llama podcast

"-ml" instead of "-mg" for specifying the llama file

.gitignore

@@ -1,7 +1,5 @@
 *.o
 *.a
-*.mlmodel
-*.mlmodelc
 .cache/
 .vs/
 .vscode/
@@ -20,6 +18,7 @@ build-sanitize-thread/
 /stream
 /command
 /talk
+/talk-llama
 /bench
 
 arm_neon.h

CMakeLists.txt

@@ -54,8 +54,6 @@ if (APPLE)
     option(WHISPER_NO_AVX              "whisper: disable AVX" OFF)
     option(WHISPER_NO_AVX2             "whisper: disable AVX2" OFF)
     option(WHISPER_NO_FMA              "whisper: disable FMA" OFF)
-
-    option(WHISPER_COREML              "whisper: enable Core ML framework" OFF)
 else()
     option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
 endif()
@@ -88,12 +86,9 @@ endif()
 
 find_package(Threads REQUIRED)
 
-# on APPLE
+# on APPLE - include Accelerate framework
-if (APPLE)
+if (APPLE AND NOT WHISPER_NO_ACCELERATE)
-    # include Accelerate framework
-    if (NOT WHISPER_NO_ACCELERATE)
     find_library(ACCELERATE_FRAMEWORK Accelerate)
-
     if (ACCELERATE_FRAMEWORK)
         message(STATUS "Accelerate framework found")
 
@@ -102,20 +97,6 @@ if (APPLE)
     else()
         message(WARNING "Accelerate framework not found")
     endif()
-    endif()
-
-    if (WHISPER_COREML)
-        find_library(FOUNDATION_FRAMEWORK Foundation)
-        find_library(COREML_FRAMEWORK CoreML)
-
-        if (COREML_FRAMEWORK)
-            message(STATUS "CoreML framework found")
-
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
-        else()
-            message(WARNING "CoreML framework not found")
-        endif()
-    endif()
 endif()
 
 if (WHISPER_SUPPORT_OPENBLAS)
@@ -202,33 +183,6 @@ if (WHISPER_PERF)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
 
-#
-# whisper.coreml - Core ML support
-#
-
-if (WHISPER_COREML)
-    set(TARGET whisper.coreml)
-
-    add_library(${TARGET}
-        coreml/whisper-encoder.h
-        coreml/whisper-encoder.mm
-        coreml/whisper-encoder-impl.h
-        coreml/whisper-encoder-impl.m
-        )
-
-    include(DefaultTargetOptions)
-
-    target_include_directories(${TARGET} PUBLIC
-        .
-        )
-
-    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
-
-    set_target_properties(${TARGET} PROPERTIES
-        COMPILE_FLAGS "-fobjc-arc"
-        )
-endif()
-
 #
 # whisper - this is the main library of the project
 #
@@ -248,10 +202,6 @@ target_include_directories(${TARGET} PUBLIC
     .
     )
 
-if (WHISPER_COREML)
-    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
-endif()
-
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 

Makefile

@@ -138,10 +138,6 @@ ifndef WHISPER_NO_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
 endif
-ifdef WHISPER_COREML
-	CXXFLAGS += -DWHISPER_USE_COREML
-	LDFLAGS  += -framework Foundation -framework CoreML
-endif
 ifdef WHISPER_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
 	LDFLAGS += -lopenblas
@@ -155,12 +151,15 @@ ifneq ($(filter aarch64%,$(UNAME_M)),)
 	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, 2, 3
+	# 32-bit Raspberry Pi 1, 2, 3
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access
 endif
 ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 4
+	# 32-bit ARM, for example on Armbian or possibly raspbian
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	
+	# 64-bit ARM, use these (TODO: auto-detect 64-bit)
+	# CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
@@ -182,7 +181,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main
+default: main bench
 
 #
 # Build library
@@ -194,26 +193,14 @@ ggml.o: ggml.c ggml.h
 whisper.o: whisper.cpp whisper.h
 	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o
 
-ifndef WHISPER_COREML
+libwhisper.a: ggml.o whisper.o
-WHISPER_OBJ = whisper.o
+	$(AR) rcs libwhisper.a ggml.o whisper.o
-else
-whisper-encoder.o: coreml/whisper-encoder.mm coreml/whisper-encoder.h
-	$(CXX) -O3 -I . -c coreml/whisper-encoder.mm -o whisper-encoder.o
 
-whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-impl.h
+libwhisper.so: ggml.o whisper.o
-	$(CXX) -O3 -I . -fobjc-arc -c coreml/whisper-encoder-impl.m -o whisper-encoder-impl.o
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o whisper.o $(LDFLAGS)
-
-WHISPER_OBJ = whisper.o whisper-encoder.o whisper-encoder-impl.o
-endif
-
-libwhisper.a: ggml.o $(WHISPER_OBJ)
-	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
-
-libwhisper.so: ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -224,21 +211,24 @@ CC_SDL=`sdl2-config --cflags --libs`
 SRC_COMMON = examples/common.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp
 
-main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o -o main $(LDFLAGS)
 	./main -h
 
-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
 
-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o command $(CC_SDL) $(LDFLAGS)
 
-bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples

README.md

@@ -313,7 +313,7 @@ whisper_print_timings:    total time = 32733.52 ms
 ## Real-time audio input example
 
 This is a naive example of performing real-time inference on audio from your microphone.
-The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continously.
+The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```java
@@ -500,6 +500,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
 | [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
 | [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
+| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
 | [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
 | [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
 | [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |

coreml/whisper-encoder-impl.h

@@ -1,142 +0,0 @@
-//
-// CoremlEncoder.h
-//
-// This file was automatically generated and should not be edited.
-//
-
-#import <Foundation/Foundation.h>
-#import <CoreML/CoreML.h>
-#include <stdint.h>
-#include <os/log.h>
-
-NS_ASSUME_NONNULL_BEGIN
-
-
-/// Model Prediction Input Type
-API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
-@interface CoremlEncoderInput : NSObject<MLFeatureProvider>
-
-/// melSegment as 1 × 80 × 3000 3-dimensional array of floats
-@property (readwrite, nonatomic, strong) MLMultiArray * melSegment;
-- (instancetype)init NS_UNAVAILABLE;
-- (instancetype)initWithMelSegment:(MLMultiArray *)melSegment NS_DESIGNATED_INITIALIZER;
-
-@end
-
-
-/// Model Prediction Output Type
-API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
-@interface CoremlEncoderOutput : NSObject<MLFeatureProvider>
-
-/// output as multidimensional array of floats
-@property (readwrite, nonatomic, strong) MLMultiArray * output;
-- (instancetype)init NS_UNAVAILABLE;
-- (instancetype)initWithOutput:(MLMultiArray *)output NS_DESIGNATED_INITIALIZER;
-
-@end
-
-
-/// Class for model loading and prediction
-API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
-@interface CoremlEncoder : NSObject
-@property (readonly, nonatomic, nullable) MLModel * model;
-
-/**
-    URL of the underlying .mlmodelc directory.
-*/
-+ (nullable NSURL *)URLOfModelInThisBundle;
-
-/**
-    Initialize CoremlEncoder instance from an existing MLModel object.
-
-    Usually the application does not use this initializer unless it makes a subclass of CoremlEncoder.
-    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
-*/
-- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
-
-/**
-    Initialize CoremlEncoder instance with the model in this bundle.
-*/
-- (nullable instancetype)init;
-
-/**
-    Initialize CoremlEncoder instance with the model in this bundle.
-
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
-- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Initialize CoremlEncoder instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
-- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Initialize CoremlEncoder instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
-- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Construct CoremlEncoder instance asynchronously with configuration.
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
-*/
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));
-
-/**
-    Construct CoremlEncoder instance asynchronously with URL of .mlmodelc directory and optional configuration.
-
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param modelURL The model URL.
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
-*/
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));
-
-/**
-    Make a prediction using the standard interface
-    @param input an instance of CoremlEncoderInput to predict from
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as CoremlEncoderOutput
-*/
-- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Make a prediction using the standard interface
-    @param input an instance of CoremlEncoderInput to predict from
-    @param options prediction options
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as CoremlEncoderOutput
-*/
-- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Make a prediction using the convenience interface
-    @param melSegment as 1 × 80 × 3000 3-dimensional array of floats:
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as CoremlEncoderOutput
-*/
-- (nullable CoremlEncoderOutput *)predictionFromMelSegment:(MLMultiArray *)melSegment error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Batch prediction
-    @param inputArray array of CoremlEncoderInput instances to obtain predictions from
-    @param options prediction options
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the predictions as NSArray<CoremlEncoderOutput *>
-*/
-- (nullable NSArray<CoremlEncoderOutput *> *)predictionsFromInputs:(NSArray<CoremlEncoderInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-@end
-
-NS_ASSUME_NONNULL_END

coreml/whisper-encoder-impl.m

@@ -1,197 +0,0 @@
-//
-// CoremlEncoder.m
-//
-// This file was automatically generated and should not be edited.
-//
-
-#if !__has_feature(objc_arc)
-#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
-#endif
-
-#import "whisper-encoder-impl.h"
-
-@implementation CoremlEncoderInput
-
-- (instancetype)initWithMelSegment:(MLMultiArray *)melSegment {
-    self = [super init];
-    if (self) {
-        _melSegment = melSegment;
-    }
-    return self;
-}
-
-- (NSSet<NSString *> *)featureNames {
-    return [NSSet setWithArray:@[@"melSegment"]];
-}
-
-- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
-    if ([featureName isEqualToString:@"melSegment"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.melSegment];
-    }
-    return nil;
-}
-
-@end
-
-@implementation CoremlEncoderOutput
-
-- (instancetype)initWithOutput:(MLMultiArray *)output {
-    self = [super init];
-    if (self) {
-        _output = output;
-    }
-    return self;
-}
-
-- (NSSet<NSString *> *)featureNames {
-    return [NSSet setWithArray:@[@"output"]];
-}
-
-- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
-    if ([featureName isEqualToString:@"output"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.output];
-    }
-    return nil;
-}
-
-@end
-
-@implementation CoremlEncoder
-
-
-/**
-    URL of the underlying .mlmodelc directory.
-*/
-+ (nullable NSURL *)URLOfModelInThisBundle {
-    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"CoremlEncoder" ofType:@"mlmodelc"];
-    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load CoremlEncoder.mlmodelc in the bundle resource"); return nil; }
-    return [NSURL fileURLWithPath:assetPath];
-}
-
-
-/**
-    Initialize CoremlEncoder instance from an existing MLModel object.
-
-    Usually the application does not use this initializer unless it makes a subclass of CoremlEncoder.
-    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
-*/
-- (instancetype)initWithMLModel:(MLModel *)model {
-    self = [super init];
-    if (!self) { return nil; }
-    _model = model;
-    if (_model == nil) { return nil; }
-    return self;
-}
-
-
-/**
-    Initialize CoremlEncoder instance with the model in this bundle.
-*/
-- (nullable instancetype)init {
-    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
-}
-
-
-/**
-    Initialize CoremlEncoder instance with the model in this bundle.
-
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
-- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
-}
-
-
-/**
-    Initialize CoremlEncoder instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
-- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
-    if (model == nil) { return nil; }
-    return [self initWithMLModel:model];
-}
-
-
-/**
-    Initialize CoremlEncoder instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
-- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
-    if (model == nil) { return nil; }
-    return [self initWithMLModel:model];
-}
-
-
-/**
-    Construct CoremlEncoder instance asynchronously with configuration.
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
-*/
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler {
-    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
-              configuration:configuration
-          completionHandler:handler];
-}
-
-
-/**
-    Construct CoremlEncoder instance asynchronously with URL of .mlmodelc directory and optional configuration.
-
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param modelURL The model URL.
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
-*/
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler {
-    [MLModel loadContentsOfURL:modelURL
-                 configuration:configuration
-             completionHandler:^(MLModel *model, NSError *error) {
-        if (model != nil) {
-            CoremlEncoder *typedModel = [[CoremlEncoder alloc] initWithMLModel:model];
-            handler(typedModel, nil);
-        } else {
-            handler(nil, error);
-        }
-    }];
-}
-
-- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
-}
-
-- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
-    if (!outFeatures) { return nil; }
-    return [[CoremlEncoderOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
-}
-
-- (nullable CoremlEncoderOutput *)predictionFromMelSegment:(MLMultiArray *)melSegment error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    CoremlEncoderInput *input_ = [[CoremlEncoderInput alloc] initWithMelSegment:melSegment];
-    return [self predictionFromFeatures:input_ error:error];
-}
-
-- (nullable NSArray<CoremlEncoderOutput *> *)predictionsFromInputs:(NSArray<CoremlEncoderInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
-    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
-    if (!outBatch) { return nil; }
-    NSMutableArray<CoremlEncoderOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
-    for (NSInteger i = 0; i < outBatch.count; i++) {
-        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
-        CoremlEncoderOutput * result = [[CoremlEncoderOutput alloc] initWithOutput:(MLMultiArray *)[resultProvider featureValueForName:@"output"].multiArrayValue];
-        [results addObject:result];
-    }
-    return results;
-}
-
-@end

coreml/whisper-encoder.h

@@ -1,22 +0,0 @@
-// Wrapper of the Core ML Whisper Encoder model
-//
-// Code is derived from the work of Github user @wangchou
-// ref: https://github.com/wangchou/callCoreMLFromCpp
-
-#if __cplusplus
-extern "C" {
-#endif
-
-struct whisper_coreml_context;
-
-struct whisper_coreml_context * whisper_coreml_init(const char * path_model);
-void whisper_coreml_free(struct whisper_coreml_context * ctx);
-
-void whisper_coreml_encode(
-        const whisper_coreml_context * ctx,
-                               float * mel,
-                               float * out);
-
-#if __cplusplus
-}
-#endif

coreml/whisper-encoder.mm

@@ -1,61 +0,0 @@
-#import "coreml/whisper-encoder.h"
-#import "coreml/whisper-encoder-impl.h"
-
-#import <CoreML/CoreML.h>
-
-#include <stdlib.h>
-
-#if __cplusplus
-extern "C" {
-#endif
-
-struct whisper_coreml_context {
-    const void * data;
-};
-
-struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
-    NSString * path_model_str = [[NSString alloc] initWithUTF8String:path_model];
-
-    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
-
-    const void * data = CFBridgingRetain([[CoremlEncoder alloc] initWithContentsOfURL:url_model error:nil]);
-
-    if (data == NULL) {
-        return NULL;
-    }
-
-    whisper_coreml_context * ctx = new whisper_coreml_context;
-
-    ctx->data = data;
-
-    return ctx;
-}
-
-void whisper_coreml_free(struct whisper_coreml_context * ctx) {
-    CFRelease(ctx->data);
-    delete ctx;
-}
-
-void whisper_coreml_encode(
-        const whisper_coreml_context * ctx,
-                               float * mel,
-                               float * out) {
-    MLMultiArray * inMultiArray = [
-        [MLMultiArray alloc] initWithDataPointer: mel
-                                           shape: @[@1, @80, @3000]
-                                        dataType: MLMultiArrayDataTypeFloat32
-                                         strides: @[@(240000), @(3000), @1]
-                                     deallocator: nil
-                                           error: nil
-    ];
-
-    CoremlEncoderOutput * outCoreML = [(__bridge id) ctx->data predictionFromMelSegment:inMultiArray error:nil];
-
-    MLMultiArray * outMA = outCoreML.output;
-
-    memcpy(out, outMA.dataPointer, outMA.count * sizeof(float));
-}
-
-#if __cplusplus
-}
-#endif

examples/CMakeLists.txt

@@ -63,5 +63,5 @@ else()
     add_subdirectory(command)
     add_subdirectory(bench)
     add_subdirectory(talk)
-    add_subdirectory(talk.llama)
+    add_subdirectory(talk-llama)
 endif()

examples/addon.node/__test__/whisper.spec.js

@@ -1,15 +1,22 @@
-const path = require('path');
+const path = require("path");
-const { whisper } = require(path.join(__dirname, '../../../build/Release/whisper-addon'));
+const { whisper } = require(path.join(
+  __dirname,
+  "../../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
 
 const whisperParamsMock = {
-    language: 'en',
+  language: "en",
-    model: path.join(__dirname, '../../../models/ggml-base.en.bin'),
+  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
-    fname_inp: path.join(__dirname, '../../../samples/jfk.wav'),
+  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
 };
 
 describe("Run whisper.node", () => {
+  test("it should receive a non-empty value", async () => {
+    let result = await whisperAsync(whisperParamsMock);
 
-    test("it should receive a non-empty value", () => {
+    expect(result.length).toBeGreaterThan(0);
-        expect(whisper(whisperParamsMock).length).toBeGreaterThan(0);
   });
 });

examples/addon.node/addon.cpp

@@ -160,22 +160,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
-
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
         const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -243,8 +227,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
-            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.initial_prompt   = params.prompt.c_str();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
 
             whisper_print_user_data user_data = { &params, &pcmf32s };
 

examples/main/main.cpp

@@ -8,6 +8,7 @@
 #include <string>
 #include <thread>
 #include <vector>
+#include <cstring>
 
 // Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
 // Lowest is red, middle is yellow, highest is green.
@@ -371,6 +372,39 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
     return true;
 }
 
+char *escape_double_quotes(const char *str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped[pos++] = '\\';
+            escaped[pos++] = '"';
+        } else {
+            escaped[pos++] = str[i];
+        }
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
 bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
     std::ofstream fout(fname);
     int indent = 0;
@@ -414,7 +448,9 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
 
     auto value_s = [&](const char *name, const char *val, bool end = false) {
         start_value(name);
-        fout << "\"" << val << (end ? "\"\n" : "\",\n");
+        char * val_escaped = escape_double_quotes(val);
+        fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
+        free(val_escaped);
     };
 
     auto end_value = [&](bool end = false) {
@@ -455,7 +491,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
                 value_i("ctx", whisper_model_n_text_ctx(ctx));
                 value_i("state", whisper_model_n_text_state(ctx));
                 value_i("head", whisper_model_n_text_head(ctx));
-                value_i("leyer", whisper_model_n_text_layer(ctx), true);
+                value_i("layer", whisper_model_n_text_layer(ctx), true);
             end_obj();
             value_i("mels", whisper_model_n_mels(ctx));
             value_i("f16", whisper_model_f16(ctx), true);
@@ -477,7 +513,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
                 const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
                 start_obj();
-                    start_obj("timestanps");
+                    start_obj("timestamps");
                         value_s("from", to_timestamp(t0, true).c_str());
                         value_s("to", to_timestamp(t1, true).c_str(), true);
                     end_obj();
@@ -639,22 +675,6 @@ int main(int argc, char ** argv) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
-
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -718,8 +738,7 @@ int main(int argc, char ** argv) {
 
             wparams.speed_up         = params.speed_up;
 
-            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.initial_prompt   = params.prompt.c_str();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;

examples/talk-llama/CMakeLists.txt

@@ -0,0 +1,16 @@
+if (WHISPER_SUPPORT_SDL2)
+    # talk-llama
+    set(TARGET talk-llama)
+    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    include(DefaultTargetOptions)
+endif ()

examples/talk-llama/README.md

@@ -0,0 +1,36 @@
+# talk-llama
+
+Talk with an LLaMA AI in your terminal
+
+[Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
+
+## Building
+
+The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+# Build the "talk-llama" executable
+make talk-llama
+
+# Run it
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+```
+
+- The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
+- The `-ml` argument specifies the LLaMA model that you would like to use. Read the instructions in https://github.com/ggerganov/llama.cpp for information about how to obtain a `ggml` compatible LLaMA model
+
+## TTS
+
+For best experience, this example needs a TTS tool to convert the generated text responses to voice.
+You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
+By default, it is configured to use MacOS's `say`, but you can use whatever you wish.
+
+## Discussion
+
+If you have any feedback, please let "us" know in the following discussion: https://github.com/ggerganov/whisper.cpp/discussions/672?converting=1

examples/talk-llama/llama.cpp

@@ -12,6 +12,19 @@
 #include <cassert>
 #include <cstring>
 
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#else
+#include <sys/types.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include <fcntl.h>
+#endif
+
+#define Min(X, Y) ((Y) > (X) ? (X) : (Y))
+#define Max(X, Y) ((Y) < (X) ? (X) : (Y))
+
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
 
@@ -142,6 +155,10 @@ struct llama_model {
     // the model memory buffer
     std::vector<uint8_t> buf;
 
+    // model memory mapped file
+    void * mm_addr = NULL;
+    uint64_t mm_length = 0;
+
     // tensors
     int n_loaded;
     std::unordered_map<std::string, struct ggml_tensor *> tensors;
@@ -165,6 +182,7 @@ struct llama_context {
 
     int64_t t_load_us = 0;
     int64_t t_start_us = 0;
+    bool has_evaluated_once = false;
 
     int64_t t_sample_us = 0;
     int64_t t_eval_us   = 0;
@@ -206,7 +224,7 @@ struct llama_context {
         }
 
         if (buf_last >= 0) {
-            buf_max_size[buf_last] = std::max(buf_max_size[buf_last], last_size);
+            buf_max_size[buf_last] = Max(buf_max_size[buf_last], last_size);
         }
 
         buf_last = i;
@@ -246,6 +264,7 @@ static bool kv_cache_init(
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size();
     params.mem_buffer = cache.buf.data();
+    params.no_alloc   = false;
 
     cache.ctx = ggml_init(params);
 
@@ -288,6 +307,58 @@ struct llama_context_params llama_context_default_params() {
 // model loading
 //
 
+static void *mmap_file(const char *fname, uint64_t *mm_length) {
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+    HANDLE hFile = CreateFileA(fname,
+                               GENERIC_READ,
+                               FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE,
+                               NULL,
+                               OPEN_EXISTING,
+                               FILE_ATTRIBUTE_NORMAL | FILE_ATTRIBUTE_NOT_CONTENT_INDEXED,
+                               NULL);
+    if (hFile == INVALID_HANDLE_VALUE) return 0;
+    LARGE_INTEGER fileSize;
+    fileSize.QuadPart = -1;
+    GetFileSizeEx(hFile, &fileSize);
+    int64_t length = fileSize.QuadPart;
+    HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+    CloseHandle(hFile);
+    if (!hMapping) return 0;
+    void *addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+    CloseHandle(hMapping);
+    if (!addr) return 0;
+#else
+    int fd = open(fname, O_RDONLY);
+    if (fd == -1) return 0;
+    int64_t length = lseek(fd, 0, SEEK_END);
+    void *addr = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, 0);
+    close(fd);
+    if (addr == MAP_FAILED) return 0;
+#endif
+    *mm_length = length;
+    return addr;
+}
+
+static void munmap_file(void * addr, size_t length) {
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+    UnmapViewOfFile(addr);
+#else
+    munmap(addr, length);
+#endif
+}
+
+static bool report_bad_magic(const char *path, uint32_t got, uint32_t want) {
+    fprintf(stderr,
+            "%s: invalid model file (bad magic [got %#x want %#x])\n"
+            "\tyou most likely need to regenerate your ggml files\n"
+            "\tthe benefit is you'll get 10-100x faster load times\n"
+            "\tsee https://github.com/ggerganov/llama.cpp/issues/91\n"
+            "\tuse convert-pth-to-ggml.py to regenerate from original pth\n"
+            "\tuse migrate-ggml-2023-03-30-pr613.py if you deleted originals\n",
+            path, got, want);
+    return false;
+}
+
 static bool llama_model_load(
         const std::string & fname,
         llama_context & lctx,
@@ -299,34 +370,35 @@ static bool llama_model_load(
         void *progress_callback_user_data) {
     fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
 
-    const int64_t t_start_us = ggml_time_us();
+    lctx.t_start_us = ggml_time_us();
-
-    lctx.t_start_us = t_start_us;
-
-    std::vector<char> f_buf(1024*1024);
 
     auto & model = lctx.model;
     auto & vocab = lctx.vocab;
 
     auto fin = std::ifstream(fname, std::ios::binary);
-    fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
     if (!fin) {
         fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
         return false;
     }
 
+    std::vector<char> f_buf(1024*1024);
+    fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
+
+    fin.seekg(0, fin.end);
+    const size_t file_size = fin.tellg();
+    fin.seekg(0);
+
     // verify magic
     {
         uint32_t magic;
         fin.read((char *) &magic, sizeof(magic));
         if (magic == LLAMA_FILE_MAGIC_UNVERSIONED) {
-            fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files!)\n",
+            fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files or convert them with convert-unversioned-ggml-to-ggml.py!)\n",
                     __func__, fname.c_str());
             return false;
         }
         if (magic != LLAMA_FILE_MAGIC) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return report_bad_magic(fname.c_str(), magic, LLAMA_FILE_MAGIC);
-            return false;
         }
 
         uint32_t format_version;
@@ -449,43 +521,24 @@ static bool llama_model_load(
                 }
     }
 
+    // map model into memory
+    char *mm_addr = NULL;
+    model.mm_addr = mmap_file(fname.c_str(), &model.mm_length);
+    if (model.mm_addr == NULL) {
+        fprintf(stderr, "%s: failed to mmap '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+    mm_addr = (char *)model.mm_addr;
+    fprintf(stderr, "%s: ggml map size = %6.2f MB\n", __func__, model.mm_length/(1024.0*1024.0));
+
     auto & ctx = model.ctx;
 
     size_t ctx_size = 0;
-
     {
-        const auto & hparams = model.hparams;
+        const auto &hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
         const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(vtype); // tok_embeddings
-
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(vtype); // output
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm
-
-        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1
-        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2
-        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3
-
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v
-
         ctx_size += (5 + 10*n_layer)*256; // object overhead
-
+        fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0);
-        fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
     }
 
     // print memory requirements
@@ -495,6 +548,7 @@ static bool llama_model_load(
         // this is the total memory required to run the inference
         const size_t mem_required =
             ctx_size +
+            model.mm_length +
             MEM_REQ_SCRATCH0.at(model.type) +
             MEM_REQ_SCRATCH1.at(model.type) +
             MEM_REQ_EVAL.at    (model.type);
@@ -514,6 +568,7 @@ static bool llama_model_load(
         struct ggml_init_params params = {
             /*.mem_size   =*/ lctx.model.buf.size(),
             /*.mem_buffer =*/ lctx.model.buf.data(),
+            /*.no_alloc   =*/ true,
         };
 
         model.ctx = ggml_init(params);
@@ -576,43 +631,19 @@ static bool llama_model_load(
         }
     }
 
-    const size_t file_offset = fin.tellg();
-
-    fin.close();
-
     std::vector<uint8_t> tmp;
 
     if (progress_callback) {
         progress_callback(0.0, progress_callback_user_data);
     }
 
-    for (int i = 0; i < n_parts; ++i) {
+    fprintf(stderr, "%s: loading tensors from '%s'\n", __func__, fname.c_str());
-        const int part_id = i;
-        //const int part_id = n_parts - i - 1;
-
-        std::string fname_part = fname;
-        if (i > 0) {
-            fname_part += "." + std::to_string(i);
-        }
-
-        fprintf(stderr, "%s: loading model part %d/%d from '%s'\n", __func__, i+1, n_parts, fname_part.c_str());
-
-        fin = std::ifstream(fname_part, std::ios::binary);
-        fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
-
-        fin.seekg(0, fin.end);
-        const size_t file_size = fin.tellg();
-
-        fin.seekg(file_offset);
 
     // load weights
     {
         size_t total_size = 0;
-
         model.n_loaded = 0;
 
-            fprintf(stderr, "%s: ", __func__);
-
         while (true) {
             int32_t n_dims;
             int32_t length;
@@ -641,155 +672,52 @@ static bool llama_model_load(
                 return false;
             }
 
-                // split_type = 0: split by columns
-                // split_type = 1: split by rows
-                int split_type = 0;
-
-                // split_type = 0:
-                // regex:
-                //   - tok_embeddings.*
-                //   - layers.*.attention.wo.weight
-                //   - layers.*.feed_forward.w2.weight
-
-                // split_type = 1:
-                // regex:
-                //   - output.*
-                //   - layers.*.attention.wq.weight
-                //   - layers.*.attention.wk.weight
-                //   - layers.*.attention.wv.weight
-                //   - layers.*.feed_forward.w1.weight
-                //   - layers.*.feed_forward.w3.weight
-                if (name.find("tok_embeddings") != std::string::npos) {
-                    split_type = 0;
-                } else if (name.find("layers") != std::string::npos) {
-                    if (name.find("attention.wo.weight") != std::string::npos) {
-                        split_type = 0;
-                    } else if (name.find("feed_forward.w2.weight") != std::string::npos) {
-                        split_type = 0;
-                    } else {
-                        split_type = 1;
-                    }
-                } else if (name.find("output") != std::string::npos) {
-                    split_type = 1;
-                }
-
             auto tensor = model.tensors[name.data()];
 
-                if (n_dims == 1) {
             if (ggml_nelements(tensor) != nelements) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
                 return false;
             }
-                } else {
-                    if (ggml_nelements(tensor)/n_parts != nelements) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                        return false;
-                    }
-                }
-
-                if (n_dims == 1) {
             if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
                 fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
                         __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
                 return false;
             }
-                } else {
-                    if (split_type == 0) {
-                        if (tensor->ne[0]/n_parts != ne[0] || tensor->ne[1] != ne[1]) {
-                            fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                                    __func__, name.data(), tensor->ne[0]/n_parts, tensor->ne[1], ne[0], ne[1]);
-                            return false;
-                        }
-                    } else {
-                        if (tensor->ne[0] != ne[0] || tensor->ne[1]/n_parts != ne[1]) {
-                            fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                                    __func__, name.data(), tensor->ne[0], tensor->ne[1]/n_parts, ne[0], ne[1]);
-                            return false;
-                        }
-                    }
-                }
-
             if (0) {
                 static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
-                    fprintf(stderr, "%24s - [%5d, %5d], type = %6s, split = %d\n", name.data(), ne[0], ne[1], ftype_str[ftype], split_type);
+                fprintf(stderr, "%24s - [%5d, %5d], type = %6s\n", name.data(), ne[0], ne[1], ftype_str[ftype]);
             }
 
-                size_t bpe = 0;
-
             switch (ftype) {
-                    case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 0:  // f32
-                    case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 1:  // f16
-                    case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                    break;
-                    case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                case 2:  // q4_0
+                case 3:  // q4_1
+                    assert(ne[0] % 64 == 0);
+                    break;
                 default:
-                            {
                     fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
                     return false;
-                            }
             };
 
-                if (n_dims == 1 || n_parts == 1) {
+            // load the tensor data into memory without copying or reading it
-                    if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+            size_t offset = fin.tellg();
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+            size_t tensor_data_size = ggml_nbytes(tensor);
-                                __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+            offset = (offset + 31) & -32;
-                        return false;
+            tensor->data = mm_addr + offset;
-                    }
+            fin.seekg(offset + tensor_data_size);
-
+            total_size += tensor_data_size;
-                    if (part_id == 0) {
-                        fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-                    } else {
-                        fin.seekg(ggml_nbytes(tensor), std::ios::cur);
-                    }
-
-                    total_size += ggml_nbytes(tensor);
-                } else {
-                    if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)/n_parts) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                                __func__, name.data(), ggml_nbytes(tensor)/n_parts, nelements*bpe);
-                        return false;
-                    }
-
-                    if (split_type == 0) {
-                        const int np0 = ne[0];
-
-                        const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
-                        assert(row_size == tensor->nb[1]);
-
-                        for (int i1 = 0; i1 < ne[1]; ++i1) {
-                            const size_t offset_row = i1*row_size;
-                            const size_t offset = offset_row + ((part_id*np0)/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
-                            fin.read(reinterpret_cast<char *>(tensor->data) + offset, row_size/n_parts);
-                        }
-                    } else {
-                        const int np1 = ne[1];
-
-                        const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
-
-                        for (int i1 = 0; i1 < ne[1]; ++i1) {
-                            const size_t offset_row = (i1 + part_id*np1)*row_size;
-                            fin.read(reinterpret_cast<char *>(tensor->data) + offset_row, row_size);
-                        }
-                    }
-
-                    total_size += ggml_nbytes(tensor)/n_parts;
-                }
-
-                //fprintf(stderr, "%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
             model.n_loaded++;
 
             // progress
             if (progress_callback) {
-                    double current_file_progress = double(size_t(fin.tellg()) - file_offset) / double(file_size - file_offset);
+                double current_progress = size_t(fin.tellg()) / double(file_size);
-                    double current_progress = (double(i) + current_file_progress) / double(n_parts);
                 progress_callback(current_progress, progress_callback_user_data);
             }
-                if (model.n_loaded % 8 == 0) {
-                    fprintf(stderr, ".");
-                    fflush(stderr);
-                }
         }
 
-            fprintf(stderr, " done\n");
+        fin.close();
 
         fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, model.n_loaded);
         if (model.n_loaded == 0) {
@@ -800,10 +728,9 @@ static bool llama_model_load(
         }
     }
 
-        fin.close();
+    // loading time will be recalculate after the first eval, so
-    }
+    // we take page faults deferred by mmap() into consideration
-
+    lctx.t_load_us = ggml_time_us() - lctx.t_start_us;
-    lctx.t_load_us = ggml_time_us() - t_start_us;
 
     if (progress_callback) {
         progress_callback(1.0, progress_callback_user_data);
@@ -849,6 +776,7 @@ static bool llama_eval_internal(
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_compute.size(),
         /*.mem_buffer =*/ buf_compute.data(),
+        /*.no_alloc   =*/ false,
     };
 
     struct ggml_context * ctx0 = ggml_init(params);
@@ -856,7 +784,7 @@ static bool llama_eval_internal(
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
     ggml_cgraph gf = {};
-    gf.n_threads = N > 255 && ggml_cpu_has_blas() ? 1 : n_threads;
+    gf.n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
@@ -922,7 +850,7 @@ static bool llama_eval_internal(
             struct ggml_tensor * KQ_scaled =
                 ggml_scale(ctx0,
                         KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)));
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
@@ -1126,7 +1054,7 @@ struct llama_tokenizer {
         size_t offs = 0;
         while (offs < text.size()) {
             llama_sp_symbol sym;
-            size_t char_len = std::min(text.size() - offs, utf8_len(text[offs]));
+            size_t char_len = Min(text.size() - offs, utf8_len(text[offs]));
             sym.text = text.c_str() + offs;
             sym.n = char_len;
             offs += char_len;
@@ -1240,12 +1168,12 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co
 // sampling
 //
 
-static void sample_top_k(std::vector<std::pair<double, llama_vocab::id>> & logits_id, int top_k) {
+static void sample_top_k(std::vector<std::pair<float, llama_vocab::id>> & logits_id, int top_k) {
     // find the top k tokens
     std::partial_sort(
             logits_id.begin(),
             logits_id.begin() + top_k, logits_id.end(),
-            [](const std::pair<double, llama_vocab::id> & a, const std::pair<double, llama_vocab::id> & b) {
+            [](const std::pair<float, llama_vocab::id> & a, const std::pair<float, llama_vocab::id> & b) {
         return a.first > b.first;
     });
 
@@ -1256,9 +1184,9 @@ static llama_vocab::id llama_sample_top_p_top_k(
         llama_context & lctx,
         const std::vector<llama_vocab::id> & last_n_tokens,
         int top_k,
-        double top_p,
+        float top_p,
-        double temp,
+        float temp,
-        double repeat_penalty) {
+        float repeat_penalty) {
     auto & rng = lctx.rng;
 
     const int n_logits = lctx.model.hparams.n_vocab;
@@ -1266,17 +1194,17 @@ static llama_vocab::id llama_sample_top_p_top_k(
     const auto & logits = lctx.logits;
     const auto * plogits = logits.data() + logits.size() - n_logits;
 
-    std::vector<std::pair<double, llama_vocab::id>> logits_id;
+    std::vector<std::pair<float, llama_vocab::id>> logits_id;
     logits_id.reserve(n_logits);
 
     {
-        const double scale = 1.0/temp;
+        const float scale = 1.0f/temp;
         for (int i = 0; i < n_logits; ++i) {
             // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
             // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
             if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) {
                 // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
-                if (plogits[i] < 0.0) {
+                if (plogits[i] < 0.0f) {
                     logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
                 } else {
                     logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
@@ -1289,18 +1217,18 @@ static llama_vocab::id llama_sample_top_p_top_k(
 
     sample_top_k(logits_id, top_k);
 
-    double maxl = -std::numeric_limits<double>::infinity();
+    float maxl = -std::numeric_limits<float>::infinity();
     for (const auto & kv : logits_id) {
-        maxl = std::max(maxl, kv.first);
+        maxl = Max(maxl, kv.first);
     }
 
     // compute probs for the top k tokens
-    std::vector<double> probs;
+    std::vector<float> probs;
     probs.reserve(logits_id.size());
 
     double sum = 0.0;
     for (const auto & kv : logits_id) {
-        double p = exp(kv.first - maxl);
+        const float p = expf(kv.first - maxl);
         probs.push_back(p);
         sum += p;
     }
@@ -1310,8 +1238,8 @@ static llama_vocab::id llama_sample_top_p_top_k(
         p /= sum;
     }
 
-    if (top_p < 1.0f) {
+    if (top_p < 1.0) {
-        double cumsum = 0.0f;
+        double cumsum = 0.0;
         for (int i = 0; i < (int) probs.size(); i++) {
             cumsum += probs[i];
             if (cumsum >= top_p) {
@@ -1345,7 +1273,7 @@ static llama_vocab::id llama_sample_top_p_top_k(
 //
 
 // TODO: reuse code from the llama_model_load() somehow
-bool llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype, int qk) {
+static bool llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype) {
     ggml_type type = GGML_TYPE_Q4_1;
 
     switch (itype) {
@@ -1385,8 +1313,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             return false;
         }
         if (magic != LLAMA_FILE_MAGIC) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return report_bad_magic(fname_inp.c_str(), magic, LLAMA_FILE_MAGIC);
-            return false;
         }
 
         fout.write((char *) &magic, sizeof(magic));
@@ -1444,7 +1371,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             return false;
         }
 
-        std::string word;
+        std::vector<char> word(32);
         vocab.id_to_token.resize(n_vocab);
         for (int i = 0; i < n_vocab; i++) {
             uint32_t len;
@@ -1452,17 +1379,17 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             fout.write((char *) &len, sizeof(len));
 
             word.resize(len);
-            finp.read ((char *) word.data(), len);
+            finp.read ((char *) &word[0], len);
-            fout.write((char *) word.data(), len);
+            fout.write((char *) &word[0], len);
 
             float score;
             finp.read ((char *) &score, sizeof(score));
             fout.write((char *) &score, sizeof(score));
 
-            vocab.token_to_id[word] = i;
+            vocab.token_to_id[word.data()] = i;
 
             auto &tok_score = vocab.id_to_token[i];
-            tok_score.tok = word;
+            tok_score.tok = word.data();
             tok_score.score = score;
         }
     }
@@ -1503,6 +1430,13 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             std::string name(length, 0);
             finp.read (&name[0], length);
 
+            {
+                // ensure tensor data is aligned
+                uint64_t offset = finp.tellg();
+                offset = (offset + 31) & -32;
+                finp.seekg(offset);
+            }
+
             {
                 static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
                 printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
@@ -1558,6 +1492,13 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             }
             fout.write(&name[0], length);
 
+            {
+                // ensure tensor data is aligned
+                uint64_t offset = fout.tellp();
+                offset = (offset + 31) & -32;
+                fout.seekp(offset);
+            }
+
             if (quantize) {
                 printf("quantizing .. ");
                 work.resize(nelements); // for quantization
@@ -1568,11 +1509,11 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
                 switch (type) {
                     case GGML_TYPE_Q4_0:
                         {
-                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], qk, hist_cur.data());
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
                         } break;
                     case GGML_TYPE_Q4_1:
                         {
-                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], qk, hist_cur.data());
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
                         } break;
                     default:
                         {
@@ -1590,7 +1531,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
                 }
 
                 for (int i = 0; i < (int) hist_cur.size(); ++i) {
-                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                    printf("%5.3f ", hist_cur[i] / float(nelements));
                 }
                 printf("\n");
             } else {
@@ -1613,7 +1554,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
 
             printf("%s: hist: ", __func__);
             for (int i = 0; i < (int) hist_all.size(); ++i) {
-                printf("%5.3f ", hist_all[i] / (float)sum_all);
+                printf("%5.3f ", hist_all[i] / float(sum_all));
             }
             printf("\n");
         }
@@ -1655,7 +1596,10 @@ struct llama_context * llama_init_from_file(
 
     if (params.use_mlock) {
         char *err;
-        if (!ggml_mlock(ctx->model.ctx, &err)) {
+        if (!ggml_mlock(ctx->model.ctx,
+                        ctx->model.mm_addr,
+                        ctx->model.mm_length,
+                        &err)) {
             fprintf(stderr, "%s\n", err);
             free(err);
             llama_free(ctx);
@@ -1705,15 +1649,18 @@ void llama_free(struct llama_context * ctx) {
         ggml_free(ctx->model.ctx);
     }
 
+    if (ctx->model.mm_addr) {
+        munmap_file(ctx->model.mm_addr, ctx->model.mm_length);
+    }
+
     delete ctx;
 }
 
 int llama_model_quantize(
         const char * fname_inp,
         const char * fname_out,
-               int   itype,
+               int   itype) {
-               int   qk) {
+    if (!llama_model_quantize_internal(fname_inp, fname_out, itype)) {
-    if (!llama_model_quantize_internal(fname_inp, fname_out, itype, qk)) {
         fprintf(stderr, "%s: failed to quantize\n", __func__);
         return 1;
     }
@@ -1731,7 +1678,11 @@ int llama_eval(
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
-
+    // get a more accurate load time, upon first eval
+    if (!ctx->has_evaluated_once) {
+        ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
+        ctx->has_evaluated_once = true;
+    }
     return 0;
 }
 
@@ -1796,9 +1747,9 @@ llama_token llama_sample_top_p_top_k(
       const llama_token * last_n_tokens_data,
                     int   last_n_tokens_size,
                     int   top_k,
-                 double   top_p,
+                  float   top_p,
-                 double   temp,
+                  float   temp,
-                 double   repeat_penalty) {
+                  float   repeat_penalty) {
     const int64_t t_start_sample_us = ggml_time_us();
 
     llama_token result = 0;
@@ -1824,21 +1775,20 @@ llama_token llama_sample_top_p_top_k(
 void llama_print_timings(struct llama_context * ctx) {
     const int64_t t_end_us = ggml_time_us();
 
-    const int32_t n_sample = std::max(1, ctx->n_sample);
+    const int32_t n_sample = Max(1, ctx->n_sample);
-    const int32_t n_eval   = std::max(1, ctx->n_eval);
+    const int32_t n_eval   = Max(1, ctx->n_eval);
-    const int32_t n_p_eval = std::max(1, ctx->n_p_eval);
+    const int32_t n_p_eval = Max(1, ctx->n_p_eval);
 
     fprintf(stderr, "\n");
-    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0f);
+    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0);
-    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3f * ctx->t_sample_us, n_sample, 1e-3f * ctx->t_sample_us / n_sample);
+    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample);
-    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3f * ctx->t_p_eval_us, n_p_eval, 1e-3f * ctx->t_p_eval_us / n_p_eval);
+    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval);
-    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3f * ctx->t_eval_us,   n_eval,   1e-3f * ctx->t_eval_us   / n_eval);
+    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval);
-    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f);
+    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0);
 }
 
 void llama_reset_timings(struct llama_context * ctx) {
     ctx->t_start_us = ggml_time_us();
-
     ctx->t_sample_us = ctx->n_sample = 0;
     ctx->t_eval_us   = ctx->n_eval   = 0;
     ctx->t_p_eval_us = ctx->n_p_eval = 0;

examples/talk-llama/llama.h

@@ -6,7 +6,7 @@
 #include <stdbool.h>
 
 #ifdef LLAMA_SHARED
-#    ifdef _WIN32
+#    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
 #            define LLAMA_API __declspec(dllexport)
 #        else
@@ -20,7 +20,7 @@
 #endif
 
 #define LLAMA_FILE_VERSION 1
-#define LLAMA_FILE_MAGIC 0x67676d66 // 'ggmf' in hex
+#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex
 #define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
 
 #ifdef __cplusplus
@@ -45,7 +45,7 @@ extern "C" {
 
     } llama_token_data;
 
-    typedef void (*llama_progress_callback)(double progress, void *ctx);
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     struct llama_context_params {
         int n_ctx;   // text context
@@ -81,8 +81,7 @@ extern "C" {
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-                   int   itype,
+                   int   itype);
-                   int   qk);
 
     // Run the llama inference to obtain the logits and probabilities for the next token.
     // tokens + n_tokens is the provided batch of new tokens to process
@@ -135,9 +134,9 @@ extern "C" {
           const llama_token * last_n_tokens_data,
                         int   last_n_tokens_size,
                         int   top_k,
-                     double   top_p,
+                      float   top_p,
-                     double   temp,
+                      float   temp,
-                     double   repeat_penalty);
+                      float   repeat_penalty);
 
     // Performance information
     LLAMA_API void llama_print_timings(struct llama_context * ctx);

examples/talk-llama/prompts/talk-alpaca.txt

@@ -0,0 +1,23 @@
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+
+Write a text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} responds with short and concise answers.
+
+### Response:
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4}

examples/talk-llama/speak.sh

@@ -10,7 +10,15 @@
 #espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
 
 # for Mac
-say "$2"
+if [ "$1" = "0" ]; then
+    say "$2"
+elif [ "$1" = "1" ]; then
+    say -v "Samantha (Enhanced)" "$2"
+elif [ "$1" = "2" ]; then
+    say -v "Daniel (Enhanced)" "$2"
+elif [ "$1" = "3" ]; then
+    say -v "Veena (Enhanced)" "$2"
+fi
 
 # Eleven Labs
 #

examples/talk-llama/talk-llama.cpp

@@ -0,0 +1,703 @@
+// Talk with AI
+//
+
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "llama.h"
+
+#include <map>
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
+    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
+    res.resize(n);
+
+    return res;
+}
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t voice_id   = 0;
+    int32_t voice_ms   = 10000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 64;
+    int32_t audio_ctx  = 0;
+
+    int32_t n_parts_llama = -1;
+
+    float vad_thold    = 0.4f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+    bool verbose_prompt = false;
+
+    std::string name_ni     = "Georgi"; // natural    intelligence
+    std::string name_ai     = "LLaMA";  // artificial intelligence
+    std::string language    = "en";
+    std::string model_wsp   = "models/ggml-base.en.bin";
+    std::string model_llama = "models/ggml-llama-7B.bin";
+    std::string speak       = "./examples/talk/speak.sh";
+    std::string prompt      = "";
+    std::string fname_out;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads      = std::stoi(argv[++i]); }
+        else if (arg == "-vid" || arg == "--voice-id")      { params.voice_id       = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms       = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id     = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens     = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx      = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold      = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold     = std::stof(argv[++i]); }
+        else if (arg == "--n-parts-llama")                  { params.n_parts_llama  = std::stoi(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up       = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate      = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special  = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy   = true; }
+        else if (arg == "--verbose-prompt")                 { params.verbose_prompt = true; }
+        else if (arg == "-nni" || arg == "--name-ni")       { params.name_ni        = argv[++i]; }
+        else if (arg == "-nai" || arg == "--name-ai")       { params.name_ai        = argv[++i]; }
+        else if (arg == "-l"   || arg == "--language")      { params.language       = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp      = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama    = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak          = argv[++i]; }
+        else if (arg == "--prompt-file")                    {
+            std::ifstream file(argv[++i]);
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vid N,   --voice-id N    [%-7d] voice ID\n",                                    params.voice_id);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -nni NAME,--name-ni NAME  [%-7s] natural intelligence name\n",                   params.name_ni.c_str());
+    fprintf(stderr, "  -nai NAME,--name-ai NAME  [%-7s] artificial intelligence name\n",                params.name_ai.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -ml FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  --n-parts-llama N         [%-7d] num parts in llama model file\n",               params.n_parts_llama);
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  --prompt-file FNAME       [%-7s] file with custom prompt to start dialog\n",     "");
+    fprintf(stderr, "  --verbose-prompt          [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string transcribe(
+        whisper_context * ctx,
+        const whisper_params & params,
+        const std::vector<float> & pcmf32,
+        const std::string prompt_text,
+        float & prob,
+        int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    std::vector<whisper_token> prompt_tokens;
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    prompt_tokens.resize(1024);
+    prompt_tokens.resize(whisper_tokenize(ctx, prompt_text.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = 2;
+
+    wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    static int iter = params.voice_id;
+    std::this_thread::sleep_for(std::chrono::milliseconds(100*iter));
+    iter = (iter + 1) % 4;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+const std::vector<std::string> k_participants = {
+    "LLaMA",
+    "GGaMA",
+    "SSaMA",
+    "RRaMA",
+};
+
+// homophones
+const std::map<std::string, std::vector<std::string>> k_homophones = {
+    { "LLaMA", { "llama", "Llama", "LLAMA", }, },
+    { "GGaMA", { "gama", "Gama", "GAMA", "gamma", "Gamma", "GAMMA", }, },
+    { "SSaMA", { "sama", "Sama", "SAMA", "samma", "Samma", "SAMMA", }, },
+    { "RRaMA", { "rama", "Rama", "RAMA", "ramma", "Ramma", "RAMMA", }, },
+};
+
+const std::string k_prompt_whisper = R"(A conversation between {1}, {10}, {11}, {12} and {13}.)";
+
+const std::map<std::string, std::string> k_prompt = {
+    {
+        k_participants.at(0),
+        R"(Text transcript of a never ending dialog, between {1}, {10}, {11}, {12} and {13}.
+There are no annotations like (30 seconds passed...) or (to himself), just what the participants say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{10}, {11}, {12} and {13} respond with short and concise answers.
+{10} is smart, objective, honest and kind. Never fails to give a meaningful and insightful answer and opinion.
+{1} is leading the conversation and asking the questions.
+
+{1}{4} Hello {10}! What is your opinion on the current state of the world?
+{10}{4} Great question {1}! I think we live in a very interesting time.
+There are many things to be concerned about, but also many things to be optimistic about.
+{1}{4} What advice would you give to a young person who is just starting out in life?
+{10}{4} I would tell them to be patient and to not be afraid to fail.
+It is important to learn from your mistakes and to keep trying.
+{1}{4})"
+    },
+    {
+        k_participants.at(1),
+        R"(Text transcript of a never ending dialog, between {1}, {10}, {11}, {12} and {13}.
+There are no annotations like (30 seconds passed...) or (to himself), just what the participants say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{10}, {11}, {12} and {13} respond with short and concise answers.
+{11} has critical thinking skills, is very knowledgeable and is a good listener. He is very humble and never arrogant.
+{1} is leading the conversation and asking the questions.
+
+{1}{4} Hello {11}! What is your opinion on the current state of the world?
+{11}{4} The world is about to experience a major change. We are on the verge of a new era.
+{1}{4} What advice would you give to a young person who is just starting out in life?
+{11}{4} My advice would be to be open minded and to be willing to learn from others.
+{1}{4})"
+    },
+    {
+        k_participants.at(2),
+        R"(Text transcript of a never ending dialog, between {1}, {10}, {11}, {12} and {13}.
+There are no annotations like (30 seconds passed...) or (to himself), just what the participants say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{10}, {11}, {12} and {13} respond with short and concise answers.
+{12} has strong leadership skills, strategic thinking, and innovative ideas. Has the ability to mentor and support young people.
+{1} is leading the conversation and asking the questions.
+
+{1}{4} Hello {12}! What is your opinion on the current state of the world?
+{12}{4} Our future is bright. We are living in a time of great opportunity.
+{1}{4} What advice would you give to a young person who is just starting out in life?
+{12}{4} I would tell them to be brave and to be willing to take risks.
+{1}{4})"
+    },
+    {
+        k_participants.at(3),
+        R"(Text transcript of a never ending dialog, between {1}, {10}, {11}, {12} and {13}.
+There are no annotations like (30 seconds passed...) or (to himself), just what the participants say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{10}, {11}, {12} and {13} respond with short and concise answers.
+{13} is rude, arrogant, and has a bad attitude. He is very opinionated and never listens to others.
+{1} is leading the conversation and asking the questions.
+
+{1}{4} Hello {13}! What is your opinion on the current state of the world?
+{13}{4} The world is a terrible place. It is full of evil and corruption.
+{1}{4} What advice would you give to a young person who is just starting out in life?
+{13}{4} I would tell them to be selfish and to never trust anyone.
+{1}{4})"
+    },
+};
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+
+    // llama init
+
+    auto lparams = llama_context_default_params();
+
+    // tune these to your liking
+    lparams.n_ctx      = 512;
+    lparams.seed       = 1;
+    lparams.f16_kv     = true;
+    lparams.n_parts    = params.n_parts_llama;
+
+    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+
+        if (!whisper_is_multilingual(ctx_wsp)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    int n_iter = 0;
+
+    bool is_running  = true;
+    bool force_speak = false;
+
+    float prob0 = 0.0f;
+
+    const std::string chat_symb = ":";
+
+    const std::string name_ni  = params.name_ni;
+    const std::string name_ai  = params.name_ai;
+
+    // the participant that was referenced last
+    std::string name_ref = name_ni;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    std::string prompt_whisper = k_prompt_whisper;
+
+    prompt_whisper = ::replace(prompt_whisper, "{1}",  name_ni);
+    prompt_whisper = ::replace(prompt_whisper, "{10}", k_participants.at(0));
+    prompt_whisper = ::replace(prompt_whisper, "{11}", k_participants.at(1));
+    prompt_whisper = ::replace(prompt_whisper, "{12}", k_participants.at(2));
+    prompt_whisper = ::replace(prompt_whisper, "{13}", k_participants.at(3));
+
+    // construct the initial prompt for LLaMA inference
+    std::string prompt_llama = params.prompt.empty() ? k_prompt.find(name_ai)->second : params.prompt;
+
+    // need to have leading ' '
+    prompt_llama.insert(0, 1, ' ');
+
+    prompt_llama = ::replace(prompt_llama, "{1}",  name_ni);
+    prompt_llama = ::replace(prompt_llama, "{10}", k_participants.at(0));
+    prompt_llama = ::replace(prompt_llama, "{11}", k_participants.at(1));
+    prompt_llama = ::replace(prompt_llama, "{12}", k_participants.at(2));
+    prompt_llama = ::replace(prompt_llama, "{13}", k_participants.at(3));
+
+    {
+        // get date string
+        std::string date_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%d/%m/%Y", now);
+            date_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{1}", date_str);
+    }
+
+    {
+        // get time string
+        std::string time_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%H:%M", now);
+            time_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{2}", time_str);
+    }
+
+    {
+        // get year string
+        std::string year_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%Y", now);
+            year_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{3}", year_str);
+    }
+
+    prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);
+
+    // evaluate the initial prompt
+
+    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_llama, true);
+
+    printf("\n");
+    printf("%s : initializing - please wait ...\n", __func__);
+
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    if (params.verbose_prompt) {
+        fprintf(stdout, "\n");
+        fprintf(stdout, "%s", prompt_whisper.c_str());
+        fprintf(stdout, "\n");
+
+        fprintf(stdout, "\n");
+        fprintf(stdout, "%s", prompt_llama.c_str());
+        fprintf(stdout, "\n");
+        fprintf(stdout, "\n");
+        fflush(stdout);
+    }
+
+    printf("%s : done! start speaking in the microphone\n", __func__);
+    printf("\n");
+    printf("%s%s", name_ni.c_str(), chat_symb.c_str());
+    fflush(stdout);
+
+    // clear audio buffer
+    audio.clear();
+
+    // text inference variables
+    const int voice_id = params.voice_id;
+    const int n_keep   = embd_inp.size();
+    const int n_ctx    = llama_n_ctx(ctx_llama);
+
+    int n_past = n_keep;
+    int n_prev = 64; // TODO arg
+
+    std::vector<llama_token> embd;
+
+    // reverse prompts for detecting when it's time to stop speaking
+    std::vector<std::string> antiprompts = {
+        name_ni + chat_symb,
+    };
+
+    for (const auto & p : k_participants) {
+        antiprompts.push_back(p + chat_symb);
+    }
+
+    std::string text_heard_all;
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        if (!is_running) {
+            break;
+        }
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        int64_t t_ms = 0;
+
+        {
+            audio.get(15000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                audio.get(params.voice_ms, pcmf32_cur);
+
+                std::string text_heard;
+
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\[.*?\\]");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\(.*?\\)");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
+                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+
+                // take first line
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
+
+                // remove leading and trailing whitespace
+                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
+                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
+
+                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
+
+                if (text_heard.empty() || tokens.empty() || force_speak) {
+                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
+                    audio.clear();
+
+                    continue;
+                }
+
+                force_speak = false;
+
+                if (text_heard[0] != ' ') {
+                    text_heard.insert(0, 1, ' ');
+                }
+
+                // replace homophones
+                for (const auto & homophone : k_homophones) {
+                    for (const auto & word : homophone.second) {
+                        text_heard = ::replace(text_heard, word, homophone.first);
+                    }
+                }
+
+                // check which participant was mentioned
+                const auto name_ref_old = name_ref;
+                for (const auto & participant : k_participants) {
+                    if (participant == name_ref) {
+                        continue;
+                    }
+
+                    if (text_heard.find(participant) != std::string::npos) {
+                        name_ref = participant;
+                        break;
+                    }
+                }
+                if (name_ref == name_ref_old && name_ref != name_ai) {
+                    name_ref = name_ni;
+                }
+
+                text_heard += "\n" + name_ref + chat_symb;
+                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
+                fflush(stdout);
+
+                text_heard_all += text_heard;
+                // keep only last 100 characters
+                if (text_heard_all.size() > 100) {
+                    text_heard_all = text_heard_all.substr(text_heard_all.size() - 100);
+                }
+
+                if (name_ref != name_ai) {
+                } else {
+                    // text inference
+                    bool done = false;
+                    std::string text_to_speak;
+
+                    embd = ::llama_tokenize(ctx_llama, text_heard_all, false);
+                    text_heard_all.clear();
+
+                    while (true) {
+                        // predict
+                        if (embd.size() > 0) {
+                            if (n_past + (int) embd.size() > n_ctx) {
+                                n_past = n_keep;
+
+                                // insert n_left/2 tokens at the start of embd from last_n_tokens
+                                embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
+
+                                //printf("\n---\n");
+                                //printf("resetting: '");
+                                //for (int i = 0; i < (int) embd.size(); i++) {
+                                //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
+                                //}
+                                //printf("'\n");
+                                //printf("\n---\n");
+                            }
+
+                            if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                                fprintf(stderr, "%s : failed to eval\n", __func__);
+                                return 1;
+                            }
+                        }
+
+                        //printf("n_iter = %d, n_past = %d, n_ctx = %d, n_keep = %d, n_prev = %d, embd.size() = %d\n", n_iter, n_past, n_ctx, n_keep, n_prev, (int) embd.size());
+
+                        embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
+                        n_past += embd.size();
+                        embd.clear();
+
+                        if (done) break;
+
+                        {
+                            // out of user input, sample next token
+                            const float top_k          = 5;
+                            const float top_p          = 0.80f;
+                            const float temp           = 0.20f;
+                            const float repeat_penalty = 1.0764f;
+
+                            const int repeat_last_n    = 256;
+
+                            llama_token id = 0;
+
+                            {
+                                auto logits = llama_get_logits(ctx_llama);
+                                logits[llama_token_eos()] = 0;
+
+                                id = llama_sample_top_p_top_k(ctx_llama,
+                                        embd_inp.data() + std::max(0, n_past - repeat_last_n),
+                                        repeat_last_n, top_k, top_p, temp, repeat_penalty);
+                            }
+
+                            if (id != llama_token_eos()) {
+                                // add it to the context
+                                embd.push_back(id);
+
+                                text_to_speak += llama_token_to_str(ctx_llama, id);
+
+                                printf("%s", llama_token_to_str(ctx_llama, id));
+                            }
+
+                            // new line
+                            if (id == 13) {
+                            }
+                        }
+
+                        {
+                            std::string last_output;
+                            for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
+                                last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
+                            }
+                            last_output += llama_token_to_str(ctx_llama, embd[0]);
+
+                            for (const std::string & antiprompt : antiprompts) {
+                                if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
+                                    done = true;
+                                    text_to_speak = ::replace(text_to_speak, antiprompt, "");
+                                    fflush(stdout);
+                                    break;
+                                }
+                            }
+                        }
+
+                        is_running = sdl_poll_events();
+
+                        if (!is_running) {
+                            break;
+                        }
+                    }
+
+                    text_to_speak = ::replace(text_to_speak, "\"", "");
+                    system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+                }
+
+                audio.clear();
+
+                ++n_iter;
+            }
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx_wsp);
+    whisper_free(ctx_wsp);
+
+    llama_print_timings(ctx_llama);
+    llama_free(ctx_llama);
+
+    return 0;
+}

examples/talk.llama/CMakeLists.txt

@@ -1,12 +0,0 @@
-if (WHISPER_SUPPORT_SDL2)
-    # talk.llama
-    set(TARGET talk-llama)
-
-    # TODO: this is temporary
-    #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk-llama.cpp llama.cpp)
-
-    include(DefaultTargetOptions)
-
-    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
-endif ()

examples/talk.llama/README.md

@@ -1,2 +0,0 @@
-# talk.llama
-

examples/talk.llama/talk-llama.cpp

@@ -1,511 +0,0 @@
-// Talk with AI
-//
-
-#include "common.h"
-#include "common-sdl.h"
-#include "whisper.h"
-#include "llama.h"
-
-#include <cassert>
-#include <cstdio>
-#include <fstream>
-#include <regex>
-#include <string>
-#include <thread>
-#include <vector>
-#include <regex>
-
-std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
-    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
-    std::vector<llama_token> res(text.size() + (int)add_bos);
-    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
-    assert(n >= 0);
-    res.resize(n);
-
-    return res;
-}
-
-// command-line parameters
-struct whisper_params {
-    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t voice_ms   = 10000;
-    int32_t capture_id = -1;
-    int32_t max_tokens = 32;
-    int32_t audio_ctx  = 0;
-
-    float vad_thold    = 0.6f;
-    float freq_thold   = 100.0f;
-
-    bool speed_up      = false;
-    bool translate     = false;
-    bool print_special = false;
-    bool print_energy  = false;
-    bool no_timestamps = true;
-
-    std::string person      = "Santa";
-    std::string language    = "en";
-    std::string model_wsp   = "models/ggml-base.en.bin";
-    std::string model_llama = "models/ggml-llama-7B.bin";
-    std::string speak       = "./examples/talk/speak.sh";
-    std::string fname_out;
-};
-
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
-
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
-    for (int i = 1; i < argc; i++) {
-        std::string arg = argv[i];
-
-        if (arg == "-h" || arg == "--help") {
-            whisper_print_usage(argc, argv, params);
-            exit(0);
-        }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
-        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
-        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
-        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
-        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
-        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
-        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
-        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
-        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
-        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
-        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
-        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
-        else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            whisper_print_usage(argc, argv, params);
-            exit(0);
-        }
-    }
-
-    return true;
-}
-
-void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
-    fprintf(stderr, "\n");
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
-    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
-    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
-    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
-    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
-    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
-    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
-    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
-    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
-    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
-    fprintf(stderr, "  -mg FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
-    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
-    fprintf(stderr, "\n");
-}
-
-std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
-    const auto t_start = std::chrono::high_resolution_clock::now();
-
-    prob = 0.0f;
-    t_ms = 0;
-
-    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
-
-    wparams.print_progress   = false;
-    wparams.print_special    = params.print_special;
-    wparams.print_realtime   = false;
-    wparams.print_timestamps = !params.no_timestamps;
-    wparams.translate        = params.translate;
-    wparams.no_context       = true;
-    wparams.single_segment   = true;
-    wparams.max_tokens       = params.max_tokens;
-    wparams.language         = params.language.c_str();
-    wparams.n_threads        = params.n_threads;
-
-    wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
-
-    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
-        return "";
-    }
-
-    int prob_n = 0;
-    std::string result;
-
-    const int n_segments = whisper_full_n_segments(ctx);
-    for (int i = 0; i < n_segments; ++i) {
-        const char * text = whisper_full_get_segment_text(ctx, i);
-
-        result += text;
-
-        const int n_tokens = whisper_full_n_tokens(ctx, i);
-        for (int j = 0; j < n_tokens; ++j) {
-            const auto token = whisper_full_get_token_data(ctx, i, j);
-
-            prob += token.p;
-            ++prob_n;
-        }
-    }
-
-    if (prob_n > 0) {
-        prob /= prob_n;
-    }
-
-    const auto t_end = std::chrono::high_resolution_clock::now();
-    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
-
-    return result;
-}
-
-// need to have leading ' '
-//const std::string k_prompt = R"( Transcript of a dialog, where {1} interacts with an Assistant named Bob. Bob is helpful, kind, honest, good at writing, and never fails to answer {1}'s requests immediately and with precision.
-//
-//{0}: Hello, Bob.
-//{1}: Hello {0}. How may I help you today?
-//{0}:)";
-
-const std::string k_prompt = R"( Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
-{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
-There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
-The transcript only includes text, it does not include markup like HTML and Markdown.
-{1} answers responds with short and concise answers.
-
-{0}{4} Hello, {1}!
-{1}{4} Hello {0}! How may I help you today?
-{0}{4} What time is it?
-{1}{4} It is {2} o'clock.
-{0}{4} What year is it?
-{1}{4} We are in {3}.
-{0}{4} What is a cat?
-{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
-{0}{4} Name a color.
-{1}{4} Blue
-{0}{4})";
-
-int main(int argc, char ** argv) {
-    whisper_params params;
-
-    if (whisper_params_parse(argc, argv, params) == false) {
-        return 1;
-    }
-
-    if (whisper_lang_id(params.language.c_str()) == -1) {
-        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
-        whisper_print_usage(argc, argv, params);
-        exit(0);
-    }
-
-    // whisper init
-
-    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
-
-    // llama init
-
-    auto lparams = llama_context_default_params();
-
-    lparams.n_ctx      = 512;
-    lparams.n_parts    = 2; // TODO fix
-    lparams.seed       = 1; // TODO fix
-    lparams.f16_kv     = true;
-
-    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
-
-    // print some info about the processing
-    {
-        fprintf(stderr, "\n");
-        if (!whisper_is_multilingual(ctx_wsp)) {
-            if (params.language != "en" || params.translate) {
-                params.language = "en";
-                params.translate = false;
-                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
-            }
-        }
-        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
-                __func__,
-                params.n_threads,
-                params.language.c_str(),
-                params.translate ? "translate" : "transcribe",
-                params.no_timestamps ? 0 : 1);
-
-        fprintf(stderr, "\n");
-    }
-
-
-    // init audio
-
-    audio_async audio(30*1000);
-    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
-        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
-        return 1;
-    }
-
-    audio.resume();
-
-    int n_iter = 0;
-
-    bool is_running  = true;
-    bool force_speak = false;
-
-    float prob0 = 0.0f;
-
-    const std::string chat_symb = ":";
-    const std::string bot_name = "LLAMA";
-
-    std::vector<float> pcmf32_cur;
-    std::vector<float> pcmf32_prompt;
-
-    std::string prompt_org = k_prompt;
-    prompt_org = ::replace(prompt_org, "{0}", params.person);
-    prompt_org = ::replace(prompt_org, "{1}", bot_name);
-
-    {
-        // get time string
-        std::string time_str;
-        {
-            time_t t = time(0);
-            struct tm * now = localtime(&t);
-            char buf[128];
-            strftime(buf, sizeof(buf), "%H:%M", now);
-            time_str = buf;
-        }
-        prompt_org = ::replace(prompt_org, "{2}", time_str);
-    }
-
-    {
-        // get year string
-        std::string year_str;
-        {
-            time_t t = time(0);
-            struct tm * now = localtime(&t);
-            char buf[128];
-            strftime(buf, sizeof(buf), "%Y", now);
-            year_str = buf;
-        }
-        prompt_org = ::replace(prompt_org, "{3}", year_str);
-    }
-
-    prompt_org = ::replace(prompt_org, "{4}", chat_symb);
-
-    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_org, true);
-
-    const int n_ctx = llama_n_ctx(ctx_llama);
-
-    printf("\n");
-    printf("%s : initializing - please wait ...\n", __func__);
-
-    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
-        fprintf(stderr, "%s : failed to eval\n", __func__);
-        return 1;
-    }
-
-    //fprintf(stdout, "\n");
-    //fprintf(stdout, "%s", prompt_org.c_str());
-    //fflush(stdout);
-
-    printf("%s : done! start speaking in the microphone\n", __func__);
-    printf("\n");
-    printf("%s%s", params.person.c_str(), chat_symb.c_str());
-    fflush(stdout);
-
-    audio.clear();
-
-    const int n_keep = embd_inp.size();
-    const int voice_id = 2;
-
-    int n_past = n_keep;
-    int n_prev = 64; // TODO arg
-
-    std::vector<llama_token> embd;
-
-    std::vector<std::string> antiprompts = {
-        params.person + chat_symb,
-    };
-
-    // main loop
-    while (is_running) {
-        // handle Ctrl + C
-        is_running = sdl_poll_events();
-
-        if (!is_running) {
-            break;
-        }
-
-        // delay
-        std::this_thread::sleep_for(std::chrono::milliseconds(100));
-
-        int64_t t_ms = 0;
-
-        {
-            audio.get(2000, pcmf32_cur);
-
-            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
-                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
-
-                audio.get(params.voice_ms, pcmf32_cur);
-
-                std::string text_heard;
-
-                if (!force_speak) {
-                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prob0, t_ms));
-                }
-
-                // remove text between brackets using regex
-                {
-                    std::regex re("\\[.*?\\]");
-                    text_heard = std::regex_replace(text_heard, re, "");
-                }
-
-                // remove text between brackets using regex
-                {
-                    std::regex re("\\(.*?\\)");
-                    text_heard = std::regex_replace(text_heard, re, "");
-                }
-
-                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
-                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
-
-                // take first line
-                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
-
-                // remove leading and trailing whitespace
-                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
-                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
-
-                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
-
-                if (text_heard.empty() || tokens.empty() || force_speak) {
-                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
-                    audio.clear();
-
-                    continue;
-                }
-
-                force_speak = false;
-
-                text_heard.insert(0, 1, ' ');
-                text_heard += "\n" + bot_name + chat_symb;
-                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
-                fflush(stdout);
-
-                embd = ::llama_tokenize(ctx_llama, text_heard, false);
-
-                // text inference
-                bool done = false;
-                std::string text_to_speak;
-                while (true) {
-                    // predict
-                    if (embd.size() > 0) {
-                        if (n_past + (int) embd.size() > n_ctx) {
-                            n_past = n_keep;
-
-                            // insert n_left/2 tokens at the start of embd from last_n_tokens
-                            embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
-
-                            //printf("\n---\n");
-                            //printf("resetting: '");
-                            //for (int i = 0; i < (int) embd.size(); i++) {
-                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
-                            //}
-                            //printf("'\n");
-                            //printf("\n---\n");
-                        }
-
-                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
-                            fprintf(stderr, "%s : failed to eval\n", __func__);
-                            return 1;
-                        }
-                    }
-
-                    //printf("n_iter = %d, n_past = %d, n_ctx = %d, n_keep = %d, n_prev = %d, embd.size() = %d\n", n_iter, n_past, n_ctx, n_keep, n_prev, (int) embd.size());
-
-                    embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
-                    n_past += embd.size();
-                    embd.clear();
-
-                    if (done) break;
-
-                    {
-                        // out of user input, sample next token
-                        const float top_k          = 5;
-                        const float top_p          = 0.80f;
-                        const float temp           = 0.30f;
-                        const float repeat_penalty = 1.1764f;
-
-                        const int repeat_last_n    = 256;
-
-                        llama_token id = 0;
-
-                        {
-                            //auto logits = llama_get_logits(ctx_llama);
-                            //logits[llama_token_eos()] = 0;
-
-                            id = llama_sample_top_p_top_k(ctx_llama,
-                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, top_k, top_p, temp, repeat_penalty);
-                        }
-
-                        if (id != llama_token_eos()) {
-                            // add it to the context
-                            embd.push_back(id);
-
-                            text_to_speak += llama_token_to_str(ctx_llama, id);
-
-                            printf("%s", llama_token_to_str(ctx_llama, id));
-                        } else {
-                            // TODO
-                            printf("EOS TOKEN - SHOULD NOT HAPPEN\n");
-                            exit(0);
-                        }
-                    }
-
-                    {
-                        std::string last_output;
-                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
-                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
-                        }
-                        last_output += llama_token_to_str(ctx_llama, embd[0]);
-
-                        for (std::string & antiprompt : antiprompts) {
-                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
-                                done = true;
-                                text_to_speak = ::replace(text_to_speak, antiprompt, "");
-                                fflush(stdout);
-                                break;
-                            }
-                        }
-                    }
-
-                    is_running = sdl_poll_events();
-
-                    if (!is_running) {
-                        break;
-                    }
-                }
-
-                text_to_speak = ::replace(text_to_speak, "\"", "");
-                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
-
-                audio.clear();
-
-                ++n_iter;
-            }
-        }
-    }
-
-    audio.pause();
-
-    whisper_print_timings(ctx_wsp);
-    whisper_free(ctx_wsp);
-
-    return 0;
-}

examples/talk.wasm/gpt-2.cpp

@@ -325,9 +325,12 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
     // create the ggml context
     {
-        struct ggml_init_params params;
+        struct ggml_init_params params = {
-        params.mem_size   = ctx_size;
+            /*.mem_size   =*/ ctx_size,
-        params.mem_buffer = NULL;
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ false,
+        };
+
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -528,9 +531,11 @@ bool gpt2_eval(
         }
     }
 
-    struct ggml_init_params params;
+    struct ggml_init_params params = {
-    params.mem_size   = buf_size;
+        /*.mem_size   =*/ buf_size,
-    params.mem_buffer = buf;
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
 

examples/talk/gpt-2.cpp

@@ -325,9 +325,11 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
     // create the ggml context
     {
-        struct ggml_init_params params;
+        struct ggml_init_params params = {
-        params.mem_size   = ctx_size;
+            /*.mem_size   =*/ ctx_size,
-        params.mem_buffer = nullptr;
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ false,
+        };
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -528,9 +530,11 @@ bool gpt2_eval(
         }
     }
 
-    struct ggml_init_params params;
+    struct ggml_init_params params = {
-    params.mem_size   = buf_size;
+        /*.mem_size   =*/ buf_size,
-    params.mem_buffer = buf;
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
 

examples/talk/speak.sh

@@ -7,7 +7,10 @@
 # Mac OS: brew install espeak
 # Linux: apt-get install espeak
 #
-espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# Mac OS "say" command
+say "$2"
 
 # Eleven Labs
 #

examples/whisper.swiftui/README.md

@@ -1,15 +1,18 @@
 A sample SwiftUI app using [whisper.cpp](https://github.com/ggerganov/whisper.cpp/) to do voice-to-text transcriptions.
 See also: [whisper.objc](https://github.com/ggerganov/whisper.cpp/tree/master/examples/whisper.objc).
 
-To use:
+**Usage**:
 
 1. Select a model from the [whisper.cpp repository](https://github.com/ggerganov/whisper.cpp/tree/master/models).[^1]
-2. Add the model to "whisper.swiftui.demo/Resources/models" via Xcode.
+2. Add the model to `whisper.swiftui.demo/Resources/models` **via Xcode**.
 3. Select a sample audio file (for example, [jfk.wav](https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav)).
-4. Add the model to "whisper.swiftui.demo/Resources/samples" via Xcode.
+4. Add the sample audio file to `whisper.swiftui.demo/Resources/samples` **via Xcode**.
 5. Select the "Release" [^2] build configuration under "Run", then deploy and run to your device.
 
+**Note:** Pay attention to the folder path: `whisper.swiftui.demo/Resources/models` is the appropriate directory to place resources whilst `whisper.swiftui.demo/Models` is related to actual code.
+
 [^1]: I recommend the tiny, base or small models for running on an iOS device.
+
 [^2]: The `Release` build can boost performance of transcription. In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.
 
 ![image](https://user-images.githubusercontent.com/1991296/212539216-0aef65e4-f882-480a-8358-0f816838fd52.png)

ggml.h

@@ -316,6 +316,7 @@ struct ggml_init_params {
     // memory pool
     size_t mem_size;   // bytes
     void * mem_buffer; // if NULL, memory will be allocated internally
+    bool   no_alloc;   // don't allocate memory for the tensor data
 };
 
 void    ggml_time_init(void); // call this once at the beginning of the program
@@ -344,7 +345,11 @@ size_t ggml_used_mem(const struct ggml_context * ctx);
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
 
 bool ggml_mlock_supported(void);
-bool ggml_mlock(struct ggml_context * ctx, char ** err_p);
+bool ggml_mlock(
+        struct ggml_context * ctx,
+        const void *opt_extra_addr,
+        size_t opt_extra_len,
+        char **err_p);
 
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
@@ -748,8 +753,8 @@ enum ggml_opt_result ggml_opt(
 // quantization
 //
 
-size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
 
 //
 // system info

models/download-coreml-model.sh

@@ -1,82 +0,0 @@
-#!/bin/bash
-
-# This script downloads Whisper model files that have already been converted to Core ML format.
-# This way you don't have to convert them yourself.
-
-src="https://huggingface.co/datasets/ggerganov/whisper.cpp-coreml"
-pfx="resolve/main/ggml"
-
-# get the path of this script
-function get_script_path() {
-    if [ -x "$(command -v realpath)" ]; then
-        echo "$(dirname $(realpath $0))"
-    else
-        local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
-        echo "$ret"
-    fi
-}
-
-models_path="$(get_script_path)"
-
-# Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
-
-# list available models
-function list_models {
-    printf "\n"
-    printf "  Available models:"
-    for model in "${models[@]}"; do
-        printf " $model"
-    done
-    printf "\n\n"
-}
-
-if [ "$#" -ne 1 ]; then
-    printf "Usage: $0 <model>\n"
-    list_models
-
-    exit 1
-fi
-
-model=$1
-
-if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
-    printf "Invalid model: $model\n"
-    list_models
-
-    exit 1
-fi
-
-# download Core ML model
-
-printf "Downloading Core ML model $model from '$src' ...\n"
-
-cd $models_path
-
-if [ -f "ggml-$model.mlmodel" ]; then
-    printf "Model $model already exists. Skipping download.\n"
-    exit 0
-fi
-
-if [ -x "$(command -v wget)" ]; then
-    wget --quiet --show-progress -O ggml-$model.mlmodel $src/$pfx-$model.mlmodel
-elif [ -x "$(command -v curl)" ]; then
-    curl -L --output ggml-$model.mlmodel $src/$pfx-$model.mlmodel
-else
-    printf "Either wget or curl is required to download models.\n"
-    exit 1
-fi
-
-
-if [ $? -ne 0 ]; then
-    printf "Failed to download Core ML model $model \n"
-    printf "Please try again later or download the original Whisper model files and convert them yourself.\n"
-    exit 1
-fi
-
-printf "Done! Model '$model' saved in 'models/ggml-$model.mlmodel'\n"
-printf "Run the following command to compile it:\n\n"
-printf "  $ xcrun coremlc compile ./models/ggml-$model.mlmodel ./models\n\n"
-printf "You can now use it like this:\n\n"
-printf "  $ ./main -m models/ggml-$model.bin -f samples/jfk.wav\n"
-printf "\n"

models/download-ggml-model.sh

@@ -12,7 +12,7 @@ pfx="resolve/main/ggml"
 # get the path of this script
 function get_script_path() {
     if [ -x "$(command -v realpath)" ]; then
-        echo "$(dirname $(realpath $0))"
+        echo "$(dirname "$(realpath "$0")")"
     else
         local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
         echo "$ret"

talk-ggama.sh

@@ -0,0 +1,6 @@
+./talk-llama \
+    -mw ./models/ggml-small.en.bin \
+    -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin \
+    --name-ni "Georgi" \
+    --name-ai "GGaMA" \
+    -t 8 -vid 1 --speak ./examples/talk-llama/speak.sh

talk-llama.sh

@@ -0,0 +1,6 @@
+./talk-llama \
+    -mw ./models/ggml-small.en.bin \
+    -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin \
+    --name-ni "Georgi" \
+    --name-ai "LLaMA" \
+    -t 8 -vid 0 --speak ./examples/talk-llama/speak.sh

talk-rrama.sh

@@ -0,0 +1,6 @@
+./talk-llama \
+    -mw ./models/ggml-small.en.bin \
+    -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin \
+    --name-ni "Georgi" \
+    --name-ai "RRaMA" \
+    -t 8 -vid 3 --speak ./examples/talk-llama/speak.sh

talk-ssama.sh

@@ -0,0 +1,6 @@
+./talk-llama \
+    -mw ./models/ggml-small.en.bin \
+    -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin \
+    --name-ni "Georgi" \
+    --name-ai "SSaMA" \
+    -t 8 -vid 2 --speak ./examples/talk-llama/speak.sh

whisper.cpp

@@ -1,8 +1,5 @@
 #define WHISPER_BUILD
 #include "whisper.h"
-#if WHISPER_USE_COREML
-#include "coreml/whisper-encoder.h"
-#endif
 
 #include "ggml.h"
 
@@ -221,14 +218,14 @@ static const std::map<std::string, std::pair<int, std::string>> g_lang = {
     { "su",  { 98,  "sundanese",      } },
 };
 
-static const size_t MB = 1024*1024;
+static const size_t MB = 1ull*1024*1024;
 
 static const std::map<e_model, size_t> MEM_REQ_SCRATCH0 = {
-    { MODEL_TINY,     12ull*MB },
+    { MODEL_TINY,     14ull*MB },
-    { MODEL_BASE,     15ull*MB },
+    { MODEL_BASE,     18ull*MB },
-    { MODEL_SMALL,    23ull*MB },
+    { MODEL_SMALL,    28ull*MB },
-    { MODEL_MEDIUM,   31ull*MB },
+    { MODEL_MEDIUM,   36ull*MB },
-    { MODEL_LARGE,    38ull*MB },
+    { MODEL_LARGE,    44ull*MB },
 };
 
 static const std::map<e_model, size_t> MEM_REQ_SCRATCH1 = {
@@ -589,10 +586,6 @@ struct whisper_state {
 
     int lang_id = 0; // english by default
 
-#ifdef WHISPER_USE_COREML
-    whisper_coreml_context * ctx_coreml;
-#endif
-
     // [EXPERIMENTAL] token-level timestamps data
     int64_t t_beg = 0;
     int64_t t_last = 0;
@@ -661,9 +654,11 @@ static bool kv_cache_init(
                                  int   n_ctx) {
     cache.buf.resize(mem_bytes);
 
-    struct ggml_init_params params;
+    struct ggml_init_params params = {
-    params.mem_size   = cache.buf.size();
+        /*.mem_size   =*/ cache.buf.size(),
-    params.mem_buffer = cache.buf.data();
+        /*.mem_buffer =*/ cache.buf.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     cache.ctx = ggml_init(params);
 
@@ -695,9 +690,11 @@ static bool kv_cache_reinit(struct whisper_kv_cache & cache) {
 
     WHISPER_ASSERT(cache.buf.size() >= 2*n_elements*ggml_type_size(wtype));
 
-    struct ggml_init_params params;
+    struct ggml_init_params params = {
-    params.mem_size   = cache.buf.size();
+        /*.mem_size   =*/ cache.buf.size(),
-    params.mem_buffer = cache.buf.data();
+        /*.mem_buffer =*/ cache.buf.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     cache.ctx = ggml_init(params);
 
@@ -1035,9 +1032,11 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
     // create the ggml context
     {
-        struct ggml_init_params params;
+        struct ggml_init_params params = {
-        params.mem_size   = wctx.model.buf->size();
+            /*.mem_size   =*/ wctx.model.buf->size(),
-        params.mem_buffer = wctx.model.buf->data();
+            /*.mem_buffer =*/ wctx.model.buf->data(),
+            /*.no_alloc   =*/ false,
+        };
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -1351,9 +1350,11 @@ static bool whisper_encode_internal(
     const int n_mels = hparams.n_mels;
     assert(mel_inp.n_mel == n_mels);
 
-    struct ggml_init_params params;
+    struct ggml_init_params params = {
-    params.mem_size   = wstate.buf_compute.size();
+        /*.mem_size   =*/ wstate.buf_compute.size(),
-    params.mem_buffer = wstate.buf_compute.data();
+        /*.mem_buffer =*/ wstate.buf_compute.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
 
@@ -1375,7 +1376,6 @@ static bool whisper_encode_internal(
         }
     }
 
-#ifndef WHISPER_USE_COREML
     struct ggml_tensor * cur;
 
     // convolution + gelu
@@ -1684,13 +1684,6 @@ static bool whisper_encode_internal(
 
         //ggml_graph_print(&gf);
     }
-#else
-    wstate.use_buf(ctx0, -1);
-
-    struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
-
-    whisper_coreml_encode(wstate.ctx_coreml, (float *) mel->data, (float *) cur->data);
-#endif
 
     // cur
     //{
@@ -1758,10 +1751,10 @@ static bool whisper_encode_internal(
 
     //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
     //        ggml_used_mem(ctx0)/1024.0/1024.0,
-    //        wctx.get_buf_max_mem(0)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-    //        wctx.get_buf_max_mem(1)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-    //        wctx.get_buf_max_mem(2)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-    //        wctx.get_buf_max_mem(3)/1024.0/1024.0);
+    //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
 
     ggml_free(ctx0);
 
@@ -1812,9 +1805,11 @@ static bool whisper_decode_internal(
 
     //WHISPER_PRINT_DEBUG("%s: n_past = %d, N = %d, M = %d, n_ctx = %d\n", __func__, n_past, N, M, n_ctx);
 
-    struct ggml_init_params params;
+    struct ggml_init_params params = {
-    params.mem_size   = wstate.buf_compute.size();
+        /*.mem_size   =*/ wstate.buf_compute.size(),
-    params.mem_buffer = wstate.buf_compute.data();
+        /*.mem_buffer =*/ wstate.buf_compute.data(),
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
 
@@ -2168,10 +2163,10 @@ static bool whisper_decode_internal(
     if (N > 1) {
         //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
         //        ggml_used_mem(ctx0)/1024.0/1024.0,
-        //        wctx.get_buf_max_mem(0)/1024.0/1024.0,
+        //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-        //        wctx.get_buf_max_mem(1)/1024.0/1024.0,
+        //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-        //        wctx.get_buf_max_mem(2)/1024.0/1024.0,
+        //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-        //        wctx.get_buf_max_mem(3)/1024.0/1024.0);
+        //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
     }
 
     ggml_free(ctx0);
@@ -2475,20 +2470,6 @@ static std::vector<whisper_vocab::id> tokenize(const whisper_vocab & vocab, cons
 // interface implementation
 //
 
-#ifdef WHISPER_USE_COREML
-// replace .bin with .mlmodelc
-static std::string whisper_get_coreml_path(std::string path_bin) {
-    auto pos = path_bin.rfind('.');
-    if (pos != std::string::npos) {
-        path_bin = path_bin.substr(0, pos);
-    }
-
-    path_bin += ".mlmodelc";
-
-    return path_bin;
-}
-#endif
-
 struct whisper_state * whisper_init_state(whisper_context * ctx) {
     whisper_state * state = new whisper_state;
 
@@ -2533,21 +2514,6 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
     state->rng = std::mt19937(0);
 
-#ifdef WHISPER_USE_COREML
-    const auto path_coreml = whisper_get_coreml_path(ctx->path_model);
-
-    fprintf(stderr, "%s: loading Core ML model from '%s'\n", __func__, path_coreml.c_str());
-    fprintf(stderr, "%s: first run on a device may take a while ...\n", __func__);
-
-    state->ctx_coreml = whisper_coreml_init(path_coreml.c_str());
-    if (!state->ctx_coreml) {
-        fprintf(stderr, "%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
-        return nullptr;
-    }
-
-    fprintf(stderr, "%s: Core ML model loaded\n", __func__);
-#endif
-
     return state;
 }
 
@@ -2563,7 +2529,6 @@ struct whisper_context * whisper_init_from_file_no_state(const char * path_model
     }
 
     loader.context = &fin;
-
     loader.read = [](void * ctx, void * output, size_t read_size) {
         std::ifstream * fin = (std::ifstream*)ctx;
         fin->read((char *)output, read_size);
@@ -2711,10 +2676,6 @@ void whisper_free(struct whisper_context * ctx) {
 
         whisper_free_state(ctx->state);
 
-#ifdef WHISPER_USE_COREML
-        whisper_coreml_free(ctx->state->ctx_coreml);
-        ctx->state->ctx_coreml = nullptr;
-#endif
         delete ctx;
     }
 }
@@ -3170,6 +3131,7 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
         /*.speed_up         =*/ false,
         /*.audio_ctx        =*/ 0,
 
+        /*.initial_prompt   =*/ nullptr,
         /*.prompt_tokens    =*/ nullptr,
         /*.prompt_n_tokens  =*/ 0,
 
@@ -3200,6 +3162,9 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
         /*.new_segment_callback           =*/ nullptr,
         /*.new_segment_callback_user_data =*/ nullptr,
 
+        /*.progress_callback           =*/ nullptr,
+        /*.progress_callback_user_data =*/ nullptr,
+
         /*.encoder_begin_callback           =*/ nullptr,
         /*.encoder_begin_callback_user_data =*/ nullptr,
 
@@ -3842,6 +3807,15 @@ int whisper_full_with_state(
         prompt_past.clear();
     }
 
+    // initial prompt
+    if (!params.prompt_tokens && params.initial_prompt) {
+        std::vector<whisper_token> prompt_tokens;
+        prompt_tokens.resize(1024);
+        prompt_tokens.resize(whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size()));
+        params.prompt_tokens = prompt_tokens.data();
+        params.prompt_n_tokens = prompt_tokens.size();
+    }
+
     // prepend the prompt tokens to the prompt_past
     if (params.prompt_tokens && params.prompt_n_tokens > 0) {
         // parse tokens from the pointer
@@ -3907,6 +3881,10 @@ int whisper_full_with_state(
                 fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress_prev);
             }
         }
+        if (params.progress_callback) {
+            params.progress_callback(
+                ctx, ctx->state, progress_prev, params.progress_callback_user_data);
+        }
 
         // of only 1 second left, then stop
         if (seek + 100 >= seek_end) {
@@ -4495,6 +4473,9 @@ int whisper_full_parallel(
         params_cur.new_segment_callback = nullptr;
         params_cur.new_segment_callback_user_data = nullptr;
 
+        params_cur.progress_callback = nullptr;
+        params_cur.progress_callback_user_data = nullptr;
+
         workers[i] = std::thread(whisper_full_with_state, ctx, states[i], std::move(params_cur), samples + start_samples, n_samples_cur);
     }
 
@@ -4755,6 +4736,7 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
             struct ggml_init_params gparams = {
                 /*.mem_size   =*/ buf.size(),
                 /*.mem_buffer =*/ buf.data(),
+                /*.no_alloc   =*/ false,
             };
 
             struct ggml_context * ctx0 = ggml_init(gparams);

whisper.h

@@ -306,6 +306,9 @@ extern "C" {
     // Use the whisper_full_...() functions to obtain the text segments
     typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data);
 
+    // Progress callback
+    typedef void (*whisper_progress_callback)(struct whisper_context * ctx, struct whisper_state * state, int progress, void * user_data);
+
     // Encoder begin callback
     // If not NULL, called before the encoder starts
     // If it returns false, the computation is aborted
@@ -356,6 +359,7 @@ extern "C" {
 
         // tokens to provide to the whisper decoder as initial prompt
         // these are prepended to any existing text context from a previous call
+        const char * initial_prompt;
         const whisper_token * prompt_tokens;
         int prompt_n_tokens;
 
@@ -391,6 +395,10 @@ extern "C" {
         whisper_new_segment_callback new_segment_callback;
         void * new_segment_callback_user_data;
 
+        // called on each progress update
+        whisper_progress_callback progress_callback;
+        void * progress_callback_user_data;
+
         // called each time before the encoder starts
         whisper_encoder_begin_callback encoder_begin_callback;
         void * encoder_begin_callback_user_data;