroup;
-			path = LocalModel;
-			sourceTree = "<group>";
-		};
 /* End PBXFileSystemSynchronizedRootGroup section */
 
 /* Begin PBXFrameworksBuildPhase section */
@@ -117,7 +112,6 @@
 			isa = PBXGroup;
 			children = (
 				3E94DF1F2D30CCF900BE39A7 /* MNNLLMiOS */,
-				3EAD33932D2BF38800021CF7 /* LocalModel */,
 				3E8592122D1D45090067B46F /* MNNLLMiOSTests */,
 				3E85921C2D1D45090067B46F /* MNNLLMiOSUITests */,
 				3EAD33842D2BDDC300021CF7 /* Frameworks */,
@@ -160,7 +154,6 @@
 			);
 			fileSystemSynchronizedGroups = (
 				3E94DF1F2D30CCF900BE39A7 /* MNNLLMiOS */,
-				3EAD33932D2BF38800021CF7 /* LocalModel */,
 			);
 			name = MNNLLMiOS;
 			packageProductDependencies = (

apps/iOS/MNNLLMChat/MNNLLMiOS/Chat/Services/README_Performance.md

@@ -1,214 +0,0 @@
-# UI Performance Optimization Guide
-
-## Overview
-
-This document explains the performance optimization utilities implemented in the chat system to ensure smooth AI streaming text output and overall UI responsiveness.
-
-## Core Components
-
-### 1. PerformanceMonitor
-
-A singleton utility for real-time performance monitoring and measurement.
-
-#### Implementation Principles
-
-- **Real-time FPS Monitoring**: Tracks UI update frequency and calculates actual FPS
-- **Frame Drop Detection**: Identifies when UI updates exceed the 16.67ms threshold (60 FPS)
-- **Operation Time Measurement**: Measures execution time of specific operations
-- **Automatic Statistics Reporting**: Logs performance metrics every second
-
-#### Key Features
-
-```swift
-class PerformanceMonitor {
-    static let shared = PerformanceMonitor()
-    
-    // Performance thresholds
-    private let targetFPS: Double = 60.0
-    private let frameThreshold: TimeInterval = 1.0 / 60.0 * 1.5 // 25ms threshold
-    
-    func recordUIUpdate() { /* Track UI updates */ }
-    func measureExecutionTime<T>(operation: String, block: () throws -> T) { /* Measure operations */ }
-}
-```
-
-#### Usage in the Chat System
-
-- Integrated into `LLMChatInteractor` to monitor message updates
-- Tracks UI update frequency during AI text streaming
-- Identifies performance bottlenecks in real-time
-
-### 2. UIUpdateOptimizer
-
-An actor-based utility for batching and throttling UI updates during streaming scenarios.
-
-#### Implementation Principles
-
-- **Batching Mechanism**: Groups multiple small updates into larger, more efficient ones
-- **Time-based Throttling**: Limits update frequency to prevent UI overload
-- **Actor-based Thread Safety**: Ensures safe concurrent access to update queue
-- **Automatic Flush Strategy**: Intelligently decides when to apply batched updates
-
-#### Architecture
-
-```swift
-actor UIUpdateOptimizer {
-    static let shared = UIUpdateOptimizer()
-    
-    private var pendingUpdates: [String] = []
-    private let batchSize: Int = 5          // Batch threshold
-    private let flushInterval: TimeInterval = 0.03 // 30ms throttling
-    
-    func addUpdate(_ content: String, completion: @escaping (String) -> Void)
-    func forceFlush(completion: @escaping (String) -> Void)
-}
-```
-
-#### Optimization Strategies
-
-1. **Batch Size Control**: Groups up to 5 updates before flushing
-2. **Time-based Throttling**: Flushes updates every 30ms maximum
-3. **Intelligent Scheduling**: Cancels redundant flush operations
-4. **Main Thread Delegation**: Ensures UI updates occur on the main thread
-
-#### Integration Points
-
-- **LLM Streaming**: Optimizes real-time text output from AI models
-- **Message Updates**: Batches frequent message content changes
-- **Force Flush**: Ensures final content is displayed when streaming ends
-
-## Performance Optimization Flow
-
-```
-AI Model Output → UIUpdateOptimizer → Batched Updates → UI Thread → Display
-                     ↓
-            PerformanceMonitor (Monitoring)
-                     ↓
-              Console Logs (Metrics)
-```
-
-## Testing and Validation
-
-### Performance Metrics
-
-1. **Target Performance**:
-   - Maintain 50+ FPS during streaming
-   - Keep frame drop rate below 5%
-   - Single operations under 16ms
-
-2. **Monitoring Indicators**:
-   - `📊 Performance Stats` - Real-time FPS and drop rate
-   - `⚠️ UI Update Lag detected` - Frame drop warnings
-   - `⏱️ Slow Operation` - Operation time alerts
-
-### Testing Methodology
-
-1. **Streaming Tests**:
-   - Test with long-form AI responses (articles, code)
-   - Monitor console output for performance warnings
-   - Observe visual smoothness of text animation
-
-2. **Load Testing**:
-   - Rapid successive message sending
-   - Large text blocks processing
-   - Multiple concurrent operations
-
-3. **Comparative Analysis**:
-   - Before/after optimization measurements
-   - Different device performance profiles
-   - Various content types and sizes
-
-### Debug Configuration
-
-For development and testing purposes:
-
-```swift
-// Example configuration adjustments (not implemented in production)
-// UIUpdateOptimizer.shared.batchSize = 10
-// UIUpdateOptimizer.shared.flushInterval = 0.05
-```
-
-## Implementation Details
-
-### UIUpdateOptimizer Algorithm
-
-1. **Add Update**: New content is appended to pending queue
-2. **Threshold Check**: Evaluate if immediate flush is needed
-   - Batch size reached (≥5 updates)
-   - Time threshold exceeded (≥30ms since last flush)
-3. **Scheduling**: If not immediate, schedule delayed flush
-4. **Flush Execution**: Combine all pending updates and execute on main thread
-5. **Cleanup**: Clear queue and reset timing
-
-### PerformanceMonitor Algorithm
-
-1. **Update Recording**: Track each UI update call
-2. **Timing Analysis**: Calculate time difference between updates
-3. **Frame Drop Detection**: Compare against 25ms threshold
-4. **Statistics Calculation**: Compute FPS and drop rate every second
-5. **Logging**: Output performance metrics to console
-
-## Integration Examples
-
-### In ViewModels
-
-```swift
-func updateUI() {
-    PerformanceMonitor.shared.recordUIUpdate()
-    // UI update code here
-}
-
-let result = PerformanceMonitor.shared.measureExecutionTime(operation: "Data Processing") {
-    return processLargeDataSet()
-}
-```
-
-### In Streaming Scenarios
-
-```swift
-await UIUpdateOptimizer.shared.addUpdate(newText) { batchedContent in
-    // Update UI with optimized batched content
-    updateTextView(with: batchedContent)
-}
-
-// When stream ends
-await UIUpdateOptimizer.shared.forceFlush { finalContent in
-    finalizeTextDisplay(with: finalContent)
-}
-```
-
-## Troubleshooting
-
-### Common Performance Issues
-
-1. **High Frame Drop Rate**:
-   - Check for blocking operations on main thread
-   - Verify batch size configuration
-   - Monitor memory usage
-
-2. **Slow Operation Warnings**:
-   - Profile specific operations causing delays
-   - Consider background threading for heavy tasks
-   - Optimize data processing algorithms
-
-3. **Inconsistent Performance**:
-   - Check device thermal state
-   - Monitor memory pressure
-   - Verify background app activity
-
-### Diagnostic Tools
-
-- **Console Monitoring**: Watch for performance log messages
-- **Xcode Instruments**: Use Time Profiler for detailed analysis
-- **Memory Graph**: Check for memory leaks affecting performance
-- **Energy Impact**: Monitor battery and thermal effects
-
-## Best Practices
-
-1. **Proactive Monitoring**: Always call `recordUIUpdate()` for critical UI operations
-2. **Batch When Possible**: Use `UIUpdateOptimizer` for frequent updates
-3. **Measure Critical Paths**: Wrap expensive operations with `measureExecutionTime`
-4. **Test on Real Devices**: Performance varies significantly across device types
-5. **Monitor in Production**: Keep performance logging enabled during development
-
-This performance optimization system ensures smooth user experience during AI text generation while providing developers with the tools needed to maintain and improve performance over time.

apps/iOS/MNNLLMChat/MNNLLMiOS/Chat/ViewModels/LLMChatViewModel.swift

@@ -74,6 +74,11 @@ final class LLMChatViewModel: ObservableObject {
     
     deinit {
         print("yxy:: LLMChat View Model deinit")
+        
+        llm?.cancelInference()
+        llm = nil
+        diffusion = nil
+        print("yxy:: LLMChat View Model cleanup complete")
     }
     
     func setupLLM(modelPath: String) {
@@ -148,6 +153,9 @@ final class LLMChatViewModel: ObservableObject {
     }
     
     func sendToLLM(draft: DraftMessage) {
+        
+        NotificationCenter.default.post(name: .dismissKeyboard, object: nil)
+        
         self.send(draft: draft, userType: .user)
         if isModelLoaded {
             if modelInfo.modelName.lowercased().contains("diffusion") {
@@ -258,6 +266,10 @@ final class LLMChatViewModel: ObservableObject {
                         await MainActor.run {
                             self.isProcessing = false
                             self.currentStreamingMessageId = nil
+                            
+                            DispatchQueue.main.asyncAfter(deadline: .now() + 0.3) {
+                                NotificationCenter.default.post(name: .dismissKeyboard, object: nil)
+                            }
                         }
                         await self.llmState.setProcessing(false)
                     }
@@ -342,6 +354,9 @@ final class LLMChatViewModel: ObservableObject {
         )
         
         interactor.disconnect()
+        
+        llm?.cancelInference()
+        
         llm = nil
         
         FileOperationManager.shared.cleanTempDirectories()

apps/iOS/MNNLLMChat/MNNLLMiOS/Chat/Views/LLMChatView.swift

@@ -83,7 +83,9 @@ struct LLMChatView: View {
         .disabled(viewModel.chatInputUnavilable)
         .toolbar {
             ToolbarItem(placement: .navigationBarLeading) {
-                Button { presentationMode.wrappedValue.dismiss() } label: {
+                Button { 
+                    presentationMode.wrappedValue.dismiss() 
+                } label: {
                     Image("backArrow", bundle: .current)
                 }
             }
@@ -119,6 +121,10 @@ struct LLMChatView: View {
             viewModel.onStart()
         }
         .onDisappear(perform: viewModel.onStop)
+        .onReceive(NotificationCenter.default.publisher(for: .dismissKeyboard)) { _ in
+            // 隐藏键盘
+            UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
+        }
     }
     
     // MARK: - LLM Chat Message Builder

apps/iOS/MNNLLMChat/MNNLLMiOS/ChatHistory/Services/ChatHistoryDatabase.swift

@@ -63,19 +63,33 @@ class ChatHistoryDatabase {
                 for msg in message.attachments {
                     if msg.type == .image {
                         var imageUrl = msg.full
+                        
+                        print("Processing image attachment: \(imageUrl.path)")
 
                         guard let copiedImage = ChatHistoryFileManager.shared.copyFile(from: imageUrl, for: historyId) else {
+                            print("Failed to copy image file: \(imageUrl.path)")
                             continue
                         }
                         
                         imageUrl = copiedImage
+                        print("Image copied to: \(imageUrl.path)")
                         
                         if imageUrl.isHEICImage() {
-                            guard let jpgUrl = AssetExtractor.convertHEICToJPG(heicUrl: imageUrl) else { continue }
+                            guard let jpgUrl = AssetExtractor.convertHEICToJPG(heicUrl: imageUrl) else { 
+                                print("Failed to convert HEIC to JPG: \(imageUrl.path)")
+                                continue 
+                            }
                             imageUrl = jpgUrl
+                            print("HEIC converted to JPG: \(imageUrl.path)")
                         }
                         
-                        copiedImages.append(LLMChatImage.init(id: msg.id, thumbnail: imageUrl, full: imageUrl))
+                        // 验证最终文件是否存在
+                        if ChatHistoryFileManager.shared.validateFileExists(at: imageUrl) {
+                            copiedImages.append(LLMChatImage.init(id: msg.id, thumbnail: imageUrl, full: imageUrl))
+                            print("Image successfully saved for history: \(imageUrl.path)")
+                        } else {
+                            print("Final image file validation failed: \(imageUrl.path)")
+                        }
                     }
                 }
                 
@@ -124,7 +138,10 @@ class ChatHistoryDatabase {
         do {
             for history in try db.prepare(chatHistories) {
                 let messagesData = history[messages].data(using: .utf8)!
-                let historyMessages = try JSONDecoder().decode([HistoryMessage].self, from: messagesData)
+                var historyMessages = try JSONDecoder().decode([HistoryMessage].self, from: messagesData)
+                
+                // 验证并修复图片路径
+                historyMessages = validateAndFixImagePaths(historyMessages, historyId: history[id])
                 
                 let chatHistory = ChatHistory(
                     id: history[id],
@@ -144,6 +161,38 @@ class ChatHistoryDatabase {
         return histories
     }
     
+    private func validateAndFixImagePaths(_ messages: [HistoryMessage], historyId: String) -> [HistoryMessage] {
+        return messages.map { message in
+            var updatedMessage = message
+            
+            if let images = message.images {
+                let validImages = images.compactMap { image -> LLMChatImage? in
+                    if ChatHistoryFileManager.shared.validateFileExists(at: image.full) {
+                        return image
+                    } else {
+                        let fileName = image.full.lastPathComponent
+                        if let validURL = ChatHistoryFileManager.shared.getValidFileURL(for: fileName, historyId: historyId) {
+                            return LLMChatImage(id: image.id, thumbnail: validURL, full: validURL)
+                        } else {
+                            print("Image file not found and cannot be recovered: \(image.full.path)")
+                            return nil
+                        }
+                    }
+                }
+                updatedMessage = HistoryMessage(
+                    id: message.id,
+                    content: message.content,
+                    images: validImages.isEmpty ? nil : validImages,
+                    audio: message.audio,
+                    isUser: message.isUser,
+                    createdAt: message.createdAt
+                )
+            }
+            
+            return updatedMessage
+        }
+    }
+    
     func deleteHistory(_ history: ChatHistory) {
         do {
             try db.run(chatHistories.filter(id == history.id).delete())
@@ -152,4 +201,4 @@ class ChatHistoryDatabase {
             print("Failed to delete history: \(error)")
         }
     }
-} 
+}

apps/iOS/MNNLLMChat/MNNLLMiOS/ChatHistory/Services/ChatHistoryFileManager.swift

@@ -47,10 +47,16 @@ class ChatHistoryFileManager {
         
         // Check if the file already exists at the destination
         if FileManager.default.fileExists(atPath: destinationURL.path) {
-            print("File already exists at \(destinationURL), returning original URL.")
+            print("File already exists at \(destinationURL), returning existing URL.")
             return destinationURL
         }
         
+        // Check if source file exists before copying
+        guard FileManager.default.fileExists(atPath: url.path) else {
+            print("Source file does not exist at \(url.path)")
+            return nil
+        }
+        
         do {
             try FileManager.default.copyItem(at: url, to: destinationURL)
             print("File copied to \(destinationURL)")
@@ -61,6 +67,24 @@ class ChatHistoryFileManager {
         return nil
     }
     
+    // Validate if a file exists at the given URL
+    func validateFileExists(at url: URL) -> Bool {
+        return FileManager.default.fileExists(atPath: url.path)
+    }
+    
+    // Get the correct file URL for a history file, checking if it exists
+    func getValidFileURL(for fileName: String, historyId: String) -> URL? {
+        let historyDirectory = baseDirectory.appendingPathComponent(historyId)
+        let fileURL = historyDirectory.appendingPathComponent(fileName)
+        
+        if FileManager.default.fileExists(atPath: fileURL.path) {
+            return fileURL
+        }
+        
+        print("File not found at expected path: \(fileURL.path)")
+        return nil
+    }
+    
     // Delete the history directory
     func deleteHistoryDirectory(for historyId: String) {
         let historyDirectory = baseDirectory.appendingPathComponent(historyId)

apps/iOS/MNNLLMChat/MNNLLMiOS/ChatHistory/Views/ChatHistoryItemView.swift

@@ -13,8 +13,8 @@ struct ChatHistoryItemView: View {
     var body: some View {
         VStack(alignment: .leading, spacing: 8) {
             
-            if let firstMessage = history.messages.last {
-                Text(String(firstMessage.content.prefix(200)))
+            if let lastMessage = getLastNonEmptyMessage() {
+                Text(String(lastMessage.content.prefix(200)))
                     .lineLimit(1)
                     .font(.system(size: 15, weight: .medium))
                     .foregroundColor(.primary)
@@ -42,4 +42,13 @@ struct ChatHistoryItemView: View {
         .padding(.vertical, 10)
         .padding(.horizontal, 0)
     }
+    
+    private func getLastNonEmptyMessage() -> HistoryMessage? {
+        for message in history.messages.reversed() {
+            if !message.content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
+                return message
+            }
+        }
+        return nil
+    }
 }

apps/iOS/MNNLLMChat/MNNLLMiOS/InferenceEngine/LLMInferenceEngineWrapper.mm

@@ -70,14 +70,79 @@
 #include <mutex>
 #include <thread>
 #include <chrono>
-#include <MNN/llm/llm.hpp>
+#include <fstream>
+#include <iomanip>
 #include <vector>
 #include <utility>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <errno.h>
+#include "MNN/expr/ExecutorScope.hpp"
 
 #import <Foundation/Foundation.h>
 #import "LLMInferenceEngineWrapper.h"
 
-using namespace MNN::Transformer;
+// Conditional include for MNN headers
+#ifdef __has_include
+  #if __has_include(<MNN/llm/llm.hpp>)
+    #include <MNN/llm/llm.hpp>
+    using namespace MNN::Transformer;
+  #else
+    // Fallback declarations when MNN headers are not available
+    namespace MNN {
+        namespace Transformer {
+            class Llm {
+             public:
+                 static Llm* createLLM(const std::string& config_path);
+                 virtual void set_config(const std::string& config) = 0;
+                 virtual void load() = 0;
+                 virtual void response(const std::string& input_str, std::ostream* os = nullptr, const char* end_with = nullptr) = 0;
+                 virtual void response(const std::vector<std::pair<std::string, std::string>>& history, std::ostream* os = nullptr, const char* end_with = nullptr, int max_new_tokens = 999999) = 0;
+                 virtual void response(const std::vector<int>& tokens, std::ostream* os = nullptr, const char* end_with = nullptr, int max_new_tokens = 999999) = 0;
+                 virtual void reset() = 0;
+                 virtual bool stopped() = 0;
+                 virtual int generate(int max_token_number = 0) = 0;
+                 struct LlmContext {
+                     int prompt_len;
+                     int gen_seq_len;
+                     int64_t prefill_us;
+                     int64_t decode_us;
+                 };
+                 virtual LlmContext* getContext() = 0;
+                 virtual ~Llm() = default;
+             };
+        }
+    }
+    using namespace MNN::Transformer;
+  #endif
+#else
+  // Fallback for older compilers
+  namespace MNN {
+      namespace Transformer {
+          class Llm {
+           public:
+               static Llm* createLLM(const std::string& config_path);
+               virtual void set_config(const std::string& config) = 0;
+               virtual void load() = 0;
+               virtual void response(const std::string& input_str, std::ostream* os = nullptr, const char* end_with = nullptr) = 0;
+               virtual void response(const std::vector<std::pair<std::string, std::string>>& history, std::ostream* os = nullptr, const char* end_with = nullptr, int max_new_tokens = 512) = 0;
+               virtual void response(const std::vector<int>& tokens, std::ostream* os = nullptr, const char* end_with = nullptr, int max_new_tokens = 512) = 0;
+               virtual void reset() = 0;
+               virtual bool stopped() = 0;
+               virtual int generate(int max_token_number = 0) = 0;
+               struct LlmContext {
+                   int prompt_len;
+                   int gen_seq_len;
+                   int64_t prefill_us;
+                   int64_t decode_us;
+               };
+               virtual LlmContext* getContext() = 0;
+               virtual ~Llm() = default;
+           };
+      }
+  }
+  using namespace MNN::Transformer;
+#endif
 
 using ChatMessage = std::pair<std::string, std::string>;
 
@@ -288,12 +353,13 @@ private:
 };
 
 @implementation LLMInferenceEngineWrapper {
-    std::shared_ptr<Llm> _llm;
+    std::shared_ptr<MNN::Transformer::Llm> _llm;
     std::vector<ChatMessage> _history;
     std::mutex _historyMutex;
     std::atomic<bool> _isProcessing;
     std::atomic<bool> _isBenchmarkRunning;
     std::atomic<bool> _shouldStopBenchmark;
+    std::atomic<bool> _shouldStopInference;
     NSString *_modelPath;
 }
 
@@ -314,6 +380,7 @@ private:
         _isProcessing = false;
         _isBenchmarkRunning = false;
         _shouldStopBenchmark = false;
+        _shouldStopInference = false;
         
         dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_HIGH, 0), ^{
             BOOL success = [self loadModelFromPath:modelPath];
@@ -335,13 +402,22 @@ private:
  * @return true if successful, false otherwise
  */
 bool remove_directory_safely(const std::string& path) {
-    try {
-        if (std::filesystem::exists(path)) {
-            std::filesystem::remove_all(path);
+    @try {
+        NSString *pathStr = [NSString stringWithUTF8String:path.c_str()];
+        NSFileManager *fileManager = [NSFileManager defaultManager];
+        
+        if ([fileManager fileExistsAtPath:pathStr]) {
+            NSError *error = nil;
+            BOOL success = [fileManager removeItemAtPath:pathStr error:&error];
+            if (!success && error) {
+                NSLog(@"Error removing directory %s: %@", path.c_str(), error.localizedDescription);
+                return false;
+            }
+            return success;
         }
         return true;
-    } catch (const std::filesystem::filesystem_error& e) {
-        NSLog(@"Error removing directory %s: %s", path.c_str(), e.what());
+    } @catch (NSException *exception) {
+        NSLog(@"Exception removing directory %s: %@", path.c_str(), exception.reason);
         return false;
     }
 }
@@ -394,7 +470,7 @@ bool remove_directory_safely(const std::string& path) {
         std::string model_dir = [bundleDirectory UTF8String];
         std::string config_path = model_dir + "/config.json";
         
-        _llm.reset(Llm::createLLM(config_path));
+        _llm.reset(MNN::Transformer::Llm::createLLM(config_path));
         if (!_llm) {
             NSLog(@"Error: Failed to create LLM from bundle");
             return NO;
@@ -454,31 +530,40 @@ bool remove_directory_safely(const std::string& path) {
             return NO;
         }
         
+        MNN::BackendConfig backendConfig;
+        auto executor = MNN::Express::Executor::newExecutor(MNN_FORWARD_CPU, backendConfig, 1);
+        MNN::Express::ExecutorScope s(executor);
+        
         // Get memory mapping setting with default fallback
         BOOL useMmap = configDict[@"use_mmap"] == nil ? YES : [configDict[@"use_mmap"] boolValue];
         
         // Create LLM instance with error checking
-        _llm.reset(Llm::createLLM(config_path));
+        _llm.reset(MNN::Transformer::Llm::createLLM(config_path));
         if (!_llm) {
             NSLog(@"Error: Failed to create LLM instance from config: %s", config_path.c_str());
             return NO;
         }
         
         // Setup temporary directory with improved error handling
-        std::string model_path_str([modelPath UTF8String]);
-        std::string temp_directory_path = model_path_str + "/temp";
+        // Use iOS system temporary directory instead of model path (which is read-only in Bundle)
+        NSString *tempDir = NSTemporaryDirectory();
+        NSString *modelName = [[modelPath lastPathComponent] stringByDeletingPathExtension];
+        NSString *tempDirPath = [tempDir stringByAppendingPathComponent:[NSString stringWithFormat:@"MNN_%@_temp", modelName]];
+        std::string temp_directory_path = [tempDirPath UTF8String];
         
         // Clean up existing temp directory
         if (!remove_directory_safely(temp_directory_path)) {
             NSLog(@"Warning: Failed to remove existing temp directory, continuing...");
         }
         
-        // Create new temp directory
+        // Create new temp directory in system temp location
         if (mkdir(temp_directory_path.c_str(), 0755) != 0 && errno != EEXIST) {
             NSLog(@"Error: Failed to create temp directory: %s, errno: %d", temp_directory_path.c_str(), errno);
             return NO;
         }
         
+        NSLog(@"Created temp directory at: %s", temp_directory_path.c_str());
+        
         // Configure LLM with proper error handling
         bool useMmapCpp = (useMmap == YES);
         std::string configStr = "{\"tmp_path\":\"" + temp_directory_path + "\", \"use_mmap\":" + (useMmapCpp ? "true" : "false") + "}";
@@ -585,12 +670,21 @@ bool remove_directory_safely(const std::string& path) {
     
     _isProcessing = true;
     
+    // Store reference for block execution
+    LLMInferenceEngineWrapper *blockSelf = self;
+    
     // Use high priority queue for better responsiveness
     dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_HIGH, 0), ^{
+        // Check if object is still valid before proceeding
+        if (!blockSelf || !blockSelf->_llm) {
+            NSLog(@"LLMInferenceEngineWrapper was deallocated or model unloaded during inference");
+            return;
+        }
+        
         @try {
             auto inference_start_time = std::chrono::high_resolution_clock::now();
             
-            OptimizedLlmStreamBuffer::CallBack callback = [output, self](const char* str, size_t len) {
+            OptimizedLlmStreamBuffer::CallBack callback = [output](const char* str, size_t len) {
                 if (output && str && len > 0) {
                     @autoreleasepool {
                         NSString *nsOutput = [[NSString alloc] initWithBytes:str
@@ -610,23 +704,77 @@ bool remove_directory_safely(const std::string& path) {
             
             // Thread-safe history management
             {
-                std::lock_guard<std::mutex> lock(self->_historyMutex);
-                self->_history.emplace_back(ChatMessage("user", [input UTF8String]));
+                std::lock_guard<std::mutex> lock(blockSelf->_historyMutex);
+                blockSelf->_history.emplace_back(ChatMessage("user", [input UTF8String]));
             }
             
             std::string inputStr = [input UTF8String];
             if (inputStr == "benchmark") {
-                [self performBenchmarkWithOutput:&os];
+                [blockSelf performBenchmarkWithOutput:&os];
             } else {
                 // Get initial context state for performance measurement
-                auto context = self->_llm->getContext();
+                auto context = blockSelf->_llm->getContext();
                 int initial_prompt_len = context->prompt_len;
                 int initial_decode_len = context->gen_seq_len;
                 int64_t initial_prefill_time = context->prefill_us;
                 int64_t initial_decode_time = context->decode_us;
                 
-                // Execute inference
-                self->_llm->response(self->_history, &os, "<eop>", 999999);
+                // Reset stop flag before starting inference
+                blockSelf->_shouldStopInference = false;
+                
+                // Execute inference with enhanced stopped status checking
+                @try {
+                    // Debug information for prompt
+                    std::string prompt_debug = "";
+                    for (const auto& msg : blockSelf->_history) {
+                        prompt_debug += msg.first + ": " + msg.second + "\n";
+                    }
+                    NSLog(@"submitNative prompt_string_for_debug:\n%s\nmax_new_tokens_: %d", prompt_debug.c_str(), 999999);
+                    
+                    // Start inference with initial response processing
+                    blockSelf->_llm->response(blockSelf->_history, &os, "<eop>", 1);
+                    int current_size = 1;
+                    int max_new_tokens = 999999;
+                    
+                    // Continue generation with precise token-by-token control
+                    while (!blockSelf->_shouldStopInference.load() && 
+                           !blockSelf->_llm->stoped() &&
+                           current_size < max_new_tokens) {
+                        
+                        // Generate single token for maximum control
+                        blockSelf->_llm->generate(1);
+                        current_size++;
+                        
+                        // Small delay to allow UI updates and stop signal processing
+                        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+                    }
+                    
+                    // Send appropriate end signal based on stop reason
+                    if (output) {
+                        dispatch_async(dispatch_get_main_queue(), ^{
+                            if (blockSelf->_shouldStopInference.load()) {
+                                output(@"<stopped>");
+                            } else {
+                                output(@"<eop>");
+                            }
+                        });
+                    }
+                    
+                    NSLog(@"Inference completed. Generated tokens: %d, Stopped by user: %s, Model stopped: %s", 
+                          current_size, 
+                          blockSelf->_shouldStopInference.load() ? "YES" : "NO",
+                          blockSelf->_llm->stoped() ? "YES" : "NO");
+                    
+                } @catch (NSException *exception) {
+                    NSLog(@"Exception during response generation: %@", exception.reason);
+                    
+                    // Send end signal even on error to unlock UI
+                    if (output) {
+                        dispatch_async(dispatch_get_main_queue(), ^{
+                            output(@"<eop>");
+                        });
+                    }
+                }
                 
                 // Calculate performance metrics if requested
                 if (showPerformance) {
@@ -683,7 +831,7 @@ bool remove_directory_safely(const std::string& path) {
             }
         }
         @finally {
-            self->_isProcessing = false;
+            blockSelf->_isProcessing = false;
         }
     });
 }
@@ -770,17 +918,25 @@ bool remove_directory_safely(const std::string& path) {
 }
 
 /**
- * Enhanced deallocation with proper cleanup
+ * Enhanced deallocation with proper cleanup and timeout
  */
 - (void)dealloc {
     NSLog(@"LLMInferenceEngineWrapper deallocating...");
     
-    // Stop any running benchmark
-    _shouldStopBenchmark = true;
+    // Actively cancel all operations first
+    [self cancelInference];
     
-    // Wait for any ongoing processing to complete
-    while (_isProcessing.load() || _isBenchmarkRunning.load()) {
+    // Wait for any ongoing processing to complete with timeout
+    int timeout = 100; // 1 second timeout (100 * 10ms)
+    while ((_isProcessing.load() || _isBenchmarkRunning.load()) && timeout > 0) {
         std::this_thread::sleep_for(std::chrono::milliseconds(10));
+        timeout--;
+    }
+    
+    if (timeout <= 0) {
+        NSLog(@"Warning: Dealloc timeout, forcing cleanup");
+        _isProcessing = false;
+        _isBenchmarkRunning = false;
     }
     
     {
@@ -879,11 +1035,17 @@ bool remove_directory_safely(const std::string& path) {
  * Cancel ongoing inference (if supported)
  */
 - (void)cancelInference {
-    if (_isProcessing.load()) {
-        NSLog(@"Inference cancellation requested");
-        // Note: Actual cancellation depends on MNN LLM implementation
-        // This is a placeholder for future enhancement
-    }
+    NSLog(@"Cancelling inference...");
+    
+    // Set all stop flags to true
+    _shouldStopInference = true;
+    _shouldStopBenchmark = true;
+    
+    // Force set processing states to false for immediate cleanup
+    _isProcessing = false;
+    _isBenchmarkRunning = false;
+    
+    NSLog(@"Inference cancellation completed - all flags set");
 }
 
 /**

apps/iOS/MNNLLMChat/MNNLLMiOS/Localizable.xcstrings

@@ -1041,7 +1041,7 @@
       }
     },
     "tag.deepThinking" : {
-      "extractionState" : "stale",
+      "comment" : "Deep thinking tag for local model",
       "localizations" : {
         "en" : {
           "stringUnit" : {
@@ -1091,6 +1091,23 @@
         }
       }
     },
+    "tag.localModel" : {
+      "comment" : "Local model inside the app",
+      "localizations" : {
+        "en" : {
+          "stringUnit" : {
+            "state" : "translated",
+            "value" : "Built-in Model"
+          }
+        },
+        "zh-Hans" : {
+          "stringUnit" : {
+            "state" : "translated",
+            "value" : "内置模型"
+          }
+        }
+      }
+    },
     "tag.math" : {
       "extractionState" : "stale",
       "localizations" : {
@@ -1292,6 +1309,16 @@
     },
     "Yes" : {
 
+    },
+    "搜索本地模型..." : {
+      "localizations" : {
+        "en" : {
+          "stringUnit" : {
+            "state" : "translated",
+            "value" : "Search local models …"
+          }
+        }
+      }
     }
   },
   "version" : "1.0"

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/LocalModelList/Views/LocalModelListView.swift

@@ -9,10 +9,25 @@ import SwiftUI
 
 struct LocalModelListView: View {
     @ObservedObject var viewModel: ModelListViewModel
+    @State private var localSearchText = ""
+    
+    private var filteredLocalModels: [ModelInfo] {
+        let downloadedModels = viewModel.models.filter { $0.isDownloaded }
+        
+        if localSearchText.isEmpty {
+            return downloadedModels
+        } else {
+            return downloadedModels.filter { model in
+                model.id.localizedCaseInsensitiveContains(localSearchText) ||
+                model.modelName.localizedCaseInsensitiveContains(localSearchText) ||
+                model.localizedTags.contains { $0.localizedCaseInsensitiveContains(localSearchText) }
+            }
+        }
+    }
     
     var body: some View {
         List {
-            ForEach(viewModel.filteredModels.filter { $0.isDownloaded }, id: \.id) { model in
+            ForEach(filteredLocalModels, id: \.id) { model in
                 Button(action: {
                     viewModel.selectModel(model)
                 }) {
@@ -25,6 +40,7 @@ struct LocalModelListView: View {
             }
         }
         .listStyle(.plain)
+        .searchable(text: $localSearchText, prompt: "搜索本地模型...")
         .refreshable {
             await viewModel.fetchModels()
         }

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/MainTabView.swift

@@ -7,7 +7,6 @@
 
 import SwiftUI
 
-// MainTabView is the primary view of the app, containing the tab bar and navigation for main sections.
 struct MainTabView: View {
     // MARK: - State Properties
     

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Models/ModelInfo.swift

@@ -85,8 +85,34 @@ struct ModelInfo: Codable {
     // MARK: - File System & Path Management
     
     var localPath: String {
-        let modelScopeId = "taobao-mnn/\(modelName)"
-        return HubApi.shared.localRepoLocation(HubApi.Repo.init(id: modelScopeId)).path
+        // Check if this is a local model from LocalModel folder or Bundle root
+        if let sources = sources, let localSource = sources["local"] {
+            guard let bundlePath = Bundle.main.resourcePath else {
+                return ""
+            }
+            
+            // Check if this is a flattened model (files directly in Bundle root)
+            if localSource.hasPrefix("bundle_root/") {
+                // For flattened models, return the Bundle root path
+                // The model files are directly in the Bundle root directory
+                return bundlePath
+            } else {
+                // Original LocalModel folder structure
+                let localModelPath = (bundlePath as NSString).deletingLastPathComponent + "/LocalModel"
+                
+                // If modelName is "LocalModel", return the LocalModel folder directly
+                if modelName == "LocalModel" {
+                    return localModelPath
+                } else {
+                    // For subdirectory models, append the model name
+                    return (localModelPath as NSString).appendingPathComponent(modelName)
+                }
+            }
+        } else {
+            // For downloaded models, use the original Hub API path
+            let modelScopeId = "taobao-mnn/\(modelName)"
+            return HubApi.shared.localRepoLocation(HubApi.Repo.init(id: modelScopeId)).path
+        }
     }
     
     // MARK: - Size Calculation & Formatting

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Models/ModelListViewModel.swift

@@ -11,7 +11,6 @@ import SwiftUI
 class ModelListViewModel: ObservableObject {
     // MARK: - Published Properties
     @Published var models: [ModelInfo] = []
-    @Published var searchText = ""
     @Published var quickFilterTags: [String] = []
     @Published var selectedModel: ModelInfo?
     @Published var showError = false
@@ -44,19 +43,6 @@ class ModelListViewModel: ObservableObject {
         Array(Set(models.compactMap { $0.vendor }))
     }
     
-    var filteredModels: [ModelInfo] {
-        let filtered = searchText.isEmpty ? models : models.filter { model in
-            model.id.localizedCaseInsensitiveContains(searchText) ||
-            model.modelName.localizedCaseInsensitiveContains(searchText) ||
-            model.localizedTags.contains { $0.localizedCaseInsensitiveContains(searchText) }
-        }
-        
-        let downloaded = filtered.filter { $0.isDownloaded }
-        let notDownloaded = filtered.filter { !$0.isDownloaded }
-        
-        return downloaded + notDownloaded
-    }
-    
     // MARK: - Initialization
     
     init() {
@@ -67,6 +53,118 @@ class ModelListViewModel: ObservableObject {
     
     // MARK: - Model Data Management
     
+    /// Load models from Bundle root directory (LocalModel folder structure flattened)
+    private func loadLocalModels() async -> [ModelInfo] {
+        let fileManager = FileManager.default
+        var localModels: [ModelInfo] = []
+        
+        guard let resourcePath = Bundle.main.resourcePath else {
+            return localModels
+        }
+        
+        do {
+            let contents = try fileManager.contentsOfDirectory(atPath: resourcePath)
+            
+            // Check if we have model files directly in Bundle root
+            let modelFiles = ["config.json", "llm_config.json", "llm.mnn", "tokenizer.txt"]
+            let foundModelFiles = contents.filter { modelFiles.contains($0) }
+            
+            if !foundModelFiles.isEmpty {
+                // Check if we have a complete model (at least config.json)
+                if foundModelFiles.contains("config.json") {
+                    let modelName = "Qwen3-0.6B-MNN"
+                    let localModel = ModelInfo(
+                        modelName: modelName,
+                        tags: [NSLocalizedString("tag.deepThinking", comment: "Deep thinking tag for local model"),
+                               NSLocalizedString("tag.localModel", comment: "Local model inside the app")],
+                        categories: ["Local Models"],
+                        vendor: "Local",
+                        sources: ["local": "bundle_root/\(modelName)"],
+                        isDownloaded: true
+                    )
+                    localModels.append(localModel)
+                    
+                    ModelStorageManager.shared.markModelAsDownloaded(modelName)
+                }
+            } else {
+                // Fallback: try to find LocalModel folder
+                let localModelPath = (resourcePath as NSString).appendingPathComponent("LocalModel")
+                var isDirectory: ObjCBool = false
+                
+                if fileManager.fileExists(atPath: localModelPath, isDirectory: &isDirectory), isDirectory.boolValue {
+                    localModels.append(contentsOf: await processLocalModelFolder(at: localModelPath))
+                }
+            }
+            
+        } catch {
+            // Silently handle error
+        }
+        
+        return localModels
+    }
+    
+    /// Process LocalModel folder (fallback for non-flattened structure)
+    private func processLocalModelFolder(at validPath: String) async -> [ModelInfo] {
+        let fileManager = FileManager.default
+        var localModels: [ModelInfo] = []
+        
+        // Check if this is a valid model directory (contains config.json)
+        let configPath = (validPath as NSString).appendingPathComponent("config.json")
+        if fileManager.fileExists(atPath: configPath) {
+            let modelName = "LocalModel"
+            let localModel = ModelInfo(
+                modelName: modelName,
+                tags: ["local", "bundled"],
+                categories: ["Local Models"],
+                vendor: "Local",
+                sources: ["local": "local/\(modelName)"],
+                isDownloaded: true
+            )
+            localModels.append(localModel)
+            
+            ModelStorageManager.shared.markModelAsDownloaded(modelName)
+        } else {
+            // Check for subdirectories that might contain models
+            do {
+                let contents = try fileManager.contentsOfDirectory(atPath: validPath)
+                
+                for item in contents {
+                    // Skip hidden files and common non-model files
+                    if item.hasPrefix(".") || item == "bench.txt" {
+                        continue
+                    }
+                    
+                    let itemPath = (validPath as NSString).appendingPathComponent(item)
+                    var isItemDirectory: ObjCBool = false
+                    
+                    if fileManager.fileExists(atPath: itemPath, isDirectory: &isItemDirectory),
+                       isItemDirectory.boolValue {
+                        
+                        let itemConfigPath = (itemPath as NSString).appendingPathComponent("config.json")
+                        
+                        if fileManager.fileExists(atPath: itemConfigPath) {
+                            let localModel = ModelInfo(
+                                modelName: item,
+                                tags: ["local", "bundled"],
+                                categories: ["Local Models"],
+                                vendor: "Local",
+                                sources: ["local": "local/\(item)"],
+                                isDownloaded: true
+                            )
+                            localModels.append(localModel)
+                            
+                            ModelStorageManager.shared.markModelAsDownloaded(item)
+                        }
+                    }
+                }
+            } catch {
+                // Silently handle error
+            }
+        }
+        
+        return localModels
+    }
+    
     @MainActor
     func fetchModels() async {
         do {
@@ -77,6 +175,10 @@ class ModelListViewModel: ObservableObject {
             
             var fetchedModels = info.models
             
+            // Add LocalModel folder models
+            let localModels = await loadLocalModels()
+            fetchedModels.append(contentsOf: localModels)
+            
             filterDiffusionModels(fetchedModels: &fetchedModels)
             loadCachedSizes(for: &fetchedModels)
             sortModels(fetchedModels: &fetchedModels)

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/ModelListView.swift

@@ -28,9 +28,6 @@ struct ModelListView: View {
             }
         }
         .searchable(text: $searchText, prompt: "Search models...")
-        .onChange(of: searchText) { _, newValue in
-            viewModel.searchText = newValue
-        }
         .refreshable {
             await viewModel.fetchModels()
         }
@@ -102,21 +99,31 @@ struct ModelListView: View {
         }
     }
     
-    // Filter models based on selected tags, categories and vendors
     private var filteredModels: [ModelInfo] {
-        let baseFiltered = viewModel.filteredModels
         
+        let searchFiltered = searchText.isEmpty ? viewModel.models : viewModel.models.filter { model in
+            model.id.localizedCaseInsensitiveContains(searchText) ||
+            model.modelName.localizedCaseInsensitiveContains(searchText) ||
+            model.localizedTags.contains { $0.localizedCaseInsensitiveContains(searchText) }
+        }
+        
+        let tagFiltered: [ModelInfo]
         if selectedTags.isEmpty && selectedCategories.isEmpty && selectedVendors.isEmpty {
-            return baseFiltered
+            tagFiltered = searchFiltered
+        } else {
+            tagFiltered = searchFiltered.filter { model in
+                let tagMatch = checkTagMatch(model: model)
+                let categoryMatch = checkCategoryMatch(model: model)
+                let vendorMatch = checkVendorMatch(model: model)
+                
+                return tagMatch && categoryMatch && vendorMatch
+            }
         }
         
-        return baseFiltered.filter { model in
-            let tagMatch = checkTagMatch(model: model)
-            let categoryMatch = checkCategoryMatch(model: model)
-            let vendorMatch = checkVendorMatch(model: model)
-            
-            return tagMatch && categoryMatch && vendorMatch
-        }
+        let downloaded = tagFiltered.filter { $0.isDownloaded }
+        let notDownloaded = tagFiltered.filter { !$0.isDownloaded }
+        
+        return downloaded + notDownloaded
     }
     
     // Extract tag matching logic as independent method

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/ModelRowSubviews/ActionButtonsView.swift

@@ -7,7 +7,6 @@
 
 import SwiftUI
 
-// MARK: - 操作按钮视图
 struct ActionButtonsView: View {
     let model: ModelInfo
     @ObservedObject var viewModel: ModelListViewModel

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/ModelRowSubviews/DownloadedButtonView.swift

@@ -7,7 +7,6 @@
 
 import SwiftUI
 
-// MARK: - 已下载按钮视图
 struct DownloadedButtonView: View {
     @Binding var showDeleteAlert: Bool
     

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/ModelRowSubviews/DownloadingButtonView.swift

@@ -7,7 +7,6 @@
 
 import SwiftUI
 
-// MARK: - 下载中按钮视图
 struct DownloadingButtonView: View {
     @ObservedObject var viewModel: ModelListViewModel
     let downloadProgress: Double

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/ModelRowSubviews/TagChip.swift

@@ -7,7 +7,6 @@
 
 import SwiftUI
 
-// MARK: - 标签芯片
 struct TagChip: View {
     let text: String
     
@@ -22,4 +21,4 @@ struct TagChip: View {
                     .stroke(Color.secondary.opacity(0.3), lineWidth: 0.5)
             )
     }
-}
+}

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/ModelRowSubviews/TagsView.swift

@@ -7,7 +7,6 @@
 
 import SwiftUI
 
-// MARK: - 标签视图
 struct TagsView: View {
     let tags: [String]
     
@@ -22,4 +21,4 @@ struct TagsView: View {
         }
         .frame(height: 25)
     }
-}
+}

apps/iOS/MNNLLMChat/MNNLLMiOS/MainTab/ModelList/Views/SwipeActionsView.swift

@@ -12,6 +12,12 @@ struct SwipeActionsView: View {
     let model: ModelInfo
     @ObservedObject var viewModel: ModelListViewModel
     
+    private func isBuiltInLocalModel(_ model: ModelInfo) -> Bool {
+        guard let vendor = model.vendor, vendor == "Local" else { return false }
+        guard let sources = model.sources, let localSource = sources["local"] else { return false }
+        return localSource.hasPrefix("bundle_root/")
+    }
+    
     var body: some View {
         if viewModel.pinnedModelIds.contains(model.id) {
             Button {
@@ -26,7 +32,7 @@ struct SwipeActionsView: View {
                 Label(LocalizedStringKey("button.pin"), systemImage: "pin")
             }.tint(.primaryBlue)
         }
-        if model.isDownloaded {
+        if model.isDownloaded && !isBuiltInLocalModel(model) {
             Button(role: .destructive) {
                 Task {
                     await viewModel.deleteModel(model)

apps/iOS/MNNLLMChat/MNNLLMiOS/Util/NotificationNames.swift

@@ -0,0 +1,17 @@
+//
+//  NotificationNames.swift
+//  MNNLLMiOS
+//
+//  Created by 游薪渝(揽清) on 2025/8/1.
+//
+
+import Foundation
+
+extension Notification.Name {
+    
+    static let dismissKeyboard = Notification.Name("dismissKeyboard")
+    
+    static let messageSent = Notification.Name("messageSent")
+    
+    static let onScrollToBottom = Notification.Name("onScrollToBottom")
+}

apps/iOS/MNNLLMChat/MNNLLMiOS/Util/Util.swift

@@ -154,6 +154,11 @@ extension Date {
         if result.contains("second") {
             return "Just now"
         }
+        
+        if result.contains("秒钟") {
+            return "刚刚"
+        }
+        
         return result
     }
 }

apps/iOS/MNNLLMChat/README-ZH.md

@@ -111,48 +111,23 @@ iPhone 因为内存有限，建议使用7B以及以下的模型，避免内存
 
 ## 本地调试
 
-如果我们希望直接电脑下载模型，不通过App内下载模型，进行调试，可以通过一下的方式。
+本地调试模型非常简单，只需要将模型文件拖动到LocalModel文件夹下，然后运行项目即可：
 
 1. 首先在 [huggingface](https://huggingface.co/taobao-mnn) 或者 [modelscope](https://modelscope.cn/organization/MNN) 下载 MNN 相关的模型
 
     <img width="400" alt="image" src="./assets/copyLocalModel.png" />
 
-
 2. 将下载之后的模型文件夹内的所有文件，拖动到项目中 LocalModel 文件夹下：
 
     <img width="200" alt="image" src="./assets/copyLocalModel2.png" />
 
-
 3. 确保以上文件都已经在 copy bundle resources 中
 
-
     <img width="400" alt="image" src="./assets/copyLocalMode3.png" />
 
+4. 运行项目，点击进入聊天对话页面，进行模型对话和调试。
 
-4. 注释下载相关代码
-
-    ```Swift
-    /*
-    try await modelClient.downloadModel(model: model) { progress in
-        Task { @MainActor in
-            DispatchQueue.main.async {
-                self.downloadProgress[model.modelId] = progress
-            }
-        }
-    }
-    */
-    ```
-5. 修改模型加载方式
-
-    在 LLMInferenceEngineWrapper 类中修改：
-
-    ```Swift
-    // BOOL success = [self loadModelFromPath:modelPath];
-    // MARK: Test Local Model
-    BOOL success = [self loadModel];
-    ```
-
-6. 运行项目，点击进入聊天对话页面，进行模型对话和调试。
+应用会自动检测并加载LocalModel文件夹中的模型，无需额外配置。
 
 
 ## Release Notes

apps/iOS/MNNLLMChat/README.md

@@ -125,14 +125,12 @@ Here is the professional technical translation of the provided text:
 
 ## Local Debugging
 
-If we want to directly download the models to the computer for debugging without downloading them through the app, we can follow these steps:
+For local debugging, simply drag the model files to the LocalModel folder and run the project:
 
 1. First, download the MNN-related models from [Hugging Face](https://huggingface.co/taobao-mnn) or [Modelscope](https://modelscope.cn/organization/MNN):
 
-
     <img width="400" alt="image" src="./assets/copyLocalModel.png" />
 
-
 2. After downloading, drag all the files from the model folder into the project's LocalModel folder:
 
     <img width="300" alt="image" src="./assets/copyLocalModel2.png" />
@@ -141,33 +139,9 @@ If we want to directly download the models to the computer for debugging without
 
     <img width="400" alt="image" src="./assets/copyLocalMode3.png" />
 
-4. Comment out the model download code:
+4. Run the project, navigate to the chat page, and perform model interactions and debugging.
 
-    ```Swift
-    /*
-    try await modelClient.downloadModel(model: model) { progress in
-        Task { @MainActor in
-            DispatchQueue.main.async {
-                self.downloadProgress[model.modelId] = progress
-            }
-        }
-    }
-    */
-    ```
-
-
-5. Modify the Model Loading Method
-
-    Modify the `LLMInferenceEngineWrapper` class:
-
-    ```Swift
-    // BOOL success = [self loadModelFromPath:modelPath];
-    // MARK: Test Local Model
-    BOOL success = [self loadModel];
-    ```
-
-
-6. Run the project, navigate to the chat page, and perform model interactions and debugging.
+The app will automatically detect and load models from the LocalModel folder without requiring additional configuration.
 
 ## Release Notes
 

docs/compile/cmake.md

@@ -16,11 +16,11 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_BUILD_QUANTOOLS  | 是否构建MNN的量化工具，默认为`OFF` |
 | MNN_EVALUATION       | 是否构建MNN的评估工具，默认为`OFF` |
 | MNN_BUILD_CONVERTER  | 是否构建MNN的转换工具，默认为`OFF` |
-| MNN_SUPPORT_QUNAT_EXTEND  | 是否编译非核心算子的量化版本，默认为`ON` |
-| MNN_SUPPORT_DEPRECATED_OP | 是否支持Tflite的量化算子等已经废弃的算子，用于兼容历史模型(1.1.0版本之前)，默认为`OFF` |
-| MNN_SUPPORT_DEPRECATED_OPV2 | 是否编译MNN更新到3.0之后已经废弃的算子，用于兼容历史模型(3.0.0版本之前)，比如 Convolution3D 和 ConvTranspose3D在3.0.0 版本之后改由模型转换器转化为对应2D算子，不再需要运行时支持，默认为`ON` |
-| MNN_REDUCE_SIZE  | 是否裁剪MNN库大小，去除求导相关算子，减少优化策略，默认为`OFF` ，开启时，MNN_SUPPORT_QUANT_EXTEND / MNN_SUPPORT_DEPRECATED_OP / MNN_SUPPORT_DEPRECATED_OPV2 都会设成 OFF|
-| MNN_SUPPORT_QUANT_EXTEND     | 是否开启Binary/Unary等算子的量化计算支持，默认为`ON` |
+| MNN_SUPPORT_QUANT_EXTEND    | 是否编译非核心算子的量化版本，默认为`ON` |
+| MNN_SUPPORT_DEPRECATED_OP   | 是否支持Tflite的量化算子等已经废弃的算子，用于兼容历史模型(1.1.0版本之前)，默认为`OFF` |
+| MNN_SUPPORT_DEPRECATED_OPV2 | 是否编译MNN更新到3.0之后已经废弃的算子，用于兼容历史模型(3.0.0版本之前)，比如 `Convolution3D` / `ConvTranspose3D` 在3.0.0 版本之后改由模型转换器转化为对应2D算子，不再需要运行时支持，默认为`ON` |
+| MNN_REDUCE_SIZE      | 是否裁剪MNN库大小，去除求导相关算子，减少优化策略，默认为`OFF` ，开启时，`MNN_SUPPORT_QUANT_EXTEND` / `MNN_SUPPORT_DEPRECATED_OP` / `MNN_SUPPORT_DEPRECATED_OPV2` / `MNN_USE_SPARSE_COMPUTE` 都会设成 `OFF` |
+| MNN_SUPPORT_QUANT_EXTEND    | 是否开启Binary/Unary等算子的量化计算支持，默认为`ON` |
 | MNN_DEBUG_MEMORY     | 是否开启MNN内存调试，默认为`OFF` |
 | MNN_DEBUG_TENSOR_SIZE | 是否开启MNN tensor size调试，默认为`OFF` |
 | MNN_GPU_TRACE        | 是否开启MNN GPU调试，默认为`OFF` |
@@ -30,7 +30,7 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_AAPL_FMWK        | 是否构建`MNN.framework`替代`*.dylib`，默认为`OFF` |
 | MNN_WITH_PLUGIN      | 是否支持`Plugin算子`，默认为`OFF` |
 | MNN_SKIPBUILD_GEOMETRY   | 是否跳过MNN的几何计算编译，若是，MNN引擎仅支持在模型转换工具时加上 --saveStaticModel 转换出来的固定输入形状的模型，默认为`OFF` |
-| MNN_BUILD_MINI       | 是否构建MNN的最小化版本，若是，开启 MNN_SKIPBUILD_GEOMETRY 和 MNN_REDUCE_SIZE，默认为`OFF` |
+| MNN_BUILD_MINI       | 是否构建MNN的最小化版本，若是，开启 `MNN_SKIPBUILD_GEOMETRY` 和 `MNN_REDUCE_SIZE`，默认为`OFF` |
 | MNN_USE_SSE          | 在x86上是否使用SSE指令集，默认为`OFF` |
 | MNN_BUILD_CODEGEN    | 是否构建MNN的代码生成部分，该功能提供了算子融合与代码生成能力，为实验性功能，默认为`OFF` |
 | MNN_ENABLE_COVERAGE  | 是否开启MNN的代码覆盖率，默认为`OFF` |
@@ -59,10 +59,12 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_NNAPI            | 是否构建`NNAPI`后端，默认为`OFF`  |
 | MNN_QNN              | 是否构建`QNN`后端，默认为`OFF` |
 | MNN_QNN_CONVERT_MODE | 在`MNN_QNN`开启的基础上,是否构建Convert模式的QNN后端，默认为`OFF` |
+| MNN_NPU            | 是否构建`NPU`后端，默认为`OFF`  |
+| MNN_USE_SPARSE_COMPUTE | 是否使用稀疏计算，默认为`ON` |
 | MNN_BUILD_BENCHMARK  | 是否构建MNN的性能测试，默认为`OFF` |
 | MNN_BUILD_TEST       | 是否构建MNN的单元测试，默认为`OFF` |
 | MNN_BUILD_FOR_ANDROID_COMMAND | 是否使用命令行构建`Android`，默认为`OFF` |
-| MNN_USE_LOGCAT       | 是否使用`logcat`代替`printf`输出日志，默认为`OFF` |
+| MNN_USE_LOGCAT       | 是否使用`logcat`代替`printf`输出日志，默认为`ON` |
 | MNN_USE_CPP11        | 是否使用`C++11`编译MNN，默认为`ON` |
 | MNN_SUPPORT_BF16     | 是否支持`BF16`，默认为`OFF` |
 | MNN_SSE_USE_FP16_INSTEAD | 在X86平台是否使用`FP16`替代`BF16`，默认为`OFF` |
@@ -72,7 +74,7 @@ MNN使用CMake构建项目，CMake中的宏定义列表如下：
 | MNN_METALLIB_SOURCE  | 使用Metal时是否直接使用Metal源码，该宏仅在`MNN_METAL=ON`时生效，默认为`ON` |
 | MNN_VULKAN_DEBUG     | 是否打开Vulkan的DEBUG模式，该宏仅在`MNN_VULKAN=ON`时生效，默认为`OFF` |
 | MNN_OPENGL_REGEN     | 是否重新生成OpenGL Kenel，该宏仅在`MNN_OPENGL=ON`时生效，默认为`OFF` |
-| MNN_TRT_DYNAMIC      | 是否通过dlopen的方式引入TRT的动态库，该宏仅在`MNN_TENSORRT=ON`时生效，默认为`OFF |
+| MNN_TRT_DYNAMIC      | 是否通过dlopen的方式引入TRT的动态库，该宏仅在`MNN_TENSORRT=ON`时生效，默认为`OFF` |
 | MNN_BUILD_TORCH      | 构建的`MNNConvert`是否支持`TorchScript`，该宏仅在`MNN_BUILD_CONVERTER=ON`时生效，默认为`OFF` |
 | MNN_TRAIN_DEBUG      | 构建的训练模块是否支持调试，该宏仅在`MNN_BUILD_TRAIN=ON`时生效，默认为`OFF` |
 | MNN_USE_OPENCV       | 构建的训练Demo是否使用`OpenCV`依赖，该宏仅在`MNN_BUILD_TRAIN=ON`时生效，默认为`OFF` |

docs/compile/other.md

@@ -56,11 +56,11 @@
   - `llm_bench` 大语言模型测评工具
 ## 测试工具
 - 相关编译选项
-  - `MNN_BUILD_TOOL` 是否编译测试工具
+  - `MNN_BUILD_TOOLS` 是否编译测试工具
 - 编译命令
     ```bash
     mkdir build && cd build
-    cmake .. -DMNN_BUILD_TOOL=ON
+    cmake .. -DMNN_BUILD_TOOLS=ON
     make -j4
     ```
 - 编译产物

docs/compile/pymnn.md

@@ -2,7 +2,7 @@
 ## 本地安装
 ```bash
 cd /path/to/MNN/pymnn/pip_package
-python build_deps.py {MNN依赖包组合} #internal,cuda,trt,cuda_tune,opencl,vulkan,render,no_sse,torch这几个字符串的任意组合，例如字符串可为:"cuda,reder,no_sse"
+python build_deps.py {MNN依赖包组合} #internal,cuda,trt,cuda_tune,opencl,vulkan,render,no_sse,torch,openmp,llm这几个字符串的任意组合，例如字符串可为:"cuda,reder,no_sse"
 python setup.py install --version {MNN版本} --deps {MNN依赖包组合}
 ```
 ## 构建Python Wheel包

docs/inference/module.md

@@ -92,7 +92,8 @@ rtmgr->setMode(Interpreter::Session_Debug);
 
 - Interpreter::HintMode::WINOGRAD_MEMORY_LEVEL ：使用 Winograd 算法优化卷积时，内存占用倾向，默认为 3 ，若希望降低内存占用可设为 0 
 - Interpreter::HintMode::GEOMETRY_COMPUTE_MASK ：几何计算相关优化开关，1为区域合并，2为复合区域合并，4为使用loop算子，8为支持几何计算重计算，需要多个功能开启时把对应值叠加。默认为功能全开。
-- Interpreter::HintMode::CPU_LITTLECORE_DECREASE_RATE ：对于 Android 设备存在大中小核的情况，大核算力到中核算力的衰减比例。默认为50（中核算力为大核的50%）
+- Interpreter::HintMode::CPU_LITTLECORE_DECREASE_RATE ：对于 Android 设备存在大中小核的情况，设置大核与小核之间的算力衰减比例，用于任务调度。默认值为50，表示小核的算力是大核的50%。MNN会根据这个比例来决定在大小核上分配的计算任务量。这个参数**并不直接绑定**线程到特定核心，而是影响任务分配策略。
+- Interpreter::HintMode::CPU_CORE_IDS ：直接将MNN的计算任务绑定到指定的CPU核心上。这是一个更强力的控制方式，可以精确控制MNN使用的CPU资源。详细用法请参考 [Session API使用 - CPU 核心绑定](../inference/session.md#cpu-核心绑定)。
 
 
 #### ExternalPath

docs/inference/python.md

@@ -47,7 +47,7 @@ import MNN.cv as cv
 import MNN.numpy as np
 import MNN.expr as expr
 
-# 配置执行后端，线程数，精度等信息；key-vlaue请查看API介绍
+# 配置执行后端，线程数，精度等信息；key-value请查看API介绍
 config = {}
 config['precision'] = 'low' # 当硬件支持（armv8.2）时使用fp16推理
 config['backend'] = 0       # CPU

docs/inference/session.md

@@ -200,6 +200,39 @@ struct BackendConfig {
 
 `sharedContext`用于自定义后端，用户可以根据自身需要赋值。
 
+#### CPU 核心绑定
+MNN 支持将计算任务绑定到指定的 CPU 核心上执行。这对于需要精细控制 CPU 资源，避免线程在不同核心之间切换，或者希望将 MNN 的计算任务限制在特定核心上，以减少对其他应用干扰的场景非常有用。
+
+通过 `Interpreter::setSessionHint` 方法，并使用 `HintMode::CPU_CORE_IDS`，可以指定一个或多个 CPU 核心的 ID。
+
+```cpp
+#include <MNN/Interpreter.hpp>
+
+// ...
+
+// 创建 Interpreter
+auto interpreter = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile("your_model.mnn"));
+
+// 设置 CPU 核心绑定
+// 假设我们希望将计算任务绑定到 0 号和 1 号 CPU 核心
+std::vector<int> cpu_ids = {0, 1};
+interpreter->setSessionHint(MNN::Interpreter::HintMode::CPU_CORE_IDS, cpu_ids.data(), cpu_ids.size());
+
+// 创建 Session
+MNN::ScheduleConfig config;
+config.type = MNN_FORWARD_CPU;
+config.numThread = 2; // 线程数最好和绑定的核心数一致
+auto session = interpreter->createSession(config);
+
+// ... 运行推理
+```
+
+**注意事项:**
+
+*   `CPU_CORE_IDS` 的设置必须在 `createSession` 之前完成。
+*   `numThread` 的数量最好设置为与绑定的 CPU 核心数量一致，以达到最佳的性能。
+*   如果指定的 CPU 核心 ID 不存在或无效，MNN 将会忽略该配置，并使用默认的线程调度策略。
+
 ### 创建多段路径Session
 需要对推理路径做出更为复杂的配置时，可以通过调度配置组来实现：
 ```cpp

docs/pymnn/MNN.md

@@ -11,9 +11,11 @@ MNN是Pymnn中最基础的Module，其中包含了V2 API所需要数据结构与
 - [expr](expr.md)
 - [numpy](numpy.md)
 - [cv](cv.md)
-- [nn](nn.md)
+- [nn](nn.md): 包含 `loss` 和 `compress` 子模块
 - [optim](optim.md)
 - [data](data.md)
+- [audio](audio.md)
+- [llm](llm.md)
 
 ---
 ### `MNN Types`

docs/pymnn/audio.md

@@ -0,0 +1,3 @@
+# audio
+
+This document is a placeholder.

docs/pymnn/llm.md

@@ -0,0 +1,3 @@
+# llm
+
+This document is a placeholder.

source/backend/cpu/arm/CMakeLists.txt

@@ -103,7 +103,7 @@ if (MNN_KLEIDIAI)
 	${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.c
     )
 
-    set_source_files_properties(${MNN_SOURCES_KLEIDIAI} PROPERTIES COMPILE_OPTIONS -march=armv8.2-a+i8mm+dotprod+sve+sve2+fp16)
+    set_source_files_properties(${MNN_SOURCES_KLEIDIAI} PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;-march=armv8.2-a+i8mm+dotprod+sve+sve2+fp16")
     set_source_files_properties(${KLEIDIAI_FILES_SME2}  PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;-march=armv8.2-a+sve+sve2")
 
 endif()

source/backend/cpu/arm/mnn_kleidiai.h

@@ -211,6 +211,11 @@ namespace MNN {
                 return false;
             }
         }
+        if(type == AccelType::QI4_SYM_CHNLQT_F32){
+            if(common->inputCount() % 2 != 0) {
+                return false;
+            }
+        }
         if(common->kernelX() == 1 && common->kernelY() == 1
             && common->padX() == 0 && common->padY() == 0
             && common->strideX() == 1 && common->strideY() == 1

source/backend/cpu/compute/KleidiAIConvolution.cpp

@@ -175,8 +175,11 @@ ErrorCode KleidiAIConvolution::onResize(const std::vector<Tensor *> &inputs, con
     if (outputOriginFmt != MNN_DATA_FORMAT_NHWC){
         b->onReleaseBuffer(mOutputConvertBuffer.get(), Backend::DYNAMIC);
     }
+
+    mPostParameters = getPostParameters();
     return NO_ERROR;
 }
+
 ErrorCode KleidiAIConvolution::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
     auto input  = inputs[0];
     auto output = outputs[0];
@@ -209,13 +212,12 @@ ErrorCode KleidiAIConvolution::onExecute(const std::vector<Tensor *> &inputs, co
     }
 
     auto outputDes = TensorUtils::getDescribe(outputs[0]);
-    auto postPtr = getPostParameters();
     auto outputPtr = output->host<uint8_t>();
     if(outputDes->dimensionFormat != MNN_DATA_FORMAT_NHWC){
         outputPtr = mOutputConvertBuffer->host<uint8_t>();
     }
 
-    kai.runMatmul(mAccelType, m, n, k, 0, lhsPacked, weightPtr, outputPtr, n * elementSize, elementSize, postPtr[3], postPtr[2]);
+    kai.runMatmul(mAccelType, m, n, k, 0, lhsPacked, weightPtr, outputPtr, n * elementSize, elementSize, mPostParameters[3], mPostParameters[2]);
 
     if(outputDes->dimensionFormat != MNN_DATA_FORMAT_NHWC){
         MNN_CONCURRENCY_BEGIN(tId, threadNum) {

source/backend/cpu/compute/KleidiAIConvolution.hpp

@@ -28,7 +28,7 @@ class KleidiAIConvolution : public CPUConvolution{
         std::shared_ptr<Tensor> mOutputConvertBuffer;
         std::shared_ptr<CPUConvolution::Resource> mResource;
         KleidiAI::AccelType mAccelType = KleidiAI::AccelType::ACC_TYPE_NUMBER;
-
+        std::vector<float> mPostParameters;
 };
 #endif //MNN_KLEIDIAI_ENABLED
 

source/backend/cpu/compute/KleidiAIDenseConvolution.hpp

@@ -37,7 +37,7 @@ struct ConvParams {
         int dilatedKernelWidth  = kernelSizeWithDilated(kernelWidth, dilatedWidth);
 
         int outputHeight = outputSize(inputHeight, padTop, padBottom, dilatedKernelHeight, strideHeight);
-        int outputWidth  = outputSize(inputHeight, padLeft, padRight, dilatedKernelWidth, strideWidth);
+        int outputWidth  = outputSize(inputWidth, padLeft, padRight, dilatedKernelWidth, strideWidth);
 
         return {outputHeight, outputWidth};
     }