Revert "llama on Metal (#885)"

This reverts commit c55f81b860.

.gitmodules

@@ -6,4 +6,4 @@
 	url = https://github.com/manyoso/llama.cpp.git
 [submodule "llama.cpp-mainline"]
 	path = gpt4all-backend/llama.cpp-mainline
-	url = https://github.com/nomic-ai/llama.cpp.git
+	url = https://github.com/ggerganov/llama.cpp.git

gpt4all-backend/CMakeLists.txt

@@ -39,9 +39,6 @@ endif()
 include(llama.cpp.cmake)
 
 set(BUILD_VARIANTS default avxonly)
-if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
-    set(BUILD_VARIANTS ${BUILD_VARIANTS} metal)
-endif()
 
 set(CMAKE_VERBOSE_MAKEFILE ON)
 
@@ -57,20 +54,10 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
     set(LLAMA_F16C ${GPT4ALL_ALLOW_NON_AVX})
     set(LLAMA_FMA  ${GPT4ALL_ALLOW_NON_AVX})
 
-    if (BUILD_VARIANT STREQUAL metal)
-        set(LLAMA_K_QUANTS YES)
-        set(LLAMA_METAL YES)
-    else()
-        set(LLAMA_K_QUANTS NO)
-        set(LLAMA_METAL NO)
-    endif()
-
     # Include GGML
     include_ggml(llama.cpp-mainline -mainline-${BUILD_VARIANT} ON)
-    if (NOT LLAMA_METAL)
+    include_ggml(llama.cpp-230511 -230511-${BUILD_VARIANT} ON)
-        include_ggml(llama.cpp-230511 -230511-${BUILD_VARIANT} ON)
+    include_ggml(llama.cpp-230519 -230519-${BUILD_VARIANT} ON)
-        include_ggml(llama.cpp-230519 -230519-${BUILD_VARIANT} ON)
-    endif()
 
     # Function for preparing individual implementations
     function(prepare_target TARGET_NAME BASE_LIB)
@@ -95,30 +82,29 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
         LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
     prepare_target(llamamodel-mainline llama-mainline)
 
-    if (NOT LLAMA_METAL)
+    add_library(llamamodel-230519-${BUILD_VARIANT} SHARED
-        add_library(llamamodel-230519-${BUILD_VARIANT} SHARED
+        llamamodel.cpp llmodel_shared.cpp)
-            llamamodel.cpp llmodel_shared.cpp)
+    target_compile_definitions(llamamodel-230519-${BUILD_VARIANT} PRIVATE
-        target_compile_definitions(llamamodel-230519-${BUILD_VARIANT} PRIVATE
+        LLAMA_VERSIONS===2 LLAMA_DATE=230519)
-            LLAMA_VERSIONS===2 LLAMA_DATE=230519)
+    prepare_target(llamamodel-230519 llama-230519)
-        prepare_target(llamamodel-230519 llama-230519)
-        add_library(llamamodel-230511-${BUILD_VARIANT} SHARED
-            llamamodel.cpp llmodel_shared.cpp)
-        target_compile_definitions(llamamodel-230511-${BUILD_VARIANT} PRIVATE
-            LLAMA_VERSIONS=<=1 LLAMA_DATE=230511)
-        prepare_target(llamamodel-230511 llama-230511)
 
-        add_library(gptj-${BUILD_VARIANT} SHARED
+    add_library(llamamodel-230511-${BUILD_VARIANT} SHARED
-            gptj.cpp utils.h utils.cpp llmodel_shared.cpp)
+        llamamodel.cpp llmodel_shared.cpp)
-        prepare_target(gptj ggml-230511)
+    target_compile_definitions(llamamodel-230511-${BUILD_VARIANT} PRIVATE
+        LLAMA_VERSIONS=<=1 LLAMA_DATE=230511)
+    prepare_target(llamamodel-230511 llama-230511)
 
-        add_library(mpt-${BUILD_VARIANT} SHARED
+    add_library(gptj-${BUILD_VARIANT} SHARED
-            mpt.cpp utils.h utils.cpp llmodel_shared.cpp)
+        gptj.cpp utils.h utils.cpp llmodel_shared.cpp)
-        prepare_target(mpt ggml-230511)
+    prepare_target(gptj ggml-230511)
 
-        add_library(replit-${BUILD_VARIANT} SHARED
+    add_library(mpt-${BUILD_VARIANT} SHARED
-            replit.cpp utils.h utils.cpp llmodel_shared.cpp)
+        mpt.cpp utils.h utils.cpp llmodel_shared.cpp)
-        prepare_target(replit ggml-230511)
+    prepare_target(mpt ggml-230511)
-    endif()
+
+    add_library(replit-${BUILD_VARIANT} SHARED
+        replit.cpp utils.h utils.cpp llmodel_shared.cpp)
+    prepare_target(replit ggml-230511)
 endforeach()
 
 add_library(llmodel

gpt4all-backend/llama.cpp-mainline

@@ -1 +1 @@
-Subproject commit b33dee282f5d8032b5f780152732dc45cbf2d349
+Subproject commit 5b57a5b72676540b6a45a3f527126299969ad241

gpt4all-backend/llama.cpp.cmake

@@ -34,7 +34,6 @@ endif()
 #
 # Option list
 #
-# some of the options here are commented out so they can be set "dynamically" before calling include_ggml()
 
 # general
 option(LLAMA_STATIC                 "llama: static link libraries"                          OFF)
@@ -69,7 +68,6 @@ option(LLAMA_OPENBLAS               "llama: use OpenBLAS"
 #option(LLAMA_CUBLAS                 "llama: use cuBLAS"                                     OFF)
 #option(LLAMA_CLBLAST                "llama: use CLBlast"                                    OFF)
 #option(LLAMA_METAL                  "llama: use Metal"                                      OFF)
-#option(LLAMA_K_QUANTS               "llama: use k-quants"                                   ON)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_DMMV_Y "1" CACHE STRING  "llama: y block size for dmmv CUDA kernels")
@@ -265,32 +263,10 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
     endif()
 
     set(GGML_SOURCES_QUANT_K )
-    set(GGML_METAL_SOURCES )
+    if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${DIRECTORY}/ggml-quants-k.h)
-    if (LLAMA_K_QUANTS)
         set(GGML_SOURCES_QUANT_K
-            ${DIRECTORY}/k_quants.h
+            ${DIRECTORY}/ggml-quants-k.h
-            ${DIRECTORY}/k_quants.c)
+            ${DIRECTORY}/ggml-quants-k.c)
-
-        if (LLAMA_METAL)
-            find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-            find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-            find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-            find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
-
-            set(GGML_METAL_SOURCES ${DIRECTORY}/ggml-metal.m ${DIRECTORY}/ggml-metal.h)
-            # get full path to the file
-            #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
-            # copy ggml-metal.metal to bin directory
-            configure_file(${DIRECTORY}/ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
-                ${FOUNDATION_LIBRARY}
-                ${METAL_FRAMEWORK}
-                ${METALKIT_FRAMEWORK}
-                ${METALPERFORMANCE_FRAMEWORK}
-            )
-        endif()
     endif()
 
     add_library(ggml${SUFFIX} OBJECT
@@ -298,16 +274,8 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
                 ${DIRECTORY}/ggml.h
                 ${GGML_SOURCES_QUANT_K}
                 ${GGML_SOURCES_CUDA}
-                ${GGML_METAL_SOURCES}
                 ${GGML_OPENCL_SOURCES})
 
-    if (LLAMA_K_QUANTS)
-        target_compile_definitions(ggml${SUFFIX} PUBLIC GGML_USE_K_QUANTS)
-    endif()
-
-    if (LLAMA_METAL AND GGML_METAL_SOURCES)
-        target_compile_definitions(ggml${SUFFIX} PUBLIC GGML_USE_METAL GGML_METAL_NDEBUG)
-    endif()
     target_include_directories(ggml${SUFFIX} PUBLIC ${DIRECTORY})
     target_compile_features(ggml${SUFFIX} PUBLIC c_std_11) # don't bump
 
@@ -327,9 +295,6 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
                     ${DIRECTORY}/llama.h
                     ${DIRECTORY}/${LLAMA_UTIL_SOURCE_FILE})
 
-        if (LLAMA_METAL AND GGML_METAL_SOURCES)
-            target_compile_definitions(llama${SUFFIX} PUBLIC GGML_USE_METAL GGML_METAL_NDEBUG)
-        endif()
         target_include_directories(llama${SUFFIX} PUBLIC ${DIRECTORY})
         target_compile_features(llama${SUFFIX} PUBLIC cxx_std_11) # don't bump
 
@@ -367,6 +332,32 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
         target_compile_definitions(ggml${SUFFIX} PRIVATE GGML_USE_CLBLAST)
     endif()
 
+    if (LLAMA_METAL)
+        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+        find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
+
+        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+        target_compile_definitions(llama${SUFFIX} PRIVATE
+            GGML_USE_METAL
+            GGML_METAL_NDEBUG)
+
+        # get full path to the file
+        #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+        # copy ggml-metal.metal to bin directory
+        configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
+
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+            ${FOUNDATION_LIBRARY}
+            ${METAL_FRAMEWORK}
+            ${METALKIT_FRAMEWORK}
+            ${METALPERFORMANCE_FRAMEWORK}
+            )
+    endif()
+
     if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
         message(STATUS "ARM detected")
         if (MSVC)

gpt4all-backend/llamamodel.cpp

@@ -115,12 +115,6 @@ bool LLamaModel::loadModel(const std::string &modelPath)
 #if LLAMA_DATE <= 230511
     d_ptr->params.n_parts  = params.n_parts;
 #endif
-#ifdef GGML_USE_METAL
-    std::cerr << "llama.cpp: using Metal" << std::endl;
-    // metal always runs the whole model if n_gpu_layers is not 0, at least
-    // currently
-    d_ptr->params.n_gpu_layers = 1;
-#endif
 
     d_ptr->ctx = llama_init_from_file(modelPath.c_str(), d_ptr->params);
     if (!d_ptr->ctx) {
@@ -234,30 +228,7 @@ DLL_EXPORT bool magic_match(std::istream& f) {
     // Check version
     uint32_t version = 0;
     f.read(reinterpret_cast<char*>(&version), sizeof(version));
-    if (!(version LLAMA_VERSIONS)) {
+    return version LLAMA_VERSIONS;
-        return false;
-    }
-#ifdef GGML_USE_METAL
-    // Check quant supported on metal
-    // skip fields
-    off_t offset = sizeof(uint32_t) * 6; // n_vocab, n_embd, n_mult, n_head, n_layer, n_rot
-    f.seekg(offset, std::ios_base::cur);
-    uint32_t ftype;
-    f.read(reinterpret_cast<char*>(&ftype), sizeof(ftype)); // ftype
-    switch((enum llama_ftype) ftype) {
-        // currently supported on Metal https://github.com/ggerganov/llama.cpp/blob/ae9663f1887513e152839e91f61c513075a19422/ggml-metal.m#L51-L55
-        case LLAMA_FTYPE_MOSTLY_F16:
-        case LLAMA_FTYPE_MOSTLY_Q2_K:
-        case LLAMA_FTYPE_MOSTLY_Q4_0:
-        case LLAMA_FTYPE_MOSTLY_Q6_K:
-        case LLAMA_FTYPE_MOSTLY_Q4_K_S:
-        case LLAMA_FTYPE_MOSTLY_Q4_K_M:
-            return true;
-        default: // unsupported quant-type for Metal
-            return false;
-    }
-#endif
-    return true;
 }
 
 DLL_EXPORT LLModel *construct() {

gpt4all-backend/llmodel.cpp

@@ -121,30 +121,20 @@ LLModel *LLModel::construct(const std::string &modelPath, std::string buildVaria
     if (!has_at_least_minimal_hardware())
         return nullptr;
 
+    //TODO: Auto-detect CUDA/OpenCL
+    if (buildVariant == "auto") {
+        if (requires_avxonly()) {
+            buildVariant = "avxonly";
+        } else {
+            buildVariant = "default";
+        }
+    }
     // Read magic
     std::ifstream f(modelPath, std::ios::binary);
     if (!f) return nullptr;
     // Get correct implementation
-    const LLModel::Implementation* impl = nullptr;
+    auto impl = implementation(f, buildVariant);
-
+    if (!impl) return nullptr;
-    #if defined(__APPLE__) && defined(__arm64__) // FIXME: See if metal works for intel macs
-        if (buildVariant == "auto") {
-            impl = implementation(f, "metal");
-        }
-    #endif
-
-    if (!impl) {
-        //TODO: Auto-detect CUDA/OpenCL
-        if (buildVariant == "auto") {
-            if (requires_avxonly()) {
-                buildVariant = "avxonly";
-            } else {
-                buildVariant = "default";
-            }
-        }
-        impl = implementation(f, buildVariant);
-        if (!impl) return nullptr;
-    }
     f.close();
     // Construct and return llmodel implementation
     return impl->construct();

gpt4all-chat/CMakeLists.txt

@@ -58,11 +58,6 @@ set (CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
 add_subdirectory(../gpt4all-backend llmodel)
 
-set(METAL_SHADER_FILE)
-if(${CMAKE_SYSTEM_NAME} MATCHES Darwin)
-  set(METAL_SHADER_FILE ../gpt4all-backend/llama.cpp-mainline/ggml-metal.metal)
-endif()
-
 qt_add_executable(chat
     main.cpp
     chat.h chat.cpp
@@ -77,7 +72,6 @@ qt_add_executable(chat
     server.h server.cpp
     logger.h logger.cpp
     sysinfo.h
-    ${METAL_SHADER_FILE}
 )
 
 qt_add_qml_module(chat
@@ -138,12 +132,6 @@ if(${CMAKE_SYSTEM_NAME} MATCHES Darwin)
     )
 endif()
 
-if(METAL_SHADER_FILE)
-    set_target_properties(chat PROPERTIES
-        RESOURCE ${METAL_SHADER_FILE}
-    )
-endif()
-
 target_compile_definitions(chat
     PRIVATE $<$<OR:$<CONFIG:Debug>,$<CONFIG:RelWithDebInfo>>:QT_QML_DEBUG>)
 target_link_libraries(chat