whisper : Metal and ggml-alloc support (#1270)

* metal : init

* whisper : factor out graph builds

* whisper : allocate encoder and decoder using ggml-alloc

* whisper : ggml-alloc is now supported

* whisper : CoreML support ggml-alloc

* build : fix ggml-alloc

* ios : update submodule

* extra : update sync-ggml.sh script to also sync ggml-alloc

* ci : see if this is causing the crash

* whisper : refactor ggml-alloc init

* whisper.android : try to fix build

* whisper : initial Metal version

* ci : try to debug vmem issue

* metal : decoder works on GPU!

* metal : add multi-decoder support

* ggml : fix ggml_nbytes (probably temp solution)

* metal : run "cross" step on the GPU

* whisper : remove ggml_repeat in the encoder

* whisper : offload the Encoder to Metal

* ggml : use simpler ggml_bytes() implementation

* ggml-alloc : try to make CI happy by reducing vram to 128GB

* whisper : add whisper_allocr to wrap ggml_allocr

* whisper : factor out alloc init in a function

* cmake : update to support Metal build

* whisper : add <functional> header

* objc : fix build (no Metal yet)

* ios : add Metal support

* swiftui : fix build

* metal : speed-up KQ multiplication

* metal : sync latest llama.cpp kernels

* readme : add Metal info

* ios : update submodule

* coreml : add code to toggle Core ML config (CPU, ANE, GPU)

* bench : fix timings by running a pre-heat

* bench : start benching the decoder

* whisper : add ggml_mul_mat_pad

* bench : fix uninitialized vars

* whisper : add comment for disabling mul-mat padding

* whisper : add description of ggml_mul_mat_pad

* whisper : clean-up ggml_mul_mat_pad

* metal : remove the "concurrent" flag

* bench : variable n_past

* ios : update SPM package

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required (VERSION 3.0)
+cmake_minimum_required (VERSION 3.5)
 
 project(whisper.cpp VERSION 1.4.2)
 
@@ -35,6 +35,12 @@ endif()
 
 # options
 
+if (APPLE)
+    set(WHISPER_METAL_DEFAULT ON)
+else()
+    set(WHISPER_METAL_DEFAULT OFF)
+endif()
+
 option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
 
 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
@@ -58,6 +64,8 @@ option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
 
 if (APPLE)
     option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
+    option(WHISPER_METAL                 "whisper: use Metal"                    ${WHISPER_METAL_DEFAULT})
+    option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
     option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
     option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
 else()
@@ -113,6 +121,34 @@ if (APPLE)
         endif()
     endif()
 
+    if (WHISPER_METAL)
+        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+        if (METAL_FRAMEWORK)
+            message(STATUS "Metal framework found")
+
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
+                ${FOUNDATION_LIBRARY}
+                ${METAL_FRAMEWORK}
+                ${METALKIT_FRAMEWORK}
+                )
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
+
+            if (WHISPER_METAL_NDEBUG)
+                set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
+            endif()
+        else()
+            message(WARNING "Metal framework not found")
+        endif()
+
+        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+        # copy ggml-metal.metal to bin directory
+        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+    endif()
+
     if (WHISPER_COREML)
         find_library(FOUNDATION_FRAMEWORK Foundation)
         find_library(COREML_FRAMEWORK CoreML)
@@ -177,7 +213,7 @@ if (WHISPER_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
 
@@ -228,7 +264,7 @@ if (WHISPER_CLBLAST)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_OPENCL_SOURCES ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
@@ -426,8 +462,11 @@ set(TARGET whisper)
 add_library(${TARGET}
     ggml.h
     ggml.c
-    ${GGML_CUDA_SOURCES}
+    ggml-alloc.h
-    ${GGML_OPENCL_SOURCES}
+    ggml-alloc.c
+    ${GGML_SOURCES_METAL}
+    ${GGML_SOURCES_CUDA}
+    ${GGML_SOURCES_OPENCL}
     whisper.h
     whisper.cpp
     )
@@ -468,9 +507,15 @@ if (BUILD_SHARED_LIBS)
         WHISPER_BUILD
         GGML_BUILD
         )
+
+    if (WHISPER_METAL)
+        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
+        #       but for some reason it does not work here like it does in llama.cpp
+        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    endif()
 endif()
 
-if (GGML_CUDA_SOURCES)
+if (GGML_SOURCES_CUDA)
     message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
     set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES OFF)
     set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
@@ -486,10 +531,13 @@ target_compile_definitions(${TARGET} PUBLIC
 
 set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
 
+include(GNUInstallDirs)
+
 install(TARGETS ${TARGET}
-    LIBRARY DESTINATION lib
+    LIBRARY  DESTINATION lib
-    ARCHIVE DESTINATION lib/static
+    ARCHIVE  DESTINATION lib/static
-    RUNTIME DESTINATION bin
+    RUNTIME  DESTINATION bin
+    RESOURCE DESTINATION bin
     PUBLIC_HEADER DESTINATION include
     )
 

Makefile

@@ -18,7 +18,7 @@ ifndef NVCC_VERSION
 	endif
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
+CCV  := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)
 
 # Mac OS + Arm can report x86_64
@@ -182,6 +182,15 @@ ifdef WHISPER_COREML_ALLOW_FALLBACK
 endif
 endif
 
+ifndef WHISPER_NO_METAL
+	ifeq ($(UNAME_S),Darwin)
+		WHISPER_METAL := 1
+
+		CXXFLAGS += -DGGML_USE_METAL
+		LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	endif
+endif
+
 ifdef WHISPER_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	LDFLAGS += -lopenblas
@@ -288,6 +297,11 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+WHISPER_OBJ += ggml-alloc.o
+
 whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
@@ -303,6 +317,13 @@ whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-imp
 WHISPER_OBJ += whisper.o whisper-encoder.o whisper-encoder-impl.o
 endif
 
+ifdef WHISPER_METAL
+ggml-metal.o: ggml-metal.m ggml-metal.h
+	$(CC) $(CFLAGS) -c $< -o $@
+
+WHISPER_OBJ += ggml-metal.o
+endif
+
 libwhisper.a: ggml.o $(WHISPER_OBJ)
 	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
 

README.md

@@ -11,14 +11,14 @@ Beta: [v1.4.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.2) / S
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate framework and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
 - [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
 - Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
-- Runs on the CPU
+- Support for CPU-only inference
 - [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
 - [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
 - [BLAS CPU support via OpenBLAS](https://github.com/ggerganov/whisper.cpp#blas-cpu-support-via-openblas)
@@ -50,6 +50,10 @@ You can also easily make your own offline voice assistant application: [command]
 
 https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
 
+On Apply Silicon, the inference runs fully on the GPU via Metal:
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/c82e8f86-60dc-49f2-b048-d2fdbd6b5225
+
 Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 
 ## Implementation details

bindings/ios

@@ -1 +1 @@
-Subproject commit de46d9e7817fe851c109d66080239d415812d32a
+Subproject commit 22a9eef021afc67f2154bc9811ed620b26299d1b

coreml/whisper-encoder.mm

@@ -22,7 +22,13 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
 
     NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
 
-    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model error:nil]);
+    // select which device to run the Core ML model on
+    MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
+    config.computeUnits = MLComputeUnitsCPUAndGPU;
+    //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
+    //config.computeUnits = MLComputeUnitsAll;
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 
     if (data == NULL) {
         return NULL;

examples/bench/bench.cpp

@@ -44,13 +44,13 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
     fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
-    fprintf(stderr, "                           %-7s  0 - whisper encoder\n",                         "");
+    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
     fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
     fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
     fprintf(stderr, "\n");
 }
 
-int whisper_bench_encoder(const whisper_params & params) {
+int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
     struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
@@ -69,12 +69,49 @@ int whisper_bench_encoder(const whisper_params & params) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return 3;
     }
-
+    // heat encoder
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
         fprintf(stderr, "error: failed to encode model: %d\n", ret);
         return 4;
     }
 
+    whisper_token tokens[512];
+    memset(tokens, 0, sizeof(tokens));
+
+    // prompt heat
+    if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation heat
+    if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        return 4;
+    }
+
+    whisper_reset_timings(ctx);
+
+    // actual run
+    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        return 4;
+    }
+
+    for (int i = 0; i < 16; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to encode model: %d\n", ret);
+            return 4;
+        }
+    }
+
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to encode model: %d\n", ret);
+            return 4;
+        }
+    }
+
     whisper_print_timings(ctx);
     whisper_free(ctx);
 
@@ -103,7 +140,7 @@ int main(int argc, char ** argv) {
     int ret = -1;
 
     switch (params.what) {
-        case 0: ret = whisper_bench_encoder(params);                break;
+        case 0: ret = whisper_bench_full(params);                break;
         case 1: ret = whisper_bench_memcpy(params.n_threads);       break;
         case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
         default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;

examples/talk-llama/CMakeLists.txt

@@ -7,7 +7,7 @@ if (WHISPER_SDL2)
 
     # TODO: this is temporary
     #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../ggml-alloc.c ../../whisper.cpp)
 
     target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
     target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

examples/whisper.android/app/src/main/jni/whisper/CMakeLists.txt

@@ -8,6 +8,7 @@ set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../../)
 set(
         SOURCE_FILES
         ${WHISPER_LIB_DIR}/ggml.c
+        ${WHISPER_LIB_DIR}/ggml-alloc.c
         ${WHISPER_LIB_DIR}/whisper.cpp
         ${CMAKE_SOURCE_DIR}/jni.c
 )
@@ -20,7 +21,7 @@ function(build_library target_name)
         SHARED
         ${SOURCE_FILES}
     )
-    
+
     target_link_libraries(${target_name} ${LOG_LIB} android)
 
     if (${target_name} STREQUAL "whisper_v8fp16_va")

examples/whisper.objc/README.md

@@ -28,6 +28,8 @@ This can significantly improve the performance of the transcription:
 
 <img width="1072" alt="image" src="https://user-images.githubusercontent.com/1991296/208511239-8d7cdbd1-aa48-41b5-becd-ca288d53cc07.png">
 
+## Core ML
+
 If you want to enable Core ML support, you can add the `-DWHISPER_USE_COREML -DWHISPER_COREML_ALLOW_FALLBACK` compiler flag for `whisper.cpp` in Build Phases:
 
 <img width="1072" alt="image" src="https://github.com/ggerganov/whisper.cpp/assets/3001525/103e8f57-6eb6-490d-a60c-f6cf6c319324">
@@ -35,3 +37,13 @@ If you want to enable Core ML support, you can add the `-DWHISPER_USE_COREML -DW
 Then follow the [`Core ML support` section of readme](../../README.md#core-ml-support) for convert the model.
 
 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.
+
+## Metal
+
+You can also enable Metal to make the inference run on the GPU of your device. This might or might not be more efficient
+compared to Core ML depending on the model and device that you use.
+
+To enable Metal, just add `-DGGML_USE_METAL` instead off the `-DWHISPER_USE_COREML` flag and you are ready.
+This will make both the Encoder and the Decoder run on the GPU.
+
+If you want to run the Encoder with Core ML and the Decoder with Metal then simply add both `-DWHISPER_USE_COREML -DGGML_USE_METAL` flags. That's all!

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -7,6 +7,9 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
+		1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 184447182AB211A2007D6BFE /* ggml-alloc.c */; };
+		1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 1844471B2AB21655007D6BFE /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-framework Foundation -framework Metal -framework MetalKit -fno-objc-arc"; }; };
+		184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7A29052BDF00BD2A04 /* AppDelegate.m */; };
 		18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7D29052BDF00BD2A04 /* SceneDelegate.m */; };
 		18627C8129052BDF00BD2A04 /* ViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C8029052BDF00BD2A04 /* ViewController.m */; };
@@ -14,7 +17,7 @@
 		18627C8629052BE000BD2A04 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 18627C8529052BE000BD2A04 /* Assets.xcassets */; };
 		18627C8929052BE000BD2A04 /* LaunchScreen.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = 18627C8729052BE000BD2A04 /* LaunchScreen.storyboard */; };
 		18627C8C29052BE000BD2A04 /* main.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C8B29052BE000BD2A04 /* main.m */; };
-		18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML -DWHISPER_COREML_ALLOW_FALLBACK"; }; };
+		18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML"; }; };
 		18627C9629052C5800BD2A04 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9529052C5800BD2A04 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE"; }; };
 		18627C9B29052CFF00BD2A04 /* ggml-base.en.bin in Resources */ = {isa = PBXBuildFile; fileRef = 18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
@@ -23,7 +26,24 @@
 		7FE3424F2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc in Resources */ = {isa = PBXBuildFile; fileRef = 7FE3424E2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc */; };
 /* End PBXBuildFile section */
 
+/* Begin PBXCopyFilesBuildPhase section */
+		184447202AB21B25007D6BFE /* CopyFiles */ = {
+			isa = PBXCopyFilesBuildPhase;
+			buildActionMask = 2147483647;
+			dstPath = "";
+			dstSubfolderSpec = 7;
+			files = (
+				184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXCopyFilesBuildPhase section */
+
 /* Begin PBXFileReference section */
+		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml-alloc.c"; sourceTree = "<group>"; };
+		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml-alloc.h"; sourceTree = "<group>"; };
+		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml-metal.m"; sourceTree = "<group>"; };
+		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml-metal.metal"; sourceTree = "<group>"; };
 		18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
 		18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@@ -80,6 +100,10 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
+				1844471D2AB2195F007D6BFE /* ggml-metal.metal */,
+				1844471B2AB21655007D6BFE /* ggml-metal.m */,
+				184447182AB211A2007D6BFE /* ggml-alloc.c */,
+				184447192AB211A2007D6BFE /* ggml-alloc.h */,
 				7FE3424E2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc */,
 				7FE342442A0C3FA20015A058 /* coreml */,
 				18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */,
@@ -126,6 +150,7 @@
 				18627C7229052BDF00BD2A04 /* Sources */,
 				18627C7329052BDF00BD2A04 /* Frameworks */,
 				18627C7429052BDF00BD2A04 /* Resources */,
+				184447202AB21B25007D6BFE /* CopyFiles */,
 			);
 			buildRules = (
 			);
@@ -194,8 +219,10 @@
 				18627C9629052C5800BD2A04 /* ggml.c in Sources */,
 				18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */,
 				7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */,
+				1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
 				18627C8C29052BE000BD2A04 /* main.m in Sources */,
 				18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */,
+				1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */,
 				7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -20,6 +20,7 @@
 		0AAC5DCC29539EB1003032C3 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DC929539EB0003032C3 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -Wno-shorten-64-to-32"; }; };
 		0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DCD2953A05C003032C3 /* WhisperState.swift */; };
 		0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DD02953A394003032C3 /* LibWhisper.swift */; };
+		18AED4812AB21F2B009D854F /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 18AED47F2AB21F2B009D854F /* ggml-alloc.c */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXFileReference section */
@@ -41,6 +42,8 @@
 		0AAC5DCA29539EB0003032C3 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = ggml.h; sourceTree = "<group>"; };
 		0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
 		0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
+		18AED47F2AB21F2B009D854F /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; path = "ggml-alloc.c"; sourceTree = "<group>"; };
+		18AED4802AB21F2B009D854F /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-alloc.h"; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
@@ -124,6 +127,8 @@
 		0AAC5DC529539E89003032C3 /* whisper.cpp */ = {
 			isa = PBXGroup;
 			children = (
+				18AED47F2AB21F2B009D854F /* ggml-alloc.c */,
+				18AED4802AB21F2B009D854F /* ggml-alloc.h */,
 				0AAC5DC929539EB0003032C3 /* ggml.c */,
 				0AAC5DCA29539EB0003032C3 /* ggml.h */,
 				0AAC5DC729539EB0003032C3 /* whisper.cpp */,
@@ -242,6 +247,7 @@
 				0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */,
 				0AAC5DCB29539EB1003032C3 /* whisper.cpp in Sources */,
 				0AA7514E2953D958001EE061 /* Recorder.swift in Sources */,
+				18AED4812AB21F2B009D854F /* ggml-alloc.c in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -369,7 +375,7 @@
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
 				DEVELOPMENT_ASSET_PATHS = "\"whisper.swiftui.demo/Supporting files/Preview Content\"";
-				DEVELOPMENT_TEAM = 3TZ9BM962G;
+				DEVELOPMENT_TEAM = P8JZH34X63;
 				ENABLE_HARDENED_RUNTIME = YES;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
@@ -410,7 +416,7 @@
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
 				DEVELOPMENT_ASSET_PATHS = "\"whisper.swiftui.demo/Supporting files/Preview Content\"";
-				DEVELOPMENT_TEAM = 3TZ9BM962G;
+				DEVELOPMENT_TEAM = P8JZH34X63;
 				ENABLE_HARDENED_RUNTIME = YES;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;

extra/bench-all.sh

@@ -44,27 +44,26 @@ if [ "$encoder_only" -eq 0 ]; then
     printf "\n"
 fi
 
-printf "| CPU | OS | Config | Model | Th | Load | Enc. | Commit |\n"
+printf "| %6s | %6s | %12s | %9s | %3s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "PP" "Commit"
-printf "| --- | -- | ------ | ----- | -- | ---- | ---- | ------ |\n"
+printf "| %6s | %6s | %12s | %9s | %3s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---"
 
 for model in "${models[@]}"; do
-    # run once to heat-up the cache
-    ./bench -m ./models/ggml-$model.bin -t $n_threads 2>/dev/null 1>/dev/null
-
     # actual run
     # store stderr output in a variable in order to parse it later
     output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)
     ret=$?
 
     # parse the output:
-    load_time=$(echo "$output" | grep "load time" | awk '{print $5}')
+    encode_time=$(echo "$output" | grep "encode time" | awk '{print $11}')
-    encode_time=$(echo "$output" | grep "encode time" | awk '{print $5}')
+    decode_time=$(echo "$output" | grep "decode time" | awk '{print $11}')
+    prompt_time=$(echo "$output" | grep "prompt time" | awk '{print $11}')
     system_info=$(echo "$output" | grep "system_info")
     n_threads=$(echo "$output" | grep "system_info" | awk '{print $4}')
 
     # floor to milliseconds
-    load_time=${load_time%.*}
+    #encode_time=${encode_time%.*}
-    encode_time=${encode_time%.*}
+    #decode_time=${decode_time%.*}
+    #prompt_time=${prompt_time%.*}
 
     config=""
 
@@ -87,6 +86,6 @@ for model in "${models[@]}"; do
     commit=$(git rev-parse --short HEAD)
 
     if [ $ret -eq 0 ]; then
-        printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
+        printf "| <todo> | <todo> | %12s | %9s | %3s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$prompt_time" "$commit"
     fi
 done

extra/sync-ggml.sh

@@ -1,18 +1,20 @@
 #!/bin/bash
 
-cp -rpv ../ggml/src/ggml.c               ./ggml.c
+cp -rpv ../ggml/src/ggml.c                ./ggml.c
-cp -rpv ../ggml/src/ggml-cuda.h          ./ggml-cuda.h
+cp -rpv ../ggml/src/ggml-alloc.c          ./ggml-alloc.c
-cp -rpv ../ggml/src/ggml-cuda.cu         ./ggml-cuda.cu
+cp -rpv ../ggml/src/ggml-cuda.h           ./ggml-cuda.h
-cp -rpv ../ggml/src/ggml-opencl.h        ./ggml-opencl.h
+cp -rpv ../ggml/src/ggml-cuda.cu          ./ggml-cuda.cu
-cp -rpv ../ggml/src/ggml-opencl.cpp      ./ggml-opencl.cpp
+cp -rpv ../ggml/src/ggml-opencl.h         ./ggml-opencl.h
-cp -rpv ../ggml/src/ggml-metal.h         ./ggml-metal.h
+cp -rpv ../ggml/src/ggml-opencl.cpp       ./ggml-opencl.cpp
-cp -rpv ../ggml/src/ggml-metal.m         ./ggml-metal.m
+cp -rpv ../ggml/src/ggml-metal.h          ./ggml-metal.h
-cp -rpv ../ggml/src/ggml-metal.metal     ./ggml-metal.metal
+cp -rpv ../ggml/src/ggml-metal.m          ./ggml-metal.m
-cp -rpv ../ggml/include/ggml/ggml.h      ./ggml.h
+cp -rpv ../ggml/src/ggml-metal.metal      ./ggml-metal.metal
-cp -rpv ../ggml/examples/common.h        ./examples/common.h
+cp -rpv ../ggml/include/ggml/ggml.h       ./ggml.h
-cp -rpv ../ggml/examples/common.cpp      ./examples/common.cpp
+cp -rpv ../ggml/include/ggml/ggml-alloc.h ./ggml-alloc.h
-cp -rpv ../ggml/examples/common-ggml.h   ./examples/common-ggml.h
+cp -rpv ../ggml/examples/common.h         ./examples/common.h
-cp -rpv ../ggml/examples/common-ggml.cpp ./examples/common-ggml.cpp
+cp -rpv ../ggml/examples/common.cpp       ./examples/common.cpp
+cp -rpv ../ggml/examples/common-ggml.h    ./examples/common-ggml.h
+cp -rpv ../ggml/examples/common-ggml.cpp  ./examples/common-ggml.cpp
 
 cp -rpv ../ggml/examples/whisper/whisper.h    ./whisper.h
 cp -rpv ../ggml/examples/whisper/whisper.cpp  ./whisper.cpp

ggml-alloc.c

@@ -6,6 +6,26 @@
 #include <stdlib.h>
 #include <string.h>
 
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/types.h>
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <memoryapi.h>
+#endif
+
+
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
@@ -99,15 +119,28 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-
+static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     return ggml_nbytes(tensor);
 
     UNUSED(alloc);
 }
 
+// check if a tensor is allocated by this buffer
+static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+}
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+#ifdef GGML_ALLOCATOR_DEBUG
+    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
+    GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
+#endif
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -131,14 +164,14 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
-            max_avail = MAX(max_avail, block->size);
         } else {
             fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
                     __func__, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
-        return;
+            return;
         }
     }
     struct free_block * block = &alloc->free_blocks[best_fit_block];
@@ -173,17 +206,17 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     void * ptr = tensor->data;
 
-    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+    if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
         return;
     }
 
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
 
@@ -277,17 +310,68 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-// address and size of the buffer when measuring
+// OS specific functions to allocate and free uncommitted virtual memory
-// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
+static void * alloc_vmem(size_t size) {
-static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
+#if defined(_WIN32)
-static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
+#elif defined(_POSIX_MAPPED_FILES)
+    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+    return ptr;
+#else
+    // use a fixed address for other platforms
+    uintptr_t base_addr = (uintptr_t)-size - 0x100;
+    return (void *)base_addr;
+#endif
+}
+
+static void free_vmem(void * base_addr, size_t size) {
+#if defined(_WIN32)
+    VirtualFree(base_addr, 0, MEM_RELEASE);
+    UNUSED(size);
+#elif defined(_POSIX_MAPPED_FILES)
+    munmap(base_addr, size);
+#else
+    // nothing to do
+    UNUSED(base_addr);
+    UNUSED(size);
+#endif
+}
+
+// allocate uncommitted virtual memory to measure the size of the graph
+static void alloc_measure_vmem(void ** base_addr, size_t * size) {
+    // 128GB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
+    do {
+        *base_addr = alloc_vmem(*size);
+        if (*base_addr != NULL) {
+            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
+            return;
+        }
+        // try again with half the size
+        *size /= 2;
+    } while (*size > 0);
+
+    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+}
+
+static void free_measure_vmem(void * base_addr, size_t size) {
+    free_vmem(base_addr, size);
+}
 
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
     struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
 
+    void * base_addr;
+    size_t size;
+
+    alloc_measure_vmem(&base_addr, &size);
+
     *alloc = (struct ggml_allocr){
-        /*.data          = */ MEASURE_BASE_ADDR,
+        /*.data          = */ base_addr,
-        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.size          = */ size,
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -307,6 +391,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
+    if (alloc->measure) {
+        free_measure_vmem(alloc->data, alloc->size);
+    }
     free(alloc);
 }
 
@@ -316,11 +403,6 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 
 //////////// compute graph allocator
 
-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
-           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
-}
-
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
     if (a->type != b->type) {
         return false;
@@ -336,28 +418,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
-static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
-    switch (t->op) {
-        case GGML_OP_PERMUTE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-            return t->src[0];
-        case GGML_OP_CPY:
-            return t->src[1];
-        default:
-            return NULL;
-    }
-}
-
-static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
-    struct ggml_tensor * parent = t;
-    do {
-        parent = get_view_parent(parent);
-    } while (ggml_is_view(parent));
-    return parent;
-}
-
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -365,7 +425,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -375,10 +434,8 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SET:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_CONT:
-        case GGML_OP_ADD_REL_POS:
             return true;
 
         default:
@@ -390,24 +447,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            size_t offset;
+            assert(node->view_src->data != NULL);
-            switch(node->op) {
+            node->data = (char *)node->view_src->data + node->view_offs;
-                case GGML_OP_VIEW:
-                    memcpy(&offset, node->op_params, sizeof(size_t));
-                    node->data = (char *) node->src[0]->data + offset;
-                    break;
-                case GGML_OP_PERMUTE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_TRANSPOSE:
-                    node->data = node->src[0]->data;
-                    break;
-                case GGML_OP_CPY:
-                    node->data = node->src[1]->data;
-                    break;
-                default:
-                    GGML_ASSERT(!"unknown view op");
-                    break;
-            }
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -418,8 +459,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     }
 
                     // if the node's data is external, then we cannot re-use it
-                    if ((char *) parent->data < (char *) alloc->data ||
+                    if (ggml_allocr_is_own(alloc, parent) == false) {
-                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
                         AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
                         continue;
                     }
@@ -427,7 +467,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct ggml_tensor * view_src = parent->view_src;
                             struct hash_node * view_src_hn = hash_get(ht, view_src);
                             if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                                 // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
@@ -453,7 +493,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-static size_t ggml_allocator_alloc_graph_tensors_n(
+static size_t ggml_allocr_alloc_graph_tensors_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -469,7 +509,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             struct ggml_tensor * node = gf->nodes[i];
 
             if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = get_view_source(node);
+                struct ggml_tensor * view_src = node->view_src;
                 hash_get(ht, view_src)->n_views += 1;
             }
 
@@ -531,11 +571,10 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 AT_PRINTF("\n");
             }
 
-
             // update parents
             // update immediately if there is no parse_seq
             // update only at barriers if there is parse_seq
-            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) {
+            if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) {
                 int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
                 int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
                 for (int i = update_start; i < update_end; i++) {
@@ -554,17 +593,17 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 
                         if (p_hn->n_children == 0 && p_hn->n_views == 0) {
                             if (ggml_is_view(parent)) {
-                                struct ggml_tensor * view_src = get_view_source(parent);
+                                struct ggml_tensor * view_src = parent->view_src;
                                 struct hash_node * view_src_hn = hash_get(ht, view_src);
                                 view_src_hn->n_views -= 1;
                                 AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
                                 if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, view_src);
+                                    ggml_allocr_free_tensor(alloc, view_src);
                                 }
                             }
                             else {
                                 if (parent->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, parent);
+                                    ggml_allocr_free_tensor(alloc, parent);
                                 }
                             }
                         }
@@ -581,7 +620,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             for (int i = 0; outputs[g][i] != NULL; i++) {
                 struct ggml_tensor * output = outputs[g][i];
                 AT_PRINTF("output: %s\n", output->name);
-                ggml_allocator_free_tensor(alloc, output);
+                ggml_allocr_free_tensor(alloc, output);
             }
         }
     }
@@ -590,5 +629,5 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
 }

ggml-metal.m

@@ -63,7 +63,10 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
     GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(soft_max_4);
     GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DECL_KERNEL(get_rows_f32);
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
@@ -77,6 +80,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(norm);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_l4);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
@@ -117,14 +121,17 @@ static NSString * const msl_library_source = @"see metal.metal";
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
     metal_printf("%s: allocating\n", __func__);
 
-    // Show all the Metal device instances in the system
-    NSArray * devices = MTLCopyAllDevices();
     id <MTLDevice> device;
     NSString * s;
+
+#if TARGET_OS_OSX
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
     for (device in devices) {
         s = [device name];
         metal_printf("%s: found device: %s\n", __func__, [s UTF8String]);
     }
+#endif
 
     // Pick and show default Metal device
     device = MTLCreateSystemDefaultDevice();
@@ -139,14 +146,22 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     ctx->n_buffers = 0;
     ctx->concur_list_len = 0;
 
-    ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+    ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-#if 0
+#ifdef GGML_SWIFT
-    // compile from source string and show compile log
+    // load the default.metallib file
     {
         NSError * error = nil;
 
-        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+        NSString * llamaBundlePath = [bundle pathForResource:@"llama_llama" ofType:@"bundle"];
+        NSBundle * llamaBundle = [NSBundle bundleWithPath:llamaBundlePath];
+        NSString * libPath = [llamaBundle pathForResource:@"default" ofType:@"metallib"];
+        NSURL * libURL = [NSURL fileURLWithPath:libPath];
+
+        // Load the metallib file into a Metal library
+        ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+
         if (error) {
             metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
@@ -161,7 +176,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
         //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
         NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        NSString * path   = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
         metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
 
         NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
@@ -207,7 +222,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
         GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(soft_max_4);
         GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf_8);
+        GGML_METAL_ADD_KERNEL(get_rows_f32);
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
@@ -221,6 +239,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(norm);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_l4);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
@@ -247,13 +266,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
     metal_printf("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+#if TARGET_OS_OSX
+    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
     if (ctx->device.maxTransferRate != 0) {
         metal_printf("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
         metal_printf("%s: maxTransferRate               = built-in GPU\n", __func__);
     }
+#endif
 
     return ctx;
 }
@@ -273,7 +294,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
     GGML_METAL_DEL_KERNEL(soft_max);
+    GGML_METAL_DEL_KERNEL(soft_max_4);
     GGML_METAL_DEL_KERNEL(diag_mask_inf);
+    GGML_METAL_DEL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DEL_KERNEL(get_rows_f32);
     GGML_METAL_DEL_KERNEL(get_rows_f16);
     GGML_METAL_DEL_KERNEL(get_rows_q4_0);
     GGML_METAL_DEL_KERNEL(get_rows_q4_1);
@@ -287,6 +311,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(norm);
     GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_l4);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
@@ -365,6 +390,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
     for (int i = 0; i < ctx->n_buffers; ++i) {
         const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
 
+        //metal_printf("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
         if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
@@ -454,6 +480,7 @@ bool ggml_metal_add_buffer(
             }
         }
 
+#if TARGET_OS_OSX
         metal_printf(", (%8.2f / %8.2f)",
                 ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
                 ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
@@ -463,6 +490,9 @@ bool ggml_metal_add_buffer(
         } else {
             metal_printf("\n");
         }
+#else
+        metal_printf(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
     }
 
     return true;
@@ -698,6 +728,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_ADD:
                         {
                             GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));
 
                             // utilize float4
                             GGML_ASSERT(ne00 % 4 == 0);
@@ -705,6 +736,7 @@ void ggml_metal_graph_compute(
 
                             if (ggml_nelements(src1) == ne10) {
                                 // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
                                 [encoder setComputePipelineState:ctx->pipeline_add_row];
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_add];
@@ -721,6 +753,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_MUL:
                         {
                             GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));
 
                             // utilize float4
                             GGML_ASSERT(ne00 % 4 == 0);
@@ -728,6 +761,7 @@ void ggml_metal_graph_compute(
 
                             if (ggml_nelements(src1) == ne10) {
                                 // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
                                 [encoder setComputePipelineState:ctx->pipeline_mul_row];
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_mul];
@@ -743,6 +777,8 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_SCALE:
                         {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+
                             const float scale = *(const float *) src1->data;
 
                             [encoder setComputePipelineState:ctx->pipeline_scale];
@@ -750,7 +786,7 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst)/4;
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
@@ -762,7 +798,7 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
@@ -782,7 +818,7 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
@@ -796,13 +832,16 @@ void ggml_metal_graph_compute(
                         {
                             const int nth = 32;
 
-                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            if (ne00%4 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -810,14 +849,23 @@ void ggml_metal_graph_compute(
                         {
                             const int n_past = ((int32_t *)(dst->op_params))[0];
 
-                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            if (ne00%8 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf_8];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            if (ne00%8 == 0) {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
+                            else {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
                         } break;
                     case GGML_OP_MUL_MAT:
                         {
@@ -830,8 +878,8 @@ void ggml_metal_graph_compute(
 
                             // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                             // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                            if (ggml_is_contiguous(src0) &&
+                            if (!ggml_is_transposed(src0) &&
-                                ggml_is_contiguous(src1) &&
+                                !ggml_is_transposed(src1) &&
                                 src1t == GGML_TYPE_F32 &&
                                 [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
                                 ne00%32 == 0 &&
@@ -856,14 +904,18 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
                                 [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
                                 [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
-                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:8];
+                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
-                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:9];
+                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
-                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:10];
+                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:13];
                                 [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                             } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
+                                int nrows = 1;
 
                                 // use custom matrix x vector kernel
                                 switch (src0t) {
@@ -873,8 +925,14 @@ void ggml_metal_graph_compute(
                                             nth1 = 1;
                                             if (ne11 * ne12 < 4) {
                                                 [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                            //} else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                            } else if (false) {
+                                                // TODO: with ggml_mul_mat_pad this kernel no longer seems to be needed
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_l4];
+                                                nrows = ne11;
                                             } else {
                                                 [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                                nrows = 4;
                                             }
                                         } break;
                                     case GGML_TYPE_Q4_0:
@@ -995,7 +1053,7 @@ void ggml_metal_graph_compute(
                                 else if (src0t == GGML_TYPE_Q6_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
-                                    int64_t ny = (ne11 + 3)/4;
+                                    int64_t ny = (ne11 + nrows - 1)/nrows;
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
@@ -1003,6 +1061,7 @@ void ggml_metal_graph_compute(
                     case GGML_OP_GET_ROWS:
                         {
                             switch (src0->type) {
+                                case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f32];  break;
                                 case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                 case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
@@ -1018,9 +1077,9 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
-                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:5];
 
                             const int64_t n = ggml_nelements(src1);
 

ggml-metal.metal

@@ -38,7 +38,7 @@ kernel void kernel_add_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   int64_t & nb,
+        constant    int64_t & nb,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -63,18 +63,18 @@ kernel void kernel_mul_row(
 }
 
 kernel void kernel_scale(
-        device const float * src0,
+        device const float4 * src0,
-        device       float * dst,
+        device       float4 * dst,
         constant     float & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }
 
 kernel void kernel_silu(
-        device const float * src0,
+        device const float4 * src0,
-        device       float * dst,
+        device       float4 * dst,
         uint tpig[[thread_position_in_grid]]) {
-    float x = src0[tpig];
+    device const float4 & x = src0[tpig];
     dst[tpig] = x / (1.0f + exp(-x));
 }
 
@@ -89,10 +89,10 @@ constant float GELU_COEF_A    = 0.044715f;
 constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
 
 kernel void kernel_gelu(
-    device const float * src0,
+    device const float4 * src0,
-    device       float * dst,
+    device       float4 * dst,
     uint tpig[[thread_position_in_grid]]) {
-    float x = src0[tpig];
+    device const float4 & x = src0[tpig];
 
     // BEWARE !!!
     // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
@@ -107,7 +107,6 @@ kernel void kernel_soft_max(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        threadgroup float  * buf [[threadgroup(0)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -119,64 +118,70 @@ kernel void kernel_soft_max(
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    buf[tpitg[0]] = -INFINITY;
+    float lmax = psrc0[tpitg[0]];
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
-        buf[tpitg[0]] = MAX(buf[tpitg[0]], psrc0[i00]);
+        lmax = MAX(lmax, psrc0[i00]);
     }
-
+    const float max = simd_max(lmax);
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg[0]/2; i > 0; i /= 2) {
-        if (tpitg[0] < i) {
-            buf[tpitg[0]] = MAX(buf[tpitg[0]], buf[tpitg[0] + i]);
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    //// broadcast - not needed. There is a threadgroup barrier above in the last iteration of
-    //               the loop, and when that is done, buf[0] has the correct (synchronized) value
-    //if (tpitg[0] == 0) {
-    //    buf[0] = buf[0];
-    //}
-
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    const float max = buf[0];
 
     // parallel sum
-    buf[tpitg[0]] = 0.0f;
+    float lsum = 0.0f;
     for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
         const float exp_psrc0 = exp(psrc0[i00] - max);
-        buf[tpitg[0]] += exp_psrc0;
+        lsum += exp_psrc0;
         // Remember the result of exp here. exp is expensive, so we really do not
         // whish to compute it twice.
         pdst[i00] = exp_psrc0;
     }
 
-    // reduce
+    const float sum = simd_sum(lsum);
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg[0]/2; i > 0; i /= 2) {
-        if (tpitg[0] < i) {
-            buf[tpitg[0]] += buf[tpitg[0] + i];
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    // broadcast - not needed, see above
-    //// broadcast
-    //if (tpitg[0] == 0) {
-    //    buf[0] = buf[0];
-    //}
-
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    const float sum = buf[0];
 
     for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
         pdst[i00] /= sum;
     }
 }
 
+kernel void kernel_soft_max_4(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    // parallel max
+    float4 lmax4 = psrc4[tpitg[0]];
+    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
+        lmax4 = fmax(lmax4, psrc4[i00]);
+    }
+    float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    const float max = simd_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+        const float4 exp_psrc4 = exp(psrc4[i00] - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    const float sum = simd_sum(lsum);
+
+    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+        pdst4[i00] /= sum;
+    }
+}
+
 kernel void kernel_diag_mask_inf(
         device const float * src0,
         device       float * dst,
@@ -192,6 +197,33 @@ kernel void kernel_diag_mask_inf(
         dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
     } else {
         dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+     }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        device const float4 * src0,
+        device       float4 * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne01,
+        constant        int & n_past,
+        uint3 tpig[[thread_position_in_grid]]) {
+
+    const int64_t i = 2*tpig[0];
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int64_t i4 = 4*i;
+    const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    const int64_t i01 = i4/(ne00);      i4 -= i01*ne00;
+    const int64_t i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        dst[i+1][k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            dst[i][k] = -INFINITY;
+        }
     }
 }
 
@@ -616,6 +648,49 @@ kernel void kernel_mul_mat_f16_f32(
     }
 }
 
+// Assumes row size (ne00) is a multiple of 4
+kernel void kernel_mul_mat_f16_f32_l4(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int nrows = ne11;
+    const int64_t r0 = tgpig.x;
+    const int64_t im = tgpig.z;
+
+    device const half4 * x4 = (device const half4 *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
+
+        float sumf = 0;
+        for (int i = tiisg; i < ne00/4; i += 32) {
+            for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
+        }
+
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}
+
 kernel void kernel_alibi_f32(
         device const float * src0,
         device       float * dst,
@@ -1123,31 +1198,40 @@ kernel void kernel_mul_mat_q3_K_f32(
     device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
     device const float     * yy = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
 
-    float yl[16];
+    float yl[32];
 
-    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask1 = 0x3030;
     const uint16_t kmask2 = 0x0f0f;
 
-    const int tid = tiisg/2;
+    const int tid = tiisg/4;
-    const int ix  = tiisg%2;
+    const int ix  = tiisg%4;
-    const int ip  = tid/8;          // 0 or 1
+    const int ip  = tid/4;          // 0 or 1
-    const int il  = tid/2 - 4*ip;   // 0...3
+    const int il  = 2*((tid%4)/2);  // 0 or 2
     const int ir  = tid%2;
     const int n   = 8;
     const int l0  = n*ir;
 
-    const uint16_t m1 = 1 << (4*ip + il);
+    // One would think that the Metal compiler would figure out that ip and il can only have
-    const uint16_t m2 = m1 << 8;
+    // 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it
+    // with these two tales.
+    //
+    // Possible masks for the high bit
+    const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200},  // ip = 0, il = 0
+                           {0x0004, 0x0400, 0x0008, 0x0800},  // ip = 0, il = 2
+                           {0x0010, 0x1000, 0x0020, 0x2000},  // ip = 1, il = 0
+                           {0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2
+
+    // Possible masks for the low 2 bits
+    const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}};
+
+    const ushort4 hm = mm[2*ip + il/2];
 
     const int shift = 2*il;
-    const uint16_t qm1 = 0x0003 << shift;
+    const float    v1 = il == 0 ? 4.f : 64.f;
-    const uint16_t qm2 = 0x0300 << shift;
+    const float    v2 = 4.f * v1;
-    const int32_t v1 = 4 << shift;
-    const int32_t v2 = 1024 << shift;
 
     const uint16_t s_shift1 = 4*ip;
-    const uint16_t s_shift2 = s_shift1 + 2*(il/2);
+    const uint16_t s_shift2 = s_shift1 + il;
-    const int ik = 4 + (il%2);
 
     const int q_offset = 32*ip + l0;
     const int y_offset = 128*ip + 32*il + l0;
@@ -1156,12 +1240,19 @@ kernel void kernel_mul_mat_q3_K_f32(
 
     device const float * y1 = yy + ix*QK_K + y_offset;
 
-    float sumf1[2] = {0.f}, sumf2[2] = {0.f};
+    uint32_t scales32, aux32;
-    for (int i = ix; i < nb; i += 2) {
+    thread uint16_t * scales16 = (thread uint16_t *)&scales32;
+    thread const int8_t * scales = (thread const int8_t *)&scales32;
+
+    float sumf1[2] = {0.f};
+    float sumf2[2] = {0.f};
+    for (int i = ix; i < nb; i += 4) {
 
         for (int l = 0; l < 8; ++l) {
-            yl[l+0] = y1[l+ 0];
+            yl[l+ 0] = y1[l+ 0];
-            yl[l+8] = y1[l+16];
+            yl[l+ 8] = y1[l+16];
+            yl[l+16] = y1[l+32];
+            yl[l+24] = y1[l+48];
         }
 
         device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
@@ -1172,27 +1263,43 @@ kernel void kernel_mul_mat_q3_K_f32(
         for (int row = 0; row < 2; ++row) {
 
             const float d_all = (float)dh[0];
-            const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
 
-            float s1 = 0, s2 = 0;
+            scales16[0] = a[4];
-            for (int l = 0; l < n; l += 2) {
+            scales16[1] = a[5];
-                const uint16_t qs = q[l/2];
+            aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030;
-                s1 += yl[l+0] * ((int32_t)(qs & qm1) - ((h[l/2] & m1) ? 0 : v1));
+            scales16[0] = a[il+0];
-                s2 += yl[l+1] * ((int32_t)(qs & qm2) - ((h[l/2] & m2) ? 0 : v2));
+            scales16[1] = a[il+1];
-            }
+            scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32;
-            float d = d_all * (s1 + 1.f/256.f * s2);
-            sumf1[row] += d * scales[0];
-            sumf2[row] += d;
 
-            s1 = s2 = 0;
+            float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0;
             for (int l = 0; l < n; l += 2) {
-                const uint16_t qs = q[l/2+8];
+                const int32_t qs = q[l/2];
-                s1 += yl[l+8] * ((int32_t)(qs & qm1) - ((h[l/2+8] & m1) ? 0 : v1));
+                s1 += yl[l+0] * (qs & qm[il/2][0]);
-                s2 += yl[l+9] * ((int32_t)(qs & qm2) - ((h[l/2+8] & m2) ? 0 : v2));
+                s2 += yl[l+1] * (qs & qm[il/2][1]);
+                s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]);
+                s4 += yl[l+16] * (qs & qm[il/2][2]);
+                s5 += yl[l+17] * (qs & qm[il/2][3]);
+                s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]);
             }
-            d = d_all * (s1 + 1.f/256.f * s2);
+            float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
-            sumf1[row] += d * scales[1];
+            float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
-            sumf2[row] += d;
+            sumf1[row] += d1 * (scales[0] - 32);
+            sumf2[row] += d2 * (scales[2] - 32);
+
+            s1 = s2 = s3 = s4 = s5 = s6 = 0;
+            for (int l = 0; l < n; l += 2) {
+                const int32_t qs = q[l/2+8];
+                s1 += yl[l+8] * (qs & qm[il/2][0]);
+                s2 += yl[l+9] * (qs & qm[il/2][1]);
+                s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]);
+                s4 += yl[l+24] * (qs & qm[il/2][2]);
+                s5 += yl[l+25] * (qs & qm[il/2][3]);
+                s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]);
+            }
+            d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
+            d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
+            sumf1[row] += d1 * (scales[1] - 32);
+            sumf2[row] += d2 * (scales[3] - 32);
 
             q  += step;
             h  += step;
@@ -1201,15 +1308,17 @@ kernel void kernel_mul_mat_q3_K_f32(
 
         }
 
-        y1 += 2 * QK_K;
+        y1 += 4 * QK_K;
 
     }
 
     for (int row = 0; row < 2; ++row) {
-        const float sumf = (sumf1[row] - 32.f*sumf2[row]) / (1 << shift);
+        const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
-        const float tot = simd_sum(sumf);
+        sumf1[row] = simd_sum(sumf);
-        if (tiisg == 0) {
+    }
-            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = tot;
+    if (tiisg == 0) {
+        for (int row = 0; row < 2; ++row) {
+            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = sumf1[row];
         }
     }
 }
@@ -1564,17 +1673,25 @@ kernel void kernel_mul_mat_q5_K_f32(
             sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
             sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
 
-            float4 acc = {0.f, 0.f, 0.f, 0.f};
+            float4 acc1 = {0.f};
+            float4 acc2 = {0.f};
             for (int l = 0; l < n; ++l) {
                 uint8_t h = qh[l];
-                acc[0] += yl[l+0] * ((uint16_t)(q1[l] & 0x0F) + (h & hm1 ? 16 : 0));
+                acc1[0] += yl[l+0] * (q1[l] & 0x0F);
-                acc[1] += yl[l+8] * ((uint16_t)(q1[l] & 0xF0) + (h & hm2 ? 256 : 0));
+                acc1[1] += yl[l+8] * (q1[l] & 0xF0);
-                acc[2] += yh[l+0] * ((uint16_t)(q2[l] & 0x0F) + (h & hm3 ? 16 : 0));
+                acc1[2] += yh[l+0] * (q2[l] & 0x0F);
-                acc[3] += yh[l+8] * ((uint16_t)(q2[l] & 0xF0) + (h & hm4 ? 256 : 0));
+                acc1[3] += yh[l+8] * (q2[l] & 0xF0);
+                acc2[0] += h & hm1 ? yl[l+0] : 0.f;
+                acc2[1] += h & hm2 ? yl[l+8] : 0.f;
+                acc2[2] += h & hm3 ? yh[l+0] : 0.f;
+                acc2[3] += h & hm4 ? yh[l+8] : 0.f;
             }
             const float dall = dh[0];
             const float dmin = dh[1];
-            sumf[row] += dall * (acc[0] * sc8[0] + acc[1] * sc8[1] * 1.f/16.f + acc[2] * sc8[4] + acc[3] * sc8[5] * 1.f/16.f) -
+            sumf[row] += dall * (sc8[0] * (acc1[0] +  16.f*acc2[0]) +
+                                 sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) +
+                                 sc8[4] * (acc1[2] +  16.f*acc2[2]) +
+                                 sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
                          dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
 
             q1 += step;
@@ -1747,6 +1864,15 @@ kernel void kernel_mul_mat_q6_K_f32(
 
 //============================= templates and their specializations =============================
 
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    float4x4 temp = *(((device float4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
+    }
+}
+
 template <typename type4x4>
 void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
     half4x4 temp = *(((device half4x4 *)src));
@@ -1758,28 +1884,30 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg)
 template <typename type4x4>
 void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 1);
-    const half d = il ? (xb->d / 16.h) : xb->d;
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
-    const half m = il ? ( -8.h * 16.h) : -8.h;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
     const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = il ? 0xF000 : 0x0F00;
+    const ushort mask1 = mask0 << 8;
 
     for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)]   = (((qs[i] & mask0)     ) + m) * d;
+        reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
-        reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) + m) * d;
+        reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
     }
 }
 
 template <typename type4x4>
 void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 2);
-    const half d = il ? (xb->d / 16.h) : xb->d;
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
-    const half m = xb->m;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
     const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = il ? 0xF000 : 0x0F00;
+    const ushort mask1 = mask0 << 8;
 
     for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)]   = (((qs[i] & mask0)     ) * d) + m;
+        reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) * d) + m;
+        reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
     }
 }
 
@@ -1815,7 +1943,7 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg
 
 template <typename type4x4>
 void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
-    const float d_all = (float)(xb->d);
+    const half d_all = xb->d;
     device const uint8_t * q = (device const uint8_t *)xb->qs;
     device const uint8_t * h = (device const uint8_t *)xb->hmask;
     device const int8_t * scales = (device const int8_t *)xb->scales;
@@ -1828,16 +1956,18 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
                                  ((il/4)>0 ? 12  : 3);
     uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
     uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
-    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2) : \
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
-                                 (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
-    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h);
+    const half ml = 4.h * dl;
 
-    il = (il/2)%4;
+    il = (il/2) & 3;
-    float   coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
 
     for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i] & m) ? 0 : 4.f/coef));
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
     }
 #else
     float    kcoef = il&1 ? 1.f/16.f : 1.f;
@@ -1852,26 +1982,31 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
 #endif
 }
 
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
 template <typename type4x4>
 void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
-    device const uint8_t * q = xb->qs;
+    device const uchar * q = xb->qs;
 
 #if QK_K == 256
-    const float d = (float)(xb->d);
-    const float min = (float)(xb->dmin);
     short is = (il/4) * 2;
     q = q + (il/4) * 32 + 16 * (il&1);
-    il = il%4;
+    il = il & 3;
-    const uchar4 sc = get_scale_min_k4(is, xb->scales);
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float dl = il<2 ? d * sc[0]   : d * sc[2]/16.h;
+    const half d   = il < 2 ? xb->d : xb->d / 16.h;
-    const float ml = il<2 ? min * sc[1] : min * sc[3];
+    const half min = xb->dmin;
+    const half dl = d * sc[0];
+    const half ml = min * sc[1];
 #else
     q = q + 16 * (il&1);
     device const uint8_t * s = xb->scales;
     device const half2 * dh = (device const half2 *)xb->d;
     const float2 d = (float2)dh[0];
     const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
-    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1 ]* (s[1]>>4);
+    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
 #endif
     const ushort mask = il<2 ? 0x0F : 0xF0;
     for (int i = 0; i < 16; ++i) {
@@ -1885,19 +2020,19 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     device const uint8_t * qh = xb->qh;
 
 #if QK_K == 256
-    const float d = (float)(xb->d);
-    const float min = (float)(xb->dmin);
     short is = (il/4) * 2;
     q  = q + 32 * (il/4) + 16 * (il&1);
     qh = qh + 16 * (il&1);
     uint8_t ul = 1 << (il/2);
-    il = il%4;
+    il = il & 3;
-    const uchar4 sc = get_scale_min_k4(is, xb->scales);
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float dl = il<2 ? d * sc[0]   : d * sc[2]/16.h;
+    const half d = il < 2 ? xb->d : xb->d / 16.h;
-    const float ml = il<2 ? min * sc[1] : min * sc[3];
+    const half min = xb->dmin;
+    const half dl = d * sc[0];
+    const half ml = min * sc[1];
 
-    const ushort mask   = il<2 ? 0x0F : 0xF0;
+    const ushort mask = il<2 ? 0x0F : 0xF0;
-    const float  qh_val = il<2 ? 16.f : 256.f;
+    const half qh_val = il<2 ? 16.h : 256.h;
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
     }
@@ -1916,7 +2051,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
 
 template <typename type4x4>
 void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
-    const float d_all = (float)(xb->d);
+    const half d_all = xb->d;
     device const uint8_t * ql = (device const uint8_t *)xb->ql;
     device const uint8_t * qh = (device const uint8_t *)xb->qh;
     device const int8_t * scales = (device const int8_t *)xb->scales;
@@ -1924,19 +2059,21 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
 #if QK_K == 256
     ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
     qh = qh + 32*(il/8) + 16*(il&1);
-    float sc = scales[(il%2) + 2 * ((il/2))];
+    half sc = scales[(il%2) + 2 * ((il/2))];
-    il = (il/2)%4;
+    il = (il/2) & 3;
 #else
     ql = ql + 16 * (il&1);
-    float sc = scales[il];
+    half sc = scales[il];
 #endif
+    const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
+    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
+    const half        coef = il>1 ? 1.f/16.h          : 1.h;
+    const half ml = d_all * sc * 32.h;
+    const half dl = d_all * sc * coef;
     for (int i = 0; i < 16; ++i) {
-        uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
+        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
-        uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
+                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
-        const float coef = il>1 ? 1.f/16.f          : 1.f;
+        reg[i/4][i%4] = dl * q - ml;
-        float q = il&1 ? ((ql[i]&kmask2)|((qh[i]&kmask1)<<2)) - 32.f/coef : \
-                         ((ql[i]&kmask2)|((qh[i]&kmask1)<<4)) - 32.f/coef;
-        reg[i/4][i%4] = d_all * sc * q * coef;
     }
 }
 
@@ -1976,22 +2113,25 @@ kernel void kernel_get_rows(
 // each block_q contains 16*nl weights
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 kernel void kernel_mul_mm(device const  uchar * src0,
-                           device const  float * src1,
+                          device const  uchar * src1,
-                           device        float * dst,
+                          device        float * dst,
-                           constant    int64_t & ne00,
+                          constant    int64_t & ne00,
-                           constant    int64_t & ne02,
+                          constant    int64_t & ne02,
-                           constant    int64_t & nb01,
+                          constant    int64_t & nb01,
-                           constant    int64_t & nb02,
+                          constant    int64_t & nb02,
-                           constant    int64_t & ne12,
+                          constant    int64_t & ne12,
-                           constant    int64_t & ne0,
+                          constant    int64_t & nb10,
-                           constant    int64_t & ne1,
+                          constant    int64_t & nb11,
-                           constant    uint & gqa,
+                          constant    int64_t & nb12,
-                           threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          constant    int64_t & ne0,
-                           uint3                 tgpig[[threadgroup_position_in_grid]],
+                          constant    int64_t & ne1,
-                           uint                  tiitg[[thread_index_in_threadgroup]],
+                          constant       uint & gqa,
-                           uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          uint3                 tgpig[[threadgroup_position_in_grid]],
+                          uint                  tiitg[[thread_index_in_threadgroup]],
+                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    threadgroup half * sa = ((threadgroup half *)shared_memory);
+    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
     threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
 
     const uint r0 = tgpig.y;
@@ -2004,7 +2144,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
     short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
-    simdgroup_half8x8 ma[4];
+    simdgroup_half8x8  ma[4];
     simdgroup_float8x8 mb[2];
     simdgroup_float8x8 c_res[8];
     for (int i = 0; i < 8; i++){
@@ -2012,10 +2152,15 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 
     short il = (tiitg % THREAD_PER_ROW);
-    uint offset0 = im/gqa*nb02; ushort offset1 = il/nl;
+
-    device const block_q  * x = (device const block_q  *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    uint   offset0 = im/gqa*nb02;
-    device const float * y = src1 + (r1 * BLOCK_SIZE_N + thread_col) * ne00 \
+    ushort offset1 = il/nl;
-                             + BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL) + im * ne00 * ne1;
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * im
+        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
         //load data and store to threadgroup memory
@@ -2095,6 +2240,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \
                           constant uint64_t &, constant uint64_t &, uint, uint, uint);
 
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
 template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
@@ -2105,14 +2251,27 @@ template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
 
-typedef void (mat_mm_t)(device const uchar *, device const float *, device float *, constant int64_t &,\
+typedef void (mat_mm_t)(
-                             constant int64_t &, constant int64_t &, constant int64_t &, constant int64_t &, \
+        device const  uchar * src0,
-                             constant int64_t &, constant int64_t &, constant uint &, threadgroup uchar *, uint3, uint, uint);
+        device const  uchar * src1,
+        device        float * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne02,
+        constant    int64_t & nb01,
+        constant    int64_t & nb02,
+        constant    int64_t & ne12,
+        constant    int64_t & nb10,
+        constant    int64_t & nb11,
+        constant    int64_t & nb12,
+        constant    int64_t & ne0,
+        constant    int64_t & ne1,
+        constant       uint & gqa,
+        threadgroup uchar *, uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1, dequantize_f16>;
+template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;

ggml.c

@@ -4303,10 +4303,21 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
 }
 
 size_t ggml_nbytes(const struct ggml_tensor * tensor) {
-    size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type);
+    size_t nbytes;
-    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+    size_t blck_size = ggml_blck_size(tensor->type);
-        nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+    if (blck_size == 1) {
+        nbytes = ggml_type_size(tensor->type);
+        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+            nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+        }
     }
+    else {
+        nbytes = tensor->ne[0]*tensor->nb[0]/blck_size;
+        for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+            nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+        }
+    }
+
     return nbytes;
 }
 
@@ -18345,10 +18356,11 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * node = cgraph->leafs[i];
 
-        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n",
+        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n",
                 i,
                 node->ne[0], node->ne[1],
-                ggml_op_name(node->op));
+                ggml_op_name(node->op),
+                ggml_get_name(node));
     }
 
     for (int i = 0; i < GGML_OP_COUNT; i++) {