ggml : try another fix

.devops/nix/package.nix

@@ -255,11 +255,11 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL || useVulkan) lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin) || (useVulkan && effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.github/workflows/build.yml

@@ -37,8 +37,6 @@ jobs:
 
       - name: Build
         id: make_build
-        env:
-            LLAMA_FATAL_WARNINGS: 1
         run: |
           CC=gcc-8 make -j $(nproc)
 
@@ -67,7 +65,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON
+          cmake ..
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -102,7 +100,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
@@ -246,8 +244,6 @@ jobs:
 
       - name: Build
         id: make_build
-        env:
-            LLAMA_FATAL_WARNINGS: 1
         run: |
           LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
@@ -281,7 +277,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
+          cmake -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test

.gitignore

@@ -23,13 +23,11 @@
 .clang-tidy
 .vs/
 .vscode/
-.idea/
 
 lcov-report/
 gcovr-report/
 
 build*
-cmake-build-*
 out/
 tmp/
 

CMakeLists.txt

@@ -55,9 +55,6 @@ option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"
 option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
 option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
-# build
-option(LLAMA_FATAL_WARNINGS             "llama: enable -Werror flag"                            OFF)
-
 # sanitizers
 option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
 option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
@@ -110,20 +107,22 @@ option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
-option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
-option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
+
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
+if (LLAMA_PERF)
+    add_definitions(-DGGML_PERF)
+endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -131,7 +130,6 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
-
 if (LLAMA_SYCL)
     set(CMAKE_CXX_STANDARD 17)
 else()
@@ -142,7 +140,6 @@ set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
-
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
@@ -154,17 +151,17 @@ endif()
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
+        link_libraries(-fsanitize=thread)
     endif()
 
     if (LLAMA_SANITIZE_ADDRESS)
         add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
+        link_libraries(-fsanitize=address)
     endif()
 
     if (LLAMA_SANITIZE_UNDEFINED)
         add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
+        link_libraries(-fsanitize=undefined)
     endif()
 endif()
 
@@ -202,29 +199,6 @@ if (LLAMA_METAL)
     # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
-    if (LLAMA_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-        set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
-
-        add_custom_command(
-            OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-            DEPENDS ${METALLIB_SOURCE}
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-    endif()
-
     if (LLAMA_METAL_SHADER_DEBUG)
         # custom command to do the following:
         #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
@@ -324,17 +298,14 @@ if (LLAMA_BLAS)
         endif()
 
         message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
-
         add_compile_options(${BLAS_LINKER_FLAGS})
-
         add_compile_definitions(GGML_USE_OPENBLAS)
-
         if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
             add_compile_definitions(GGML_BLAS_USE_MKL)
         endif()
-
-        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
+
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -359,6 +330,9 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+#        if (LLAMA_CUDA_CUBLAS)
+#            add_compile_definitions(GGML_CUDA_CUBLAS)
+#        endif()
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -413,20 +387,15 @@ if (LLAMA_MPI)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
-
         set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c)
-
+        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-
         if (NOT MSVC)
             add_compile_options(-Wno-cast-qual)
         endif()
-
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
-
         # Even if you're only using the C header, C++ programs may bring in MPI
         # C++ functions, so more linkage is needed
         if (MPI_CXX_FOUND)
@@ -458,28 +427,31 @@ if (LLAMA_VULKAN)
     if (Vulkan_FOUND)
         message(STATUS "Vulkan found")
 
-        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
-        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
+        add_library(ggml-vulkan OBJECT ggml-vulkan.cpp ggml-vulkan.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
+        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
         if (LLAMA_VULKAN_CHECK_RESULTS)
-            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
         endif()
 
         if (LLAMA_VULKAN_DEBUG)
-            add_compile_definitions(GGML_VULKAN_DEBUG)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
         endif()
 
         if (LLAMA_VULKAN_VALIDATE)
-            add_compile_definitions(GGML_VULKAN_VALIDATE)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
         endif()
 
         if (LLAMA_VULKAN_RUN_TESTS)
-            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
         endif()
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -491,45 +463,43 @@ if (LLAMA_HIPBLAS)
     if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
     endif()
-
     if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
     endif()
 
-    find_package(hip     REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
 
-    message(STATUS "HIP and hipBLAS found")
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        if (LLAMA_HIP_UMA)
+            add_compile_definitions(GGML_HIP_UMA)
+        endif()
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
+        endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
+        endif()
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 
-    set(GGML_HEADERS_ROCM ggml-cuda.h)
-    set(GGML_SOURCES_ROCM ggml-cuda.cu)
-
-    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
-
-    if (LLAMA_HIP_UMA)
-        add_compile_definitions(GGML_HIP_UMA)
+        if (LLAMA_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
     endif()
-
-    if (LLAMA_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
-    if (LLAMA_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-    endif()
-
-    add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-
-    set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
-
-    if (LLAMA_STATIC)
-        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-    endif()
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
@@ -539,14 +509,10 @@ if (LLAMA_SYCL)
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
-
-    message(STATUS "SYCL found")
-
-    add_compile_definitions(GGML_USE_SYCL)
-
     if (LLAMA_SYCL_F16)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
+    add_compile_definitions(GGML_USE_SYCL)
 
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
@@ -555,7 +521,7 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
 
-    set(GGML_HEADERS_SYCL ggml-sycl.h)
+    set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
@@ -574,61 +540,61 @@ if (LLAMA_KOMPUTE)
     endif()
 
     function(compile_shader)
-        set(options)
-        set(oneValueArgs)
-        set(multiValueArgs SOURCES)
-        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        foreach(source ${compile_shader_SOURCES})
-            get_filename_component(filename ${source} NAME)
-            set(spv_file ${filename}.spv)
-            add_custom_command(
-                OUTPUT ${spv_file}
-                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                COMMENT "Compiling ${source} to ${spv_file}"
-                )
+      set(options)
+      set(oneValueArgs)
+      set(multiValueArgs SOURCES)
+      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+      foreach(source ${compile_shader_SOURCES})
+        get_filename_component(filename ${source} NAME)
+        set(spv_file ${filename}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+              COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${spv_file}"
+        )
 
-            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-            set(FILE_NAME "shader${RAW_FILE_NAME}")
-            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-            if(CMAKE_GENERATOR MATCHES "Visual Studio")
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-                    )
-            else()
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                    )
-            endif()
-        endforeach()
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+            )
+        else()
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+            )
+        endif()
+      endforeach()
     endfunction()
 
     if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
@@ -638,66 +604,66 @@ if (LLAMA_KOMPUTE)
 
         # Compile our shaders
         compile_shader(SOURCES
-            kompute-shaders/op_scale.comp
-            kompute-shaders/op_scale_8.comp
-            kompute-shaders/op_add.comp
-            kompute-shaders/op_addrow.comp
-            kompute-shaders/op_mul.comp
-            kompute-shaders/op_silu.comp
-            kompute-shaders/op_relu.comp
-            kompute-shaders/op_gelu.comp
-            kompute-shaders/op_softmax.comp
-            kompute-shaders/op_norm.comp
-            kompute-shaders/op_rmsnorm.comp
-            kompute-shaders/op_diagmask.comp
-            kompute-shaders/op_mul_mat_mat_f32.comp
-            kompute-shaders/op_mul_mat_f16.comp
-            kompute-shaders/op_mul_mat_q8_0.comp
-            kompute-shaders/op_mul_mat_q4_0.comp
-            kompute-shaders/op_mul_mat_q4_1.comp
-            kompute-shaders/op_mul_mat_q6_k.comp
-            kompute-shaders/op_getrows_f16.comp
-            kompute-shaders/op_getrows_q4_0.comp
-            kompute-shaders/op_getrows_q4_1.comp
-            kompute-shaders/op_getrows_q6_k.comp
-            kompute-shaders/op_rope_f16.comp
-            kompute-shaders/op_rope_f32.comp
-            kompute-shaders/op_cpy_f16_f16.comp
-            kompute-shaders/op_cpy_f16_f32.comp
-            kompute-shaders/op_cpy_f32_f16.comp
-            kompute-shaders/op_cpy_f32_f32.comp
+          kompute-shaders/op_scale.comp
+          kompute-shaders/op_scale_8.comp
+          kompute-shaders/op_add.comp
+          kompute-shaders/op_addrow.comp
+          kompute-shaders/op_mul.comp
+          kompute-shaders/op_silu.comp
+          kompute-shaders/op_relu.comp
+          kompute-shaders/op_gelu.comp
+          kompute-shaders/op_softmax.comp
+          kompute-shaders/op_norm.comp
+          kompute-shaders/op_rmsnorm.comp
+          kompute-shaders/op_diagmask.comp
+          kompute-shaders/op_mul_mat_mat_f32.comp
+          kompute-shaders/op_mul_mat_f16.comp
+          kompute-shaders/op_mul_mat_q8_0.comp
+          kompute-shaders/op_mul_mat_q4_0.comp
+          kompute-shaders/op_mul_mat_q4_1.comp
+          kompute-shaders/op_mul_mat_q6_k.comp
+          kompute-shaders/op_getrows_f16.comp
+          kompute-shaders/op_getrows_q4_0.comp
+          kompute-shaders/op_getrows_q4_1.comp
+          kompute-shaders/op_getrows_q6_k.comp
+          kompute-shaders/op_rope_f16.comp
+          kompute-shaders/op_rope_f32.comp
+          kompute-shaders/op_cpy_f16_f16.comp
+          kompute-shaders/op_cpy_f16_f32.comp
+          kompute-shaders/op_cpy_f32_f16.comp
+          kompute-shaders/op_cpy_f32_f32.comp
         )
 
         # Create a custom target for our generated shaders
         add_custom_target(generated_shaders DEPENDS
-            shaderop_scale.h
-            shaderop_scale_8.h
-            shaderop_add.h
-            shaderop_addrow.h
-            shaderop_mul.h
-            shaderop_silu.h
-            shaderop_relu.h
-            shaderop_gelu.h
-            shaderop_softmax.h
-            shaderop_norm.h
-            shaderop_rmsnorm.h
-            shaderop_diagmask.h
-            shaderop_mul_mat_mat_f32.h
-            shaderop_mul_mat_f16.h
-            shaderop_mul_mat_q8_0.h
-            shaderop_mul_mat_q4_0.h
-            shaderop_mul_mat_q4_1.h
-            shaderop_mul_mat_q6_k.h
-            shaderop_getrows_f16.h
-            shaderop_getrows_q4_0.h
-            shaderop_getrows_q4_1.h
-            shaderop_getrows_q6_k.h
-            shaderop_rope_f16.h
-            shaderop_rope_f32.h
-            shaderop_cpy_f16_f16.h
-            shaderop_cpy_f16_f32.h
-            shaderop_cpy_f32_f16.h
-            shaderop_cpy_f32_f32.h
+          shaderop_scale.h
+          shaderop_scale_8.h
+          shaderop_add.h
+          shaderop_addrow.h
+          shaderop_mul.h
+          shaderop_silu.h
+          shaderop_relu.h
+          shaderop_gelu.h
+          shaderop_softmax.h
+          shaderop_norm.h
+          shaderop_rmsnorm.h
+          shaderop_diagmask.h
+          shaderop_mul_mat_mat_f32.h
+          shaderop_mul_mat_f16.h
+          shaderop_mul_mat_q8_0.h
+          shaderop_mul_mat_q4_0.h
+          shaderop_mul_mat_q4_1.h
+          shaderop_mul_mat_q6_k.h
+          shaderop_getrows_f16.h
+          shaderop_getrows_q4_0.h
+          shaderop_getrows_q4_1.h
+          shaderop_getrows_q6_k.h
+          shaderop_rope_f16.h
+          shaderop_rope_f32.h
+          shaderop_cpy_f16_f16.h
+          shaderop_cpy_f16_f32.h
+          shaderop_cpy_f32_f16.h
+          shaderop_cpy_f32_f32.h
         )
 
         # Create a custom command that depends on the generated_shaders
@@ -710,10 +676,8 @@ if (LLAMA_KOMPUTE)
 
         # Add the stamp to the main sources to ensure dependency tracking
         set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ggml-kompute.h   ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-
+        set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
         add_compile_definitions(GGML_USE_KOMPUTE)
-
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
     else()
@@ -721,18 +685,6 @@ if (LLAMA_KOMPUTE)
     endif()
 endif()
 
-if (LLAMA_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
-
-    add_compile_definitions(GGML_USE_CPU_HBM)
-
-    target_link_libraries(ggml PUBLIC memkind)
-endif()
-
-if (LLAMA_PERF)
-    add_compile_definitions(GGML_PERF)
-endif()
-
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -757,30 +709,28 @@ function(get_flags CCID CCVER)
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
+    elseif (CCID MATCHES "Intel")
+        if (NOT LLAMA_SYCL)
+            # enable max optimization level when using Intel compiler
+            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            add_link_options(-fuse-ld=lld -static-intel)
+        endif()
     endif()
 
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
 
-if (LLAMA_FATAL_WARNINGS)
-    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        list(APPEND C_FLAGS   -Werror)
-        list(APPEND CXX_FLAGS -Werror)
-    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
-        add_compile_options(/WX)
-    endif()
-endif()
-
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                                  -Werror=implicit-int -Werror=implicit-function-declaration)
-        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                          -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
 
-        list(APPEND C_FLAGS   ${WARNING_FLAGS})
-        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -796,10 +746,9 @@ endif()
 set(CUDA_CXX_FLAGS "")
 
 if (LLAMA_CUBLAS)
-    set(CUDA_FLAGS -use_fast_math)
-
-    if (LLAMA_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
+    if (NOT MSVC)
+        list(APPEND CUDA_FLAGS -Wno-pedantic)
     endif()
 
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
@@ -833,11 +782,7 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
-    endif()
-
-    if (NOT MSVC)
-        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
     endif()
 endif()
 
@@ -876,7 +821,6 @@ execute_process(
     ERROR_VARIABLE output
     OUTPUT_QUIET
 )
-
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
 endif()
@@ -886,10 +830,10 @@ endif()
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 if (MSVC)
-    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
-    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
 else ()
-    set(CMAKE_GENERATOR_PLATFORM_LWR "")
+  set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
 
 if (NOT MSVC)
@@ -1083,6 +1027,11 @@ endif()
 
 # ggml
 
+if (GGML_USE_CPU_HBM)
+    add_definitions(-DGGML_USE_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+endif()
+
 add_library(ggml OBJECT
             ggml.c
             ggml.h
@@ -1099,17 +1048,16 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
             ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
-            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
-target_compile_features   (ggml PUBLIC c_std_11) # don't bump
-
+target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+if (GGML_USE_CPU_HBM)
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
-
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
     add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
@@ -1125,8 +1073,7 @@ add_library(llama
             )
 
 target_include_directories(llama PUBLIC .)
-target_compile_features   (llama PUBLIC cxx_std_11) # don't bump
-
+target_compile_features(llama PUBLIC cxx_std_11) # don't bump
 target_link_libraries(llama PRIVATE
     ggml
     ${LLAMA_EXTRA_LIBS}
@@ -1177,7 +1124,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
         "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

Makefile

@@ -97,10 +97,9 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -I. -Icommon
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -173,7 +172,7 @@ ifdef LLAMA_DEBUG
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -216,11 +215,6 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
-ifeq ($(LLAMA_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
-	MK_CXXFLAGS += -Werror
-endif
-
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -385,9 +379,6 @@ ifdef LLAMA_CUBLAS
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	MK_NVCCFLAGS += -use_fast_math
-ifdef LLAMA_FATAL_WARNINGS
-	MK_NVCCFLAGS += -Werror all-warnings
-endif # LLAMA_FATAL_WARNINGS
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
@@ -446,9 +437,9 @@ ifdef LLAMA_CUDA_CCBIN
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -533,29 +524,11 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
-ifdef LLAMA_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJS        += ggml-metal-embed.o
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal
-	@echo "Embedding Metal library"
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -567,10 +540,9 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
 override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -578,7 +550,7 @@ override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
+CUDA_CXXFLAGS := $(GF_CXXFLAGS)
 endif
 
 #
@@ -719,7 +691,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
@@ -890,7 +862,3 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

README-sycl.md

@@ -272,7 +272,7 @@ Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact
 
 a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
 
-Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
+Recommend to install to default folder: **/opt/intel/oneapi**.
 
 Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
 

README.md

@@ -10,8 +10,13 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631
-- Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590
+- Remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD: https://github.com/ggerganov/llama.cpp/pull/5240
+- Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
+  - [SYCL backend](README-sycl.md) is ready (1/28/2024), support Linux/Windows in Intel GPUs (iGPU, Arc/Flex/Max series)
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
+- Collecting Apple Silicon performance stats:
+  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
+  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -56,7 +61,7 @@ variety of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
-- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
+- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
 - Vulkan, SYCL, and (partial) OpenCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@@ -102,7 +107,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
 - [x] [InternLM2](https://huggingface.co/models?search=internlm2)
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
-- [x] [Gemma](https://ai.google.dev/gemma)
 
 **Multimodal models:**
 
@@ -152,7 +156,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
 
 ---
 
@@ -765,7 +768,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 
 #### How to run
 
-1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
 2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```

build.zig

@@ -123,7 +123,6 @@ pub fn build(b: *std.build.Builder) !void {
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
 
     _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
     _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
@@ -132,7 +131,7 @@ pub fn build(b: *std.build.Builder) !void {
     _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
     _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -33,7 +33,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
+CMAKE_EXTRA=""
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
@@ -219,7 +219,7 @@ function gg_run_open_llama_3b_v2 {
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
     head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
@@ -401,7 +401,7 @@ function gg_run_open_llama_7b_v2 {
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
     path_models="../models-mnt/open-llama/7B-v2"

common/common.cpp

@@ -1704,7 +1704,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
-    fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
@@ -1742,7 +1741,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 
     fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
     fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
-    fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
@@ -1751,7 +1750,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
 
     fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
-    fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
@@ -1802,8 +1801,7 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
             if (cs_curr[j] < 0) { continue; }
             if (seqs.find(cs_curr[j]) == seqs.end()) {
                 if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                const size_t sz = seqs.size();
-                seqs[cs_curr[j]] = sz;
+                seqs[cs_curr[j]] = seqs.size();
             }
         }
         if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }

common/sampling.cpp

@@ -121,7 +121,7 @@ static void sampler_queue(
                    struct llama_context * ctx_main,
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
-                                 size_t   min_keep) {
+                                 size_t & min_keep) {
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
@@ -249,7 +249,7 @@ static llama_token llama_sampling_sample_impl(
             id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
             // temperature sampling
-            size_t min_keep = std::max(1, params.min_keep);
+            size_t min_keep = std::max(1, params.n_probs);
 
             sampler_queue(ctx_main, params, cur_p, min_keep);
 

common/sampling.h

@@ -22,7 +22,6 @@ enum class llama_sampler_type : char {
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
     int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
     int32_t     top_k                 = 40;       // <= 0 to use vocab size
     float       top_p                 = 0.95f;    // 1.0 = disabled
     float       min_p                 = 0.05f;    // 0.0 = disabled

examples/baby-llama/baby-llama.cpp

@@ -1533,17 +1533,16 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        struct ggml_cgraph * gf = NULL;
-        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
+        ggml_cgraph gf = {};
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
-        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
+        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
         // struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
-        ggml_build_forward_expand(gf, e);
-        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(&gf, e);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
@@ -1553,8 +1552,8 @@ int main(int argc, char ** argv) {
         opt_params_lbfgs.lbfgs.n_iter = 16;
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
-        ggml_build_forward_expand(gf, e);
-        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(&gf, e);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1601,14 +1600,13 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            struct ggml_cgraph * gf = NULL;
-            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
+            ggml_cgraph gf = {};
 
             int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
+            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
 
-            ggml_build_forward_expand(gf, logits);
-            ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+            ggml_build_forward_expand(&gf, logits);
+            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

examples/batched-bench/batched-bench.cpp

@@ -159,7 +159,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");

examples/batched/batched.cpp

@@ -92,7 +92,7 @@ int main(int argc, char ** argv) {
 
     const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -325,14 +325,14 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %u\n", __func__, params->n_mult);
-    printf("%s: n_head:  %u\n", __func__, params->n_head);
-    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
-    printf("%s: n_layer: %u\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
+    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
+    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
+    printf("%s: n_head:  %d\n", __func__, params->n_head);
+    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
+    printf("%s: n_layer: %d\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -350,25 +350,25 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
+    printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
 
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
+    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
 
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
 
     // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
 
-    printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
 
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");

examples/export-lora/export-lora.cpp

@@ -7,6 +7,8 @@
 #include <string>
 #include <thread>
 
+static const size_t tensor_alignment = 32;
+
 struct lora_info {
     std::string filename;
     float scale;

examples/json-schema-to-grammar.py

@@ -87,21 +87,7 @@ class SchemaConverter:
         elif schema_type == 'array' and 'items' in schema:
             # TODO `prefixItems` keyword
             item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
-            list_item_operator = f'("," space {item_rule_name})'
-            successive_items = ""
-            min_items = schema.get("minItems", 0)
-            if min_items > 0:
-               first_item = f"({item_rule_name})"
-               successive_items = list_item_operator * (min_items - 1)
-               min_items -= 1
-            else:
-               first_item = f"({item_rule_name})?"
-            max_items = schema.get("maxItems")
-            if max_items is not None and max_items > min_items:
-                successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-            else:
-                successive_items += list_item_operator + "*"
-            rule = f'"[" space {first_item} {successive_items} "]" space'
+            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
             return self._add_rule(rule_name, rule)
 
         else:

examples/llava/README.md

@@ -53,45 +53,20 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-pa
 5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b --skip-unknown
+python ./convert.py ../llava-v1.5-7b
 ```
 
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
 ## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
+
+1) Backup your pth/safetensor model files as llava-surgery modifies them
+2) Use `python llava-surgery-v2.py -C -m /path/to/hf-model` which also supports llava-1.5 variants pytorch as well as safetensor models:
 - you will find a llava.projector and a llava.clip file in your model directory
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-4) Create the visual gguf model:
-```console
-python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
+3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory (https://huggingface.co/cmp-nct/llava-1.6-gguf/blob/main/config_vit.json) and rename it to config.json.
+4) Create the visual gguf model: `python ./examples/llava/convert-image-encoder-to-gguf.py -m ../path/to/vit --llava-projector ../path/to/llava.projector --output-dir ../path/to/output --clip-model-is-vision`
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-5) Then convert the model to gguf format:
-```console
-python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-6) And finally we can run the llava-cli using the 1.6 model version:
-```console
-./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
-```
-
+5) Everything else as usual: convert.py the hf model, quantize as needed
 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
 **note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
 

examples/llava/clip.cpp

@@ -616,9 +616,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             KQ = ggml_soft_max_inplace(ctx0, KQ);
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
             KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size);
-            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            KQV = ggml_cont(ctx0, ggml_permute(ctx0, KQV, 0, 2, 1, 3));
 
-            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
+            cur = ggml_cpy(ctx0, KQV, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size));
         }
 
         // attention output

examples/llava/llava-surgery-v2.py

@@ -65,7 +65,9 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
         for name in clip_tensors:
             del checkpoint[name]
 
+        # Save the updated checkpoint
         checkpoint_path = checkpoint_path
+        save_model(checkpoint, checkpoint_path, file_type)
         return True
     return False
 
@@ -150,6 +152,16 @@ for name in first_mm_tensors:
 if len(projector) > 0:
     save_model(projector, f"{args.model}/llava.projector", 'pytorch')
 
+for name in mm_tensors:
+    del last_checkpoint[name]
+for name in first_mm_tensors:
+    del first_checkpoint[name]
+
+if len(mm_tensors) > 0:
+    save_model(last_checkpoint, projector_checkpoint_path, file_type)
+if len(first_mm_tensors) > 0:
+    save_model(first_checkpoint, newline_checkpoint_path, file_type)
+
 print("Done!")
 print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava-surgery.py

@@ -19,12 +19,19 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_project
 projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
+# remove these tensors from the checkpoint and save it again
+for name in mm_tensors:
+    del checkpoint[name]
+
 # BakLLaVA models contain CLIP tensors in it
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
 if len(clip_tensors) > 0:
     clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
     torch.save(clip, f"{args.model}/llava.clip")
 
+    # remove these tensors
+    for name in clip_tensors:
+        del checkpoint[name]
 
     # added tokens should be removed to be able to convert Mistral models
     if os.path.exists(f"{args.model}/added_tokens.json"):
@@ -32,6 +39,7 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
+torch.save(checkpoint, path)
 
 print("Done!")
 print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")

examples/llava/llava.cpp

@@ -311,7 +311,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
     if (!image_embd) {
         fprintf(stderr, "Unable to allocate memory for image embeddings\n");

examples/llava/llava.h

@@ -31,8 +31,6 @@ struct llava_image_embed {
 /** sanity check for clip <-> llava embed size match */
 LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
 
-LLAVA_API bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
-
 /** build an image embed from image file bytes */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
 /** build an image embed from a path to an image filename */

examples/perplexity/perplexity.cpp

@@ -309,7 +309,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
 }
 
 static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
-    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
@@ -447,7 +447,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         return perplexity_v2(ctx, params);
     }
 
-    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
@@ -1623,7 +1623,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     uint32_t n_ctx;
     in.read((char *)&n_ctx, sizeof(n_ctx));
     if (n_ctx > llama_n_ctx(ctx)) {
-        fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
+        fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
                 __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
     }
 

examples/quantize/quantize.cpp

@@ -23,7 +23,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
     { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
     { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
-    { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
@@ -32,7 +31,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
     { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
-    { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.25 bpw non-linear quantization", },
     { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
     { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
     { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
@@ -289,10 +287,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
+    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
         fprintf(stderr, "\n===============================================================================================\n");
-        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
         fprintf(stderr, "===============================================================================================\n\n\n");
         return 1;
     }

examples/server/README.md

@@ -39,8 +39,6 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 - `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
-- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
-- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
 
 ## Build
 
@@ -134,11 +132,9 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - 503 -> `{"status": "loading model"}` if the model is still being loaded.
-  - 500 -> `{"status": "error"}` if the model failed to load.
-  - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
+  - `{"status": "loading model"}` if the model is still being loaded.
+  - `{"status": "error"}` if the model failed to load.
+  - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
 
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -200,8 +196,6 @@ node index.js
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
-    `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum (default: 0)
-
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
@@ -385,69 +379,6 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
-
-### Result JSON
-
-```json
-[
-    {
-        "dynatemp_exponent": 1.0,
-        "dynatemp_range": 0.0,
-        "frequency_penalty": 0.0,
-        "grammar": "",
-        "id": 0,
-        "ignore_eos": false,
-        "logit_bias": [],
-        "min_p": 0.05000000074505806,
-        "mirostat": 0,
-        "mirostat_eta": 0.10000000149011612,
-        "mirostat_tau": 5.0,
-        "model": "llama-2-7b-32k-instruct.Q2_K.gguf",
-        "n_ctx": 2048,
-        "n_keep": 0,
-        "n_predict": 100000,
-        "n_probs": 0,
-        "next_token": {
-            "has_next_token": true,
-            "n_remain": -1,
-            "num_tokens_predicted": 0,
-            "stopped_eos": false,
-            "stopped_limit": false,
-            "stopped_word": false,
-            "stopping_word": ""
-        },
-        "penalize_nl": true,
-        "penalty_prompt_tokens": [],
-        "presence_penalty": 0.0,
-        "prompt": "Say hello to llama.cpp",
-        "repeat_last_n": 64,
-        "repeat_penalty": 1.100000023841858,
-        "samplers": [
-            "top_k",
-            "tfs_z",
-            "typical_p",
-            "top_p",
-            "min_p",
-            "temperature"
-        ],
-        "seed": 42,
-        "state": 1,
-        "stop": [
-            "\n"
-        ],
-        "stream": false,
-        "task_id": 0,
-        "temperature": 0.0,
-        "tfs_z": 1.0,
-        "top_k": 40,
-        "top_p": 0.949999988079071,
-        "typical_p": 1.0,
-        "use_penalty_prompt_tokens": false
-    }
-]
-```
-
 ## More examples
 
 ### Change system prompt on runtime

examples/server/oai.hpp

@@ -15,11 +15,13 @@
 using json = nlohmann::json;
 
 inline static json oaicompat_completion_params_parse(
-    const struct llama_model * model,
     const json &body, /* openai api json semantics */
     const std::string &chat_template)
 {
     json llama_params;
+    std::string formatted_prompt = chat_template == "chatml"
+        ? format_chatml(body["messages"])  // OpenAI 'messages' to chatml (with <|im_start|>,...)
+        : format_llama2(body["messages"]); // OpenAI 'messages' to llama2 (with [INST],...)
 
     llama_params["__oaicompat"] = true;
 
@@ -32,7 +34,7 @@ inline static json oaicompat_completion_params_parse(
     // https://platform.openai.com/docs/api-reference/chat/create
     llama_sampling_params default_sparams;
     llama_params["model"]             = json_value(body, "model", std::string("unknown"));
-    llama_params["prompt"]            = format_chat(model, chat_template, body["messages"]);
+    llama_params["prompt"]            = formatted_prompt;
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.0);
     llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);

examples/server/public/index.html

@@ -234,7 +234,6 @@
       mirostat_eta: 0.1, // learning rate
       grammar: '',
       n_probs: 0, // no completion_probabilities,
-      min_keep: 0, // min probs from each sampler,
       image_data: [],
       cache_prompt: true,
       api_key: ''
@@ -792,9 +791,6 @@
             <fieldset>
               ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
-            <fieldset>
-              ${IntField({ label: "Min Probabilities from each Sampler", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
-            </fieldset>
             <fieldset>
               <label for="api_key">API Key</label>
               <input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />

examples/server/server.cpp

@@ -5,7 +5,6 @@
 #include "oai.hpp"
 
 #include "../llava/clip.h"
-#include "../llava/llava.h"
 
 #include "stb_image.h"
 
@@ -29,7 +28,6 @@
 #include <chrono>
 #include <condition_variable>
 #include <atomic>
-#include <signal.h>
 
 using json = nlohmann::json;
 
@@ -38,11 +36,10 @@ struct server_params
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
-    std::string chat_template = "";
+    std::string chat_template = "chatml";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
-    bool slots_endpoint = true;
 };
 
 bool server_verbose = false;
@@ -161,7 +158,6 @@ struct llama_client_slot
     int32_t n_decoded   = 0;
     int32_t n_remaining = -1;
     int32_t i_batch     = -1;
-    int32_t n_predict   = -1;
 
     int32_t num_prompt_tokens           = 0;
     int32_t num_prompt_tokens_processed = 0;
@@ -413,7 +409,6 @@ struct llama_server_context
 
             slot.id = i;
             slot.n_ctx = n_ctx_slot;
-            slot.n_predict = params.n_predict;
 
             LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
 
@@ -549,16 +544,6 @@ struct llama_server_context
         slot->params.seed               = json_value(data, "seed",              default_params.seed);
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
-        slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
-
-        if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
-            // Might be better to reject the request with a 400 ?
-            LOG_WARNING("Max tokens to predict exceeds server configuration", {
-                {"params.n_predict", slot->params.n_predict},
-                {"slot.n_predict", slot->n_predict},
-            });
-            slot->params.n_predict = slot->n_predict;
-        }
 
         // infill
         if (data.count("input_prefix") != 0)
@@ -998,12 +983,43 @@ struct llama_server_context
             {
                 continue;
             }
-
-            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
+            clip_image_f32_batch img_res_v;
+            img_res_v.size = 0;
+            img_res_v.data = nullptr;
+            if (!clip_image_preprocess(clp_ctx, img.img_data, img_res_v))
+            {
+                LOG_TEE("Error processing the given image");
+                clip_free(clp_ctx);
+                clip_image_f32_batch_free(img_res_v);
+                return false;
+            }
+            if (img_res_v.size == 0)
+            {
                 LOG_TEE("Error processing the given image");
                 return false;
             }
 
+            // note: assumes only one image was returned by clip_image_preprocess
+            clip_image_f32 * img_res = img_res_v.data;
+
+            img.image_tokens = clip_n_patches(clp_ctx);
+            img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx));
+            if (!img.image_embedding)
+            {
+                LOG_TEE("Unable to allocate memory for image embeddings\n");
+                clip_image_f32_batch_free(img_res_v);
+                clip_free(clp_ctx);
+                return false;
+            }
+            LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id);
+            if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
+            {
+                LOG_TEE("Unable to encode image\n");
+                clip_image_f32_batch_free(img_res_v);
+                return false;
+            }
+
+            clip_image_f32_batch_free(img_res_v);
 
             img.request_encode_image = false;
         }
@@ -1036,7 +1052,6 @@ struct llama_server_context
 
         return json {
             {"n_ctx",             slot.n_ctx},
-            {"n_predict",         slot.n_predict},
             {"model",             params.model_alias},
             {"seed",              slot.params.seed},
             {"temperature",       slot.sparams.temp},
@@ -1064,7 +1079,6 @@ struct llama_server_context
             {"stream",            slot.params.stream},
             {"logit_bias",        slot.sparams.logit_bias},
             {"n_probs",           slot.sparams.n_probs},
-            {"min_keep",          slot.sparams.min_keep},
             {"grammar",           slot.sparams.grammar},
             {"samplers",          samplers_sequence}
         };
@@ -1899,17 +1913,14 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  --mmproj MMPROJ_FILE      path to a multimodal projector file for LLaVA.\n");
     printf("  --log-disable             disables logging to a file.\n");
-    printf("  --slots-endpoint-disable  disables slots monitoring endpoint.\n");
     printf("\n");
-    printf("  -n, --n-predict           maximum tokens to predict (default: %d)\n", params.n_predict);
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                            advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                            types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("  -gan N, --grp-attn-n N    set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`");
     printf("  -gaw N, --grp-attn-w N    set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
-    printf("  --chat-template JINJA_TEMPLATE\n");
-    printf("                            set custom jinja chat template (default: template taken from model's metadata)\n");
-    printf("                            Note: only commonly used templates are accepted, since we don't have jinja parser\n");
+    printf("  --chat-template FORMAT_NAME");
+    printf("                            set chat template, possible valus is: llama2, chatml (default %s)", sparams.chat_template.c_str());
     printf("\n");
 }
 
@@ -2349,10 +2360,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             log_set_target(stdout);
             LOG_INFO("logging to file is disabled.", {});
         }
-        else if (arg == "--slots-endpoint-disable")
-        {
-            sparams.slots_endpoint = false;
-        }
         else if (arg == "--chat-template")
         {
             if (++i >= argc)
@@ -2360,13 +2367,13 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            if (!verify_custom_template(argv[i])) {
-                fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
-                fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
+            std::string value(argv[i]);
+            if (value != "chatml" && value != "llama2") {
+                fprintf(stderr, "error: chat template can be \"llama2\" or \"chatml\", but got: %s\n", value.c_str());
                 invalid_param = true;
                 break;
             }
-            sparams.chat_template = argv[i];
+            sparams.chat_template = value;
         }
         else if (arg == "--override-kv")
         {
@@ -2504,9 +2511,6 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
     }
 }
 
-std::function<void(int)> shutdown_handler;
-inline void signal_handler(int signal) { shutdown_handler(signal); }
-
 int main(int argc, char **argv)
 {
 #if SERVER_VERBOSE != 1
@@ -2553,40 +2557,13 @@ int main(int argc, char **argv)
         res.set_header("Access-Control-Allow-Headers", "*");
     });
 
-    svr.Get("/health", [&](const httplib::Request& req, httplib::Response& res) {
+    svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) {
         server_state current_state = state.load();
         switch(current_state) {
-            case SERVER_STATE_READY: {
-                int available_slots  = 0;
-                int processing_slots = 0;
-                for (llama_client_slot &slot: llama.slots) {
-                    if (slot.available()) {
-                        available_slots++;
-                    } else {
-                        processing_slots++;
-                    }
-                }
-                if (available_slots > 0) {
-                    json health = {
-                            {"status",           "ok"},
-                            {"slots_idle",       available_slots},
-                            {"slots_processing", processing_slots}};
-                    res.set_content(health.dump(), "application/json");
-                    res.status = 200; // HTTP OK
-                } else {
-                    json health = {
-                            {"status",           "no slot available"},
-                            {"slots_idle",       available_slots},
-                            {"slots_processing", processing_slots}};
-                    res.set_content(health.dump(), "application/json");
-                    if (req.has_param("fail_on_no_slot")) {
-                        res.status = 503; // HTTP Service Unavailable
-                    } else {
-                        res.status = 200; // HTTP OK
-                    }
-                }
+            case SERVER_STATE_READY:
+                res.set_content(R"({"status": "ok"})", "application/json");
+                res.status = 200; // HTTP OK
                 break;
-            }
             case SERVER_STATE_LOADING_MODEL:
                 res.set_content(R"({"status": "loading model"})", "application/json");
                 res.status = 503; // HTTP Service Unavailable
@@ -2598,32 +2575,6 @@ int main(int argc, char **argv)
         }
     });
 
-    if (sparams.slots_endpoint) {
-        svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
-            json slots;
-            for (llama_client_slot & slot : llama.slots) {
-                json slot_data = llama.get_formated_generation(slot);
-                slot_data["id"] = slot.id;
-                slot_data["task_id"] = slot.task_id;
-                slot_data["state"] = slot.state;
-                slot_data["prompt"] = slot.prompt;
-                slot_data["next_token"] = {
-                        {"has_next_token", slot.has_next_token},
-                        {"n_remain", slot.n_remaining},
-                        {"num_tokens_predicted", slot.n_decoded},
-                        {"stopped_eos", slot.stopped_eos},
-                        {"stopped_word", slot.stopped_word},
-                        {"stopped_limit", slot.stopped_limit},
-                        {"stopping_word", slot.stopping_word},
-                };
-
-                slots.push_back(slot_data);
-            }
-            res.set_content(slots.dump(), "application/json");
-            res.status = 200; // HTTP OK
-        });
-    }
-
     svr.set_logger(log_server_request);
 
     svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
@@ -2884,7 +2835,7 @@ int main(int argc, char **argv)
                 if (!validate_api_key(req, res)) {
                     return;
                 }
-                json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template);
+                json data = oaicompat_completion_params_parse(json::parse(req.body), sparams.chat_template);
 
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
@@ -3177,25 +3128,8 @@ int main(int argc, char **argv)
         std::placeholders::_2,
         std::placeholders::_3
     ));
-
-    shutdown_handler = [&](int) {
-        llama.queue_tasks.terminate();
-    };
-
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-    struct sigaction sigint_action;
-    sigint_action.sa_handler = signal_handler;
-    sigemptyset (&sigint_action.sa_mask);
-    sigint_action.sa_flags = 0;
-    sigaction(SIGINT, &sigint_action, NULL);
-#elif defined (_WIN32)
-    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-    };
-    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
-#endif
     llama.queue_tasks.start_loop();
-    svr.stop();
+
     t.join();
 
     llama_backend_free();

examples/server/utils.hpp

@@ -167,47 +167,50 @@ static T json_value(const json &body, const std::string &key, const T &default_v
         : default_value;
 }
 
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-inline bool verify_custom_template(const std::string & tmpl) {
-    llama_chat_message chat[] = {{"user", "test"}};
-    std::vector<char> buf(1);
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size());
-    return res >= 0;
+inline std::string format_llama2(std::vector<json> messages)
+{
+    std::ostringstream output;
+    bool is_inside_turn = false;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        if (!is_inside_turn) {
+            output << "[INST] ";
+        }
+        std::string role    = json_value(*it, "role", std::string("user"));
+        std::string content = json_value(*it, "content", std::string(""));
+        if (role == "system") {
+            output << "<<SYS>>\n" << content << "\n<<SYS>>\n\n";
+            is_inside_turn = true;
+        } else if (role == "user") {
+            output << content << " [/INST]";
+            is_inside_turn = true;
+        } else {
+            output << " " << content << " </s>";
+            is_inside_turn = false;
+        }
+    }
+
+    LOG_VERBOSE("format_llama2", {{"text", output.str()}});
+
+    return output.str();
 }
 
-// Format given chat. If tmpl is empty, we take the template from model metadata
-inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages)
+inline std::string format_chatml(std::vector<json> messages)
 {
-    size_t alloc_size = 0;
-    // vector holding all allocated string to be passed to llama_chat_apply_template
-    std::vector<std::string> str(messages.size() * 2);
-    std::vector<llama_chat_message> chat(messages.size());
+    std::ostringstream chatml_msgs;
 
-    for (size_t i = 0; i < messages.size(); ++i) {
-        auto &curr_msg = messages[i];
-        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
-        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
-        alloc_size     += str[i*2 + 1].length();
-        chat[i].role    = str[i*2 + 0].c_str();
-        chat[i].content = str[i*2 + 1].c_str();
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
     }
 
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
-    std::vector<char> buf(alloc_size * 2);
+    chatml_msgs << "<|im_start|>assistant" << '\n';
 
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+    LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
 
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-    }
-
-    std::string formatted_chat(buf.data(), res);
-    LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
-
-    return formatted_chat;
+    return chatml_msgs.str();
 }
 
 //
@@ -217,7 +220,6 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
 struct llama_server_queue {
     int id = 0;
     std::mutex mutex_tasks;
-    bool running;
     // queues
     std::vector<task_server> queue_tasks;
     std::vector<task_server> queue_tasks_deferred;
@@ -276,18 +278,9 @@ struct llama_server_queue {
         queue_tasks_deferred.clear();
     }
 
-    // end the start_loop routine
-    void terminate() {
-        {
-            std::unique_lock<std::mutex> lock(mutex_tasks);
-            running = false;
-        }
-        condition_tasks.notify_all();
-    }
-
-    // Start the main loop.
+    // Start the main loop. This call is blocking
+    [[noreturn]]
     void start_loop() {
-        running = true;
         while (true) {
             // new task arrived
             LOG_VERBOSE("have new task", {});
@@ -331,12 +324,8 @@ struct llama_server_queue {
             {
                 std::unique_lock<std::mutex> lock(mutex_tasks);
                 if (queue_tasks.empty()) {
-                    if (!running) {
-                        LOG_VERBOSE("ending start_loop", {});
-                        return;
-                    }
                     condition_tasks.wait(lock, [&]{
-                        return (!queue_tasks.empty() || !running);
+                        return !queue_tasks.empty();
                     });
                 }
             }

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -111,13 +111,13 @@ static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
 static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
-    printf("%s: n_head:  %u\n", __func__, params->n_head);
-    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
-    printf("%s: n_layer: %u\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
+    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
+    printf("%s: n_head:  %d\n", __func__, params->n_head);
+    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
+    printf("%s: n_layer: %d\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }
 
 static void set_param_model(struct my_llama_model * model) {

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1708118438,
-        "narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
+        "lastModified": 1707268954,
+        "narHash": "sha256-2en1kvde3cJVc3ZnTy8QeD2oKcseLFjYPLKhIGDanQ0=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
+        "rev": "f8e2ebd66d097614d51a56a755450d4ae1632df1",
         "type": "github"
       },
       "original": {

flake.nix

@@ -150,7 +150,6 @@
             packages =
               {
                 default = config.legacyPackages.llamaPackages.llama-cpp;
-                vulkan = config.packages.default.override { useVulkan = true; };
               }
               // lib.optionalAttrs pkgs.stdenv.isLinux {
                 opencl = config.packages.default.override { useOpenCL = true; };
@@ -158,6 +157,7 @@
 
                 mpi-cpu = config.packages.default.override { useMpi = true; };
                 mpi-cuda = config.packages.default.override { useMpi = true; };
+                vulkan = config.packages.default.override { useVulkan = true; };
               }
               // lib.optionalAttrs (system == "x86_64-linux") {
                 rocm = config.legacyPackages.llamaPackagesRocm.llama-cpp;

ggml-alloc.c

@@ -377,9 +377,6 @@ struct ggml_gallocr {
 
     struct node_alloc * node_allocs; // [n_nodes]
     int n_nodes;
-
-    struct tensor_alloc * leaf_allocs; // [n_leafs]
-    int n_leafs;
 };
 
 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
@@ -430,7 +427,6 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
     free(galloc->buffers);
     free(galloc->buf_tallocs);
     free(galloc->node_allocs);
-    free(galloc->leaf_allocs);
     free(galloc);
 }
 
@@ -468,7 +464,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
             for (int i = 0; i < GGML_MAX_SRC; i++) {
                 struct ggml_tensor * parent = node->src[i];
                 if (parent == NULL) {
-                    continue;
+                    break;
                 }
 
                 // if the node's data is external, then we cannot re-use it
@@ -548,8 +544,22 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
     memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
 
+    // allocate all graph inputs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (graph->nodes[i]->flags & GGML_TENSOR_FLAG_INPUT) {
+            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            if (graph->nodes[i]->src[j] == NULL) {
+                break;
+            }
+            if (graph->nodes[i]->src[j]->flags & GGML_TENSOR_FLAG_INPUT) {
+                ggml_gallocr_allocate_node(galloc, graph->nodes[i]->src[j], get_node_buffer_id(node_buffer_ids, i));
+            }
+        }
+    }
+
     // count number of children and views
-    // allocate all graph inputs and leafs first to avoid overwriting them
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
@@ -558,37 +568,14 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
 
-        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
-            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
-        }
-
         for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-
-            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
-
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
-                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                break;
             }
+            ggml_gallocr_hash_get(galloc, parent)->n_children += 1;
         }
-    }
-
-    // allocate the remaining leafs that are unused on the graph
-    // these are effectively static tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-
-        if (hn->n_children == 0) {
-            assert(!hn->allocated);
-            // since buffer ids are only given for nodes, these leafs are always allocated in the first buffer
-            ggml_gallocr_allocate_node(galloc, leaf, 0);
-        }
-    }
+   }
 
     // allocate tensors
     for (int i = 0; i < graph->n_nodes; i++) {
@@ -599,7 +586,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             ggml_gallocr_allocate_node(galloc, parent, buffer_id);
         }
@@ -611,7 +598,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             AT_PRINTF("%s", parent->name);
             if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
@@ -624,7 +611,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
             p_hn->n_children -= 1;
@@ -709,18 +696,6 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
             }
         }
     }
-    if (galloc->n_leafs < graph->n_leafs) {
-        free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(struct tensor_alloc), graph->n_leafs);
-        GGML_ASSERT(galloc->leaf_allocs != NULL);
-    }
-    galloc->n_leafs = graph->n_leafs;
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].offset = hn->offset;
-        galloc->leaf_allocs[i].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
-    }
 
     // reallocate buffers if needed
     for (int i = 0; i < galloc->n_buffers; i++) {
@@ -747,8 +722,8 @@ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
     return ggml_gallocr_reserve_n(galloc, graph, NULL);
 }
 
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, struct node_alloc * node_alloc, struct tensor_alloc * tensor_alloc) {
+    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[node_alloc->buffer_id], node) <= tensor_alloc->size_max);
 
     if (node->view_src != NULL) {
         if (node->buffer == NULL) {
@@ -757,20 +732,29 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[node_alloc->buffer_id], node);
         }
     } else {
         if (node->data == NULL) {
             assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
-            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[node_alloc->buffer_id], node) <= tensor_alloc->size_max);
+            void * base = ggml_backend_buffer_get_base(galloc->buffers[node_alloc->buffer_id]);
             void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[node_alloc->buffer_id], node, addr);
         } else {
             if (node->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
+
+#ifndef NDEBUG
+            size_t offset =
+                (char *)node->data -
+                (char *)ggml_backend_buffer_get_base(node->buffer);
+            size_t size = ggml_backend_buffer_get_alloc_size(node->buffer, node);
+            assert(tensor_alloc->offset == SIZE_MAX || offset == tensor_alloc->offset);
+            assert(tensor_alloc->offset == SIZE_MAX || size <= tensor_alloc->size_max);
+#endif
         }
     }
 }
@@ -789,13 +773,6 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
         return true;
     }
 
-    if (galloc->n_leafs != graph->n_leafs) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
-#endif
-        return true;
-    }
-
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
@@ -810,7 +787,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                continue;
+                break;
             }
             if (!ggml_gallocr_node_needs_realloc(galloc, src, node_alloc, &node_alloc->src[j])) {
 #ifndef NDEBUG
@@ -850,24 +827,17 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
     }
 
     // allocate the graph tensors from the previous assignments
-    // nodes
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                continue;
+                break;
             }
-            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
+            ggml_gallocr_init_tensor(galloc, src, node_alloc, &node_alloc->src[j]);
         }
-        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
-    }
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct tensor_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, 0, leaf_alloc);
+        ggml_gallocr_init_tensor(galloc, node, node_alloc, &node_alloc->dst);
     }
 
     return true;

ggml-backend.c

@@ -756,7 +756,7 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CPY:
-            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
+            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
         default:
@@ -1006,7 +1006,6 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, gg
         }
     }
     GGML_ASSERT(false && "tensor buffer type not supported by any backend");
-    return -1; // silence warning
 }
 
 #if 0
@@ -1041,7 +1040,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
         if (src == NULL) {
-            continue;
+            break;
         }
         if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend = ggml_backend_sched_backend_from_buffer(sched, src->buffer);
@@ -1088,7 +1087,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                continue;
+                break;
             }
             ggml_backend_t src_backend = tensor_backend(src);
             fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
@@ -1144,7 +1143,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                continue;
+                break;
             }
             if (tensor_backend_id(src) == -1) {
                 tensor_backend_id(src) = ggml_backend_sched_backend_id_from_cur(sched, src);
@@ -1256,7 +1255,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                continue;
+                break;
             }
             int src_backend_id = tensor_backend_id(src);
             if (src_backend_id == -1) {
@@ -1315,7 +1314,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
-                    continue;
+                    break;
                 }
                 int src_backend_id = tensor_backend_id(src);
                 assert(src_backend_id != -1); // all inputs should be assigned by now
@@ -1362,7 +1361,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                continue;
+                break;
             }
             ggml_backend_t src_backend = tensor_backend(src);
             if (src_backend != tensor_backend /* && src_backend != NULL */) {
@@ -1668,7 +1667,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         struct ggml_tensor * s = src->src[i];
         if (s == NULL) {
-            continue;
+            break;
         }
         dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
     }
@@ -1697,7 +1696,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         struct ggml_tensor * s = src->src[i];
         if (s == NULL) {
-            continue;
+            break;
         }
         graph_copy_init_tensor(hash_set, node_copies, node_init, s);
     }

ggml-cuda.cu

@@ -54,8 +54,6 @@
 #define cudaDeviceProp hipDeviceProp_t
 #define cudaDeviceSynchronize hipDeviceSynchronize
 #define cudaError_t hipError_t
-#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
-#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
 #define cudaEventCreateWithFlags hipEventCreateWithFlags
 #define cudaEventDisableTiming hipEventDisableTiming
 #define cudaEventRecord hipEventRecord
@@ -519,24 +517,6 @@ typedef struct {
 } block_iq3_xxs;
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
-#define QR1_S 8
-#define QI1_S (QK_K / (4*QR1_S))
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/8];
-    uint8_t scales[QK_K/16];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-#define QK4_NL 32
-#define QR4_NL 2
-#define QI4_NL (QK4_NL / (4*QR4_NL))
-typedef struct {
-    half d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -662,18 +642,18 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
     return a;
 }
 
-//static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-//#pragma unroll
-//    for (int mask = 16; mask > 0; mask >>= 1) {
-//        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
-//    }
-//    return a;
-//#else
-//    (void) a;
-//    NO_DEVICE_CODE;
-//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-//}
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    }
+    return a;
+#else
+    (void) a;
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
 
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -683,18 +663,18 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
-//static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
-//#pragma unroll
-//    for (int mask = 16; mask > 0; mask >>= 1) {
-//        x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-//    }
-//    return x;
-//#else
-//    (void) x;
-//    NO_DEVICE_CODE;
-//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
-//}
+static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+#else
+    (void) x;
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+}
 
 static __device__ __forceinline__ float op_repeat(const float a, const float b) {
     return b;
@@ -1701,137 +1681,6 @@ static const __device__ uint32_t iq3xxs_grid[256] = {
     0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
 };
 
-static const __device__ uint64_t iq1s_grid[512] = {
-    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
-    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
-    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
-    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
-    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
-    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
-    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
-    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
-    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
-    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
-    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
-    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
-    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
-    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
-    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
-    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
-    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
-    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
-    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
-    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
-    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
-    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
-    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
-    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
-    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
-    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
-    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
-    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
-    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
-    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
-    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
-    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
-    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
-    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
-    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
-    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
-    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
-    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
-    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
-    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
-    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
-    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
-    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
-    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
-    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
-    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
-    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
-    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
-    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
-    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
-    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
-    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
-    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
-    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
-    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
-    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
-    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
-    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
-    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
-    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
-    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
-    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
-    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
-    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
-    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
-    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
-    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
-    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
-    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
-    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
-    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
-    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
-    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
-    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
-    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
-    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
-    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
-    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
-    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
-    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
-    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
-    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
-    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
-    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
-    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
-    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
-    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
-    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
-    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
-    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
-    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
-    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
-    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
-    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
-    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
-    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
-    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
-    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
-    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
-    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
-    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
-    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
-    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
-    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
-    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
-    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
-    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
-    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
-    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
-    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
-    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
-    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
-    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
-    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
-    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
-    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
-    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
-    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
-    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
-    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
-    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
-    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
-    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
-    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
-    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
-    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
-    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
-    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
-};
-
 static const __device__ uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
     144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
@@ -1974,49 +1823,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
 
 }
 
-template<typename dst_t>
-static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int i   = blockIdx.x;
-    const block_iq1_s * x = (const block_iq1_s  *) vx;
-
-    const int tid = threadIdx.x;
-#if QK_K == 256
-    const int il = tid/8; // 0...3
-    const int ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const int i8 = 4*ib+il;
-    uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
-    const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
-    const float d = (float)x[i].d * (2*(h & 7) + 1);
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
-#else
-    assert(false);
-#endif
-
-}
-
-static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int i   = blockIdx.x;
-    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
-
-    const int tid = threadIdx.x;
-    const int il = tid/8; // 0...3
-    const int ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
-    const uint8_t  * q4 = x[ib].qs + 4*il;
-    const float d = (float)x[ib].d;
-    for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
-        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
-    }
-
-}
-
 static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
@@ -4672,12 +4478,10 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
     return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
 #else
-    (void) ksigns64;
     assert(false);
     return 0.f;
 #endif
 #else
-    (void) ksigns64;
     assert(false);
     return 0.f;
 #endif
@@ -4718,99 +4522,6 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #endif
 }
 
-static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
-    const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
-
-    const int ib32 = iqs;
-    int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
-    const uint8_t h1 = bq1->scales[2*ib32+0];
-    const uint8_t h2 = bq1->scales[2*ib32+1];
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const int * q8 = (const int *)bq8_1[ib32].qs;
-    const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
-    const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
-    const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
-    const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
-    for (int j = 0; j < 2; ++j) {
-        sumi1 = __dp4a(q8[j+0], grid1[j], sumi1);
-        sumi2 = __dp4a(q8[j+2], grid2[j], sumi2);
-        sumi3 = __dp4a(q8[j+4], grid3[j], sumi3);
-        sumi4 = __dp4a(q8[j+6], grid4[j], sumi4);
-    }
-#else
-    const int8_t   * q8 = bq8_1[ib32].qs;
-    const int8_t * grid1 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
-    const int8_t * grid2 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
-    const int8_t * grid3 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
-    const int8_t * grid4 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
-    for (int j = 0; j < 8; ++j) {
-        sumi1 += q8[j+ 0] * grid1[j];
-        sumi2 += q8[j+ 8] * grid2[j];
-        sumi3 += q8[j+16] * grid3[j];
-        sumi4 += q8[j+24] * grid4[j];
-    }
-#endif
-    const float d = (float)bq1->d * __low2float(bq8_1[ib32].ds);
-    return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
-                sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
-#else
-    assert(false);
-    return 0.f;
-#endif
-}
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values,
-        int & val1, int & val2) {
-
-    uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32;
-    aux32 = q4 & 0x0f0f0f0f;
-    uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8);
-    uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8);
-    val1 = v1 | (v2 << 16);
-    aux32 = (q4 >> 4) & 0x0f0f0f0f;
-    v1 = values[q8[0]] | (values[q8[1]] << 8);
-    v2 = values[q8[2]] | (values[q8[3]] << 8);
-    val2 = v1 | (v2 << 16);
-}
-#endif
-
-static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_iq4_nl * bq = (const block_iq4_nl *) vbq;
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs;
-    const int32_t  * q8 = (const int32_t  *)bq8_1->qs + iqs;
-
-    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
-
-    int v1, v2;
-    int sumi1 = 0, sumi2 = 0;
-    for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
-        const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16);
-        get_int_from_table_16(aux, values, v1, v2);
-        sumi1 = __dp4a(v1, q8[l+0], sumi1);
-        sumi2 = __dp4a(v2, q8[l+4], sumi2);
-    }
-
-#else
-    const uint8_t * q4 = bq->qs + 4*iqs;
-    const int8_t  * q8 = bq8_1->qs + 4*iqs;
-
-    int sumi1 = 0, sumi2 = 0;
-    for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) {
-        sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf];
-        sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >>  4];
-    }
-#endif
-    const float d = (float)bq->d * __low2float(bq8_1->ds);
-    return d * (sumi1 + sumi2);
-}
-
 template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
               allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
 static __device__ __forceinline__ void mul_mat_q(
@@ -6245,31 +5956,149 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
-template <bool vals_smem, int ncols_template, int block_size_template>
-static __global__ void soft_max_f32(const float * x, const float * mask, const float * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
-    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+template <bool vals_smem, int ncols_template, int block_size_template, bool need_check>
+static __global__ void soft_max_f16(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+    const int ncols_data = ncols_template == 0 ? ncols_par : ncols_template;
+    const int ncols_smem = GGML_PAD(ncols_data, 2*WARP_SIZE)/2;
 
     const int tid  = threadIdx.x;
     const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
 
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
 
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    float slope = 0.0f;
+    extern __shared__ half data_soft_max_f16[];
+    half  * buf_iw = data_soft_max_f16 + 0; // shared memory buffer for inter-warp communication
+    // (shared memory) buffer to cache values between iterations:
+    half2 * vals = vals_smem ? (half2 *) (buf_iw + WARP_SIZE) : (half2 *) (dst + rowx*ncols_data);
+    // if the buffer is larger than max. shared memory per block, use dst as temp. buffer instead
+    // in that case col_smem == col_data must be enforced to avoid race conditions
 
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = rowx/nrows_y; // head index
+    half2 max_val = make_half2(-INFINITY, -INFINITY);
 
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+#pragma unroll
+    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
+        const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
+        const int col_smem = vals_smem ? col0 + tid : col_data;
 
-        slope = powf(base, exp);
+        const int ix = rowx*ncols_data + col_data;
+        const int iy = rowy*ncols_data + col_data;
+
+        half2 val;
+        if (need_check && col_data + 0 >= ncols_data) {
+            val.x = -INFINITY;
+        } else {
+            val.x = x[ix + 0]*scale + (y ? y[iy + 0] : 0.0f);
+        }
+        if (need_check && col_data + WARP_SIZE >= ncols_data) {
+            val.y = -INFINITY;
+        } else {
+            val.y = x[ix + WARP_SIZE]*scale + (y ? y[iy + WARP_SIZE] : 0.0f);
+        }
+        if (!need_check || col_smem < (vals_smem ? ncols_smem : ncols_data)) {
+            vals[col_smem] = val;
+        }
+        max_val = __hmax2(max_val, val);
     }
 
+    // find the max value in the block
+    max_val = warp_reduce_max(max_val);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf_iw[lane_id] = -INFINITY;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = __hmax(max_val.x, max_val.y);
+        }
+        __syncthreads();
+
+        max_val = __half2half2(buf_iw[lane_id]);
+        max_val = warp_reduce_max(max_val);
+    } else {
+        max_val = __half2half2(__hmax(max_val.x, max_val.y));
+    }
+
+    half2 tmp = make_half2(0.0f, 0.0f); // partial sums
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
+        const int col_smem = vals_smem ? col0 + tid : 2*col0 + 2*warp_id*WARP_SIZE + lane_id;
+
+        if (ncols_template == 0 && col_smem >= (vals_smem ? ncols_smem : ncols_data)) {
+            break;
+        }
+
+        const half2 val = h2exp(vals[col_smem] - max_val);
+
+        tmp += val;
+        vals[col_smem] = val;
+    }
+
+    // find the sum of exps in the block
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf_iw[lane_id] = 0.0f;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = tmp.x + tmp.y;
+        }
+        __syncthreads();
+
+        tmp = __half2half2(buf_iw[lane_id]);
+        tmp = warp_reduce_sum(tmp);
+    } else {
+        tmp = __half2half2(tmp.x + tmp.y);
+    }
+
+    const half2 inv_sum = make_half2(1.0f, 1.0f) / tmp;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
+        const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
+        const int col_smem = vals_smem ? col0 + tid : col_data;
+
+        const int idst = rowx*ncols_data + col_data;
+        const half2 result = vals[col_smem] * inv_sum;
+
+        if (need_check && col_data + 0 >= ncols_data) {
+            return;
+        }
+        dst[idst] = result.x;
+
+        if (need_check && col_data + WARP_SIZE >= ncols_data) {
+            return;
+        }
+
+        dst[idst + WARP_SIZE] = result.y;
+    }
+#else
+    (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale;
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+}
+
+template <bool vals_smem, int ncols_template, int block_size_template>
+static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
+    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+
+    const int tid  = threadIdx.x;
+    const int rowx = blockIdx.x;
+    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
+
+    const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
+
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
     // shared memory buffer to cache values between iterations:
@@ -6288,8 +6117,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
 
-        const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
-
+        const float val = x[ix]*scale + (y ? y[iy] : 0.0f);
         vals[col] = val;
         max_val = max(max_val, val);
     }
@@ -6850,18 +6678,6 @@ static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
-template<typename dst_t>
-static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = (k + QK_K - 1) / QK_K;
-    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
-}
-
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
@@ -6901,10 +6717,6 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq2_xs_cuda;
         case GGML_TYPE_IQ3_XXS:
             return dequantize_row_iq3_xxs_cuda;
-        case GGML_TYPE_IQ1_S:
-            return dequantize_row_iq1_s_cuda;
-        case GGML_TYPE_IQ4_NL:
-            return dequantize_row_iq4_nl_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
         default:
@@ -6940,10 +6752,6 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq2_xs_cuda;
         case GGML_TYPE_IQ3_XXS:
             return dequantize_row_iq3_xxs_cuda;
-        case GGML_TYPE_IQ1_S:
-            return dequantize_row_iq1_s_cuda;
-        case GGML_TYPE_IQ4_NL:
-            return dequantize_row_iq4_nl_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half>;
         default:
@@ -7781,53 +7589,89 @@ static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols
     diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
 }
 
-static void soft_max_f32_cuda(const float * x, const float * mask, const float * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+static void soft_max_f16_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
+    int nth = WARP_SIZE;
+    while (nth < ncols_x/2 && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
+    const dim3 block_dims(nth,     1, 1);
+    const dim3 block_nums(nrows_x, 1, 1);
+    const size_t shmem = (GGML_PAD(ncols_x, 2*WARP_SIZE) + WARP_SIZE)*sizeof(half);
+    static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
+    if (shmem <= g_device_caps[g_main_device].smpb) {
+        switch (ncols_x) {
+            case 32:
+                soft_max_f16<true, 32, 32, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 64:
+                soft_max_f16<true, 64, 32, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 128:
+                soft_max_f16<true, 128, 64, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 256:
+                soft_max_f16<true, 256, 128, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 512:
+                soft_max_f16<true, 512, 256, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 1024:
+                soft_max_f16<true, 1024, 512, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 2048:
+                soft_max_f16<true, 2048, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 4096:
+                soft_max_f16<true, 4096, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            default:
+                soft_max_f16<true, 0, 0, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+        }
+    } else {
+        const size_t shmem_low = WARP_SIZE*sizeof(half);
+        soft_max_f16<false, 0, 0, true><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+    }
+}
+
+static void soft_max_f32_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
     int nth = WARP_SIZE;
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
     const dim3 block_nums(nrows_x, 1, 1);
     const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
-
-    const uint32_t n_head_kv   = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
     if (shmem < g_device_caps[g_main_device].smpb) {
         switch (ncols_x) {
             case 32:
-                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 64:
-                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 128:
-                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 256:
-                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 512:
-                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
             default:
-                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
                 break;
         }
     } else {
         const size_t shmem_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
     }
 }
 
@@ -8687,8 +8531,6 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ4_NL:
             return max_compute_capability >= CC_RDNA2 ? 128 : 64;
         default:
             GGML_ASSERT(false);
@@ -8712,8 +8554,6 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ4_NL:
             return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
@@ -8811,14 +8651,6 @@ static void ggml_cuda_op_mul_mat_vec_q(
             mul_mat_vec_q_cuda<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
                 (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;
-        case GGML_TYPE_IQ1_S:
-            mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_s, 1, vec_dot_iq1_s_q8_1>
-                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_NL:
-            mul_mat_vec_q_cuda<QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
-                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -9258,36 +9090,30 @@ static void ggml_cuda_op_soft_max(
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
 
-    const int64_t ne00    = src0->ne[0];
+    const int64_t ne00 = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
-    const int64_t nrows_y = src0->ne[1];
+    const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1;
 
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
+    float scale = 1.0f;
+    memcpy(&scale, dst->op_params, sizeof(float));
 
-    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION >= CUDART_HMAX
+#ifdef GGML_CUDA_F16
+    const bool use_f16_soft_max = true;
+#else
+    const bool use_f16_soft_max = false;
+#endif // GGML_CUDA_F16
+#else
+    const bool use_f16_soft_max = false;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && CUDART_VERSION >= CUDART_HMAX
 
-    // positions tensor
-    float * src2_dd = nullptr;
-    cuda_pool_alloc<float> src2_f;
-
-    ggml_tensor * src2 = dst->src[2];
-    const bool use_src2 = src2 != nullptr;
-
-    if (use_src2) {
-        const bool src2_on_device = src2->backend == GGML_BACKEND_GPU;
-
-        if (src2_on_device) {
-            ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
-            src2_dd = (float *) src2_extra->data_device[g_main_device];
-        } else {
-            src2_dd = src2_f.alloc(ggml_nelements(src2));
-            CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
-        }
+    if (use_f16_soft_max) {
+        soft_max_f16_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
+    } else {
+        soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
     }
 
-    soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream);
+    (void) dst;
 }
 
 static void ggml_cuda_op_scale(
@@ -9422,15 +9248,9 @@ static void ggml_cuda_set_peer_access(const int n_tokens) {
             CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
             if (can_access_peer) {
                 if (enable_peer_access) {
-                    cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
-                    if (err != cudaErrorPeerAccessAlreadyEnabled) {
-                        CUDA_CHECK(err);
-                    }
+                    CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
                 } else {
-                    cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
-                    if (err != cudaErrorPeerAccessNotEnabled) {
-                        CUDA_CHECK(err);
-                    }
+                    CUDA_CHECK(cudaDeviceDisablePeerAccess(id_other));
                 }
             }
         }
@@ -11102,10 +10922,10 @@ GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backe
     UNUSED(buffer);
 }
 
-static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
-    UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
-}
+// unused at the moment
+//static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
+//    return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
+//}
 
 GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
@@ -11493,7 +11313,7 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             if (node->src[j] != nullptr) {
                 assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT);
-                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
+                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
                 assert(node->src[j]->extra != nullptr);
             }
         }
@@ -11541,8 +11361,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                     return false;
                 }
                 ggml_type a_type = a->type;
-                if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
-                    a_type == GGML_TYPE_IQ1_S   || a_type == GGML_TYPE_IQ4_NL) {
+                if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }

ggml-metal.m

@@ -61,8 +61,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@@ -85,8 +83,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
   //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@@ -105,8 +101,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@@ -122,8 +116,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@@ -139,8 +131,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
     GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@@ -186,7 +176,7 @@ struct ggml_metal_context {
 // MSL code
 // TODO: move the contents here when ready
 //       for now it is easier to work in a separate file
-// static NSString * const msl_library_source = @"see metal.metal";
+//static NSString * const msl_library_source = @"see metal.metal";
 
 // Here to assist with NSBundle Path Hack
 @interface GGMLMetalClass : NSObject
@@ -282,14 +272,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 return NULL;
             }
         } else {
-#if GGML_METAL_EMBED_LIBRARY
-            GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__);
-
-            extern const char ggml_metallib_start[];
-            extern const char ggml_metallib_end[];
-
-            NSString * src  = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath;
@@ -312,7 +294,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
             }
-#endif
 
             @autoreleasepool {
                 // dictionary of preprocessor macros
@@ -452,8 +433,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,       true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,           get_rows_iq4_nl,        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,               ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,             ctx->support_simdgroup_reduction);
@@ -476,8 +455,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,     ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,      ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,         mul_mv_iq4_nl_f32,      ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,      ctx->support_simdgroup_reduction);
       //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,      ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,      ctx->support_simdgroup_reduction);
@@ -496,8 +473,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,  ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,   ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,      mul_mv_id_iq4_nl_f32,   ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,         ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,         ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,        ctx->support_simdgroup_mm);
@@ -513,8 +488,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,     ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,         mul_mm_iq4_nl_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,     ctx->support_simdgroup_mm);
@@ -530,8 +503,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,  ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,   ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,      mul_mm_id_iq4_nl_f32,   ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,              true);
@@ -757,7 +728,6 @@ static bool ggml_metal_graph_compute(
 
         size_t offs_src0 = 0;
         size_t offs_src1 = 0;
-        size_t offs_src2 = 0;
         size_t offs_dst  = 0;
 
         id<MTLCommandBuffer> command_buffer  = command_buffers[cb_idx];
@@ -776,7 +746,6 @@ static bool ggml_metal_graph_compute(
 
             struct ggml_tensor * src0 = gf->nodes[i]->src[0];
             struct ggml_tensor * src1 = gf->nodes[i]->src[1];
-            struct ggml_tensor * src2 = gf->nodes[i]->src[2];
             struct ggml_tensor * dst  = gf->nodes[i];
 
             switch (dst->op) {
@@ -838,7 +807,6 @@ static bool ggml_metal_graph_compute(
 
             id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
             id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
-            id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
             id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
             //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
@@ -1220,16 +1188,7 @@ static bool ggml_metal_graph_compute(
                             pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline;
                         }
 
-                        const float scale    = ((float *) dst->op_params)[0];
-                        const float max_bias = ((float *) dst->op_params)[1];
-
-                        const int64_t nrows_x = ggml_nrows(src0);
-                        const int64_t nrows_y = src0->ne[1];
-                        const uint32_t n_head_kv   = nrows_x/nrows_y;
-                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
-
-                        const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+                        const float scale = ((float *) dst->op_params)[0];
 
                         [encoder setComputePipelineState:pipeline];
                         [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
@@ -1238,20 +1197,11 @@ static bool ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
                         }
-                        if (id_src2) {
-                            [encoder setBuffer:id_src2 offset:offs_src2   atIndex:2];
-                        } else {
-                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:2];
-                        }
-                        [encoder setBuffer:id_dst   offset:offs_dst          atIndex:3];
-                        [encoder setBytes:&ne00     length:sizeof(ne00)      atIndex:4];
-                        [encoder setBytes:&ne01     length:sizeof(ne01)      atIndex:5];
-                        [encoder setBytes:&ne02     length:sizeof(ne02)      atIndex:6];
-                        [encoder setBytes:&scale    length:sizeof(scale)     atIndex:7];
-                        [encoder setBytes:&max_bias length:sizeof(max_bias)  atIndex:8];
-                        [encoder setBytes:&m0       length:sizeof(m0)        atIndex:9];
-                        [encoder setBytes:&m1       length:sizeof(m1)        atIndex:10];
-                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:11];
+                        [encoder setBuffer:id_dst  offset:offs_dst    atIndex:2];
+                        [encoder setBytes:&ne00  length:sizeof(ne00)  atIndex:3];
+                        [encoder setBytes:&ne01  length:sizeof(ne01)  atIndex:4];
+                        [encoder setBytes:&ne02  length:sizeof(ne02)  atIndex:5];
+                        [encoder setBytes:&scale length:sizeof(scale) atIndex:6];
                         [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -1347,8 +1297,6 @@ static bool ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                             }
 
@@ -1483,18 +1431,6 @@ static bool ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ1_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
-                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1529,7 +1465,7 @@ static bool ggml_metal_graph_compute(
 
                             if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                 src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
-                                src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
+                                src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -1542,11 +1478,6 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src0t == GGML_TYPE_IQ4_NL) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
                             else if (src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1583,6 +1514,8 @@ static bool ggml_metal_graph_compute(
                         // max size of the src1ids array in the kernel stack
                         GGML_ASSERT(ne11 <= 512);
 
+                        struct ggml_tensor * src2 = gf->nodes[i]->src[2];
+
                         const int64_t  ne20 = src2 ? src2->ne[0] : 0;
                         const int64_t  ne21 = src2 ? src2->ne[1] : 0;
                         const int64_t  ne22 = src2 ? src2->ne[2] : 0;
@@ -1640,8 +1573,6 @@ static bool ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                             }
 
@@ -1779,18 +1710,6 @@ static bool ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ1_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
-                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@@ -1841,7 +1760,7 @@ static bool ggml_metal_graph_compute(
 
                             if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
                                 src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
-                                src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
+                                src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
@@ -1854,11 +1773,6 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ4_NL) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
                             else if (src2t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1900,8 +1814,6 @@ static bool ggml_metal_graph_compute(
                             case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
                             case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
                             case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
-                            case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S  ].pipeline; break;
-                            case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
                             case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
                             default: GGML_ASSERT(false && "not implemented");
                         }

ggml-metal.metal

@@ -351,17 +351,12 @@ kernel void kernel_sum_rows(
 kernel void kernel_soft_max(
         device const float * src0,
         device const float * src1,
-        device const float * src2,
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint32_t & n_head_log2,
-        threadgroup  float * buf [[threadgroup(0)]],
+        threadgroup float  * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -373,26 +368,13 @@ kernel void kernel_soft_max(
 
     device const float * psrc0 =         src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device const float * pmask = src1 != src0 ? src1                               + i01*ne00 : nullptr;
-    device const float * ppos  = src2 != src0 ? src2                                          : nullptr;
     device       float * pdst  =         dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
-    float slope = 0.0f;
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int64_t h = i02;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = pow(base, exp);
-    }
-
     // parallel max
     float lmax = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f));
     }
 
     // find the max value in the block
@@ -417,7 +399,7 @@ kernel void kernel_soft_max(
     // parallel sum
     float lsum = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
     }
@@ -455,17 +437,12 @@ kernel void kernel_soft_max(
 kernel void kernel_soft_max_4(
         device const float * src0,
         device const float * src1,
-        device const float * src2,
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint32_t & n_head_log2,
-        threadgroup  float * buf [[threadgroup(0)]],
+        threadgroup float  * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -477,25 +454,13 @@ kernel void kernel_soft_max_4(
 
     device const float4 * psrc4 =                (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
-    device const float4 * ppos  = src2 != src0 ? (device const float4 *)(src2)                                                 : nullptr;
     device       float4 * pdst4 =                (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
-    float slope = 0.0f;
-
-    if (max_bias > 0.0f) {
-        const int64_t h = i02;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = pow(base, exp);
-    }
-
     // parallel max
     float4 lmax4 = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f));
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@@ -521,7 +486,7 @@ kernel void kernel_soft_max_4(
     // parallel sum
     float4 lsum4 = 0.0f;
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
@@ -2525,19 +2490,6 @@ typedef struct {
 } block_iq3_xxs;
 // 98 bytes / block for QK_K = 256, so 3.0625 bpw
 
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/8];
-    uint8_t scales[QK_K/16];
-} block_iq1_s;
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    half    d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3795,137 +3747,6 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
     0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
 };
 
-#define NGRID_IQ1S 512
-constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = {
-    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
-    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
-    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
-    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
-    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
-    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
-    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
-    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
-    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
-    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
-    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
-    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
-    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
-    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
-    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
-    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
-    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
-    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
-    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
-    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
-    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
-    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
-    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
-    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
-    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
-    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
-    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
-    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
-    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
-    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
-    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
-    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
-    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
-    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
-    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
-    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
-    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
-    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
-    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
-    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
-    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
-    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
-    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
-    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
-    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
-    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
-    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
-    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
-    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
-    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
-    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
-    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
-    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
-    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
-    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
-    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
-    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
-    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
-    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
-    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
-    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
-    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
-    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
-    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
-    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
-    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
-    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
-    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
-    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
-    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
-    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
-    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
-    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
-    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
-    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
-    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
-    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
-    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
-    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
-    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
-    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
-    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
-    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
-    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
-    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
-    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
-    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
-    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
-    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
-    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
-    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
-    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
-    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
-    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
-    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
-    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
-    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
-    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
-    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
-    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
-    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
-    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
-    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
-    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
-    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
-    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
-    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
-    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
-    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
-    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
-    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
-    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
-    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
-    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
-    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
-    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
-    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
-    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
-    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
-    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
-    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
-    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
-    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
-    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
-    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
-    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
-    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
-    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
-};
 
 constexpr constant static uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
@@ -4033,10 +3854,7 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    (void) x;
-    (void) y;
-    (void) yl;
-    (void) nb32;
+    // TODO
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4179,10 +3997,7 @@ void kernel_mul_mv_iq2_xs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    (void) x;
-    (void) y;
-    (void) yl;
-    (void) nb32;
+    // TODO
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4318,10 +4133,7 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    (void) x;
-    (void) y;
-    (void) yl;
-    (void) nb32;
+    // TODO
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4361,250 +4173,6 @@ kernel void kernel_mul_mv_iq3_xxs_f32(
     kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
-void kernel_mul_mv_iq1_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-    device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
-    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[16];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int nb32 = nb * (QK_K / 32);
-
-#if QK_K == 256
-    const int ix = tiisg/2;
-    const int il = tiisg%2;
-
-    device const float * y4 = y + 32 * ix + 16 * il;
-
-    for (int ib32 = ix; ib32 < nb32; ib32 += 16) {
-
-        for (int i = 0; i < 16; ++i) {
-            yl[i] = y4[i];
-        }
-
-        const int ibl = ib32 / (QK_K / 32);
-        const int ib  = ib32 % (QK_K / 32);
-
-        device const block_iq1_s * xr = x + ibl;
-        device const uint8_t * qs = xr->qs + 4 * ib + 2 * il;
-        device const uint8_t * sc = xr->scales + 2 * ib + il;
-        device const half    * dh = &xr->d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
-            constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
-
-            float2 sum = {0};
-            for (int j = 0; j < 8; ++j) {
-                sum[0] += yl[j+ 0] * grid1[j];
-                sum[1] += yl[j+ 8] * grid2[j];
-            }
-            sumf[row] += (float)dh[0] * (sum[0] * (2*(sc[0] & 7) + 1) + sum[1] * (2*((sc[0] >> 4) & 7) + 1));
-
-            dh += nb*sizeof(block_iq1_s)/2;
-            qs += nb*sizeof(block_iq1_s);
-            sc += nb*sizeof(block_iq1_s);
-        }
-
-        y4 += 16 * 32;
-    }
-#else
-    (void) x;
-    (void) y;
-    (void) yl;
-    (void) nb32;
-#endif
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-constexpr constant static float kvalues_iq4nl_f[16] = {
-    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
-};
-
-void kernel_mul_mv_iq4_nl_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup float  * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int nb = ne00/QK4_NL;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    const int first_row = (r0 * 2 + sgitg) * 2;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-    device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
-    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
-
-    const int ix = tiisg/2;  // 0...15
-    const int it = tiisg%2;  // 0 or 1
-
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    float4 yl[4];
-    float sumf[2]={0.f}, all_sum;
-
-    device const float * yb = y + ix * QK4_NL + it * 8;
-
-    uint32_t aux32[2];
-    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
-
-    float4 qf1, qf2;
-
-    for (int ib = ix; ib < nb; ib += 16) {
-
-        device const float4 * y4 = (device const float4 *)yb;
-        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
-
-        for (int row = 0; row < 2; ++row) {
-
-            device const block_iq4_nl & xb = x[row*nb + ib];
-            device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
-
-            float4 acc1 = {0.f}, acc2 = {0.f};
-
-            aux32[0] = q4[0] | (q4[1] << 16);
-            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
-            aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
-            acc1 += yl[0] * qf1;
-            acc2 += yl[1] * qf2;
-
-            aux32[0] = q4[2] | (q4[3] << 16);
-            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
-            aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
-            acc1 += yl[2] * qf1;
-            acc2 += yl[3] * qf2;
-
-            acc1 += acc2;
-
-            sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
-
-        }
-
-        yb += 16 * QK4_NL;
-    }
-
-    for (int row = 0; row < 2; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq1_s_f32")]]
-kernel void kernel_mul_mv_iq1_s_f32(
-        device const  void * src0,
-        device const float * 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,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
-}
-
-[[host_name("kernel_mul_mv_iq4_nl_f32")]]
-kernel void kernel_mul_mv_iq4_nl_f32(
-        device const  void * src0,
-        device const float * 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,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup float * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-}
 
 //============================= templates and their specializations =============================
 
@@ -4801,8 +4369,6 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     const float dl = d * sc[0];
     const float ml = min * sc[1];
 #else
-    (void) get_scale_min_k4_just2;
-
     q = q + 16 * (il&1);
     device const uint8_t * s = xb->scales;
     device const half2 * dh = (device const half2 *)xb->d;
@@ -4952,37 +4518,6 @@ void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x
     }
 }
 
-template <typename type4x4>
-void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
-    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
-    const float d = xb->d;
-    device const uint8_t * qs = xb->qs + 2*il;
-    device const uint8_t * sc = xb->scales + il;
-    const float dl1 = d * (2*(sc[0] & 7) + 1);
-    const float dl2 = d * (2*((sc[0] >> 4) & 7) + 1);
-    constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
-    constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
-    for (int i = 0; i < 8; ++i) {
-        reg[i/4+0][i%4] = dl1 * grid1[i];
-        reg[i/4+2][i%4] = dl2 * grid2[i];
-    }
-}
-
-template <typename type4x4>
-void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
-    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
-    const float d = xb->d;
-    uint32_t aux32;
-    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
-    for (int i = 0; i < 4; ++i) {
-        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
-        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
-        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
-        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
-        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
-    }
-}
-
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -5525,8 +5060,6 @@ template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // matrix-matrix multiplication
@@ -5566,8 +5099,6 @@ template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // indirect matrix-matrix multiplication
@@ -5619,8 +5150,6 @@ template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mu
 template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // matrix-vector multiplication
@@ -6588,131 +6117,3 @@ kernel void kernel_mul_mv_id_iq3_xxs_f32(
         tiisg,
         sgitg);
 }
-
-[[host_name("kernel_mul_mv_id_iq1_s_f32")]]
-kernel void kernel_mul_mv_id_iq1_s_f32(
-        device const    char * ids,
-        device const    char * src1,
-        device         float * dst,
-        constant    uint64_t & nbi1,
-        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     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        constant        uint & r2,
-        constant        uint & r3,
-        constant         int & idx,
-        device const    char * src00,
-        device const    char * src01,
-        device const    char * src02,
-        device const    char * src03,
-        device const    char * src04,
-        device const    char * src05,
-        device const    char * src06,
-        device const    char * src07,
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   tiisg[[thread_index_in_simdgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
-
-    const int64_t bid = tgpig.z/(ne12*ne13);
-
-    tgpig.z = tgpig.z%(ne12*ne13);
-
-    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
-
-    kernel_mul_mv_iq1_s_f32_impl(
-        src0[id],
-        (device const float *) (src1 + bid*nb11),
-        dst + bid*ne0,
-        ne00,
-        ne01,
-        ne02,
-        ne10,
-        ne12,
-        ne0,
-        ne1,
-        r2,
-        r3,
-        tgpig,
-        tiisg,
-        sgitg);
-}
-
-[[host_name("kernel_mul_mv_id_iq4_nl_f32")]]
-kernel void kernel_mul_mv_id_iq4_nl_f32(
-        device const    char * ids,
-        device const    char * src1,
-        device         float * dst,
-        constant    uint64_t & nbi1,
-        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     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        constant        uint & r2,
-        constant        uint & r3,
-        constant         int & idx,
-        device const    char * src00,
-        device const    char * src01,
-        device const    char * src02,
-        device const    char * src03,
-        device const    char * src04,
-        device const    char * src05,
-        device const    char * src06,
-        device const    char * src07,
-        threadgroup float    * shared_values [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   tiisg[[thread_index_in_simdgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
-
-    const int64_t bid = tgpig.z/(ne12*ne13);
-
-    tgpig.z = tgpig.z%(ne12*ne13);
-
-    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
-
-    kernel_mul_mv_iq4_nl_f32_impl(
-        src0[id],
-        (device const float *) (src1 + bid*nb11),
-        dst + bid*ne0,
-        ne00,
-        ne01,
-        ne02,
-        ne10,
-        ne12,
-        ne0,
-        ne1,
-        r2,
-        r3,
-        shared_values,
-        tgpig,
-        tiisg,
-        sgitg);
-}

ggml-quants.h

@@ -191,21 +191,6 @@ typedef struct {
 } block_iq3_xxs;
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/8];
-    uint8_t scales[QK_K/16];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -225,7 +210,6 @@ void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGM
 void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int k);
 void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k);
 void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int k);
 
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
@@ -241,7 +225,6 @@ void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 
 // Dequantization
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
@@ -260,8 +243,6 @@ void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRI
 void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 
 // Dot product
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -278,8 +259,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 //
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
@@ -287,8 +266,6 @@ void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const
 size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t quantize_iq1_s  (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q3_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q4_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
@@ -299,8 +276,8 @@ size_t quantize_q4_1   (const float * src, void * dst, int nrows, int n_per_row,
 size_t quantize_q5_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q5_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
+void iq2xs_init_impl(int grid_size);
+void iq2xs_free_impl(int grid_size);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);
 

ggml-sycl.cpp

@@ -9188,22 +9188,174 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
     }
 }
 
-template <int qk, int qi, typename block_q_t, int vdr,
-          vec_dot_q_sycl_t vec_dot_q_sycl>
-static void mul_mat_vec_q_sycl_submitter(const void *vx, const void *vy,
-                                         float *dst, const int ncols,
-                                         const int nrows,
-                                         dpct::queue_ptr stream) {
-  GGML_ASSERT(ncols % QK4_0 == 0);
-  const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-  const sycl::range<3> block_nums(1, 1, block_num_y);
-  const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-  stream->parallel_for(
-      sycl::nd_range<3>(block_nums * block_dims, block_dims), [=
-  ](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-        mul_mat_vec_q<qk, qi, block_q_t, vdr, vec_dot_q_sycl>(
-            vx, vy, dst, ncols, nrows, item_ct1);
-      });
+static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ,
+                          vec_dot_q4_0_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_1 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ,
+                          vec_dot_q4_1_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK5_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ,
+                          vec_dot_q5_0_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK5_1 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ,
+                          vec_dot_q5_1_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK8_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ,
+                          vec_dot_q8_0_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ,
+                          vec_dot_q2_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ,
+                          vec_dot_q3_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ,
+                          vec_dot_q4_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ,
+                          vec_dot_q5_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ,
+                          vec_dot_q6_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
 }
 
 int get_device_index_by_id(int id){
@@ -11943,63 +12095,37 @@ inline void ggml_sycl_op_mul_mat_vec_q(
     const int64_t ne00 = src0->ne[0];
     const int64_t row_diff = row_high - row_low;
 
-    // TODO: support these quantization types
-    GGML_ASSERT(!(src0->type == GGML_TYPE_IQ2_XXS ||
-                  src0->type == GGML_TYPE_IQ2_XS ||
-                  src0->type == GGML_TYPE_IQ3_XXS ||
-                  src0->type == GGML_TYPE_IQ1_S));
-
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-          mul_mat_vec_q_sycl_submitter<QK4_0, QI4_0, block_q4_0,
-                                       VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q4_1:
-          mul_mat_vec_q_sycl_submitter<QK4_1, QI4_1, block_q4_1,
-                                       VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q5_0:
-          mul_mat_vec_q_sycl_submitter<QK5_0, QI5_0, block_q5_0,
-                                       VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q5_1:
-          mul_mat_vec_q_sycl_submitter<QK5_1, QI5_1, block_q5_1,
-                                       VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q8_0:
-          mul_mat_vec_q_sycl_submitter<QK8_0, QI8_0, block_q8_0,
-                                       VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q2_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI2_K, block_q2_K,
-                                       VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q3_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI3_K, block_q3_K,
-                                       VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q4_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI4_K, block_q4_K,
-                                       VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q5_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI5_K, block_q5_K,
-                                       VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q6_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI6_K, block_q6_K,
-                                       VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -12019,7 +12145,7 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec(
     const int64_t src1_ncols, const int64_t src1_padded_row_size,
     const dpct::queue_ptr &stream) {
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t row_diff = row_high - row_low;
 
@@ -14642,8 +14768,7 @@ GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_bu
 static ggml_backend_buffer_t
 ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                            size_t size) try {
-    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
-    int device = (int) buft_ctx->device;
+    int device = (int) (intptr_t) buft->context;
 
     ggml_sycl_set_device(device);
     int device_index = get_device_index_by_id(device);
@@ -14721,7 +14846,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
         for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
             ggml_backend_sycl_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
-                /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i)},
+                /* .context  = */ (ggml_backend_buffer_type_context_t) (intptr_t) i,
             };
         }
         ggml_backend_sycl_buffer_type_initialized = true;
@@ -14783,6 +14908,10 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
 
 // backend
 
+struct ggml_backend_context_sycl {
+    int device;
+};
+
 static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
     return GGML_SYCL_NAME;
 
@@ -14790,14 +14919,14 @@ static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
 }
 
 static void ggml_backend_sycl_free(ggml_backend_t backend) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     delete sycl_ctx;
     delete backend;
 }
 
 static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     return ggml_backend_sycl_buffer_type(sycl_ctx->device);
 }
@@ -14806,7 +14935,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                ggml_tensor *tensor,
                                                const void *data, size_t offset,
                                                size_t size) try {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
@@ -14824,7 +14953,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                const ggml_tensor *tensor,
                                                void *data, size_t offset,
                                                size_t size) try {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
@@ -14839,7 +14968,7 @@ catch (sycl::exception const &exc) {
 }
 
 static void ggml_backend_sycl_synchronize(ggml_backend_t backend) try {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->wait()));
 
@@ -14875,7 +15004,7 @@ static void ggml_backend_sycl_graph_plan_compute(ggml_backend_t backend, ggml_ba
 }
 
 static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     ggml_sycl_set_main_device(sycl_ctx->device);
 
@@ -14964,12 +15093,6 @@ static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_ten
                     return false;
                 }
 
-                if (a->type == GGML_TYPE_IQ1_S) {
-                    return false;
-                }
-                if (a->type == GGML_TYPE_IQ3_XXS) {
-                  return false;
-                }
                 if (a->type == GGML_TYPE_IQ2_XXS) {
                     return false;
                 }
@@ -15089,9 +15212,8 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     // not strictly necessary, but it may reduce the overhead of the first graph_compute
     ggml_sycl_set_main_device(device);
 
-    ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context {
-        /* .device = */ device,
-        /* .name   = */ GGML_SYCL_NAME + std::to_string(device),
+    ggml_backend_context_sycl * ctx = new ggml_backend_context_sycl {
+        /* .device = */ device
     };
 
     ggml_backend_t sycl_backend = new ggml_backend {

ggml-vulkan.cpp

@@ -1091,9 +1091,6 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 }
 
-static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
-static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
-
 void ggml_vk_instance_init() {
     if (vk_instance_initialized) {
         return;
@@ -1103,42 +1100,28 @@ void ggml_vk_instance_init() {
 #endif
 
     vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION };
+    const std::vector<const char*> layers = {
+#ifdef GGML_VULKAN_VALIDATE
+        "VK_LAYER_KHRONOS_validation",
+#endif
+    };
+    const std::vector<const char*> extensions = {
+#ifdef GGML_VULKAN_VALIDATE
+        "VK_EXT_validation_features",
+#endif
+    };
+    vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags(), &app_info, layers, extensions);
+#ifdef GGML_VULKAN_VALIDATE
+    const std::vector<vk::ValidationFeatureEnableEXT> features_enable = { vk::ValidationFeatureEnableEXT::eBestPractices };
+    vk::ValidationFeaturesEXT validation_features = {
+        features_enable,
+        {},
+    };
+    validation_features.setPNext(nullptr);
+    instance_create_info.setPNext(&validation_features);
 
-    const std::vector<vk::ExtensionProperties> instance_extensions = vk::enumerateInstanceExtensionProperties();
-    const bool validation_ext = ggml_vk_instance_validation_ext_available(instance_extensions);
-    const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions);
-
-    std::vector<const char*> layers;
-
-    if (validation_ext) {
-        layers.push_back("VK_LAYER_KHRONOS_validation");
-    }
-    std::vector<const char*> extensions;
-    if (validation_ext) {
-        extensions.push_back("VK_EXT_validation_features");
-    }
-    if (portability_enumeration_ext) {
-        extensions.push_back("VK_KHR_portability_enumeration");
-    }
-    vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions);
-    if (portability_enumeration_ext) {
-        instance_create_info.flags |= vk::InstanceCreateFlagBits::eEnumeratePortabilityKHR;
-    }
-
-    std::vector<vk::ValidationFeatureEnableEXT> features_enable;
-    vk::ValidationFeaturesEXT validation_features;
-
-    if (validation_ext) {
-        features_enable = { vk::ValidationFeatureEnableEXT::eBestPractices };
-        validation_features = {
-            features_enable,
-            {},
-        };
-        validation_features.setPNext(nullptr);
-        instance_create_info.setPNext(&validation_features);
-
-        std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
-    }
+    std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
+#endif
     vk_instance.instance = vk::createInstance(instance_create_info);
 
     memset(vk_instance.initialized, 0, sizeof(bool) * GGML_VK_MAX_DEVICES);
@@ -1167,7 +1150,7 @@ void ggml_vk_instance_init() {
     vk_instance_initialized = true;
 }
 
-static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
+void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     GGML_ASSERT(idx < vk_instance.device_indices.size());
     size_t dev_num = vk_instance.device_indices[idx];
 #ifdef GGML_VULKAN_DEBUG
@@ -1185,12 +1168,12 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     vk_instance.devices[idx] = std::make_shared<vk_device>();
     ctx->device = vk_instance.devices[idx];
     ctx->device.lock()->physical_device = devices[dev_num];
-    const std::vector<vk::ExtensionProperties> ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties();
+    std::vector<vk::ExtensionProperties> ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties();
 
     bool maintenance4_support = false;
 
     // Check if maintenance4 is supported
-    for (const auto& properties : ext_props) {
+    for (auto properties : ext_props) {
         if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
             maintenance4_support = true;
         }
@@ -1221,7 +1204,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     bool fp16_storage = false;
     bool fp16_compute = false;
 
-    for (const auto& properties : ext_props) {
+    for (auto properties : ext_props) {
         if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
             fp16_storage = true;
         } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
@@ -4573,13 +4556,13 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     }
 }
 
-GGML_CALL static int ggml_vk_get_device_count() {
+GGML_CALL int ggml_vk_get_device_count() {
     ggml_vk_instance_init();
 
     return vk_instance.device_indices.size();
 }
 
-GGML_CALL static void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
+GGML_CALL void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
     ggml_vk_instance_init();
 
     std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
@@ -4597,7 +4580,7 @@ void ggml_vk_init_cpu_assist() {
 
     std::cerr << "ggml_vulkan: Found " << ggml_vk_get_device_count() << " Vulkan devices:" << std::endl;
 
-    for (int i = 0; i < ggml_vk_get_device_count(); i++) {
+    for (size_t i = 0; i < ggml_vk_get_device_count(); i++) {
         ggml_vk_print_gpu_info(i);
     }
     // Initialize the first backend to make sure CPU matrix multiplications can be offloaded.
@@ -5284,7 +5267,7 @@ GGML_CALL void ggml_backend_vk_get_device_description(int device, char * descrip
 }
 
 GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
-    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+    GGML_ASSERT(device < vk_instance.device_indices.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
 
@@ -5318,42 +5301,6 @@ GGML_CALL int ggml_backend_vk_reg_devices() {
     return vk_instance.device_indices.size();
 }
 
-// Extension availability
-static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
-#ifdef GGML_VULKAN_VALIDATE
-    bool portability_enumeration_ext = false;
-    // Check for portability enumeration extension for MoltenVK support
-    for (const auto& properties : instance_extensions) {
-        if (strcmp("VK_KHR_portability_enumeration", properties.extensionName) == 0) {
-            return true;
-        }
-    }
-    if (!portability_enumeration_ext) {
-        std::cerr << "ggml_vulkan: WARNING: Instance extension VK_KHR_portability_enumeration not found." << std::endl;
-    }
-#endif
-    return false;
-
-    UNUSED(instance_extensions);
-}
-static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
-#ifdef __APPLE__
-    bool portability_enumeration_ext = false;
-    // Check for portability enumeration extension for MoltenVK support
-    for (const auto& properties : instance_extensions) {
-        if (strcmp("VK_KHR_portability_enumeration", properties.extensionName) == 0) {
-            return true;
-        }
-    }
-    if (!portability_enumeration_ext) {
-        std::cerr << "ggml_vulkan: WARNING: Instance extension VK_KHR_portability_enumeration not found." << std::endl;
-    }
-#endif
-    return false;
-
-    UNUSED(instance_extensions);
-}
-
 // checks
 
 #ifdef GGML_VULKAN_CHECK_RESULTS

ggml.h

@@ -354,8 +354,6 @@ extern "C" {
         GGML_TYPE_IQ2_XXS = 16,
         GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_IQ3_XXS = 18,
-        GGML_TYPE_IQ1_S   = 19,
-        GGML_TYPE_IQ4_NL  = 20,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -393,8 +391,6 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
-        GGML_FTYPE_MOSTLY_IQ1_S   = 18, // except 1d tensors
-        GGML_FTYPE_MOSTLY_IQ4_NL  = 19, // except 1d tensors
     };
 
     // available tensor operations:
@@ -1387,17 +1383,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
+    // fused soft_max(a*scale + mask)
     // mask is optional
-    // pos is required when max_bias > 0.0f
-    // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * mask,
-            struct ggml_tensor  * pos,
-            float                 scale,
-            float                 max_bias);
+            float                 scale);
 
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
@@ -1499,13 +1491,12 @@ extern "C" {
 
     // alibi position embedding
     // in-place, returns view(a)
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
+    GGML_API struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_head,
-            float                 bias_max),
-        "use ggml_soft_max_ext instead (will be removed in Mar 2024)");
+            float                 bias_max);
 
     // clamp
     // in-place, returns view(a)

gguf-py/gguf/constants.py

@@ -111,7 +111,6 @@ class MODEL_ARCH(IntEnum):
     ORION      = auto()
     INTERNLM2  = auto()
     MINICPM    = auto()
-    GEMMA      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -168,7 +167,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.ORION:          "orion",
     MODEL_ARCH.INTERNLM2:      "internlm2",
     MODEL_ARCH.MINICPM:        "minicpm",
-    MODEL_ARCH.GEMMA:          "gemma",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -513,19 +511,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
-    MODEL_ARCH.GEMMA: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_Q,
-        MODEL_TENSOR.ATTN_K,
-        MODEL_TENSOR.ATTN_V,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_GATE,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_UP,
-        MODEL_TENSOR.FFN_NORM,
-    ],
     # TODO
 }
 

llama.cpp

@@ -208,7 +208,6 @@ enum llm_arch {
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
     LLM_ARCH_MINICPM,
-    LLM_ARCH_GEMMA,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -235,7 +234,6 @@ static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ORION,           "orion"      },
     { LLM_ARCH_INTERNLM2,       "internlm2"  },
     { LLM_ARCH_MINICPM,         "minicpm"    },
-    { LLM_ARCH_GEMMA,           "gemma"      },
 };
 
 enum llm_kv {
@@ -762,22 +760,6 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
         },
     },
-    {
-        LLM_ARCH_GEMMA,
-        {
-            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
-            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-        },
-    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1575,14 +1557,13 @@ struct llama_hparams {
     uint32_t n_yarn_orig_ctx;
     int32_t  rope_scaling_type_train;
 
-    float f_clamp_kqv      = 0.0f;
-    float f_max_alibi_bias = 0.0f;
+    float f_clamp_kqv;
+    float f_max_alibi_bias;
 
     bool causal_attn = true;
-    bool need_kq_pos = false;
-
     uint32_t pooling_type = LLAMA_POOLING_NONE;
 
+
     bool operator!=(const llama_hparams & other) const {
         if (this->vocab_only    != other.vocab_only)    return true;
         if (this->n_vocab       != other.n_vocab)       return true;
@@ -1942,7 +1923,6 @@ struct llama_context {
     struct ggml_tensor * inp_embd;      // F32 [n_embd, n_batch]
     struct ggml_tensor * inp_pos;       // I32 [n_batch]
     struct ggml_tensor * inp_KQ_mask;   // F32 [n_ctx, n_batch]
-    struct ggml_tensor * inp_KQ_pos;    // F32 [n_ctx]
     struct ggml_tensor * inp_K_shift;   // I32 [n_ctx]
     struct ggml_tensor * inp_mean;      // F32 [n_batch, n_batch]
     struct ggml_tensor * inp_cls;       // I32 [n_batch]
@@ -2544,8 +2524,6 @@ struct llama_model_loader {
                 case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
                 case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
                 case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
-                case GGML_TYPE_IQ1_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ1_S;   break;
-                case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2895,8 +2873,6 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
         case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small";
         case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw";
-        case LLAMA_FTYPE_MOSTLY_IQ1_S  :return "IQ1_S - 1.5625 bpw";
-        case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw";
 
         default: return "unknown, may not work";
     }
@@ -3078,11 +3054,6 @@ static void llm_load_hparams(
                     case 40: model.type = e_model::MODEL_13B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
-
-                if (model.type == e_model::MODEL_13B) {
-                    // TODO: become GGUF KV parameter
-                    hparams.f_max_alibi_bias = 8.0f;
-                }
             } break;
         case LLM_ARCH_STARCODER:
             {
@@ -3110,9 +3081,6 @@ static void llm_load_hparams(
                     case 32: model.type = e_model::MODEL_1B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
-
-                // TODO: become GGUF KV parameter
-                hparams.f_max_alibi_bias = 8.0f;
             } break;
         case LLM_ARCH_BERT:
             {
@@ -3158,12 +3126,11 @@ static void llm_load_hparams(
                             case 4096: model.type = e_model::MODEL_7B; break;
                         } break;
                 }
-
-                // TODO: become GGUF KV parameter
-                hparams.f_max_alibi_bias = 8.0f;
             } break;
         case LLM_ARCH_MPT:
             {
+                hparams.f_clamp_kqv = 0.0f;
+
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,      hparams.f_clamp_kqv, false);
                 ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias);
@@ -3261,24 +3228,10 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_GEMMA:
-            {
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-
-                switch (hparams.n_layer) {
-                    case 18: model.type = e_model::MODEL_2B; break;
-                    case 28: model.type = e_model::MODEL_7B; break;
-                    default: model.type = e_model::MODEL_UNKNOWN;
-               }
-            } break;
         default: (void)0;
     }
 
     model.ftype = ml.ftype;
-
-    if (hparams.f_max_alibi_bias > 0.0f) {
-        hparams.need_kq_pos = true;
-    }
 }
 
 // TODO: This should probably be in llama.h
@@ -4388,37 +4341,6 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
-            case LLM_ARCH_GEMMA:
-                {
-                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-
-                    // output
-                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-
-                    const int64_t n_ff          = hparams.n_ff;
-                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
-                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
-                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
-
-                    for (uint32_t i = 0; i < n_layer; ++i) {
-                        ggml_context * ctx_layer = ctx_for_layer(i);
-                        ggml_context * ctx_split = ctx_for_layer_split(i);
-
-                        auto & layer = model.layers[i];
-
-                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-
-                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
-                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
-                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
-                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
-
-                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
-                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-                    }
-                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -4852,10 +4774,10 @@ static struct ggml_tensor * llm_build_kqv(
          struct ggml_tensor * wo_b,
          struct ggml_tensor * q_cur,
          struct ggml_tensor * kq_mask,
-         struct ggml_tensor * kq_pos,
                     int64_t   n_ctx,
                     int32_t   n_tokens,
                     int32_t   n_kv,
+                    float     max_alibi_bias,
                     float     kq_scale,
          const llm_build_cb & cb,
                     int       il) {
@@ -4885,26 +4807,26 @@ static struct ggml_tensor * llm_build_kqv(
         ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
     }
 
-#if defined(GGML_USE_VULKAN) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_SYCL)
-#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Vulkan, Kompute, and SYCL")
-#pragma message("      Falling back to ggml_alibi(). Will become an error in Mar 2024")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5488")
-    if (hparams.f_max_alibi_bias > 0.0f) {
+    if (max_alibi_bias > 0.0f) {
+        // temporary branch until we figure out how to handle ggml_alibi through ggml_add
         kq = ggml_scale(ctx, kq, kq_scale);
         cb(kq, "kq_scaled", il);
 
-        kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, hparams.f_max_alibi_bias);
-        cb(kq, "kq_scaled_alibi", il);
+        if (max_alibi_bias > 0.0f) {
+            // TODO: n_head or n_head_kv
+            // TODO: K-shift is likely not working
+            // TODO: change to ggml_add
+            kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias);
+            cb(kq, "kq_scaled_alibi", il);
+        }
 
         kq = ggml_add(ctx, kq, kq_mask);
         cb(kq, "kq_masked", il);
 
         kq = ggml_soft_max(ctx, kq);
         cb(kq, "kq_soft_max", il);
-    } else
-#endif
-    {
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_pos, kq_scale, hparams.f_max_alibi_bias);
+    } else {
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale);
         cb(kq, "kq_soft_max_ext", il);
     }
 
@@ -4952,11 +4874,11 @@ static struct ggml_tensor * llm_build_kv(
          struct ggml_tensor * v_cur,
          struct ggml_tensor * q_cur,
          struct ggml_tensor * kq_mask,
-         struct ggml_tensor * kq_pos,
                     int64_t   n_ctx,
                     int32_t   n_tokens,
                     int32_t   kv_head,
                     int32_t   n_kv,
+                    float     max_alibi_bias,
                     float     kq_scale,
          const llm_build_cb & cb,
                     int       il) {
@@ -4970,8 +4892,9 @@ static struct ggml_tensor * llm_build_kv(
     llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il);
 
     struct ggml_tensor * cur;
-    cur  = llm_build_kqv(ctx, model, hparams, kv, graph, wo, wo_b,
-            q_cur, kq_mask, kq_pos, n_ctx, n_tokens, n_kv, kq_scale, cb, il);
+    cur  = llm_build_kqv(ctx, model, hparams, kv, graph,
+            wo, wo_b,
+            q_cur, kq_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, kq_scale, cb, il);
     cb(cur, "kqv_out", il);
 
     return cur;
@@ -5139,7 +5062,7 @@ struct llm_build_context {
                 }
 
                 Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
                     hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
@@ -5154,7 +5077,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5284,10 +5207,6 @@ struct llm_build_context {
         struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = ggml_view_1d(ctx0, lctx.inp_KQ_pos, n_kv, 0);
-        cb(KQ_pos, "KQ_pos", -1);
-
         // shift the entire K-cache if needed
         if (do_rope_shift) {
             llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
@@ -5336,9 +5255,12 @@ struct llm_build_context {
                 cb(Kcur, "Kcur", il);
 
 
+                // apply ALiBi for 13B model
+                const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f;
+
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5462,7 +5384,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5561,7 +5483,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5766,7 +5688,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Q, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Q, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5828,10 +5750,6 @@ struct llm_build_context {
         struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = ggml_view_1d(ctx0, lctx.inp_KQ_pos, n_kv, 0);
-        cb(KQ_pos, "KQ_pos", -1);
-
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
 
@@ -5859,7 +5777,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5960,7 +5878,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             } else {
                 // compute Q and K and RoPE them
@@ -5991,7 +5909,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6067,10 +5985,6 @@ struct llm_build_context {
         struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = ggml_view_1d(ctx0, lctx.inp_KQ_pos, n_kv, 0);
-        cb(KQ_pos, "KQ_pos", -1);
-
         inpL = llm_build_norm(ctx0, inpL, hparams,
                 model.tok_norm,
                 model.tok_norm_b,
@@ -6104,7 +6018,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6164,10 +6078,6 @@ struct llm_build_context {
         struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = ggml_view_1d(ctx0, lctx.inp_KQ_pos, n_kv, 0);
-        cb(KQ_pos, "KQ_pos", -1);
-
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * attn_norm;
 
@@ -6201,7 +6111,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6323,7 +6233,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6438,7 +6348,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6559,7 +6469,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6686,7 +6596,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f, cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6789,7 +6699,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
             struct ggml_tensor * sa_out = cur;
@@ -6888,7 +6798,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -6997,7 +6907,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -7115,7 +7025,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -7234,7 +7144,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -7366,7 +7276,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -7425,113 +7335,6 @@ struct llm_build_context {
 
         return gf;
     }
-
-    struct ggml_cgraph * build_gemma() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
-
-        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
-        cb(inpL, "inp_embd", -1);
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
-        cb(inp_pos, "inp_pos", -1);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
-        cb(KQ_mask, "KQ_mask", -1);
-
-        // shift the entire K-cache if needed
-        if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
-        }
-
-        for (int il = 0; il < n_layer; ++il) {
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos,
-                        n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Qcur, "Qcur", il);
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
-                cb(Qcur, "Qcur_scaled", il);
-
-                Kcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos,
-                        n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-                cb(cur, "kqv_out", il);
-            }
-            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = llm_build_norm(ctx0, sa_out, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_ffn(ctx0, cur,
-                        model.layers[il].ffn_up, NULL,
-                        model.layers[il].ffn_gate, NULL,
-                        model.layers[il].ffn_down, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, sa_out);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
 };
 
 static struct ggml_cgraph * llama_build_graph(
@@ -7640,10 +7443,6 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_minicpm();
             } break;
-        case LLM_ARCH_GEMMA:
-            {
-                result = llm.build_gemma();
-            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -7708,18 +7507,6 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
         }
     }
 
-    if (hparams.need_kq_pos) {
-        const int64_t n_kv = kv_self.n;
-
-        assert(ggml_backend_buffer_is_host(lctx.inp_KQ_pos->buffer));
-
-        float * data = (float *) lctx.inp_KQ_pos->data;
-
-        for (int i = 0; i < n_kv; ++i) {
-            data[i] = float(lctx.kv_self.cells[i].pos);
-        }
-    }
-
     if (kv_self.has_shift) {
         const int64_t n_ctx = cparams.n_ctx;
 
@@ -10486,20 +10273,20 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
             new_type = GGML_TYPE_Q8_0;
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
             new_type = GGML_TYPE_Q5_K;
         }
         else if (new_type != GGML_TYPE_Q8_0) {
             new_type = GGML_TYPE_Q6_K;
         }
     } else if (name == "token_embd.weight") {
-        if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
             new_type = GGML_TYPE_Q2_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
             new_type = GGML_TYPE_Q4_K;
         }
-    } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+    } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
         if (name.find("attn_v.weight") != std::string::npos) {
             if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
             else new_type = GGML_TYPE_Q2_K;
@@ -10509,9 +10296,6 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K;
             ++qs.i_ffn_down;
         }
-        else if (name.find("attn_output.weight") != std::string::npos) {
-            if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) new_type = GGML_TYPE_IQ2_XXS;
-        }
     } else if (name.find("attn_v.weight") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
             new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
@@ -10526,9 +10310,6 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL && qs.model.hparams.n_gqa() >= 4) {
-            new_type = GGML_TYPE_Q5_K;
-        }
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
@@ -10581,9 +10362,6 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
                 if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
             }
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL && !qs.has_imatrix) {
-            if (i_layer < n_layer/8) new_type = GGML_TYPE_Q5_K;
-        }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
             new_type = GGML_TYPE_Q5_K;
@@ -10600,7 +10378,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         if (arch != LLM_ARCH_FALCON) {
             if (qs.model.hparams.n_expert == 8) {
                 if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
-                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL  ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
                     new_type = GGML_TYPE_Q5_K;
                 }
@@ -10651,7 +10429,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
     if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
         new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K ||
         new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS ||
-        new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+        new_type == GGML_TYPE_IQ3_XXS) {
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
@@ -10666,9 +10444,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             case GGML_TYPE_IQ2_XXS:
             case GGML_TYPE_IQ2_XS:
             case GGML_TYPE_IQ3_XXS:
-            case GGML_TYPE_IQ1_S:
-            case GGML_TYPE_Q2_K:
-            case GGML_TYPE_Q3_K: new_type = GGML_TYPE_IQ4_NL; break;
+            case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break;
+            case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break;
             case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
             case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q5_1; break;
             case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break;
@@ -10709,8 +10486,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS: quantized_type = GGML_TYPE_IQ2_XXS; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XS:  quantized_type = GGML_TYPE_IQ2_XS;  break;
         case LLAMA_FTYPE_MOSTLY_IQ3_XXS: quantized_type = GGML_TYPE_IQ3_XXS; break;
-        case LLAMA_FTYPE_MOSTLY_IQ1_S:   quantized_type = GGML_TYPE_IQ1_S;   break;
-        case LLAMA_FTYPE_MOSTLY_IQ4_NL:  quantized_type = GGML_TYPE_IQ4_NL;  break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -10884,7 +10659,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             }
             if ((new_type == GGML_TYPE_IQ2_XXS ||
                  new_type == GGML_TYPE_IQ2_XS  ||
-                 new_type == GGML_TYPE_IQ1_S   ||
                 (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
                 LLAMA_LOG_ERROR("\n\n============================================================\n");
                 LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
@@ -11660,7 +11434,7 @@ struct llama_context * llama_new_context_with_model(
         // graph inputs
         {
             ggml_init_params init_params = {
-                /* .mem_size   */ ggml_tensor_overhead()*8,
+                /* .mem_size   */ ggml_tensor_overhead()*7,
                 /* .mem_buffer */ nullptr,
                 /* .no_alloc   */ true,
             };
@@ -11670,7 +11444,6 @@ struct llama_context * llama_new_context_with_model(
             ctx->inp_embd    = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, hparams.n_embd, cparams.n_batch);
             ctx->inp_pos     = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch);
             ctx->inp_KQ_mask = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_ctx, cparams.n_batch);
-            ctx->inp_KQ_pos  = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_ctx);
             ctx->inp_K_shift = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_ctx);
             ctx->inp_mean    = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_batch, cparams.n_batch);
             ctx->inp_cls     = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch);
@@ -11679,7 +11452,6 @@ struct llama_context * llama_new_context_with_model(
             ggml_set_name(ctx->inp_embd,    "inp_embd");
             ggml_set_name(ctx->inp_pos,     "inp_pos");
             ggml_set_name(ctx->inp_KQ_mask, "inp_KQ_mask");
-            ggml_set_name(ctx->inp_KQ_pos,  "inp_KQ_pos");
             ggml_set_name(ctx->inp_K_shift, "inp_K_shift");
             ggml_set_name(ctx->inp_mean,    "inp_mean");
             ggml_set_name(ctx->inp_cls,     "inp_cls");
@@ -12687,123 +12459,6 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
     return 0;
 }
 
-// trim whitespace from the beginning and end of a string
-static std::string trim(const std::string & str) {
-    size_t start = 0;
-    size_t end = str.size();
-    while (start < end && isspace(str[start])) {
-        start += 1;
-    }
-    while (end > start && isspace(str[end - 1])) {
-        end -= 1;
-    }
-    return str.substr(start, end - start);
-}
-
-// Simple version of "llama_apply_chat_template" that only works with strings
-// This function uses heuristic checks to determine commonly used template. It is not a jinja parser.
-static int32_t llama_chat_apply_template_internal(
-    const std::string & tmpl,
-    const std::vector<const llama_chat_message *> & chat,
-    std::string & dest, bool add_ass) {
-    // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
-    std::stringstream ss;
-    if (tmpl.find("<|im_start|>") != std::string::npos) {
-        // chatml template
-        for (auto message : chat) {
-            ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";
-        }
-        if (add_ass) {
-            ss << "<|im_start|>assistant\n";
-        }
-    } else if (tmpl.find("[INST]") != std::string::npos) {
-        // llama2 template and its variants
-        // [variant] support system message
-        bool support_system_message = tmpl.find("<<SYS>>") != std::string::npos;
-        // [variant] space before + after response
-        bool space_around_response = tmpl.find("' ' + eos_token") != std::string::npos;
-        // [variant] add BOS inside history
-        bool add_bos_inside_history = tmpl.find("bos_token + '[INST]") != std::string::npos;
-        // [variant] trim spaces from the input message
-        bool strip_message = tmpl.find("content.strip()") != std::string::npos;
-        // construct the prompt
-        bool is_inside_turn = true; // skip BOS at the beginning
-        ss << "[INST] ";
-        for (auto message : chat) {
-            std::string content = strip_message ? trim(message->content) : message->content;
-            std::string role(message->role);
-            if (!is_inside_turn) {
-                is_inside_turn = true;
-                ss << (add_bos_inside_history ? "<s>[INST] " : "[INST] ");
-            }
-            if (role == "system") {
-                if (support_system_message) {
-                    ss << "<<SYS>>\n" << content << "\n<</SYS>>\n\n";
-                } else {
-                    // if the model does not support system message, we still include it in the first message, but without <<SYS>>
-                    ss << content << "\n";
-                }
-            } else if (role == "user") {
-                ss << content << " [/INST]";
-            } else {
-                ss << (space_around_response ? " " : "") << content << (space_around_response ? " " : "") << "</s>";
-                is_inside_turn = false;
-            }
-        }
-        // llama2 templates seem to not care about "add_generation_prompt"
-    } else if (tmpl.find("<|user|>") != std::string::npos) {
-        // zephyr template
-        for (auto message : chat) {
-            ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";
-        }
-        if (add_ass) {
-            ss << "<|assistant|>\n";
-        }
-    } else {
-        // template not supported
-        return -1;
-    }
-    dest = ss.str();
-    return dest.size();
-}
-
-LLAMA_API int32_t llama_chat_apply_template(
-                const struct llama_model * model,
-                              const char * tmpl,
-         const struct llama_chat_message * chat,
-                                  size_t   n_msg,
-                                    bool   add_ass,
-                                    char * buf,
-                                 int32_t   length) {
-    std::string curr_tmpl(tmpl == nullptr ? "" : tmpl);
-    if (tmpl == nullptr) {
-        GGML_ASSERT(model != nullptr);
-        // load template from model
-        std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
-        std::string template_key = "tokenizer.chat_template";
-        int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
-        if (res < 0) {
-            // worst case: there is no information about template, we will use chatml by default
-            curr_tmpl = "<|im_start|>"; // see llama_chat_apply_template_internal
-        } else {
-            curr_tmpl = std::string(model_template.data(), model_template.size());
-        }
-    }
-    // format the chat to string
-    std::vector<const llama_chat_message *> chat_vec;
-    chat_vec.resize(n_msg);
-    for (size_t i = 0; i < n_msg; i++) {
-        chat_vec[i] = &chat[i];
-    }
-    std::string formatted_chat;
-    int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass);
-    if (res < 0) {
-        return res;
-    }
-    strncpy(buf, formatted_chat.c_str(), length);
-    return res;
-}
-
 struct llama_timings llama_get_timings(struct llama_context * ctx) {
     struct llama_timings result = {
         /*.t_start_ms  =*/ 1e-3 * ctx->t_start_us,

llama.h

@@ -100,8 +100,6 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q3_K_XS       = 22, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ3_XXS       = 23, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ1_S         = 24, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ4_NL        = 25, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -306,12 +304,6 @@ extern "C" {
         int32_t n_eval;
     };
 
-    // used in chat template
-    typedef struct llama_chat_message {
-        const char * role;
-        const char * content;
-    } llama_chat_message;
-
     // Helpers for getting default parameters
     LLAMA_API struct llama_model_params llama_model_default_params(void);
     LLAMA_API struct llama_context_params llama_context_default_params(void);
@@ -706,25 +698,6 @@ extern "C" {
                                   char * buf,
                                int32_t   length);
 
-    /// Apply chat template. Inspired by hf apply_chat_template() on python.
-    /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function only support some known jinja templates. It is not a jinja parser.
-    /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
-    /// @param chat Pointer to a list of multiple llama_chat_message
-    /// @param n_msg Number of llama_chat_message in this chat
-    /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
-    /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
-    /// @param length The size of the allocated buffer
-    /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
-    LLAMA_API int32_t llama_chat_apply_template(
-              const struct llama_model * model,
-                            const char * tmpl,
-       const struct llama_chat_message * chat,
-                                size_t   n_msg,
-                                  bool   add_ass,
-                                  char * buf,
-                               int32_t   length);
-
     //
     // Grammar
     //

scripts/compare-commits.sh

@@ -1,37 +0,0 @@
-#!/bin/bash
-
-if [ $# -lt 2 ]; then
-    echo "usage: ./scripts/compare-commits.sh <commit1> <commit2> [additional llama-bench arguments]"
-    exit 1
-fi
-
-set -e
-set -x
-
-bench_args="${@:3}"
-
-rm -f llama-bench.sqlite
-
-backend="cpu"
-
-if [[ "$OSTYPE" == "darwin"* ]]; then
-    backend="metal"
-elif command -v nvcc &> /dev/null; then
-    backend="cuda"
-fi
-
-make_opts=""
-
-if [[ "$backend" == "cuda" ]]; then
-    make_opts="LLAMA_CUBLAS=1"
-fi
-
-git checkout $1
-make clean && make -j32 $make_opts llama-bench
-./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
-
-git checkout $2
-make clean && make -j32 $make_opts llama-bench
-./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
-
-./scripts/compare-llama-bench.py -b $1 -c $2

scripts/get-flags.mk

@@ -1,6 +1,6 @@
 ifeq '' '$(findstring clang,$(shell $(GF_CC) --version))'
 	GF_CC_IS_GCC = 1
-	GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null; echo; $(GF_CC) -dumpversion; } | awk -F. '/./ { printf("%02d%02d%02d", $$1, $$2, $$3); exit }')
+	GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null || $(GF_CC) -dumpversion; } | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
 else
 	GF_CC_IS_CLANG = 1
 	ifeq '' '$(findstring Apple,$(shell $(GF_CC) --version))'

scripts/get-wikitext-2.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-wget https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
 
 echo "Usage:"
 echo ""

scripts/sync-ggml.last

@@ -1 +1 @@
-4712fd12d7acb9971f850b1b98588f934cb39444
+5070f078a67c18c11736e78316ab715ca9afde16

tests/CMakeLists.txt

@@ -28,7 +28,6 @@ endfunction()
 llama_build_and_test_executable(test-quantize-fns.cpp)
 llama_build_and_test_executable(test-quantize-perf.cpp)
 llama_build_and_test_executable(test-sampling.cpp)
-llama_build_and_test_executable(test-chat-template.cpp)
 
 llama_build_executable(test-tokenizer-0-llama.cpp)
 llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)

tests/test-backend-ops.cpp

@@ -1085,32 +1085,24 @@ struct test_diag_mask_inf : public test_case {
 struct test_soft_max : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
-    const bool mask;
     const float scale;
-    const float max_bias;
+    const bool mask;
 
     std::string vars() override {
-        return VARS_TO_STR5(type, ne, mask, scale, max_bias);
+        return VARS_TO_STR4(type, ne, scale, mask);
     }
 
     test_soft_max(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 10},
-            bool mask = false,
             float scale = 1.0f,
-            float max_bias = 0.0f)
-        : type(type), ne(ne), mask(mask), scale(scale), max_bias(max_bias) {}
+            bool mask = false)
+        : type(type), ne(ne), scale(scale), mask(mask) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * mask = nullptr;
-        if (this->mask) {
-            mask = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]);
-        }
-        ggml_tensor * pos = nullptr;
-        if (max_bias > 0.0f) {
-            pos = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ne[0]);
-        }
-        ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, pos, scale, max_bias);
+        ggml_tensor * b = nullptr;
+        if (mask) { b = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]); }
+        ggml_tensor * out = ggml_soft_max_ext(ctx, a, b, scale);
         return out;
     }
 };
@@ -1155,6 +1147,30 @@ struct test_rope : public test_case {
     }
 };
 
+// GGML_OP_ALIBI
+struct test_alibi : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    int n_past;
+    int n_head;
+    float bias_max;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, ne, n_past, n_head, bias_max);
+    }
+
+    test_alibi(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            int n_past = 512, int n_head = 10, float bias_max = 0.5f)
+        : type(type), ne(ne), n_past(n_past), n_head(n_head), bias_max(bias_max) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_alibi(ctx, a, n_past, n_head, bias_max);
+        return out;
+    }
+};
+
 // GGML_OP_POOL2D
 struct test_pool2d : public test_case {
     enum ggml_op_pool pool_type;
@@ -1472,7 +1488,7 @@ struct test_moe : public test_case {
         ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
 
         ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur);
-        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, nullptr, 1.0f/sqrtf(n_embd), 0.0f);
+        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, 1.0f/sqrtf(n_embd));
 
         // select experts
         ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok);
@@ -1601,6 +1617,7 @@ public:
         ggml_cpy(ctx, v_cur_t, v_cache_view);
     }
 
+    // if max_alibi_bias > 0 then apply ALiBi
     struct ggml_tensor * llm_build_kqv(
             struct ggml_context * ctx,
              struct ggml_tensor * k_l,
@@ -1619,7 +1636,7 @@ public:
 
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
 
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, nullptr, kq_scale, 0.0f);
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale);
 
         // split cached v into n_head heads
         struct ggml_tensor * v =
@@ -1917,8 +1934,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
         GGML_TYPE_Q6_K,
         GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
-        GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S,
-        GGML_TYPE_IQ4_NL,
+        GGML_TYPE_IQ3_XXS,
     };
 
     // unary ops
@@ -2067,7 +2083,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10,  1}, 5));
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5));
 
-#if 0
     std::uniform_int_distribution<> dist_ne1(1, 50);
     int exponent = 1;
     while (exponent < (1 << 17)) {
@@ -2076,29 +2091,14 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         for (int n = 0; n < 10; ++n) {
             int64_t ne0 = dist_ne0(rng);
             int64_t ne1 = dist_ne1(rng);
-            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, n/2 == 0, 0.1f, ne0 < 1000 ? 4.0f : 0.0f));
+            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}));
         }
 
         exponent <<= 1;
     }
-#endif
-    for (bool mask : {false, true}) {
-        for (float max_bias : {0.0f, 8.0f}) {
-            for (float scale : {1.0f, 0.1f}) {
-                for (int64_t ne0 : {16, 1024}) {
-                    for (int64_t ne1 : {16, 1024}) {
-                        test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, scale, max_bias));
-                        test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, scale, max_bias));
-                    }
-                }
-            }
-        }
-    }
 
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 8.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 8.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, 0.1f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, 0.1f, true));
 
     for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
         test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512)); // llama 7B
@@ -2113,6 +2113,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  32, 2, 512)); // neox (phi-2)
     }
 
+    test_cases.emplace_back(new test_alibi());
     test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
     test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
 

tests/test-chat-template.cpp

@@ -1,64 +0,0 @@
-#include <iostream>
-#include <string>
-#include <vector>
-#include <sstream>
-
-#undef NDEBUG
-#include <cassert>
-
-#include "llama.h"
-
-int main(void) {
-    llama_chat_message conversation[] = {
-        {"system", "You are a helpful assistant"},
-        {"user", "Hello"},
-        {"assistant", "Hi there"},
-        {"user", "Who are you"},
-        {"assistant", "   I am an assistant   "},
-        {"user", "Another question"},
-    };
-    size_t message_count = 6;
-    std::vector<std::string> templates = {
-        // teknium/OpenHermes-2.5-Mistral-7B
-        "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
-        // mistralai/Mistral-7B-Instruct-v0.2
-        "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
-        // TheBloke/FusionNet_34Bx2_MoE-AWQ
-        "{%- for idx in range(0, messages|length) -%}\\n{%- if messages[idx]['role'] == 'user' -%}\\n{%- if idx > 1 -%}\\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\\n{%- else -%}\\n{{- messages[idx]['content'] + ' [/INST]' -}}\\n{%- endif -%}\\n{% elif messages[idx]['role'] == 'system' %}\\n{{- '[INST] <<SYS>>\\\\n' + messages[idx]['content'] + '\\\\n<</SYS>>\\\\n\\\\n' -}}\\n{%- elif messages[idx]['role'] == 'assistant' -%}\\n{{- ' '  + messages[idx]['content'] + ' ' + eos_token -}}\\n{% endif %}\\n{% endfor %}",
-        // bofenghuang/vigogne-2-70b-chat
-        "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\\\n' + system_message + '\\\\n<</SYS>>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>\\\\n' + content.strip() + '\\\\n<</SYS>>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' '  + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}",
-    };
-    std::vector<std::string> expected_substr = {
-        "<|im_start|>assistant\n   I am an assistant   <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant",
-        "[/INST]Hi there</s>[INST] Who are you [/INST]   I am an assistant   </s>[INST] Another question [/INST]",
-        "</s><s>[INST] Who are you [/INST]    I am an assistant    </s><s>[INST] Another question [/INST]",
-        "[/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
-    };
-    std::vector<char> formatted_chat(1024);
-    int32_t res;
-
-    // test invalid chat template
-    res = llama_chat_apply_template(nullptr, "INVALID TEMPLATE", conversation, message_count, true, formatted_chat.data(), formatted_chat.size());
-    assert(res < 0);
-
-    for (size_t i = 0; i < templates.size(); i++) {
-        std::string custom_template = templates[i];
-        std::string substr = expected_substr[i];
-        formatted_chat.resize(1024);
-        res = llama_chat_apply_template(
-            nullptr,
-            custom_template.c_str(),
-            conversation,
-            message_count,
-            true,
-            formatted_chat.data(),
-            formatted_chat.size()
-        );
-        formatted_chat.resize(res);
-        std::string output(formatted_chat.data(), formatted_chat.size());
-        std::cout << output << "\n-------------------------\n";
-        // expect the "formatted_chat" to contain pre-defined strings
-        assert(output.find(substr) != std::string::npos);
-    }
-    return 0;
-}

tests/test-grammar-parser.cpp

@@ -38,8 +38,8 @@ term  ::= [0-9]+)""";
         // pretty print error message before asserting
         if (expected_pair.first != key || expected_pair.second != value)
         {
-            fprintf(stderr, "expected_pair: %s, %u\n", expected_pair.first.c_str(), expected_pair.second);
-            fprintf(stderr, "actual_pair: %s, %u\n", key.c_str(), value);
+            fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
+            fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
             fprintf(stderr, "expected_pair != actual_pair\n");
         }
 
@@ -96,9 +96,9 @@ term  ::= [0-9]+)""";
             // pretty print error message before asserting
             if (expected_element.type != element.type || expected_element.value != element.value)
             {
-                fprintf(stderr, "index: %u\n", index);
-                fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %u\n", element.type, element.value);
+                fprintf(stderr, "index: %d\n", index);
+                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
                 fprintf(stderr, "expected_element != actual_element\n");
             }
 
@@ -144,8 +144,8 @@ term  ::= [0-9]+)""";
         // pretty print error message before asserting
         if (expected_pair.first != key || expected_pair.second != value)
         {
-            fprintf(stderr, "expected_pair: %s, %u\n", expected_pair.first.c_str(), expected_pair.second);
-            fprintf(stderr, "actual_pair: %s, %u\n", key.c_str(), value);
+            fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
+            fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
             fprintf(stderr, "expected_pair != actual_pair\n");
         }
 
@@ -235,9 +235,9 @@ term  ::= [0-9]+)""";
             // pretty print error message before asserting
             if (expected_element.type != element.type || expected_element.value != element.value)
             {
-                fprintf(stderr, "index: %u\n", index);
-                fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %u\n", element.type, element.value);
+                fprintf(stderr, "index: %d\n", index);
+                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
                 fprintf(stderr, "expected_element != actual_element\n");
             }
 

tests/test-llama-grammar.cpp

@@ -180,8 +180,8 @@ int main()
             if (expected_element.type != element->type || expected_element.value != element->value)
             {
                 fprintf(stderr, "index: %d\n", index);
-                fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %u\n", element->type, element->value);
+                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %d\n", element->type, element->value);
                 fprintf(stderr, "expected_element != actual_element\n");
             }