common : fix ub (#5530)

This commit updates the surgery script to not remove the tensors from the
model file. For this to work the `--skip-unknown` flag is added as an
argument to the convert.py script in README.md.

The motivation for this change is that the surgery script currently
removes the projector tensors from the model file. If the model was
checked out from a repository, the model file will have been updated
and have to be checked out again to reset this effect. If this can be
avoided I think it would be preferable.

I did not perform this change for BakLLaVA models as I am not sure
how that part works.

.flake8

@@ -1,2 +1,3 @@
 [flake8]
 max-line-length = 125
+ignore = W503

.github/workflows/build.yml

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

.github/workflows/python-lint.yml

@@ -16,5 +16,5 @@ jobs:
       - name: flake8 Lint
         uses: py-actions/flake8@v2
         with:
-            ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704"
+            ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503"
             exclude: "examples/*,examples/*/**,*/**/__init__.py"

.gitignore

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

CMakeLists.txt

@@ -55,6 +55,9 @@ 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)
@@ -112,17 +115,14 @@ option(LLAMA_MPI                             "llama: use MPI"
 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(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -130,6 +130,7 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
+
 if (LLAMA_SYCL)
     set(CMAKE_CXX_STANDARD 17)
 else()
@@ -140,9 +141,18 @@ 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)
 
+if (LLAMA_FATAL_WARNINGS)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+        add_compile_options(-Werror)
+    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/WX)
+    endif()
+endif()
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -151,17 +161,17 @@ 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()
 
@@ -298,14 +308,17 @@ 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_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
+        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"
@@ -330,9 +343,6 @@ 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()
@@ -387,15 +397,20 @@ 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 ggml-mpi.h)
+        set(GGML_SOURCES_MPI ggml-mpi.c)
+
         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)
@@ -427,31 +442,28 @@ if (LLAMA_VULKAN)
     if (Vulkan_FOUND)
         message(STATUS "Vulkan found")
 
-        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)
+        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
+        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
         if (LLAMA_VULKAN_CHECK_RESULTS)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
+            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
         endif()
 
         if (LLAMA_VULKAN_DEBUG)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
+            add_compile_definitions(GGML_VULKAN_DEBUG)
         endif()
 
         if (LLAMA_VULKAN_VALIDATE)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
+            add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()
 
         if (LLAMA_VULKAN_RUN_TESTS)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
+            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
         endif()
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -463,43 +475,45 @@ 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)
-    find_package(hipblas)
-    find_package(rocblas)
+    find_package(hip     REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
 
-    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)
+    message(STATUS "HIP and hipBLAS found")
 
-        if (LLAMA_STATIC)
-            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-        endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
-    else()
-        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    set(GGML_HEADERS_ROCM ggml-cuda.h)
+    set(GGML_SOURCES_ROCM ggml-cuda.cu)
+
+    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+
+    if (LLAMA_HIP_UMA)
+        add_compile_definitions(GGML_HIP_UMA)
     endif()
+
+    if (LLAMA_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (LLAMA_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    add_compile_definitions(GGML_CUDA_DMMV_X=${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)
@@ -509,10 +523,14 @@ if (LLAMA_SYCL)
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
+
+    message(STATUS "SYCL found")
+
+    add_compile_definitions(GML_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/${SYCL_INCLUDE_DIR})
@@ -521,7 +539,7 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
 
-    set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
+    set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
@@ -540,61 +558,61 @@ 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")
@@ -604,66 +622,66 @@ 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
@@ -676,8 +694,10 @@ 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()
@@ -685,6 +705,18 @@ 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 "")
@@ -821,6 +853,7 @@ execute_process(
     ERROR_VARIABLE output
     OUTPUT_QUIET
 )
+
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
 endif()
@@ -830,10 +863,10 @@ 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)
@@ -855,11 +888,21 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
      CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
     message(STATUS "ARM detected")
     if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
-        add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
@@ -1017,11 +1060,6 @@ 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
@@ -1038,16 +1076,17 @@ 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>)
@@ -1063,7 +1102,8 @@ 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}
@@ -1114,7 +1154,7 @@ 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

@@ -215,6 +215,35 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
+ifeq ($(LLAMA_FATAL_WARNINGS),1)
+	MK_CFLAGS += -Werror
+	MK_CXXFLAGS += -Werror
+endif
+
+ifeq ($(CC_IS_CLANG), 1)
+	# clang options
+	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
+	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+
+	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+else
+	# gcc options
+	MK_CFLAGS        += -Wdouble-promotion
+	MK_HOST_CXXFLAGS += -Wno-array-bounds
+
+	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
+		MK_HOST_CXXFLAGS += -Wno-format-truncation
+	endif
+	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
+		MK_HOST_CXXFLAGS += -Wextra-semi
+	endif
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -569,6 +598,14 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
 endif # LLAMA_CUBLAS
 $(info )
 

Package.swift

@@ -13,17 +13,31 @@ let package = Package(
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
-    dependencies: [
-        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
-    ],
     targets: [
         .target(
             name: "llama",
-            dependencies: ["ggml"],
             path: ".",
-            exclude: ["ggml-metal.metal"],
+            exclude: [
+               "cmake",
+               "examples",
+               "scripts",
+               "models",
+               "tests",
+               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
+               "Makefile"
+            ],
             sources: [
+                "ggml.c",
                 "llama.cpp",
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "ggml-quants.c",
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
             ],
             publicHeadersPath: "spm-headers",
             cSettings: [

README.md

@@ -958,7 +958,7 @@ We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 

ci/run.sh

@@ -33,7 +33,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA=""
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
@@ -568,6 +568,54 @@ function gg_sum_open_llama_7b_v2 {
     #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
+# bge-small
+
+function gg_run_embd_bge_small {
+    cd ${SRC}
+
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+
+    gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/bge-small"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+
+    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+
+    set +e
+}
+
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -591,6 +639,8 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    test $ret -eq 0 && gg_run embd_bge_small
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

common/common.cpp

@@ -340,13 +340,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.samplers_sequence = parse_samplers_input(argv[i]);
+            const auto sampler_names = string_split(argv[i], ';');
+            sparams.samplers_sequence = sampler_types_from_names(sampler_names, true);
         } else if (arg == "--sampling-seq") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.samplers_sequence = argv[i];
+            sparams.samplers_sequence = sampler_types_from_chars(argv[i]);
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -670,7 +671,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
-            params.numa = true;
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+            else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+            else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+            else { invalid_param = true; break; }
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "--no-display-prompt") {
@@ -906,6 +915,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     const llama_sampling_params & sparams = params.sparams;
 
+    std::string sampler_type_chars;
+    std::string sampler_type_names;
+    for (const auto sampler_type : sparams.samplers_sequence) {
+        sampler_type_chars += static_cast<char>(sampler_type);
+        sampler_type_names += sampler_type_to_name_string(sampler_type) + ";";
+    }
+    sampler_type_names.pop_back();
+
     printf("\n");
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
@@ -926,7 +943,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -tb N, --threads-batch N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
     printf("  -td N, --threads-draft N");
-    printf("                        number of threads to use during generation (default: same as --threads)");
+    printf("                        number of threads to use during generation (default: same as --threads)\n");
     printf("  -tbd N, --threads-batch-draft N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads-draft)\n");
     printf("  -p PROMPT, --prompt PROMPT\n");
@@ -947,8 +964,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
-    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
+    printf("                        (default: %s)\n", sampler_type_names.c_str());
+    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
     printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
@@ -996,7 +1014,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --winogrande-tasks N  number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks);
     printf("  --multiple-choice     compute multiple choice score over random tasks from datafile supplied with -f\n");
     printf("  --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks);
-    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base");
+    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base\n");
     printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
@@ -1013,7 +1031,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_supports_mmap()) {
         printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa TYPE           attempt optimizations that help on some NUMA systems\n");
+    printf("                          - distribute: spread execution evenly over all nodes\n");
+    printf("                          - isolate: only spawn threads on CPUs on the node that execution started on\n");
+    printf("                          - numactl: use the CPU map provided by numactl\n");
     printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
     printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
     if (llama_supports_gpu_offload()) {
@@ -1097,45 +1118,101 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
 }
 
 //
-// String parsing
+// String utils
 //
 
-std::string parse_samplers_input(std::string input) {
-    std::string output = "";
+std::vector<std::string> string_split(std::string input, char separator) {
+    std::vector<std::string> parts;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(0, separator_pos);
+        parts.emplace_back(part);
+        input = input.substr(separator_pos + 1);
+        separator_pos = input.find(separator);
+    }
+    parts.emplace_back(input);
+    return parts;
+}
+
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
+    };
+
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, char> samplers_symbols {
-        {"top_k",      'k'},
-        {"top-k",      'k'},
-        {"top_p",      'p'},
-        {"top-p",      'p'},
-        {"nucleus",    'p'},
-        {"typical_p",  'y'},
-        {"typical-p",  'y'},
-        {"typical",    'y'},
-        {"min_p",      'm'},
-        {"min-p",      'm'},
-        {"tfs_z",      'f'},
-        {"tfs-z",      'f'},
-        {"tfs",        'f'},
-        {"temp",       't'},
-        {"temperature",'t'}
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
+        {"top-k",       llama_sampler_type::TOP_K},
+        {"top-p",       llama_sampler_type::TOP_P},
+        {"nucleus",     llama_sampler_type::TOP_P},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
+        {"typical",     llama_sampler_type::TYPICAL_P},
+        {"min-p",       llama_sampler_type::MIN_P},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
+        {"tfs",         llama_sampler_type::TFS_Z},
+        {"temp",        llama_sampler_type::TEMPERATURE}
     };
-    // expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
-    size_t separator = input.find(';');
-    while (separator != input.npos) {
-        std::string name = input.substr(0,separator);
-        input = input.substr(separator+1);
-        separator = input.find(';');
 
-        if (samplers_symbols.find(name) != samplers_symbols.end()) {
-            output += samplers_symbols[name];
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names.size());
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
+            sampler_types.push_back(sampler_item->second);
+        }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
         }
     }
-    if (samplers_symbols.find(input) != samplers_symbols.end()) {
-        output += samplers_symbols[input];
+    return sampler_types;
+}
+
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string) {
+    std::unordered_map<char, llama_sampler_type> sampler_name_map {
+        {'k', llama_sampler_type::TOP_K},
+        {'p', llama_sampler_type::TOP_P},
+        {'y', llama_sampler_type::TYPICAL_P},
+        {'m', llama_sampler_type::MIN_P},
+        {'f', llama_sampler_type::TFS_Z},
+        {'t', llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names_string.size());
+    for (const auto & c : names_string) {
+        const auto sampler_item = sampler_name_map.find(c);
+        if (sampler_item != sampler_name_map.end()) {
+            sampler_types.push_back(sampler_item->second);
+        }
+    }
+    return sampler_types;
+}
+
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type) {
+    switch (sampler_type) {
+        case llama_sampler_type::TOP_K:       return "top_k";
+        case llama_sampler_type::TFS_Z:       return "tfs_z";
+        case llama_sampler_type::TYPICAL_P:   return "typical_p";
+        case llama_sampler_type::TOP_P:       return "top_p";
+        case llama_sampler_type::MIN_P:       return "min_p";
+        case llama_sampler_type::TEMPERATURE: return "temperature";
+        default : return "";
     }
-    return output;
 }
 
 //
@@ -1550,6 +1627,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
     fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
     fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");
 
 #ifdef NDEBUG
     fprintf(stream, "debug: false\n");
@@ -1639,7 +1717,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
     fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
     fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
-    fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
     fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
@@ -1664,7 +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: %d # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "seed: %u # 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);
@@ -1673,7 +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: %d\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "threads: %d # default: %u\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);
@@ -1724,7 +1801,8 @@ 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; }
-                seqs[cs_curr[j]] = seqs.size();
+                const size_t sz = seqs.size();
+                seqs[cs_curr[j]] = sz;
             }
         }
         if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }

common/common.h

@@ -76,6 +76,7 @@ struct gpt_params {
     float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
     int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED;
+    ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -134,7 +135,6 @@ struct gpt_params {
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
@@ -162,10 +162,13 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 void process_escapes(std::string& input);
 
 //
-// String parsing
+// String utils
 //
 
-std::string parse_samplers_input(std::string input);
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
+std::vector<std::string> string_split(std::string input, char separator);
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
 
 //
 // Model utils

common/sampling.cpp

@@ -103,15 +103,10 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
 std::string llama_sampling_order_print(const llama_sampling_params & params) {
     std::string result = "CFG -> Penalties ";
     if (params.mirostat == 0) {
-        for (auto s : params.samplers_sequence) {
-            switch (s) {
-                case 'k': result += "-> top_k "; break;
-                case 'f': result += "-> tfs_z "; break;
-                case 'y': result += "-> typical_p "; break;
-                case 'p': result += "-> top_p "; break;
-                case 'm': result += "-> min_p "; break;
-                case 't': result += "-> temp "; break;
-                default : break;
+        for (auto sampler_type : params.samplers_sequence) {
+            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            if (!sampler_type_name.empty()) {
+                result += "-> " + sampler_type_name + " ";
             }
         }
     } else {
@@ -127,8 +122,6 @@ static void sampler_queue(
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
                                  size_t & min_keep) {
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
@@ -137,16 +130,16 @@ static void sampler_queue(
     const float         min_p             = params.min_p;
     const float         tfs_z             = params.tfs_z;
     const float         typical_p         = params.typical_p;
-    const std::string & samplers_sequence = params.samplers_sequence;
+    const std::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
 
-    for (auto s : samplers_sequence) {
-        switch (s){
-            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
-            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
-            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
-            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
-            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't':
+    for (auto sampler_type : samplers_sequence) {
+        switch (sampler_type) {
+            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case llama_sampler_type::TEMPERATURE:
                 if (dynatemp_range > 0) {
                     float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
                     float dynatemp_max = std::max(0.0f, temp + dynatemp_range);

common/sampling.h

@@ -8,6 +8,16 @@
 #include <vector>
 #include <unordered_map>
 
+// sampler types
+enum class llama_sampler_type : char {
+    TOP_K       = 'k',
+    TOP_P       = 'p',
+    MIN_P       = 'm',
+    TFS_Z       = 'f',
+    TYPICAL_P   = 'y',
+    TEMPERATURE = 't'
+};
+
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
@@ -28,7 +38,15 @@ typedef struct llama_sampling_params {
     float       mirostat_tau          = 5.00f;    // target entropy
     float       mirostat_eta          = 0.10f;    // learning rate
     bool        penalize_nl           = true;     // consider newlines as a repeatable token
-    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
+
+    std::vector<llama_sampler_type> samplers_sequence = {
+        llama_sampler_type::TOP_K,
+        llama_sampler_type::TFS_Z,
+        llama_sampler_type::TYPICAL_P,
+        llama_sampler_type::TOP_P,
+        llama_sampler_type::MIN_P,
+        llama_sampler_type::TEMPERATURE
+    };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 

convert-hf-to-gguf.py

@@ -10,7 +10,7 @@ import re
 import sys
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
 
 import numpy as np
 import torch
@@ -25,15 +25,6 @@ import gguf
 from convert import HfVocab
 
 
-# check for any of the given keys in the dictionary and return the value of the first key found
-def get_key_opts(d, keys):
-    for k in keys:
-        if k in d:
-            return d[k]
-    print(f"Could not find any of {keys}")
-    sys.exit()
-
-
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -58,6 +49,15 @@ class Model:
         self.hparams = Model.load_hparams(self.dir_model)
         self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
+
+    def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
+        key = next((k for k in keys if k in self.hparams), None)
+        if key is not None:
+            return self.hparams[key]
+        if optional:
+            return None
+        raise KeyError(f"could not find any of: {keys}")
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -79,28 +79,33 @@ class Model:
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_block_count(self.hparams.get(
-            "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
-        ))
-        if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
+        self.gguf_writer.add_block_count(self.block_count)
+
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
-        if (n_embd := self.hparams.get("hidden_size")) is not None:
-            self.gguf_writer.add_embedding_length(n_embd)
-        if (n_ff := self.hparams.get("intermediate_size")) is not None:
+
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        self.gguf_writer.add_embedding_length(n_embd)
+
+        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_heads")) is not None:
-            self.gguf_writer.add_head_count(n_head)
+
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_head_count(n_head)
+
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
 
-        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
-            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
+            self.gguf_writer.add_layer_norm_eps(f_norm_eps)
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
 
-        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+        self.gguf_writer.add_file_type(self.ftype)
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@@ -209,6 +214,10 @@ class Model:
             return InternLM2Model
         if model_architecture == "MiniCPMForCausalLM":
             return MiniCPMModel
+        if model_architecture == "BertModel":
+            return BertModel
+        if model_architecture == "NomicBertModel":
+            return NomicBertModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -264,6 +273,10 @@ class Model:
             return gguf.MODEL_ARCH.INTERNLM2
         if arch == "MiniCPMForCausalLM":
             return gguf.MODEL_ARCH.MINICPM
+        if arch == "BertModel":
+            return gguf.MODEL_ARCH.BERT
+        if arch == "NomicBertModel":
+            return gguf.MODEL_ARCH.NOMIC_BERT
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -1293,21 +1306,21 @@ class GPT2Model(Model):
 
 class Phi2Model(Model):
     def set_gguf_parameters(self):
-        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
-        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
-        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
-        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
+        rot_pct = self.find_hparam(["partial_rotary_factor"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
         self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
@@ -1629,6 +1642,149 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
+class BertModel(Model):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.vocab_size = None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_causal_attention(False)
+
+        # get pooling path
+        with open(self.dir_model / "modules.json", encoding="utf-8") as f:
+            modules = json.load(f)
+        pooling_path = None
+        for mod in modules:
+            if mod["type"] == "sentence_transformers.models.Pooling":
+                pooling_path = mod["path"]
+                break
+
+        # get pooling type
+        pooling_type = gguf.PoolingType.NONE
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            else:
+                raise NotImplementedError("Only MEAN and CLS pooling types supported")
+
+        self.gguf_writer.add_pooling_type(pooling_type.value)
+
+    def set_vocab(self):
+        path = self.dir_model
+        added_tokens_path = self.dir_model if self.dir_model.exists() else None
+
+        # use huggingface vocab to get all tokens
+        vocab = HfVocab(path, added_tokens_path)
+        tokens, scores, toktypes = zip(*vocab.all_tokens())
+        assert len(tokens) == vocab.vocab_size
+        self.vocab_size = vocab.vocab_size
+
+        # we need this to validate the size of the token_type embeddings
+        # though currently we are passing all zeros to the token_type embeddings
+        n_token_types = len(set(toktypes))
+        self.gguf_writer.add_token_type_count(n_token_types)
+
+        # convert to phantom space vocab
+        def phantom(tok, typ):
+            if tok.startswith(b"[") and tok.endswith(b"]"):
+                return tok
+            if tok.startswith(b"##"):
+                return tok[2:]
+            return b"\xe2\x96\x81" + tok
+        tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
+
+        # set up bos and eos tokens (cls and sep)
+        self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
+        self.gguf_writer.add_eos_token_id(vocab.tokenizer.sep_token_id)
+
+        # add vocab to gguf
+        self.gguf_writer.add_tokenizer_model("bert")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # handle special tokens
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def write_tensors(self):
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+        tensors = dict(self.get_tensors())
+        for name, data_torch in tensors.items():
+            # we are only using BERT for embeddings so we don't need the pooling layer
+            if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
+                continue  # we don't need these
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            data = data_torch.squeeze().numpy()
+            n_dims = len(data.shape)
+            new_dtype: type[np.floating[Any]]
+
+            if (
+                self.ftype == 1 and name.endswith(".weight") and n_dims == 2
+                and name != "embeddings.token_type_embeddings.weight"  # not used with get_rows, must be F32
+            ):
+                # if f16 desired, convert any float32 2-dim weight tensors to float16
+                new_dtype = np.float16
+            else:
+                # if f32 desired, convert any float16 to float32
+                new_dtype = np.float32
+
+            print(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}")
+
+            if data.dtype != new_dtype:
+                data = data.astype(new_dtype)
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+class NomicBertModel(BertModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors
+        assert self.hparams["qkv_proj_bias"] is False
+        assert self.hparams["mlp_fc1_bias"] is False
+        assert self.hparams["mlp_fc2_bias"] is False
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+
+    def get_tensors(self):
+        assert self.vocab_size is not None
+        for name, data in super().get_tensors():
+            # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
+            if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
+                rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
+                assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
+                data = data[:self.vocab_size, :]
+            yield name, data
+
+
 ###### CONVERSION LOGIC ######
 
 

convert-persimmon-to-gguf.py

@@ -88,7 +88,8 @@ def main():
     gguf_writer.add_embedding_length(hidden_size)
     gguf_writer.add_block_count(block_count)
     gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
-    gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+    # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
+    gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
     gguf_writer.add_head_count(head_count)
     gguf_writer.add_head_count_kv(head_count_kv)
     gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)

convert.py

@@ -1173,7 +1173,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 
-def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
+def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
     should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
@@ -1199,7 +1199,11 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     for name, lazy_tensor in model.items():
         tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
         if name_new is None:
-            raise Exception(f"Unexpected tensor name: {name}")
+            if skip_unknown:
+                print(f"Unexpected tensor name: {name} - skipping")
+                continue
+            else:
+                raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1377,19 +1381,20 @@ def main(args_in: list[str] | None = None) -> None:
         output_choices.append("q8_0")
     vocab_types = ["spm", "bpe", "hfft"]
     parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
-    parser.add_argument("--awq-path",    type=Path,              help="Path to scale awq cache file", default=None)
-    parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
-    parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
-    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
-    parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--vocab-type",  choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
-    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
-    parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
-    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
-    parser.add_argument("--big-endian",  action="store_true",    help="model is executed on big endian machine")
-    parser.add_argument("--pad-vocab",   action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--awq-path",     type=Path,              help="Path to scale awq cache file", default=None)
+    parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
+    parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
+    parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
+    parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
+    parser.add_argument("--vocab-type",   choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
+    parser.add_argument("--outfile",      type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency",  type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
+    parser.add_argument("--big-endian",   action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
     if args.awq_path:
@@ -1461,7 +1466,7 @@ def main(args_in: list[str] | None = None) -> None:
     print(f"Special vocab info: {special_vocab}")
 
     model   = model_plus.model
-    model   = convert_model_names(model, params)
+    model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
     outfile = args.outfile or default_outfile(model_plus.paths, ftype)

examples/CMakeLists.txt

@@ -38,6 +38,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
+    add_subdirectory(gguf)
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)

examples/batched-bench/batched-bench.cpp

@@ -82,7 +82,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -158,7 +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 = %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("%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("\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.swift/Sources/main.swift

@@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu
 let n_len: Int = 32
 
 // init LLM
-llama_backend_init(false)
+llama_backend_init()
 defer {
     llama_backend_free()
 }

examples/batched/batched.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -91,7 +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 = %d, 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 = %u, 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/beam-search/beam-search.cpp

@@ -119,7 +119,8 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * 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: %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);
+    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);
 }
 
 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 [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, 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);
 
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
+    printf("[%s:GG] Allocating [%u] 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 [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, 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);
 
     // printing the per-layer allocations here so we dont print in the for loop.
-    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] 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] float space for layer.ffn_norm for [%d] layers\n",__func__,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] 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);
+    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);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");

examples/embedding/embedding.cpp

@@ -7,6 +7,51 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static std::vector<std::string> split_lines(const std::string & s) {
+    std::string line;
+    std::vector<std::string> lines;
+    std::stringstream ss(s);
+    while (std::getline(ss, line)) {
+        lines.push_back(line);
+    }
+    return lines;
+}
+
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+    }
+}
+
+static void normalize(float * vec, float * out, int n) {
+    float norm = 0;
+    for (int i = 0; i < n; i++) {
+        norm += vec[i] * vec[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n; i++) {
+        out[i] = vec[i] / norm;
+    }
+}
+
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+    // clear previous kv_cache values (irrelevant for embeddings)
+    llama_kv_cache_clear(ctx);
+
+    // run model
+    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
+    }
+
+    // normalize on copy
+    for (int k = 0; k < n_seq; k++) {
+        float * emb = llama_get_embeddings_ith(ctx, k);
+        float * out = output + k * n_embd;
+        normalize(emb, out, n_embd);
+    }
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -29,7 +74,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -55,49 +101,84 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    int n_past = 0;
+    // split the prompt into lines
+    std::vector<std::string> prompts = split_lines(params.prompt);
 
-    // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch == params.n_ctx);
 
+    // tokenize the prompts and trim
+    std::vector<std::vector<int32_t>> inputs;
+    for (const auto & prompt : prompts) {
+        auto inp = ::llama_tokenize(ctx, prompt, true);
+        if (inp.size() > n_batch) {
+            inp.resize(n_batch);
+        }
+        inputs.push_back(inp);
+    }
+
+    // tokenization stats
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-        for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        for (int i = 0; i < (int) inputs.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            for (int j = 0; j < (int) inputs[i].size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
         }
-        fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)n_ctx) {
-        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), n_ctx);
-        return 1;
-    }
-
-    while (!embd_inp.empty()) {
-        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return 1;
-        }
-        n_past += n_tokens;
-        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
-    }
+    // initialize batch
+    const int n_prompts = prompts.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
 
+    // allocate output
     const int n_embd = llama_n_embd(model);
-    const auto * embeddings = llama_get_embeddings(ctx);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    float * emb = embeddings.data();
 
-    for (int i = 0; i < n_embd; i++) {
-        printf("%f ", embeddings[i]);
+    // break into batches
+    int p = 0; // number of prompts processed already
+    int s = 0; // number of prompts in current batch
+    for (int k = 0; k < n_prompts; k++) {
+        // clamp to n_batch tokens
+        auto & inp = inputs[k];
+        const uint64_t n_toks = inp.size();
+
+        // encode if at capacity
+        if (batch.n_tokens + n_toks > n_batch) {
+            float * out = emb + p * n_embd;
+            batch_decode(ctx, batch, out, s, n_embd);
+            llama_batch_clear(batch);
+            p += s;
+            s = 0;
+        }
+
+        // add to batch
+        batch_add_seq(batch, inp, s);
+        s += 1;
     }
-    printf("\n");
 
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
+
+    // print first 3 embeddings
+    for (int j = 0; j < std::min(3, n_prompts); j++) {
+        fprintf(stderr, "embedding %d: ", j);
+        for (int i = 0; i < n_embd; i++) {
+            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+        }
+        fprintf(stderr, "\n\n");
+    }
+    fprintf(stderr, "\n");
+
+    // clean up
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/export-lora/export-lora.cpp

@@ -7,8 +7,6 @@
 #include <string>
 #include <thread>
 
-static const size_t tensor_alignment = 32;
-
 struct lora_info {
     std::string filename;
     float scale;
@@ -337,24 +335,14 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
     params.mem_buffer = NULL;
     params.no_alloc   = true;
     struct ggml_context * ctx = NULL;
-    struct ggml_allocr * alloc = NULL;
-    struct ggml_cgraph * gf = NULL;
+    struct ggml_gallocr * alloc = NULL;
+    struct ggml_cgraph  * gf = NULL;
 
     ctx   = ggml_init(params);
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
-    ggml_allocr_free(alloc);
-    ggml_free(ctx);
 
-    static std::vector<uint8_t> data_compute;
-    data_compute.resize(alloc_size + tensor_alignment);
-
-    ctx   = ggml_init(params);
-    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
-    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-    ggml_allocr_alloc_graph(alloc, gf);
-    ggml_allocr_free(alloc);
+    ggml_gallocr_alloc_graph(alloc, gf);
 
     struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
     static std::vector<uint8_t> data_work;
@@ -363,6 +351,7 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
 
     ggml_graph_compute(gf, &cplan);
 
+    ggml_gallocr_free(alloc);
     ggml_free(ctx);
     return true;
 }

examples/finetune/README.md

@@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
   --rank-wk N                LORA rank for wk tensor (default 4)
   --rank-wv N                LORA rank for wv tensor (default 4)
   --rank-wo N                LORA rank for wo tensor (default 4)
-  --rank-w1 N                LORA rank for w1 tensor (default 4)
-  --rank-w2 N                LORA rank for w2 tensor (default 4)
-  --rank-w3 N                LORA rank for w3 tensor (default 4)
+  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
+  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
+  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
 ```
 
 The LORA rank of 'norm' tensors should always be 1.

examples/finetune/finetune.cpp

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 #include "llama.h"
 #include "common.h"
 #include "train.h"
@@ -13,8 +14,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const size_t tensor_alignment = 32;
-
 struct my_llama_hparams {
     uint32_t n_vocab    = 32000;
     uint32_t n_ctx      = 512;
@@ -61,9 +60,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -86,9 +85,9 @@ struct my_llama_lora_hparams {
     uint32_t n_rank_wv = 4;
     uint32_t n_rank_wo = 4;
     uint32_t n_rank_ffn_norm = 1;
-    uint32_t n_rank_w1 = 4;
-    uint32_t n_rank_w2 = 4;
-    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_ffn_gate = 4;
+    uint32_t n_rank_ffn_down = 4;
+    uint32_t n_rank_ffn_up = 4;
     uint32_t n_rank_tok_embeddings = 4;
     uint32_t n_rank_norm = 1;
     uint32_t n_rank_output = 4;
@@ -118,17 +117,17 @@ struct my_llama_lora_layer {
     struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * w1_a;
-    struct ggml_tensor * w1_b;
-    struct ggml_tensor * w2_a;
-    struct ggml_tensor * w2_b;
-    struct ggml_tensor * w3_a;
-    struct ggml_tensor * w3_b;
+    struct ggml_tensor * ffn_gate_a;
+    struct ggml_tensor * ffn_gate_b;
+    struct ggml_tensor * ffn_down_a;
+    struct ggml_tensor * ffn_down_b;
+    struct ggml_tensor * ffn_up_a;
+    struct ggml_tensor * ffn_up_b;
 };
 
 struct my_llama_lora {
     struct ggml_context * ctx = NULL;
-    std::vector<uint8_t> data;
+    ggml_backend_buffer_t data;
 
     my_llama_lora_hparams hparams;
 
@@ -209,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
     printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
     printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
     printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
-    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
-    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
-    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_ffn_gate       : %u\n", __func__, params->n_rank_ffn_gate);
+    printf("%s: n_rank_ffn_down       : %u\n", __func__, params->n_rank_ffn_down);
+    printf("%s: n_rank_ffn_up         : %u\n", __func__, params->n_rank_ffn_up);
     printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
     printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
     printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
@@ -320,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
         layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
         layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
-        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
-        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
-        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+        layer.ffn_gate       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.ffn_down       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.ffn_up         = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
 
         assert_shape_1d(layer.attention_norm, hparams.n_embd);
         assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
@@ -330,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
         assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
         assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
-        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
-        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
-        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_gate,       hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_down,       hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.ffn_up,         hparams.n_embd, hparams.n_ff);
     }
 }
 
@@ -363,69 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
         ggml_set_param(ctx, layer.wo_b);
         ggml_set_param(ctx, layer.ffn_norm_a);
         ggml_set_param(ctx, layer.ffn_norm_b);
-        ggml_set_param(ctx, layer.w1_a);
-        ggml_set_param(ctx, layer.w1_b);
-        ggml_set_param(ctx, layer.w2_a);
-        ggml_set_param(ctx, layer.w2_b);
-        ggml_set_param(ctx, layer.w3_a);
-        ggml_set_param(ctx, layer.w3_b);
-    }
-}
-
-static void alloc_lora(struct ggml_allocr * alloc, struct my_llama_lora * lora) {
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_a);
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_b);
-    ggml_allocr_alloc(alloc, lora->norm_a);
-    ggml_allocr_alloc(alloc, lora->norm_b);
-    ggml_allocr_alloc(alloc, lora->output_a);
-    ggml_allocr_alloc(alloc, lora->output_b);
-    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
-        auto & layer = lora->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm_a);
-        ggml_allocr_alloc(alloc, layer.attention_norm_b);
-        ggml_allocr_alloc(alloc, layer.wq_a);
-        ggml_allocr_alloc(alloc, layer.wq_b);
-        ggml_allocr_alloc(alloc, layer.wk_a);
-        ggml_allocr_alloc(alloc, layer.wk_b);
-        ggml_allocr_alloc(alloc, layer.wv_a);
-        ggml_allocr_alloc(alloc, layer.wv_b);
-        ggml_allocr_alloc(alloc, layer.wo_a);
-        ggml_allocr_alloc(alloc, layer.wo_b);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_a);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_b);
-        ggml_allocr_alloc(alloc, layer.w1_a);
-        ggml_allocr_alloc(alloc, layer.w1_b);
-        ggml_allocr_alloc(alloc, layer.w2_a);
-        ggml_allocr_alloc(alloc, layer.w2_b);
-        ggml_allocr_alloc(alloc, layer.w3_a);
-        ggml_allocr_alloc(alloc, layer.w3_b);
-    }
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_a->grad);
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_b->grad);
-    ggml_allocr_alloc(alloc, lora->norm_a->grad);
-    ggml_allocr_alloc(alloc, lora->norm_b->grad);
-    ggml_allocr_alloc(alloc, lora->output_a->grad);
-    ggml_allocr_alloc(alloc, lora->output_b->grad);
-    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
-        auto & layer = lora->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm_a->grad);
-        ggml_allocr_alloc(alloc, layer.attention_norm_b->grad);
-        ggml_allocr_alloc(alloc, layer.wq_a->grad);
-        ggml_allocr_alloc(alloc, layer.wq_b->grad);
-        ggml_allocr_alloc(alloc, layer.wk_a->grad);
-        ggml_allocr_alloc(alloc, layer.wk_b->grad);
-        ggml_allocr_alloc(alloc, layer.wv_a->grad);
-        ggml_allocr_alloc(alloc, layer.wv_b->grad);
-        ggml_allocr_alloc(alloc, layer.wo_a->grad);
-        ggml_allocr_alloc(alloc, layer.wo_b->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_a->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_b->grad);
-        ggml_allocr_alloc(alloc, layer.w1_a->grad);
-        ggml_allocr_alloc(alloc, layer.w1_b->grad);
-        ggml_allocr_alloc(alloc, layer.w2_a->grad);
-        ggml_allocr_alloc(alloc, layer.w2_b->grad);
-        ggml_allocr_alloc(alloc, layer.w3_a->grad);
-        ggml_allocr_alloc(alloc, layer.w3_b->grad);
+        ggml_set_param(ctx, layer.ffn_gate_a);
+        ggml_set_param(ctx, layer.ffn_gate_b);
+        ggml_set_param(ctx, layer.ffn_down_a);
+        ggml_set_param(ctx, layer.ffn_down_b);
+        ggml_set_param(ctx, layer.ffn_up_a);
+        ggml_set_param(ctx, layer.ffn_up_b);
     }
 }
 
@@ -493,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
         layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
 
-        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
-        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
-        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
-        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
-        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
-        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+        layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd);
+        layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff);
+        layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff);
+        layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd);
+        layer.ffn_up_a   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_embd);
+        layer.ffn_up_b   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_ff);
 
         ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
         ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
@@ -512,28 +454,18 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
         ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
         ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
-        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
-        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
-        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
-        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
-        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
-        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
     }
 
     set_param_lora(lora);
 
-    // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
-
-    // allocate data
-    struct ggml_allocr * alloc = NULL;
-    lora->data.resize(size + tensor_alignment);
-    alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
-    alloc_lora(alloc, lora);
-    ggml_allocr_free(alloc);
+    // allocate data for lora tensors
+    lora->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type());
 }
 
 static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) {
@@ -565,12 +497,12 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
         ggml_set_zero(layer.ffn_norm_b);
 
-        randomize_tensor_normal(layer.w1_a, rnd);
-        ggml_set_zero(layer.w1_b);
-        randomize_tensor_normal(layer.w2_a, rnd);
-        ggml_set_zero(layer.w2_b);
-        randomize_tensor_normal(layer.w3_a, rnd);
-        ggml_set_zero(layer.w3_b);
+        randomize_tensor_normal(layer.ffn_gate_a, rnd);
+        ggml_set_zero(layer.ffn_gate_b);
+        randomize_tensor_normal(layer.ffn_down_a, rnd);
+        ggml_set_zero(layer.ffn_down_b);
+        randomize_tensor_normal(layer.ffn_up_a, rnd);
+        ggml_set_zero(layer.ffn_up_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -579,7 +511,7 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
 static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct my_llama_model * model,
         struct my_llama_lora  * lora,
-        struct ggml_allocr    * alloc,
+        ggml_gallocr_t          alloc,
         struct ggml_context   * ctx,
         struct ggml_cgraph    * gf,
         struct ggml_cgraph    * gb,
@@ -590,7 +522,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const  int              n_tokens,
         const  int              n_batch,
         const  bool             enable_flash_attn,
-        const  bool             enable_checkpointing) {
+        const  bool             enable_checkpointing,
+        const  bool             measure_only) {
 
     ggml_set_scratch(ctx, { 0, 0, nullptr, });
     const int n_past = 0;
@@ -622,13 +555,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    ggml_allocr_alloc(alloc, KQ_pos);
-    if (!ggml_allocr_is_measure(alloc)) {
-        int * data = (int *) KQ_pos->data;
-        for (int i = 0; i < N; ++i) {
-            data[i] = n_past + i;
-        }
-    }
+    ggml_set_input(KQ_pos);
 
     // rope has so much parameters that we make a custom function for it
     auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
@@ -683,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
         struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
         struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
-        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
-        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
-        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
-        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
-        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
-        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
-        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+        struct ggml_tensor * wq       = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk       = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv       = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo       = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b));
+        struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b));
+        struct ggml_tensor * ffn_up   = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b));
 
         struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
         struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
@@ -732,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, ffn_up, t24);                             set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, ffn_gate, t24);                           set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, ffn_down, t28);                           set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         if (enable_checkpointing) {
@@ -780,7 +707,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     // input gradient
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
-    ggml_allocr_alloc(alloc, t36->grad);
+    ggml_set_input(t36->grad);
     // KQ_pos
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
 
@@ -796,20 +723,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
     // note: they will be freed in reverse order
     for (unsigned int i = 0; i < checkpoints.size(); ++i) {
         if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
-            ggml_allocr_alloc(alloc, checkpoints[i]);
+            ggml_set_input(checkpoints[i]);
         }
     }
 
-    ggml_allocr_alloc_graph(alloc, gb);
+    if (measure_only) {
+        ggml_gallocr_reserve(alloc, gb);
+    } else {
+        ggml_gallocr_alloc_graph(alloc, gb);
+
+        // set KQ_pos
+        {
+            int * data = (int *) KQ_pos->data;
+            for (int i = 0; i < N; ++i) {
+                data[i] = n_past + i;
+            }
+        }
+    }
 
     // remove the additional nodes and leafs
     for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
@@ -859,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
 
     init_lora(model, lora);
 
@@ -886,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
         copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
         copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
         copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
-        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
-        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
-        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
-        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
-        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
-        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+        copy_tensor_by_name(layer.ffn_gate_a,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_a));
+        copy_tensor_by_name(layer.ffn_gate_b,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_b));
+        copy_tensor_by_name(layer.ffn_down_a,       f_ggml_ctx, ggml_get_name(layer.ffn_down_a));
+        copy_tensor_by_name(layer.ffn_down_b,       f_ggml_ctx, ggml_get_name(layer.ffn_down_b));
+        copy_tensor_by_name(layer.ffn_up_a,         f_ggml_ctx, ggml_get_name(layer.ffn_up_a));
+        copy_tensor_by_name(layer.ffn_up_b,         f_ggml_ctx, ggml_get_name(layer.ffn_up_b));
     }
 }
 
@@ -929,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_ffn_gate);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_ffn_down);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_ffn_up);
 
     gguf_add_tensor(fctx, lora->tok_embeddings_a);
     gguf_add_tensor(fctx, lora->tok_embeddings_b);
@@ -955,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
         gguf_add_tensor(fctx, layer.wo_b);
         gguf_add_tensor(fctx, layer.ffn_norm_a);
         gguf_add_tensor(fctx, layer.ffn_norm_b);
-        gguf_add_tensor(fctx, layer.w1_a);
-        gguf_add_tensor(fctx, layer.w1_b);
-        gguf_add_tensor(fctx, layer.w2_a);
-        gguf_add_tensor(fctx, layer.w2_b);
-        gguf_add_tensor(fctx, layer.w3_a);
-        gguf_add_tensor(fctx, layer.w3_b);
+        gguf_add_tensor(fctx, layer.ffn_gate_a);
+        gguf_add_tensor(fctx, layer.ffn_gate_b);
+        gguf_add_tensor(fctx, layer.ffn_down_a);
+        gguf_add_tensor(fctx, layer.ffn_down_b);
+        gguf_add_tensor(fctx, layer.ffn_up_a);
+        gguf_add_tensor(fctx, layer.ffn_up_b);
     }
 }
 
@@ -1165,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
         write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
         write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
-        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
     }
 }
 
@@ -1200,9 +1139,9 @@ struct train_params {
     uint32_t n_rank_wv;
     uint32_t n_rank_wo;
     uint32_t n_rank_ffn_norm;
-    uint32_t n_rank_w1;
-    uint32_t n_rank_w2;
-    uint32_t n_rank_w3;
+    uint32_t n_rank_ffn_gate;
+    uint32_t n_rank_ffn_down;
+    uint32_t n_rank_ffn_up;
     uint32_t n_rank_tok_embeddings;
     uint32_t n_rank_norm;
     uint32_t n_rank_output;
@@ -1213,9 +1152,9 @@ struct train_params {
     bool custom_n_rank_wv;
     bool custom_n_rank_wo;
     bool custom_n_rank_ffn_norm;
-    bool custom_n_rank_w1;
-    bool custom_n_rank_w2;
-    bool custom_n_rank_w3;
+    bool custom_n_rank_ffn_gate;
+    bool custom_n_rank_ffn_down;
+    bool custom_n_rank_ffn_up;
     bool custom_n_rank_tok_embeddings;
     bool custom_n_rank_norm;
     bool custom_n_rank_output;
@@ -1247,9 +1186,9 @@ static struct train_params get_default_train_params() {
     params.n_rank_wv             = 4;
     params.n_rank_wo             = 4;
     params.n_rank_ffn_norm       = 1;
-    params.n_rank_w1             = 4;
-    params.n_rank_w2             = 4;
-    params.n_rank_w3             = 4;
+    params.n_rank_ffn_gate       = 4;
+    params.n_rank_ffn_down       = 4;
+    params.n_rank_ffn_up         = 4;
     params.n_rank_tok_embeddings = 4;
     params.n_rank_norm           = 1;
     params.n_rank_output         = 4;
@@ -1260,9 +1199,9 @@ static struct train_params get_default_train_params() {
     params.custom_n_rank_wv             = false;
     params.custom_n_rank_wo             = false;
     params.custom_n_rank_ffn_norm       = false;
-    params.custom_n_rank_w1             = false;
-    params.custom_n_rank_w2             = false;
-    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_ffn_gate       = false;
+    params.custom_n_rank_ffn_down       = false;
+    params.custom_n_rank_ffn_up         = false;
     params.custom_n_rank_tok_embeddings = false;
     params.custom_n_rank_norm           = false;
     params.custom_n_rank_output         = false;
@@ -1293,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
     fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_gate N          LORA rank for ffn_gate tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_down N          LORA rank for ffn_down tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_up N            LORA rank for ffn_up tensor, overrides default rank.\n");
 
     print_common_train_usage(argc, argv, &params->common);
 }
@@ -1430,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_wo = std::stoi(argv[i]);
             params->custom_n_rank_wo = true;
-        } else if (arg == "--rank-w1") {
+        } else if (arg == "--rank-ffn_gate") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w1 = std::stoi(argv[i]);
-            params->custom_n_rank_w1 = true;
-        } else if (arg == "--rank-w2") {
+            params->n_rank_ffn_gate = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_gate = true;
+        } else if (arg == "--rank-ffn_down") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w2 = std::stoi(argv[i]);
-            params->custom_n_rank_w2 = true;
-        } else if (arg == "--rank-w3") {
+            params->n_rank_ffn_down = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_down = true;
+        } else if (arg == "--rank-ffn_up") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w3 = std::stoi(argv[i]);
-            params->custom_n_rank_w3 = true;
+            params->n_rank_ffn_up = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_up = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1513,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
         nx += ggml_nelements(layer.wo_b);
         nx += ggml_nelements(layer.ffn_norm_a);
         nx += ggml_nelements(layer.ffn_norm_b);
-        nx += ggml_nelements(layer.w1_a);
-        nx += ggml_nelements(layer.w1_b);
-        nx += ggml_nelements(layer.w2_a);
-        nx += ggml_nelements(layer.w2_b);
-        nx += ggml_nelements(layer.w3_a);
-        nx += ggml_nelements(layer.w3_b);
+        nx += ggml_nelements(layer.ffn_gate_a);
+        nx += ggml_nelements(layer.ffn_gate_b);
+        nx += ggml_nelements(layer.ffn_down_a);
+        nx += ggml_nelements(layer.ffn_down_b);
+        nx += ggml_nelements(layer.ffn_up_a);
+        nx += ggml_nelements(layer.ffn_up_b);
     }
     return nx;
 }
@@ -1572,9 +1511,9 @@ int main(int argc, char ** argv) {
     uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
     uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
     uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
-    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
-    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
-    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_ffn_gate           = params.custom_n_rank_ffn_gate       ? params.n_rank_ffn_gate       : params.lora_r;
+    uint32_t n_rank_ffn_down           = params.custom_n_rank_ffn_down       ? params.n_rank_ffn_down       : params.lora_r;
+    uint32_t n_rank_ffn_up             = params.custom_n_rank_ffn_up         ? params.n_rank_ffn_up         : params.lora_r;
     uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
     uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
     uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
@@ -1584,9 +1523,9 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_wv             = n_rank_wv;
     lora.hparams.n_rank_wo             = n_rank_wo;
     lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
-    lora.hparams.n_rank_w1             = n_rank_w1;
-    lora.hparams.n_rank_w2             = n_rank_w2;
-    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_ffn_gate       = n_rank_ffn_gate;
+    lora.hparams.n_rank_ffn_down       = n_rank_ffn_down;
+    lora.hparams.n_rank_ffn_up         = n_rank_ffn_up;
     lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
     lora.hparams.n_rank_norm           = n_rank_norm;
     lora.hparams.n_rank_output         = n_rank_output;
@@ -1627,9 +1566,9 @@ int main(int argc, char ** argv) {
         || (lora.hparams.n_rank_wv             != n_rank_wv)
         || (lora.hparams.n_rank_wo             != n_rank_wo)
         || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
-        || (lora.hparams.n_rank_w1             != n_rank_w1)
-        || (lora.hparams.n_rank_w2             != n_rank_w2)
-        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_ffn_gate       != n_rank_ffn_gate)
+        || (lora.hparams.n_rank_ffn_down       != n_rank_ffn_down)
+        || (lora.hparams.n_rank_ffn_up         != n_rank_ffn_up)
         || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
         || (lora.hparams.n_rank_norm           != n_rank_norm)
         || (lora.hparams.n_rank_output         != n_rank_output)
@@ -1663,7 +1602,7 @@ int main(int argc, char ** argv) {
     printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
     printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
     printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
-    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + lora.data.size()), (float) (ggml_used_mem(lora.ctx) + lora.data.size()) / (1024.0f*1024.0f));
+    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)), (float) (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)) / (1024.0f*1024.0f));
 
     if (params.only_write_lora) {
         save_train_files_data save_data;
@@ -1690,10 +1629,6 @@ int main(int argc, char ** argv) {
     int n_vocab  = model.hparams.n_vocab;
     int n_batch  = params.common.n_batch;
 
-
-    std::vector<uint8_t> mem_input_data;
-    std::vector<uint8_t> mem_compute_data;
-
     // context for input tensors without their data
     struct ggml_init_params ctx_input_params = {
         ggml_tensor_overhead() * 2, // mem_size
@@ -1706,17 +1641,11 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * tokens_input  = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch);
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
-    // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
-    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
-
     // allocate input tensors
-    mem_input_data.resize(max_input_size);
-    ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc_inps, tokens_input);
-    ggml_allocr_alloc(alloc_inps, target_probs);
+    // measure required memory for input tensors
+    ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type());
+    size_t max_input_size = ggml_backend_buffer_get_size(input_data);
+    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // context for compute tensors without their data
     const size_t estimated_compute_size_wo_data = (
@@ -1743,7 +1672,7 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment);
+        ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
         gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
         gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1756,14 +1685,15 @@ int main(int argc, char ** argv) {
             &logits, tokens_input, target_probs,
             n_tokens, n_batch,
             params.common.use_flash,
-            params.common.use_checkpointing
+            params.common.use_checkpointing,
+            true
         );
-        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer
         if (max_compute_size < best_compute_size) {
             best_compute_size = max_compute_size;
             best_order = gf->order;
         }
-        ggml_allocr_free(alloc);
+        ggml_gallocr_free(alloc);
         ggml_free(ctx_compute);
     }
     size_t max_compute_size = best_compute_size;
@@ -1774,9 +1704,8 @@ int main(int argc, char ** argv) {
         "invalid");
 
     // allocate compute tensors
-    mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
     gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1789,11 +1718,9 @@ int main(int argc, char ** argv) {
         &logits, tokens_input, target_probs,
         n_tokens, n_batch,
         params.common.use_flash,
-        params.common.use_checkpointing
+        params.common.use_checkpointing,
+        false
     );
-    ggml_allocr_free(alloc);
-    ggml_allocr_free(alloc_inps);
-
 
     // tokenize data
     std::vector<llama_token> train_tokens;
@@ -1908,6 +1835,8 @@ int main(int argc, char ** argv) {
     ggml_free(ctx_work);
     ggml_free(ctx_compute);
     ggml_free(ctx_input);
+    ggml_gallocr_free(alloc);
+
 
     int64_t t1 = ggml_time_ms();
     printf("%s: total training time: ", __func__);

examples/imatrix/imatrix.cpp

@@ -568,7 +568,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model_params mparams = llama_model_params_from_gpt_params(params);
 

examples/infill/infill.cpp

@@ -202,7 +202,8 @@ int main(int argc, char ** argv) {
     std::mt19937 rng(params.seed);
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/llama-bench/llama-bench.cpp

@@ -1151,8 +1151,7 @@ int main(int argc, char ** argv) {
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
     }
-    bool numa = false;
-    llama_backend_init(numa);
+    llama_backend_init();
 
     // initialize printer
     std::unique_ptr<printer> p;

examples/llama.android/app/src/main/cpp/llama-android.cpp

@@ -274,8 +274,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
-    llama_backend_init(numa);
+Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+    llama_backend_init();
 }
 
 extern "C"

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -51,7 +51,7 @@ actor LlamaContext {
     }
 
     static func create_context(path: String) throws -> LlamaContext {
-        llama_backend_init(false)
+        llama_backend_init()
         var model_params = llama_model_default_params()
 
 #if targetEnvironment(simulator)

examples/llava/README.md

@@ -1,10 +1,12 @@
 # LLaVA
 
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
+as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
 
 The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
 and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
 models are available.
+For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
 
 After API is confirmed, more models will be supported / uploaded.
 
@@ -18,10 +20,11 @@ After building, run: `./llava-cli` to see the usage. For example:
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
 
-## Model conversion
+## LLaVA 1.5
 
-- Clone `llava-v15-7b` and `clip-vit-large-patch14-336` locally:
+- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
@@ -29,28 +32,75 @@ git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
 git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 ```
 
-2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+2. Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
 python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
+4. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
 
 ```sh
 python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```
 
-4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
+5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b
+python ./convert.py ../llava-v1.5-7b --skip-unknown
 ```
 
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
+## LLaVA 1.6 gguf conversion
+
+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 (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) 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)
+
+## llava-cli templating and llava-1.6 prompting
+
+llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
+For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
+
+**For Mistral and using llava-cli binary:**
+Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
+The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
+
+**For the 34B this should work:**
+Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
+
+
+## How to know if you are running in llava-1.5 or llava-1.6 mode
+
+When running llava-cli you will see a visual information right before the prompt is being processed:
+
+**Llava-1.5:**
+`encode_image_with_clip: image embedding created: 576 tokens`
+
+**Llava-1.6 (anything above 576):**
+`encode_image_with_clip: image embedding created: 2880 tokens`
+
+
+Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
+
+
+
+
 ## TODO
 
-- [ ] Support non-CPU backend for the image encoding part.
+- [x] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.h

@@ -24,25 +24,7 @@ struct clip_ctx;
 extern "C" {
 #endif
 
-struct clip_vision_hparams {
-    int32_t image_size;
-    int32_t patch_size;
-    int32_t hidden_size;
-    int32_t n_intermediate;
-    int32_t projection_dim;
-    int32_t n_head;
-    int32_t n_layer;
-    float eps;
-};
-
-CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
-
-CLIP_API void clip_free(struct clip_ctx * ctx);
-
-CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-
-CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+struct clip_ctx;
 
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
@@ -54,18 +36,43 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
+CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
+
+CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
-CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  & batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch & batch);
 
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-CLIP_API bool clip_image_preprocess  (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
+/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
 CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
 CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
 

examples/llava/convert-image-encoder-to-gguf.py

@@ -71,25 +71,26 @@ def bytes_to_unicode():
     return dict(zip(bs, cs))
 
 
-ap = argparse.ArgumentParser(prog="convert_hf_to_gguf.py")
+ap = argparse.ArgumentParser()
 ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
 ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
 ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
                 help="Save a vision-only model. It can't be used to encode texts")
-ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
                 help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
-ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
-ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
 default_image_mean = [0.48145466, 0.4578275, 0.40821073]
 default_image_std = [0.26862954, 0.26130258, 0.27577711]
-ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
-ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
 
 # with proper
 args = ap.parse_args()
@@ -105,7 +106,7 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
-if args.clip_model_is_vision:
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
     vocab = None
     tokens = None
 else:
@@ -133,7 +134,7 @@ ftype = 1
 if args.use_f32:
     ftype = 0
 
-if args.clip_model_is_vision:
+if args.clip_model_is_vision or args.clip_model_is_openclip:
     model = CLIPVisionModel.from_pretrained(dir_model)
     processor = None
 else:
@@ -202,6 +203,57 @@ if has_vision_encoder:
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
     block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
+                            #     /**
+                            #      "image_grid_pinpoints": [
+                            #         [
+                            #         336,
+                            #         672
+                            #         ],
+                            #         [
+                            #         672,
+                            #         336
+                            #         ],
+                            #         [
+                            #         672,
+                            #         672
+                            #         ],
+                            #         [
+                            #         1008,
+                            #         336
+                            #         ],
+                            #         [
+                            #         336,
+                            #         1008
+                            #         ]
+                            #     ],
+                            #     Flattened:
+                            #     [
+                            #         336, 672,
+                            #         672, 336,
+                            #         672, 672,
+                            #         1008, 336,
+                            #         336, 1008
+                            #     ]
+                            #  *
+                            #  */
+    if "image_grid_pinpoints" in v_hparams:
+        # flatten it
+        image_grid_pinpoints = []
+        for pinpoint in v_hparams["image_grid_pinpoints"]:
+            for p in pinpoint:
+                image_grid_pinpoints.append(p)
+        fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
+    if "image_crop_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
+    if "image_aspect_ratio" in v_hparams:
+        fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
+    if "image_split_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
+    if "mm_patch_merge_type" in v_hparams:
+        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
+    if "mm_projector_type" in v_hparams:
+        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
+
 
     if processor is not None:
         image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean

examples/llava/llava-cli.cpp

@@ -155,11 +155,29 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
         printf("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
         printf("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     } else {
         // llava-1.5 native mode
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     }
 
     eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
@@ -171,13 +189,17 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     fprintf(stderr, "\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-
+    std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
-
         printf("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
         fflush(stdout);
     }
 
@@ -196,7 +218,8 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
-    llama_backend_init(params->numa);
+    llama_backend_init();
+    llama_numa_init(params->numa);
 
     llama_model_params model_params = llama_model_params_from_gpt_params(*params);
 

examples/llava/llava-surgery-v2.py

@@ -0,0 +1,167 @@
+import argparse
+import glob
+import os
+import torch
+from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file
+
+# Function to determine if file is a SafeTensor file
+def is_safetensor_file(file_path):
+    return file_path.endswith('.safetensors')
+
+
+# Unified loading function
+def load_model(file_path):
+    if is_safetensor_file(file_path):
+        tensors = {}
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            for key in f.keys():
+                tensors[key] = f.get_tensor(key).clone()
+                # output shape
+                print(f"{key} : {tensors[key].shape}")
+        return tensors, 'safetensor'
+    else:
+        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
+
+
+# Unified saving function
+def save_model(model, file_path, file_type):
+    if file_type == 'safetensor':
+        # safe_save(model, file_path)
+        save_file(model, file_path)
+    else:
+        torch.save(model, file_path)
+
+
+# Adapted function to clean vision tower from checkpoint
+def clean_vision_tower_from_checkpoint(checkpoint_path):
+    checkpoint, file_type = load_model(checkpoint_path)
+    # file_type = 'pytorch'
+    model_path = os.path.dirname(checkpoint_path)
+    print(f"Searching for vision tower tensors in {checkpoint_path}")
+    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+
+    if len(clip_tensors) > 0:
+        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
+        # Adapted for file type
+        clip_path = os.path.join(model_path, "llava.clip")
+
+        if os.path.exists(clip_path):
+            print(f"Loading existing llava.clip from {clip_path}")
+            existing_clip, _ = load_model(clip_path)
+        else:
+            print(f"Creating new llava.clip at {clip_path}")
+            existing_clip = {}
+        # Update existing_clip with new tensors, avoid duplicates
+        for name in clip_tensors:
+            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
+            print(f"Adding {simple_name} to llava.clip")
+            if simple_name not in existing_clip:
+                existing_clip[simple_name] = checkpoint[name]
+
+        # Save the updated clip tensors back to llava.clip
+        save_model(existing_clip, clip_path, 'pytorch')
+
+        # Remove the tensors from the original checkpoint
+        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
+
+def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
+    newline_checkpoint_path = None
+    projector_checkpoint_path = None
+
+    for path in checkpoint_paths:
+        checkpoint, _ = load_model(path)
+        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
+            newline_checkpoint_path = path
+        if projector(checkpoint):
+            projector_checkpoint_path = path
+
+    return newline_checkpoint_path, projector_checkpoint_path
+
+def newline_criteria(checkpoint):
+    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
+
+
+# Command-line interface setup
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
+ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
+args = ap.parse_args()
+
+if args.clean_vision_tower:
+    # Generalized to handle both PyTorch and SafeTensors models
+    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
+    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+    for projector_checkpoint_path in checkpoint_paths:
+        print(f"Cleaning {projector_checkpoint_path}")
+        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
+            print(f"No vision tower found in {projector_checkpoint_path}")
+            # we break once none is found, so far all models append them at the end
+            # break
+    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
+
+# Now we look for the projector in the last checkpoint
+model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+# last_checkpoint_path = checkpoint_paths[0]
+# first_checkpoint_path = checkpoint_paths[-1]
+newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
+
+print(f"Taking projector from {projector_checkpoint_path}")
+first_mm_tensors = []
+first_checkpoint = None
+if newline_checkpoint_path is not None:
+    print(f"Taking newline from {newline_checkpoint_path}")
+    first_checkpoint, file_type = load_model(newline_checkpoint_path)
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+
+# Load the checkpoint
+mm_tensors = []
+last_checkpoint = None
+if projector_checkpoint_path is not None:
+    last_checkpoint, file_type = load_model(projector_checkpoint_path)
+    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+
+if len(mm_tensors) == 0:
+    if last_checkpoint is not None:
+        for k, v in last_checkpoint.items():
+            print(k)
+    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.")
+    print("No tensors found. Is this a LLaVA model?")
+    exit()
+
+print(f"Found {len(mm_tensors)} tensors to extract.")
+print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
+# projector = {name: checkpoint.[name].float() for name in mm_tensors}
+projector = {}
+for name in mm_tensors:
+    projector[name] = last_checkpoint[name].float()
+for name in first_mm_tensors:
+    projector[name] = first_checkpoint[name].float()
+
+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

@@ -19,10 +19,6 @@ 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:
@@ -39,8 +35,8 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
-torch.save(checkpoint, path)
+    torch.save(checkpoint, path)
 
 print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava.cpp

@@ -2,32 +2,296 @@
 #include "common.h"
 #include "llama.h"
 #include "llava.h"
+#include "base64.hpp"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
+#include <numeric>
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+struct clip_image_grid_shape {
+    int first;
+    int second;
+};
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
+    int original_width  = original_size.first;
+    int original_height = original_size.second;
+
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width  = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+/**
+ * @brief Get the anyres image grid shape object
+ *
+ * @param image_size
+ * @param grid_pinpoints
+ * @param image_patch_size
+ * @return <int, int>
+ */
+static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
+    /**
+        Conversion from gguf flat array to vector:
+        std::vector<std::pair<int, int>> possible_resolutions;
+        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+        }
+     */
+    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
+    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
+}
+
+// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+    struct {
+        struct ggml_tensor * newline;
+        struct ggml_context * ctx;
+    } model;
+
+    const int32_t image_size = clip_image_size(ctx_clip);
+    const int32_t patch_size = clip_patch_size(ctx_clip);
+
+    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
+
+    int num_patches_width  = grid_shape.first;  // grid 1-4
+    int num_patches_height = grid_shape.second; // grid 1-4
+
+    const size_t num_images = num_patches_width * num_patches_height + 1;
+
+    // TODO: size calculation is not calculated - it's only tens of MB
+    size_t ctx_size = 0;
+
+    {
+        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+    }
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
+    };
+
+    // Python reference code for full unpad:
+    /*
+        base_image_feature = image_feature[0]
+        image_feature = image_feature[1:]
+        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+        image_feature = unpad_image(image_feature, image_sizes[image_idx])
+        image_feature = torch.cat((
+            image_feature,
+            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
+        ), dim=-1)
+        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+    */
+    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
+    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
+    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
+    // Once all images are processed to prepended the base_image_features without any changes.
+
+    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
+    /*
+        image_feature = image_feature.view(2, 2, 24, 24, 4096)
+        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+        image_feature = image_feature.view(2, 24, 2, 24, 4096)
+        image_feature = image_feature.flatten(0, 3)
+
+        // Reshape to 4D tensor by merging the last two dimensions
+        image_feature = image_feature.view(2, 2, 24, 24*4096)
+        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
+        image_feature = image_feature.view(-1, 4096)
+    */
+
+    model.ctx = ggml_init(params);
+
+    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
+    model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
+    if (newline_tmp->backend != GGML_BACKEND_CPU) {
+        if (newline_tmp->buffer == NULL) {
+            printf("newline_tmp tensor buffer is NULL\n");
+        }
+        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
+    } else {
+        model.newline->data = newline_tmp->data;
+        if (model.newline->data == NULL) {
+            printf("newline_tmp tensor data is NULL\n");
+        }
+    }
+
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
+    // fill it with the image embeddings, ignoring the base
+    for (size_t i = 1; i < num_images; i++) {
+        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
+        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
+    }
+
+    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
+    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
+
+    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
+                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                num_patches_per_side,
+                                                                num_patches_width,
+                                                                num_patches_height,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
+    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
+    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
+    /**
+     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
+         image_feature = torch.cat((
+        image_feature,
+        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+    ), dim=-1)
+     *
+     */
+
+    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
+    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
+    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
+    ggml_build_forward_expand(gf, flatten);
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+
+    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
+    // append without newline tokens (default behavior in llava_arch when not using unpad ):
+    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+
+    // Debug: Test single segments
+    // Current findings: sending base image, sending a segment embedding all works similar to python
+    // However, permuted embeddings do not work yet (stride issue?)
+    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
+    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
+    // *n_img_pos_out=576;
+
+    ggml_free(model.ctx);
+    return true;
+}
 
-#include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    clip_image_f32 * img_res = clip_image_f32_init();
-    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
+    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
+    clip_image_f32_batch img_res_v;
+    img_res_v.size = 0;
+    img_res_v.data = nullptr;
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
-        clip_image_f32_free(img_res);
+        delete[] img_res_v.data;
         return false;
     }
 
-    *n_img_pos = clip_n_patches(ctx_clip);
-
     const int64_t t_img_enc_start_us = ggml_time_us();
-    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
-    clip_image_f32_free(img_res);
-    if (!encoded) {
-        fprintf(stderr, "Unable to encode image\n");
 
-        return false;
+    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
+
+    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+        // flat / default llava-1.5 type embedding
+        *n_img_pos = clip_n_patches(ctx_clip);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
+        delete[] img_res_v.data;
+        if (!encoded) {
+            fprintf(stderr, "Unable to encode image\n");
+
+            return false;
+        }
+    } else {
+        // spatial_unpad llava-1.6 type embedding
+        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            if (!encoded) {
+                fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        const int32_t * image_grid = clip_image_grid(ctx_clip);
+
+        std::vector<std::pair<int, int>> grid_pinpoints;
+        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
+        }
+
+        // free all img_res_v - not needed anymore
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
+
+        const int32_t image_size = clip_image_size(ctx_clip);
+
+        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
+
+        int n_img_pos_out;
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        *n_img_pos = n_img_pos_out;
+
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+
+        // debug image/segment/normalization content:
+        // clip_image_u8 * tmp = clip_image_u8_init();
+        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
+        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
+    printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
@@ -48,10 +312,9 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }
 
 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));
+    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");
-        free(image_embd);
         return false;
     }
 
@@ -85,7 +348,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
     return true;
 }
 
-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) {
+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) {
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
@@ -142,7 +405,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     return true;
 }
 
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
     unsigned char* image_bytes;
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
@@ -151,13 +414,13 @@ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct
         return NULL;
     }
 
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
     free(image_bytes);
 
     return embed;
 }
 
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
+void llava_image_embed_free(struct llava_image_embed * embed) {
     free(embed->embed);
     free(embed);
 }

examples/llava/llava.h

@@ -3,7 +3,6 @@
 
 #include "ggml.h"
 
-
 #ifdef LLAMA_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
@@ -42,7 +41,6 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
 
-
 #ifdef __cplusplus
 }
 #endif

examples/llava/requirements.txt

@@ -0,0 +1,3 @@
+-r ../../requirements/requirements-convert.txt
+pillow~=10.2.0
+torch~=2.1.1

examples/lookahead/lookahead.cpp

@@ -54,7 +54,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/lookup/lookup.cpp

@@ -1,7 +1,9 @@
 #include "common.h"
+#include "ggml.h"
 #include "llama.h"
 
 #include <cmath>
+#include <cstdint>
 #include <cstdio>
 #include <string>
 #include <vector>
@@ -29,7 +31,8 @@ int main(int argc, char ** argv){
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;
@@ -73,6 +76,8 @@ int main(int argc, char ** argv){
     int n_drafted = 0;
     int n_accept  = 0;
 
+    int64_t t_draft_us = 0;
+
     int n_past = inp.size();
 
     bool has_eos = false;
@@ -160,7 +165,7 @@ int main(int argc, char ** argv){
 
         // generate n_pred tokens through prompt lookup
         auto prompt_lookup = [&]() -> void {
-            int inp_size = inp.size();
+            const int inp_size = inp.size();
             for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
                 const llama_token * ngram = &inp[inp_size - ngram_size];
 
@@ -191,8 +196,12 @@ int main(int argc, char ** argv){
             return;
         };
 
+        const int64_t t_start_draft_us = ggml_time_us();
+
         prompt_lookup();
 
+        t_draft_us += ggml_time_us() - t_start_draft_us;
+
         llama_decode(ctx, batch_tgt);
         ++n_past;
 
@@ -210,6 +219,8 @@ int main(int argc, char ** argv){
     LOG_TEE("n_draft   = %d\n", n_draft);
     LOG_TEE("n_predict = %d\n", n_predict);
     LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("t_draft   = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
     LOG_TEE("n_accept  = %d\n", n_accept);
     LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
 

examples/main/README.md

@@ -283,7 +283,11 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### NUMA support
 
--   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
+-   `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
+-   `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
+-   `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitraty core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
+
+ These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
 ### Memory Float 32
 

examples/main/main.cpp

@@ -98,7 +98,7 @@ static void write_logfile(
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
-        if (!is_interacting) {
+        if (!is_interacting && g_params->interactive) {
             is_interacting = true;
         } else {
             console::cleanup();
@@ -185,7 +185,8 @@ int main(int argc, char ** argv) {
     }
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -392,7 +393,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
     }
 
-    if (params.interactive) {
+    // ctrl+C handling
+    {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
         struct sigaction sigint_action;
         sigint_action.sa_handler = sigint_handler;
@@ -405,7 +407,9 @@ int main(int argc, char ** argv) {
         };
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
+    }
 
+    if (params.interactive) {
         LOG_TEE("%s: interactive mode on.\n", __func__);
 
         if (!params.antiprompt.empty()) {

examples/parallel/parallel.cpp

@@ -122,7 +122,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/passkey/passkey.cpp

@@ -71,7 +71,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 

examples/perplexity/perplexity.cpp

@@ -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 %d, while the current context is %d. Increase it with -c and retry\n",
+        fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
                 __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
     }
 
@@ -1809,7 +1809,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/quantize/quantize.cpp

@@ -23,6 +23,7 @@ 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",            },
@@ -237,7 +238,7 @@ int main(int argc, char ** argv) {
         params.imatrix = &imatrix_data;
     }
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // parse command line arguments
     const std::string fname_inp = argv[arg_idx];
@@ -287,9 +288,10 @@ 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) && 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 || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
         fprintf(stderr, "\n===============================================================================================\n");
-        fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\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, "===============================================================================================\n\n\n");
         return 1;
     }

examples/server/README.md

@@ -16,6 +16,13 @@ Command line options:
 - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+- `--numa STRATEGY`: Attempt one of the below optimization strategies  that help on some NUMA systems
+- `--numa distribute`: Spread execution evenly over all nodes
+- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
+- `--numa numactl`: Use the CPU map provided by numactl
+if run without this previously, it is recommended to drop the system page cache before using this
+see https://github.com/ggerganov/llama.cpp/issues/1437
+
 - `--numa`: Attempt optimizations that help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
@@ -185,7 +192,7 @@ node index.js
 
     `ignore_eos`: Ignore end of stream token and continue generating (default: false).
 
-    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
+    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. (default: []).
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
@@ -197,6 +204,8 @@ node index.js
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. (default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values)
+
 ### Result JSON
 
 - Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.

examples/server/oai.hpp

@@ -15,9 +15,13 @@
 using json = nlohmann::json;
 
 inline static json oaicompat_completion_params_parse(
-    const json &body /* openai api json semantics */)
+    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;
 
@@ -30,7 +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_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    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/server.cpp

@@ -36,6 +36,7 @@ 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 = "chatml";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
@@ -435,10 +436,6 @@ struct llama_server_context
         default_generation_settings_for_props["seed"] = -1;
 
         batch = llama_batch_init(n_ctx, 0, params.n_parallel);
-
-        // empty system prompt
-        system_prompt = "";
-        system_tokens.clear();
     }
 
     std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
@@ -625,18 +622,36 @@ struct llama_server_context
             const int n_vocab = llama_n_vocab(model);
             for (const auto &el : *logit_bias)
             {
-                if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
+                if (el.is_array() && el.size() == 2)
                 {
-                    llama_token tok = el[0].get<llama_token>();
-                    if (tok >= 0 && tok < n_vocab)
+                    float bias;
+                    if (el[1].is_number())
                     {
-                        if (el[1].is_number())
+                        bias = el[1].get<float>();
+                    }
+                    else if (el[1].is_boolean() && !el[1].get<bool>())
+                    {
+                        bias = -INFINITY;
+                    }
+                    else
+                    {
+                        continue;
+                    }
+
+                    if (el[0].is_number_integer())
+                    {
+                        llama_token tok = el[0].get<llama_token>();
+                        if (tok >= 0 && tok < n_vocab)
                         {
-                            slot->sparams.logit_bias[tok] = el[1].get<float>();
+                            slot->sparams.logit_bias[tok] = bias;
                         }
-                        else if (el[1].is_boolean() && !el[1].get<bool>())
+                    }
+                    else if (el[0].is_string())
+                    {
+                        auto toks = llama_tokenize(model, el[0].get<std::string>(), false);
+                        for (auto tok : toks)
                         {
-                            slot->sparams.logit_bias[tok] = -INFINITY;
+                            slot->sparams.logit_bias[tok] = bias;
                         }
                     }
                 }
@@ -657,6 +672,24 @@ struct llama_server_context
             }
         }
 
+        const auto &samplers_sequence = data.find("samplers");
+        if (samplers_sequence != data.end() && samplers_sequence->is_array())
+        {
+            std::vector<std::string> sampler_names;
+            for (const auto &sampler_name : *samplers_sequence)
+            {
+                if (sampler_name.is_string())
+                {
+                    sampler_names.emplace_back(sampler_name);
+                }
+            }
+            slot->sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+        }
+        else
+        {
+            slot->sparams.samplers_sequence = default_sparams.samplers_sequence;
+        }
+
         if (multimodal)
         {
             const auto &images_data = data.find("image_data");
@@ -746,27 +779,30 @@ struct llama_server_context
     }
 
     void update_system_prompt() {
-        system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
-
-        llama_batch_clear(batch);
-
         kv_cache_clear();
+        system_tokens.clear();
 
-        for (int i = 0; i < (int) system_tokens.size(); ++i)
-        {
-            llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
-        }
+        if (!system_prompt.empty()) {
+            system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
 
-        if (llama_decode(ctx, batch) != 0)
-        {
-            LOG_TEE("%s: llama_decode() failed\n", __func__);
-            return;
-        }
+            llama_batch_clear(batch);
 
-        // assign the system KV cache to all parallel sequences
-        for (int32_t i = 1; i < params.n_parallel; ++i)
-        {
-            llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
+            for (int i = 0; i < (int)system_tokens.size(); ++i)
+            {
+                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
+            }
+
+            if (llama_decode(ctx, batch) != 0)
+            {
+                LOG_TEE("%s: llama_decode() failed\n", __func__);
+                return;
+            }
+
+            // assign the system KV cache to all parallel sequences
+            for (int32_t i = 1; i < params.n_parallel; ++i)
+            {
+                llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
+            }
         }
 
         LOG_TEE("system prompt updated\n");
@@ -788,10 +824,8 @@ struct llama_server_context
         name_user      = sys_props.value("anti_prompt", "");
         name_assistant = sys_props.value("assistant_name", "");
 
-        if (slots.size() > 0)
-        {
-            notify_system_prompt_changed();
-        }
+
+        notify_system_prompt_changed();
     }
 
     static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
@@ -949,18 +983,31 @@ struct llama_server_context
             {
                 continue;
             }
-            clip_image_f32 * img_res = clip_image_f32_init();
-            if (!clip_image_preprocess(clp_ctx, img.img_data, img_res, /*pad2square =*/ true))
+            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;
             }
@@ -968,9 +1015,12 @@ struct llama_server_context
             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_free(img_res);
+
+            clip_image_f32_batch_free(img_res_v);
+
             img.request_encode_image = false;
         }
 
@@ -994,6 +1044,12 @@ struct llama_server_context
         const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model));
         const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() &&
                                 eos_bias->second < 0.0f && std::isinf(eos_bias->second);
+        std::vector<std::string> samplers_sequence;
+        for (const auto &sampler_type : slot.sparams.samplers_sequence)
+        {
+            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+        }
+
         return json {
             {"n_ctx",             slot.n_ctx},
             {"model",             params.model_alias},
@@ -1024,6 +1080,7 @@ struct llama_server_context
             {"logit_bias",        slot.sparams.logit_bias},
             {"n_probs",           slot.sparams.n_probs},
             {"grammar",           slot.sparams.grammar},
+            {"samplers",          samplers_sequence}
         };
     }
 
@@ -1592,10 +1649,6 @@ struct llama_server_context
                         LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed);
                     }
 
-                    LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
-
-                    llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
-
                     slot.cache_tokens = prompt_tokens;
 
                     if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
@@ -1609,6 +1662,10 @@ struct llama_server_context
                         }
                     }
 
+                    LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
+
+                    llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
+
                     LOG_VERBOSE("prompt ingested", {
                                                     {"n_past",  slot.n_past},
                                                     {"cached",  tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)},
@@ -1820,7 +1877,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     {
         printf("  --no-mmap                 do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    printf("  --numa                    attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa TYPE               attempt optimizations that help on some NUMA systems\n");
+    printf("                              - distribute: spread execution evenly over all nodes\n");
+    printf("                              - isolate: only spawn threads on CPUs on the node that execution started on\n");
+    printf("                              - numactl: use the CPU map provided my numactl\n");
     if (llama_supports_gpu_offload()) {
         printf("  -ngl N, --n-gpu-layers N\n");
         printf("                            number of layers to store in VRAM\n");
@@ -1859,6 +1919,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     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 FORMAT_NAME");
+    printf("                            set chat template, possible valus is: llama2, chatml (default %s)", sparams.chat_template.c_str());
     printf("\n");
 }
 
@@ -2227,9 +2289,17 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.use_mmap = false;
         }
-        else if (arg == "--numa")
-        {
-            params.numa = true;
+        else if (arg == "--numa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            } else {
+                std::string value(argv[i]);
+                /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+                else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+                else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+                else { invalid_param = true; break; }
+            }
         }
         else if (arg == "--embedding")
         {
@@ -2290,6 +2360,21 @@ 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 == "--chat-template")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            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 = value;
+        }
         else if (arg == "--override-kv")
         {
             if (++i >= argc) {
@@ -2445,7 +2530,8 @@ int main(int argc, char **argv)
         params.model_alias = params.model;
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER},
                             {"commit", LLAMA_COMMIT}});
@@ -2743,13 +2829,13 @@ int main(int argc, char **argv)
 
 
     // TODO: add mount point without "/v1" prefix -- how?
-    svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/v1/chat/completions", [&llama, &validate_api_key, &sparams](const httplib::Request &req, httplib::Response &res)
             {
                 res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
-                json data = oaicompat_completion_params_parse(json::parse(req.body));
+                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);

examples/server/utils.hpp

@@ -167,6 +167,34 @@ static T json_value(const json &body, const std::string &key, const T &default_v
         : default_value;
 }
 
+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();
+}
+
 inline std::string format_chatml(std::vector<json> messages)
 {
     std::ostringstream chatml_msgs;
@@ -180,6 +208,8 @@ inline std::string format_chatml(std::vector<json> messages)
 
     chatml_msgs << "<|im_start|>assistant" << '\n';
 
+    LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
+
     return chatml_msgs.str();
 }
 

examples/simple/simple.cpp

@@ -31,7 +31,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 

examples/speculative/speculative.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model_tgt = NULL;
     llama_model * model_dft = NULL;

examples/tokenize/tokenize.cpp

@@ -17,7 +17,7 @@ int main(int argc, char ** argv) {
 
     const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();
     model_params.vocab_only = true;

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

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 #include "common.h"
 #include "train.h"
 #include "llama.h"
@@ -19,8 +20,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const size_t tensor_alignment = 32;
-
 struct my_llama_hparams {
     uint32_t n_vocab = 32000;
     uint32_t n_ctx   = 512;
@@ -51,14 +50,14 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
     struct ggml_context * ctx = NULL;
-    std::vector<uint8_t> data;
+    ggml_backend_buffer_t data = NULL;
 
     my_llama_hparams hparams;
 
@@ -112,13 +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: %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);
+    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);
 }
 
 static void set_param_model(struct my_llama_model * model) {
@@ -141,42 +140,9 @@ static void set_param_model(struct my_llama_model * model) {
         ggml_set_param(ctx, layer.wv);
         ggml_set_param(ctx, layer.wo);
         ggml_set_param(ctx, layer.ffn_norm);
-        ggml_set_param(ctx, layer.w1);
-        ggml_set_param(ctx, layer.w2);
-        ggml_set_param(ctx, layer.w3);
-    }
-}
-
-static void alloc_model(struct ggml_allocr * alloc, struct my_llama_model * model) {
-    ggml_allocr_alloc(alloc, model->tok_embeddings);
-    ggml_allocr_alloc(alloc, model->norm);
-    ggml_allocr_alloc(alloc, model->output);
-    for (uint32_t i = 0; i < model->layers.size(); ++i) {
-        auto & layer = model->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm);
-        ggml_allocr_alloc(alloc, layer.wq);
-        ggml_allocr_alloc(alloc, layer.wk);
-        ggml_allocr_alloc(alloc, layer.wv);
-        ggml_allocr_alloc(alloc, layer.wo);
-        ggml_allocr_alloc(alloc, layer.ffn_norm);
-        ggml_allocr_alloc(alloc, layer.w1);
-        ggml_allocr_alloc(alloc, layer.w2);
-        ggml_allocr_alloc(alloc, layer.w3);
-    }
-    ggml_allocr_alloc(alloc, model->tok_embeddings->grad);
-    ggml_allocr_alloc(alloc, model->norm->grad);
-    ggml_allocr_alloc(alloc, model->output->grad);
-    for (uint32_t i = 0; i < model->layers.size(); ++i) {
-        auto & layer = model->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm->grad);
-        ggml_allocr_alloc(alloc, layer.wq->grad);
-        ggml_allocr_alloc(alloc, layer.wk->grad);
-        ggml_allocr_alloc(alloc, layer.wv->grad);
-        ggml_allocr_alloc(alloc, layer.wo->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm->grad);
-        ggml_allocr_alloc(alloc, layer.w1->grad);
-        ggml_allocr_alloc(alloc, layer.w2->grad);
-        ggml_allocr_alloc(alloc, layer.w3->grad);
+        ggml_set_param(ctx, layer.ffn_gate);
+        ggml_set_param(ctx, layer.ffn_down);
+        ggml_set_param(ctx, layer.ffn_up);
     }
 }
 
@@ -232,9 +198,9 @@ static void init_model(struct my_llama_model * model) {
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
-        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
-        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
-        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_gate = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
+        layer.ffn_up   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
 
         ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
 
@@ -245,24 +211,15 @@ static void init_model(struct my_llama_model * model) {
 
         ggml_set_name(layer.ffn_norm,       tni(LLM_TENSOR_FFN_NORM, i));
 
-        ggml_set_name(layer.w1,             tni(LLM_TENSOR_FFN_GATE, i));
-        ggml_set_name(layer.w2,             tni(LLM_TENSOR_FFN_DOWN, i));
-        ggml_set_name(layer.w3,             tni(LLM_TENSOR_FFN_UP, i));
+        ggml_set_name(layer.ffn_gate,       tni(LLM_TENSOR_FFN_GATE, i));
+        ggml_set_name(layer.ffn_down,       tni(LLM_TENSOR_FFN_DOWN, i));
+        ggml_set_name(layer.ffn_up,         tni(LLM_TENSOR_FFN_UP, i));
     }
 
     set_param_model(model);
 
-    // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
-
     // allocate data
-    struct ggml_allocr * alloc = NULL;
-    model->data.resize(size + tensor_alignment);
-    alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
-    alloc_model(alloc, model);
+    model->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type());
 }
 
 static void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) {
@@ -287,9 +244,9 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
 
         randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, rnd);
-        randomize_tensor_normal(layer.w2, rnd);
-        randomize_tensor_normal(layer.w3, rnd);
+        randomize_tensor_normal(layer.ffn_gate, rnd);
+        randomize_tensor_normal(layer.ffn_down, rnd);
+        randomize_tensor_normal(layer.ffn_up,   rnd);
     }
 
     free_random_normal_distribution(rnd);
@@ -297,7 +254,7 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
 
 static struct ggml_tensor * llama_build_train_graphs(
         struct my_llama_model * model,
-        struct ggml_allocr    * alloc,
+        ggml_gallocr_t          alloc,
         struct ggml_context   * ctx,
         struct ggml_cgraph    * gf,
         struct ggml_cgraph    * gb,
@@ -308,7 +265,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         const  int              n_tokens,
         const  int              n_batch,
         const  bool             enable_flash_attn,
-        const  bool             enable_checkpointing) {
+        const  bool             enable_checkpointing,
+        const  bool             measure_only) {
 
     ggml_set_scratch(ctx, { 0, 0, nullptr, });
     const int n_past = 0;
@@ -334,13 +292,7 @@ static struct ggml_tensor * llama_build_train_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    ggml_allocr_alloc(alloc, KQ_pos);
-    if (!ggml_allocr_is_measure(alloc)) {
-        int * data = (int *) KQ_pos->data;
-        for (int i = 0; i < N; ++i) {
-            data[i] = n_past + i;
-        }
-    }
+    ggml_set_input(KQ_pos);
 
     // rope has so much parameters that we make a custom function for it
     auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
@@ -404,11 +356,11 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, f_norm_rms_eps);                    set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, layer.ffn_norm, t22);                    set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                               set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.w3, t24);                          set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.w1, t24);                          set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.ffn_up, t24);                      set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.ffn_gate, t24);                    set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                    set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                               set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.w2, t28);                          set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.ffn_down, t28);                    set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                               set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         checkpoints.push_back(cur);
@@ -448,21 +400,31 @@ static struct ggml_tensor * llama_build_train_graphs(
         // KQ_pos
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
         GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
-
-        ggml_allocr_alloc(alloc, t36->grad);
+        ggml_set_input(t36->grad);
 
         // allocating checkpoints in one block to reduce memory fragmentation
         // note: they will be freed in reverse order
         for (int i = 0; i < (int) checkpoints.size(); ++i) {
             if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
-                ggml_allocr_alloc(alloc, checkpoints[i]);
+                ggml_set_input(checkpoints[i]);
             }
         }
 
         //int n_leafs_after = gb->n_leafs;
         //int n_nodes_after = gb->n_nodes;
+        if (measure_only) {
+            // FIXME: will still allocate
+            ggml_gallocr_reserve(alloc, gb);
+        } else {
+            ggml_gallocr_alloc_graph(alloc, gb);
 
-        ggml_allocr_alloc_graph(alloc, gb);
+            if (!measure_only) {
+                int * data = (int *) KQ_pos->data;
+                for (int i = 0; i < N; ++i) {
+                    data[i] = n_past + i;
+                }
+            }
+        }
 
         // remove the additional nodes and leafs
         for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
@@ -559,9 +521,9 @@ static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_contex
         copy_tensor_by_name(layer.wv,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
         copy_tensor_by_name(layer.wo,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
         copy_tensor_by_name(layer.ffn_norm,       f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
-        copy_tensor_by_name(layer.w1,             f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
-        copy_tensor_by_name(layer.w2,             f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
-        copy_tensor_by_name(layer.w3,             f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
+        copy_tensor_by_name(layer.ffn_gate,       f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
+        copy_tensor_by_name(layer.ffn_down,       f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
+        copy_tensor_by_name(layer.ffn_up,         f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
     }
 }
 
@@ -702,9 +664,9 @@ static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vo
         gguf_add_tensor(fctx, layer.wv);
         gguf_add_tensor(fctx, layer.wo);
         gguf_add_tensor(fctx, layer.ffn_norm);
-        gguf_add_tensor(fctx, layer.w1);
-        gguf_add_tensor(fctx, layer.w2);
-        gguf_add_tensor(fctx, layer.w3);
+        gguf_add_tensor(fctx, layer.ffn_gate);
+        gguf_add_tensor(fctx, layer.ffn_down);
+        gguf_add_tensor(fctx, layer.ffn_up);
     }
 }
 
@@ -953,9 +915,9 @@ static int64_t get_parameter_count(struct my_llama_model* model) {
         nx += ggml_nelements(layer.wv);
         nx += ggml_nelements(layer.wo);
         nx += ggml_nelements(layer.ffn_norm);
-        nx += ggml_nelements(layer.w1);
-        nx += ggml_nelements(layer.w2);
-        nx += ggml_nelements(layer.w3);
+        nx += ggml_nelements(layer.ffn_gate);
+        nx += ggml_nelements(layer.ffn_down);
+        nx += ggml_nelements(layer.ffn_up);
     }
     return nx;
 }
@@ -1046,7 +1008,7 @@ int main(int argc, char ** argv) {
     printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
     printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
     printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
-    printf("%s: model_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(model.ctx) + model.data.size()), (float) (ggml_used_mem(model.ctx) + model.data.size()) / (1024.0f*1024.0f));
+    printf("%s: model_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(model.ctx) + ggml_backend_buffer_get_size(model.data)), (float) (ggml_used_mem(model.ctx) + ggml_backend_buffer_get_size(model.data)) / (1024.0f*1024.0f));
 
     if (params.only_write_model) {
         save_train_files_data save_data;
@@ -1073,11 +1035,6 @@ int main(int argc, char ** argv) {
     int n_vocab  = model.hparams.n_vocab;
     int n_batch  = params.common.n_batch;
 
-    std::vector<uint8_t> mem_input_data;
-    std::vector<uint8_t> mem_compute_data;
-
-    ggml_allocr * alloc = NULL;
-
     // context for input tensors without their data
     struct ggml_init_params ctx_input_params = {
         ggml_tensor_overhead() * 2, // mem_size
@@ -1091,16 +1048,10 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
-    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
-
     // allocate input tensors
-    mem_input_data.resize(max_input_size);
-    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc, tokens_input);
-    ggml_allocr_alloc(alloc, target_probs);
+    ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type());
+    size_t max_input_size = ggml_backend_buffer_get_size(input_data);
+    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // context for compute tensors without their data
     const size_t estimated_compute_size_wo_data = (
@@ -1127,7 +1078,7 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        alloc = ggml_allocr_new_measure(tensor_alignment);
+        ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
         gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
         gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1140,9 +1091,10 @@ int main(int argc, char ** argv) {
             &logits, tokens_input, target_probs,
             n_tokens, n_batch,
             params.common.use_flash,
-            params.common.use_checkpointing
+            params.common.use_checkpointing,
+            true
         );
-        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer
         if (max_compute_size < best_compute_size) {
             best_compute_size = max_compute_size;
             best_order = gf->order;
@@ -1157,9 +1109,8 @@ int main(int argc, char ** argv) {
         "invalid");
 
     // allocate compute tensors
-    mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
     gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1172,7 +1123,8 @@ int main(int argc, char ** argv) {
         &logits, tokens_input, target_probs,
         n_tokens, n_batch,
         params.common.use_flash,
-        params.common.use_checkpointing
+        params.common.use_checkpointing,
+        false
     );
 
     std::vector<llama_token> train_tokens;

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1706732774,
-        "narHash": "sha256-hqJlyJk4MRpcItGYMF+3uHe8HvxNETWvlGtLuVpqLU0=",
+        "lastModified": 1708118438,
+        "narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "b8b232ae7b8b144397fdb12d20f592e5e7c1a64d",
+        "rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
         "type": "github"
       },
       "original": {

ggml-alloc.h

@@ -6,88 +6,62 @@
 extern "C" {
 #endif
 
-struct ggml_backend;
-struct ggml_backend_buffer;
-struct ggml_backend_buffer_type;
-
-//
-// Legacy API
-//
-
-typedef struct ggml_allocr * ggml_allocr_t;
-
-// initialize allocator for use with CPU backend only
-GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
-
-// initialize allocator for use with ggml-backend
-GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
-
-// tell the allocator to parse nodes following the order described in the list
-// you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
-
-GGML_API void   ggml_allocr_free       (ggml_allocr_t alloc);
-GGML_API bool   ggml_allocr_is_measure (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_reset      (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_alloc      (ggml_allocr_t alloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_allocr_max_size   (ggml_allocr_t alloc);
-
-GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
-
-//
-// ggml-backend v2 API
-//
-
-// Separate tensor and graph allocator objects
-// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
-// The original API is kept as a wrapper around the new API
+typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct ggml_backend * ggml_backend_t;
 
 // Tensor allocator
 typedef struct ggml_tallocr * ggml_tallocr_t;
 
-GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
-
-GGML_API void   ggml_tallocr_free       (ggml_tallocr_t talloc);
-GGML_API bool   ggml_tallocr_is_measure (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_reset      (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_alloc      (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_tallocr_max_size   (ggml_tallocr_t talloc);
-
+GGML_API ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void           ggml_tallocr_free(ggml_tallocr_t talloc);
+GGML_API void           ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor);
 
 // Graph allocator
+/*
+  Example usage:
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct ggml_cgraph * graph = build_graph(batch);
+    ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   ggml_set_output(): output tensors are never freed and never overwritten
+
 typedef struct ggml_gallocr * ggml_gallocr_t;
 
-GGML_API ggml_gallocr_t ggml_gallocr_new(void);
-GGML_API void   ggml_gallocr_free(ggml_gallocr_t galloc);
+GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
+GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
+GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 
-GGML_API void   ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
-GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids);
 
-// Allocate tensors from the allocators given by the hash table
-GGML_API void   ggml_gallocr_alloc_graph_n(
-                    ggml_gallocr_t galloc,
-                    struct ggml_cgraph * graph,
-                    struct ggml_hash_set hash_set,
-                    ggml_tallocr_t * hash_node_talloc);
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
+GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
 
+GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
 
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
 
 #ifdef  __cplusplus
 }

ggml-backend.h

@@ -83,8 +83,9 @@ extern "C" {
 
     GGML_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    GGML_API GGML_CALL bool ggml_backend_is_cpu           (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
     // Create a backend buffer from an existing pointer
     GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
@@ -129,11 +130,7 @@ extern "C" {
 
         // in build_graph:
         build_graph(...) {
-            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
-            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
-            ggml_allocr_alloc(alloc_cpu, tensor);
-
-            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            // manually assign nodes to a backend (optional, should not be needed in most cases)
             struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
             ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
         }
@@ -163,20 +160,19 @@ extern "C" {
     GGML_API ggml_backend_sched_t  ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
     GGML_API void                  ggml_backend_sched_free(ggml_backend_sched_t sched);
     // Initialize backend buffers from a measure graph
-    GGML_API void                  ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API bool                  ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
     // Get the number of splits of the last graph
     GGML_API int                   ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
 
-    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API size_t                ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
 
     GGML_API void                  ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
     GGML_API ggml_backend_t        ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
     // Allocate and compute graph on the backend scheduler
-    GGML_API void                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API bool                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
 
-    // Reset all assignments and allocators - must be called before using the sched allocators to allocate inputs
+    // Reset all assignments and allocators - must be called before changing the node backends
     GGML_API void                  ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute

ggml-cuda.cu

@@ -150,8 +150,8 @@
 #define CUDA_USE_TENSOR_CORES
 #endif
 
-// max batch size to use MMQ kernels when tensor cores are available
-#define MMQ_MAX_BATCH_SIZE 32
+#define MMVQ_MAX_BATCH_SIZE  8 // max batch size to use MMVQ kernels
+#define  MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available
 
 #if defined(GGML_USE_HIPBLAS)
 #define __CUDA_ARCH__ 1300
@@ -517,6 +517,15 @@ 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 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
 
@@ -1681,6 +1690,137 @@ 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,
@@ -1823,6 +1963,29 @@ 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 __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");
@@ -4522,6 +4685,49 @@ 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
+}
+
 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(
@@ -5310,51 +5516,59 @@ template <bool need_check> static __global__ void
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
-#define MMVQ_NWARPS_NVIDIA    4
-#define MMVQ_NWARPS_AMD_RDNA2 1
-#define MMVQ_NWARPS_AMD_OLD   4
-
-template <int nwarps, int ncols_y_template, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
+template <int ncols_y, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1) // tells the compiler to use as many registers as it wants
+// tell the compiler to use as many registers as it wants, see nwarps definition below
+__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void mul_mat_vec_q(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y_par, const int nrows_dst) {
+    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
 
-    const int ncols_y = ncols_y_template != 0 ? ncols_y_template : ncols_y_par;
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+    constexpr int nwarps              = 1;
+    constexpr int rows_per_cuda_block = 1;
+#else
+    constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+    constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
 
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    const int row = blockIdx.x;
-
-    const int blocks_per_row_x = ncols_x / qk;
-    const int blocks_per_col_y = nrows_y / QK8_1;
-    const int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
+    const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    const     int row0 = rows_per_cuda_block*blockIdx.x;
+    const     int blocks_per_row_x = ncols_x / qk;
+    const     int blocks_per_col_y = nrows_y / QK8_1;
+    constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
 
 // partial sum for each thread
-    float tmp[ncols_y_template != 0 ? ncols_y_template : 8] = {0.0f};
+    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
 
     const block_q_t  * x = (const block_q_t  *) vx;
     const block_q8_1 * y = (const block_q8_1 *) vy;
 
-    for (int i = tid / (qi/vdr); i < blocks_per_row_x; i += blocks_per_iter) {
-        const int ibx = row*blocks_per_row_x + i; // x block index
+    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
 
-        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
-
-        const int iqs  = vdr * (tid % (qi/vdr)); // x block quant index when casting the quants to int
+        // x block quant index when casting the quants to int
+        const int kqs = vdr * (tid % (qi/vdr));
 
 #pragma unroll
         for (int j = 0; j < ncols_y; ++j) {
-            tmp[j] += vec_dot_q_cuda(&x[ibx], &y[j*blocks_per_col_y + iby], iqs);
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp[j][i] += vec_dot_q_cuda(
+                    &x[kbx + (row0 + i)*blocks_per_row_x], &y[j*blocks_per_col_y + kby], kqs);
+            }
         }
     }
 
-    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y_template != 0 ? ncols_y_template : 8][WARP_SIZE];
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
     if (threadIdx.y > 0) {
 #pragma unroll
         for (int j = 0; j < ncols_y; ++j) {
-            tmp_shared[threadIdx.y-1][j][threadIdx.x] = tmp[j];
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+            }
         }
     }
     __syncthreads();
@@ -5366,13 +5580,16 @@ static __global__ void mul_mat_vec_q(
 #pragma unroll
     for (int j = 0; j < ncols_y; ++j) {
 #pragma unroll
-        for (int i = 0; i < nwarps-1; ++i) {
-            tmp[j] += tmp_shared[i][j][threadIdx.x];
+        for (int i = 0; i < rows_per_cuda_block; ++i) {
+#pragma unroll
+            for (int l = 0; l < nwarps-1; ++l) {
+                tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+            }
+            tmp[j][i] = warp_reduce_sum(tmp[j][i]);
         }
-        tmp[j] = warp_reduce_sum(tmp[j]);
 
-        if (threadIdx.x == 0) {
-            dst[j*nrows_dst + row] = tmp[j];
+        if (threadIdx.x < rows_per_cuda_block) {
+            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }
 }
@@ -5945,149 +6162,31 @@ 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, 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 (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__ 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
-
-    half2 max_val = make_half2(-INFINITY, -INFINITY);
-
-#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 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) {
+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;
 
     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 rowy = rowx % nrows_y; // broadcast the mask 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;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = rowx/nrows_y; // head index
+
+        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 = powf(base, exp);
+    }
+
     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:
@@ -6106,7 +6205,8 @@ static __global__ void soft_max_f32(const float * x, const float * y, float * ds
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
 
-        const float val = x[ix]*scale + (y ? y[iy] : 0.0f);
+        const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + slope*pos[col];
+
         vals[col] = val;
         max_val = max(max_val, val);
     }
@@ -6667,6 +6767,12 @@ 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 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;
@@ -6706,6 +6812,8 @@ 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_F32:
             return convert_unary_cuda<float>;
         default:
@@ -6741,6 +6849,8 @@ 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_F16:
             return convert_unary_cuda<half>;
         default:
@@ -6851,65 +6961,75 @@ static void mul_mat_vec_q_cuda(
     const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % qk == 0);
-    GGML_ASSERT(ncols_y <= 4);
+    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
 
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
 
-    int nwarps;
-    if (g_device_caps[id].cc >= CC_OFFSET_AMD) {
-        nwarps = g_device_caps[id].cc >= CC_RDNA2 ? MMVQ_NWARPS_AMD_RDNA2 : MMVQ_NWARPS_AMD_OLD;
-    } else {
-        nwarps = MMVQ_NWARPS_NVIDIA;
-    }
+    int64_t nwarps = 1;
+    int64_t rows_per_cuda_block = 1;
 
-    const dim3 block_nums(nrows_x, 1, 1);
+    if (g_device_caps[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
+        switch(ncols_y) {
+            case 1:
+                nwarps = 4;
+                rows_per_cuda_block = 1;
+                break;
+            case 2:
+            case 3:
+            case 4:
+                nwarps = 4;
+                rows_per_cuda_block = 2;
+                break;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                nwarps = 2;
+                rows_per_cuda_block = 2;
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+    }
+    const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
+    const dim3 block_nums(nblocks, 1, 1);
     const dim3 block_dims(WARP_SIZE, nwarps, 1);
 
-    switch (nwarps) {
-        case 1: switch(ncols_y) {
-            case 1:
-                mul_mat_vec_q<1, 1, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            case 2:
-                mul_mat_vec_q<1, 2, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            case 3:
-                mul_mat_vec_q<1, 3, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            case 4:
-                mul_mat_vec_q<1, 4, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        } break;
-        case 4: switch(ncols_y) {
-            case 1:
-                mul_mat_vec_q<4, 1, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            case 2:
-                mul_mat_vec_q<4, 2, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            case 3:
-                mul_mat_vec_q<4, 3, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            case 4:
-                mul_mat_vec_q<4, 4, qk, qi, block_q_t, vdr, vec_dot>
-                    <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        } break;
-
+    switch (ncols_y) {
+        case 1:
+            mul_mat_vec_q<1, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 2:
+            mul_mat_vec_q<2, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 3:
+            mul_mat_vec_q<3, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 4:
+            mul_mat_vec_q<4, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 5:
+            mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 6:
+            mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 7:
+            mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 8:
+            mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -7568,89 +7688,53 @@ 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_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) {
+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) {
     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, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 64:
-                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 128:
-                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 256:
-                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 512:
-                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 break;
             default:
-                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                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);
                 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, y, dst, ncols_x, nrows_y, scale);
+        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);
     }
 }
 
@@ -7922,6 +8006,7 @@ GGML_CALL void ggml_init_cublas() {
         if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
             initialized = true;
             g_cublas_loaded = false;
+            fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
             return;
         }
 
@@ -8509,6 +8594,7 @@ 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:
             return max_compute_capability >= CC_RDNA2 ? 128 : 64;
         default:
             GGML_ASSERT(false);
@@ -8532,6 +8618,7 @@ 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:
             return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
@@ -8629,6 +8716,10 @@ 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;
         default:
             GGML_ASSERT(false);
             break;
@@ -9068,30 +9159,36 @@ 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 = src1 ? ggml_nrows(src1) : 1;
+    const int64_t nrows_y = src0->ne[1];
 
-    float scale = 1.0f;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
 
-#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
+    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
-    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);
+    // positions tensor
+    float * src2_dd = dst_dd; // default to avoid null checks in the kernel
+    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 = use_src2 && src2->backend == GGML_BACKEND_GPU;
+        ggml_tensor_extra_gpu * src2_extra = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr;
+
+        if (src2_on_device) {
+            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));
+        }
     }
 
-    (void) dst;
+    soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream);
 }
 
 static void ggml_cuda_op_scale(
@@ -9735,7 +9832,7 @@ static __global__ void k_compute_batched_ptrs(
     ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
 }
 
-static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
@@ -9893,39 +9990,69 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
 
     int64_t min_compute_capability = INT_MAX;
 
+    bool any_pascal_with_slow_fp16 = false;
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
         auto & tensor_split = buft_ctx->tensor_split;
         for (int id = 0; id < g_device_count; ++id) {
-            if (min_compute_capability > g_device_caps[id].cc && tensor_split[id] < (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) {
+            // skip devices that are not going to do any work:
+            if (tensor_split[id] >= (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) {
+                continue;
+            }
+
+            if (min_compute_capability > g_device_caps[id].cc) {
                 min_compute_capability = g_device_caps[id].cc;
             }
+            if (g_device_caps[id].cc == 610) {
+                any_pascal_with_slow_fp16 = true;
+            }
         }
     } else {
-        min_compute_capability = g_device_caps[g_main_device].cc;
+        min_compute_capability    = g_device_caps[g_main_device].cc;
+        any_pascal_with_slow_fp16 = g_device_caps[g_main_device].cc == 610;
     }
 
+    // check data types and tensor shapes for custom matrix multiplication kernels:
+    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->ne[1] == 1;
+
+    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    bool              use_mul_mat_q =  ggml_cuda_supports_mmq(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
     const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
-    bool               use_mul_mat_q = ggml_is_quantized(src0->type);
+
 #ifdef CUDA_USE_TENSOR_CORES
     use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
 #endif // CUDA_USE_TENSOR_CORES
 
 #else
 
-    const bool fp16_performance_good = min_compute_capability >= CC_VOLTA;
-    bool               use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+    // fp16 performance is good on Volta or newer and on P100 (compute capability 6.0)
+    const bool fp16_performance_good = min_compute_capability >= CC_PASCAL && !any_pascal_with_slow_fp16;
+
+    // mmvq and mmq need the __dp4a instruction which on NVIDIA is only available for CC >= 6.1
+    use_mul_mat_vec_q = use_mul_mat_vec_q && min_compute_capability >= MIN_CC_DP4A;
+    use_mul_mat_q     = use_mul_mat_q     && min_compute_capability >= MIN_CC_DP4A;
+
 #ifdef CUDA_USE_TENSOR_CORES
     // when tensor cores are available, use them for large batch size
     // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-    use_mul_mat_q = use_mul_mat_q && !(fp16_performance_good && src1->ne[1] > MMQ_MAX_BATCH_SIZE);
+    use_mul_mat_q     = use_mul_mat_q     && (!fp16_performance_good || src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
 #endif // CUDA_USE_TENSOR_CORES
 
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
-    use_mul_mat_q = use_mul_mat_q && ggml_cuda_supports_mmq(src0->type);
+    // if mmvq is available it's a better choice than dmmv:
+#ifndef GGML_CUDA_FORCE_DMMV
+    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
+#endif // GGML_CUDA_FORCE_DMMV
 
     // debug helpers
     //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
@@ -9943,33 +10070,15 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
     } else if (!split && all_on_device && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
-        ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
-    } else if (src0->type == GGML_TYPE_F32) {
-        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
-    } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
-        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->type == GGML_TYPE_F32) {
-#ifdef GGML_CUDA_FORCE_DMMV
-            const bool use_mul_mat_vec_q = false;
-#else
-            const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
-#endif // GGML_CUDA_FORCE_DMMV
-
-            if (use_mul_mat_vec_q) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
-            } else {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
-            }
-        } else {
-            if (src1->ne[1] <= 4 && min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type) && src1->type == GGML_TYPE_F32) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
-            } else if (use_mul_mat_q) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
-            } else {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
-            }
-        }
+        ggml_cuda_mul_mat_batched_cublas(src0, src1, dst);
+    } else if (use_dequantize_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
+    } else if (use_mul_mat_vec_q) {
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
+    } else if (use_mul_mat_q) {
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
     } else {
-        GGML_ASSERT(false);
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     }
 }
 
@@ -11327,7 +11436,7 @@ 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) {
+                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) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }

ggml-metal.m

@@ -61,6 +61,7 @@ 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_I32,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@@ -83,6 +84,7 @@ 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_ID_F32_F32,
   //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@@ -101,6 +103,7 @@ 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_MM_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@@ -116,6 +119,7 @@ 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_ID_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@@ -131,6 +135,7 @@ 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_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
     GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@@ -176,7 +181,7 @@ 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
@@ -433,6 +438,7 @@ 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_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);
@@ -455,6 +461,7 @@ 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_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);
@@ -473,6 +480,7 @@ 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_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);
@@ -488,6 +496,7 @@ 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_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);
@@ -503,6 +512,7 @@ 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_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);
@@ -687,6 +697,7 @@ static bool ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
+    @autoreleasepool {
     MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
     edesc.dispatchType = MTLDispatchTypeSerial;
 
@@ -727,6 +738,7 @@ 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];
@@ -745,6 +757,7 @@ 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) {
@@ -806,6 +819,7 @@ 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));
@@ -1187,7 +1201,16 @@ 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 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);
 
                         [encoder setComputePipelineState:pipeline];
                         [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
@@ -1196,11 +1219,20 @@ static bool ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
                         }
-                        [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];
+                        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 setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -1296,6 +1328,7 @@ 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;
                                 default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                             }
 
@@ -1430,6 +1463,12 @@ 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;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1464,7 +1503,7 @@ 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_Q4_K) {
+                                src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || 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) {
@@ -1513,8 +1552,6 @@ 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;
@@ -1572,6 +1609,7 @@ 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;
                                 default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                             }
 
@@ -1709,6 +1747,12 @@ 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;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@@ -1759,7 +1803,7 @@ 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_Q4_K) {
+                                src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || 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) {
@@ -1813,6 +1857,7 @@ 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_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
                             default: GGML_ASSERT(false && "not implemented");
                         }
@@ -2272,6 +2317,7 @@ static bool ggml_metal_graph_compute(
         [[MTLCaptureManager sharedCaptureManager] stopCapture];
     }
 
+    }
     return true;
 }
 

ggml-metal.metal

@@ -351,12 +351,17 @@ 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,
-        threadgroup float  * buf [[threadgroup(0)]],
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -368,13 +373,26 @@ 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));
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]);
     }
 
     // find the max value in the block
@@ -399,7 +417,7 @@ 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)) - max_val);
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
     }
@@ -437,12 +455,17 @@ 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,
-        threadgroup float  * buf [[threadgroup(0)]],
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -454,13 +477,25 @@ 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));
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]);
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@@ -486,7 +521,7 @@ 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)) - max_val);
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
@@ -2490,6 +2525,13 @@ 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;
+
+
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3747,6 +3789,137 @@ 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,
@@ -4173,6 +4346,123 @@ 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
+    // TODO
+#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;
+        }
+    }
+}
+
+[[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);
+}
+
 
 //============================= templates and their specializations =============================
 
@@ -4518,6 +4808,22 @@ 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 block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -5060,6 +5366,7 @@ 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>;
 
 //
 // matrix-matrix multiplication
@@ -5099,6 +5406,7 @@ 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>;
 
 //
 // indirect matrix-matrix multiplication
@@ -5150,6 +5458,7 @@ 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>;
 
 //
 // matrix-vector multiplication
@@ -6117,3 +6426,66 @@ 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);
+}

ggml-quants.h

@@ -191,6 +191,13 @@ 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");
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -243,22 +250,24 @@ 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);
 
 // Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+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);
+void ggml_vec_dot_q4_1_q8_1(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_q5_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);
+void ggml_vec_dot_q5_1_q8_1(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_q8_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);
 
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q2_K_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_q3_K_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_q4_K_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_q5_K_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_q6_K_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_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);
 
 //
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
@@ -266,6 +275,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, const void * GGML
 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_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);
@@ -276,8 +286,8 @@ 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(int grid_size);
-void iq2xs_free_impl(int grid_size);
+void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);
 

ggml-sycl.cpp

@@ -11578,11 +11578,8 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
     }
     char * dst_ptr = (char *) dst;
 
-    const int64_t ne0 = src->ne[0];
-    const int64_t nb0 = src->nb[0];
-    const int64_t nb1 = src->nb[1];
-    const int64_t nb2 = src->nb[2];
-    const int64_t nb3 = src->nb[3];
+    GGML_TENSOR_LOCALS_1(int64_t, ne, src, ne);
+    GGML_TENSOR_LOCALS(int64_t, nb, src, nb);
     const enum ggml_type type = src->type;
     const int64_t ts = ggml_type_size(type);
     const int64_t bs = ggml_blck_size(type);
@@ -12426,9 +12423,7 @@ inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1,
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
+    GGML_TENSOR_LOCALS_3(int64_t, ne0, src0, ne);
     const int64_t nrows = ggml_nrows(src0);
 
     //const int n_past = ((int32_t *) dst->op_params)[0];
@@ -12758,15 +12753,9 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
                                  ggml_sycl_op_mul_mat_t op,
                                  const bool convert_src1_to_q8_1) try {
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
     const int64_t nrows1 = ggml_nrows(src1);
 
     GGML_ASSERT(ne03 == ne13);
@@ -13337,23 +13326,13 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0,
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
 
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03);
+    GGML_TENSOR_LOCALS(int64_t, nb0, src0, nb);
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
 
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
+    GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb);
 
     const int64_t ne1 = ggml_nelements(src1);
     const int64_t ne  = ggml_nelements(dst);
@@ -13655,23 +13634,15 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
     GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src00->ne[1];
-    const int64_t ne02 = src00->ne[2];
-    const int64_t ne03 = src00->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne0, src00, ne);
 
     //const int64_t nb01 = src00->nb[1];
-    const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
+    GGML_TENSOR_LOCALS(int64_t, nb0, src00, nb);
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
 
+    GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb);
     //const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
 
     const int64_t ne1 = ggml_nelements(src1);
     const int64_t ne  = ggml_nelements(dst);
@@ -13940,25 +13911,7 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1,
     GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
     GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-
-    const int64_t nb00 = src0->nb[0];
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-    const int64_t nb03 = src0->nb[3];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-
-
-    const int64_t nb10 = src1->nb[0];
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2];
-    const int64_t nb13 = src1->nb[3];
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
     SYCL_CHECK(ggml_sycl_set_device(g_main_device));
     dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0];

ggml-vulkan.cpp

@@ -27,6 +27,7 @@
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 
 #define VK_VENDOR_ID_AMD 0x1002
+#define VK_VENDOR_ID_APPLE 0x106b
 #define VK_VENDOR_ID_INTEL 0x8086
 #define VK_VENDOR_ID_NVIDIA 0x10de
 
@@ -706,9 +707,21 @@ static void ggml_vk_queue_cleanup(ggml_backend_vk_context * ctx, vk_queue& q) {
     q.cmd_buffer_idx = 0;
 }
 
-static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags) {
+static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
+    for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
+        vk::MemoryType memory_type = mem_props->memoryTypes[i];
+        if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
+            (flags & memory_type.propertyFlags) == flags &&
+            mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
+            return static_cast<int32_t>(i);
+        }
+    }
+    return UINT32_MAX;
+}
+
+static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ")" << std::endl;
+    std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")" << std::endl;
 #endif
     vk_buffer buf = std::make_shared<vk_buffer_struct>();
 
@@ -735,15 +748,15 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
 
     uint32_t memory_type_index = UINT32_MAX;
 
-    for (uint32_t i = 0; i < mem_props.memoryTypeCount; ++i) {
-        vk::MemoryType memory_type = mem_props.memoryTypes[i];
-        if ((mem_req.memoryTypeBits & ((uint64_t)1 << i)) && (req_flags & memory_type.propertyFlags) == req_flags && mem_props.memoryHeaps[memory_type.heapIndex].size >= mem_req.size) {
-            memory_type_index = i;
-            break;
-        }
+    memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
+    buf->memory_property_flags = req_flags;
+
+    if (memory_type_index == UINT32_MAX && fallback_flags) {
+        memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
+        buf->memory_property_flags = fallback_flags;
     }
 
-    if (memory_type_index >= mem_props.memoryTypeCount) {
+    if (memory_type_index == UINT32_MAX) {
         ctx->device.lock()->device.destroyBuffer(buf->buffer);
         buf->size = 0;
         throw vk::OutOfDeviceMemoryError("No suitable memory type found");
@@ -757,10 +770,9 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
         buf->size = 0;
         throw e;
     }
-    buf->memory_property_flags = req_flags;
     buf->ptr = nullptr;
 
-    if (req_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+    if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
         buf->ptr = ctx->device.lock()->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE);
     }
 
@@ -777,9 +789,9 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
     return buf;
 }
 
-static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags) {
+static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
     try {
-        return ggml_vk_create_buffer(ctx, size, req_flags);
+        return ggml_vk_create_buffer(ctx, size, req_flags, fallback_flags);
     } catch (const vk::SystemError& e) {
         std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
         std::cerr << "ggml_vulkan: " << e.what() << std::endl;
@@ -790,16 +802,16 @@ static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size
 static vk_buffer ggml_vk_create_buffer_device(ggml_backend_vk_context * ctx, size_t size) {
     vk_buffer buf;
     try {
-        buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
-    } catch (const vk::SystemError& e) {
         if (ctx->device.lock()->uma) {
             // Fall back to host memory type
-            buf = ggml_vk_create_buffer_check(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+            buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
         } else {
-            std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
-            std::cerr << "ggml_vulkan: " << e.what() << std::endl;
-            throw e;
+            buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
         }
+    } catch (const vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        throw e;
     }
 
     return buf;
@@ -1079,7 +1091,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 }
 
-void ggml_vk_instance_init() {
+static void ggml_vk_instance_init() {
     if (vk_instance_initialized) {
         return;
     }
@@ -1138,7 +1150,7 @@ void ggml_vk_instance_init() {
     vk_instance_initialized = true;
 }
 
-void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
+static 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
@@ -1421,7 +1433,9 @@ static void * ggml_vk_host_malloc(ggml_backend_vk_context * ctx, size_t size) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_host_malloc(" << size << ")" << std::endl;
 #endif
-    vk_buffer buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+    vk_buffer buf = ggml_vk_create_buffer(ctx, size,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
 
     if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
         fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
@@ -1567,7 +1581,9 @@ static void deferred_memcpy(void * dst, const void * src, size_t size, std::vect
 static void ggml_vk_ensure_sync_staging_buffer(ggml_backend_vk_context * ctx, size_t size) {
     if (ctx->sync_staging == nullptr || ctx->sync_staging->size < size) {
         ggml_vk_destroy_buffer(ctx->sync_staging);
-        ctx->sync_staging = ggml_vk_create_buffer_check(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+        ctx->sync_staging = ggml_vk_create_buffer_check(ctx, size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     }
 }
 
@@ -2034,18 +2050,100 @@ static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ct
     return ctx->pipeline_matmul_f32_aligned_l.align;
 }
 
-static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
-#ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
-#endif
+static vk_pipeline* ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
     if (bit16_x && bit16_y) {
-        if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) {
+        if (m <= 32 || n <= 32) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " S" << std::endl;
 #endif
             return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
         }
-        if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+        return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m;
+    }
+    if (bit16_x && !bit16_y) {
+        if (m <= 32 || n <= 32) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+            return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
+        }
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+        return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m;
+    }
+    if (!bit16_x && bit16_y) {
+        GGML_ASSERT(false);
+    }
+
+    if (m <= 32 || n <= 32) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+        return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
+    }
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+    return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m;
+}
+
+static vk_pipeline* ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+    if (bit16_x && bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m;
+    }
+    if (bit16_x && !bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m;
+    }
+    if (!bit16_x && bit16_y) {
+        GGML_ASSERT(false);
+    }
+    return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m;
+}
+
+static vk_pipeline* ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+    if (bit16_x && bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
+    }
+    if (bit16_x && !bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
+    }
+    if (!bit16_x && bit16_y) {
+        GGML_ASSERT(false);
+    }
+    return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
+}
+
+static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
+#endif
+    switch (ctx->device.lock()->vendor_id) {
+    case VK_VENDOR_ID_AMD:
+        return ggml_vk_guess_matmul_pipeline_amd(ctx, bit16_x, bit16_y, m, n, aligned);
+    case VK_VENDOR_ID_APPLE:
+        return ggml_vk_guess_matmul_pipeline_apple(ctx, bit16_x, bit16_y, aligned);
+    case VK_VENDOR_ID_INTEL:
+        return ggml_vk_guess_matmul_pipeline_intel(ctx, bit16_x, bit16_y, aligned);
+    }
+
+    if (bit16_x && bit16_y) {
+        if (m <= 32 || n <= 32) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+            return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
+        }
+        if (m <= 64 || n <= 64) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " M" << std::endl;
 #endif
@@ -2057,13 +2155,13 @@ static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
         return aligned ? &ctx->pipeline_matmul_f16_aligned_l : &ctx->pipeline_matmul_f16_l;
     }
     if (bit16_x && !bit16_y) {
-        if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) {
+        if (m <= 32 || n <= 32) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " S" << std::endl;
 #endif
             return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
         }
-        if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) {
+        if (m <= 64 || n <= 64) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " M" << std::endl;
 #endif
@@ -2078,13 +2176,13 @@ static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
         GGML_ASSERT(false);
     }
 
-    if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) {
+    if (m <= 32 || n <= 32) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " S" << std::endl;
 #endif
         return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
     }
-    if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) {
+    if (m <= 64 || n <= 64) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " M" << std::endl;
 #endif
@@ -3999,7 +4097,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     std::cerr << "ggml_vk_preallocate_buffers(qx_size: " << ctx->prealloc_size_qx << " qy_size: " << ctx->prealloc_size_qy << " x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << ")" << std::endl;
 #endif
 #if defined(GGML_VULKAN_RUN_TESTS)
-    ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+    ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     ggml_vk_test_transfer(ctx, 8192 * 1000, false);
     ggml_vk_test_transfer(ctx, 8192 * 1000, true);
 
@@ -4091,7 +4191,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         if (ctx->staging != nullptr) {
             ggml_vk_destroy_buffer(ctx->staging);
         }
-        ctx->staging = ggml_vk_create_buffer_check(ctx, ctx->staging_size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+        ctx->staging = ggml_vk_create_buffer_check(ctx, ctx->staging_size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     }
 }
 
@@ -4454,13 +4556,13 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     }
 }
 
-GGML_CALL int ggml_vk_get_device_count() {
+GGML_CALL static int ggml_vk_get_device_count() {
     ggml_vk_instance_init();
 
     return vk_instance.device_indices.size();
 }
 
-GGML_CALL void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
+GGML_CALL static 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();
@@ -4478,7 +4580,7 @@ void ggml_vk_init_cpu_assist() {
 
     std::cerr << "ggml_vulkan: Found " << ggml_vk_get_device_count() << " Vulkan devices:" << std::endl;
 
-    for (size_t i = 0; i < ggml_vk_get_device_count(); i++) {
+    for (int 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.
@@ -5165,7 +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 < vk_instance.device_indices.size());
+    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
 

ggml.h

@@ -354,6 +354,7 @@ extern "C" {
         GGML_TYPE_IQ2_XXS = 16,
         GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_IQ3_XXS = 18,
+        GGML_TYPE_IQ1_S   = 19,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -391,6 +392,7 @@ 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
     };
 
     // available tensor operations:
@@ -505,11 +507,17 @@ extern "C" {
 
     enum ggml_log_level {
         GGML_LOG_LEVEL_ERROR = 2,
-        GGML_LOG_LEVEL_WARN = 3,
-        GGML_LOG_LEVEL_INFO = 4,
+        GGML_LOG_LEVEL_WARN  = 3,
+        GGML_LOG_LEVEL_INFO  = 4,
         GGML_LOG_LEVEL_DEBUG = 5
     };
 
+    enum ggml_tensor_flag {
+        GGML_TENSOR_FLAG_INPUT  = 1,
+        GGML_TENSOR_FLAG_OUTPUT = 2,
+        GGML_TENSOR_FLAG_PARAM  = 4,
+    };
+
     // ggml object
     struct ggml_object {
         size_t offs;
@@ -543,7 +551,7 @@ extern "C" {
         // op params - allocated as int32_t for alignment
         int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 
-        bool is_param;
+        int32_t flags;
 
         struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
@@ -567,6 +575,11 @@ extern "C" {
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
 
+    // Abort callback
+    // If not NULL, called before ggml computation
+    // If it returns true, the computation is aborted
+    typedef bool (*ggml_abort_callback)(void * data);
+
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -576,8 +589,8 @@ extern "C" {
         int n_threads;
 
         // abort ggml_graph_compute when true
-        bool (*abort_callback)(void * data);
-        void * abort_callback_data;
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
     };
 
     enum ggml_cgraph_eval_order {
@@ -647,6 +660,16 @@ extern "C" {
         void * wdata;
     };
 
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
+    };
+
     // misc
 
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
@@ -657,7 +680,7 @@ extern "C" {
 
     GGML_API void    ggml_print_backtrace(void);
 
-    GGML_API void    ggml_numa_init(void); // call once for better performance on NUMA systems
+    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
     GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
@@ -1362,13 +1385,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask)
+    // fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
     // 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,
-            float                 scale);
+            struct ggml_tensor  * pos,
+            float                 scale,
+            float                 max_bias);
 
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
@@ -1470,12 +1497,13 @@ extern "C" {
 
     // alibi position embedding
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_alibi(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_head,
-            float                 bias_max);
+            float                 bias_max),
+        "use ggml_soft_max_ext instead (will be removed in Mar 2024)");
 
     // clamp
     // in-place, returns view(a)
@@ -2087,6 +2115,12 @@ extern "C" {
             ggml_opt_callback callback,
             void * callback_data);
 
+    //
+    // tensor flags
+    //
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+
     //
     // quantization
     //
@@ -2273,6 +2307,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_ssse3      (void);
     GGML_API int ggml_cpu_has_sycl       (void);
     GGML_API int ggml_cpu_has_vsx        (void);
+    GGML_API int ggml_cpu_has_matmul_int8(void);
 
     //
     // Internal types and functions exposed for tests and benchmarks
@@ -2286,7 +2321,8 @@ extern "C" {
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int k);
-    typedef void (*ggml_vec_dot_t)   (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+    typedef void (*ggml_vec_dot_t)   (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                      const void * GGML_RESTRICT y, size_t by, int nrc);
 
     typedef struct {
         const char      * type_name;
@@ -2298,6 +2334,7 @@ extern "C" {
         ggml_from_float_t from_float_reference;
         ggml_vec_dot_t    vec_dot;
         enum ggml_type    vec_dot_type;
+        int64_t           nrows; // number of rows to process simultaneously;
     } ggml_type_traits_t;
 
     GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);

ggml_vk_generate_shaders.py

@@ -2067,6 +2067,8 @@ type_names = {
 
 K_QUANTS_PER_ITERATION = 2
 
+ASYNCIO_CONCURRENCY = 64
+
 output_dir = gettempdir()
 
 lock = asyncio.Lock()
@@ -2291,7 +2293,14 @@ async def main():
     tasks.append(string_to_spv("rope_neox_f32", rope_neox_src, {"A_TYPE": "float", "D_TYPE": "float"}))
     tasks.append(string_to_spv("rope_neox_f16", rope_neox_src, {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
 
-    await asyncio.gather(*tasks)
+    # Helper to decorate tasks with semaphore acquisition.
+    async def withSemaphore(sem, task):
+        async with sem:
+            return await task
+
+    # Run tasks concurrently guarded by a concurrency limit.
+    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
+    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
 
     with open("ggml-vulkan-shaders.hpp", "w") as f:
         f.write("#include <cstdint>\n\n")

gguf-py/examples/reader.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+import sys
+from pathlib import Path
+from gguf.gguf_reader import GGUFReader
+
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+
+def read_gguf_file(gguf_file_path):
+    """
+    Reads and prints key-value pairs and tensor information from a GGUF file in an improved format.
+
+    Parameters:
+    - gguf_file_path: Path to the GGUF file.
+    """
+
+    reader = GGUFReader(gguf_file_path)
+
+    # List all key-value pairs in a columnized format
+    print("Key-Value Pairs:")
+    max_key_length = max(len(key) for key in reader.fields.keys())
+    for key, field in reader.fields.items():
+        value = field.parts[field.data[0]]
+        print(f"{key:{max_key_length}} : {value}")
+    print("----")
+
+    # List all tensors
+    print("Tensors:")
+    tensor_info_format = "{:<30} | Shape: {:<15} | Size: {:<12} | Quantization: {}"
+    print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization"))
+    print("-" * 80)
+    for tensor in reader.tensors:
+        shape_str = "x".join(map(str, tensor.shape))
+        size_str = str(tensor.n_elements)
+        quantization_str = tensor.tensor_type.name
+        print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str))
+
+
+if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print("Usage: reader.py <path_to_gguf_file>")
+        sys.exit(1)
+    gguf_file_path = sys.argv[1]
+    read_gguf_file(gguf_file_path)

gguf-py/gguf/constants.py

@@ -40,6 +40,7 @@ class Keys:
         TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
         EXPERT_COUNT          = "{arch}.expert_count"
         EXPERT_USED_COUNT     = "{arch}.expert_used_count"
+        POOLING_TYPE          = "{arch}.pooling_type"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -50,6 +51,7 @@ class Keys:
         VALUE_LENGTH      = "{arch}.attention.value_length"
         LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
         LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
+        CAUSAL            = "{arch}.attention.causal"
 
     class Rope:
         DIMENSION_COUNT      = "{arch}.rope.dimension_count"
@@ -60,22 +62,25 @@ class Keys:
         SCALING_FINETUNED    = "{arch}.rope.scaling.finetuned"
 
     class Tokenizer:
-        MODEL         = "tokenizer.ggml.model"
-        LIST          = "tokenizer.ggml.tokens"
-        TOKEN_TYPE    = "tokenizer.ggml.token_type"
-        SCORES        = "tokenizer.ggml.scores"
-        MERGES        = "tokenizer.ggml.merges"
-        BOS_ID        = "tokenizer.ggml.bos_token_id"
-        EOS_ID        = "tokenizer.ggml.eos_token_id"
-        UNK_ID        = "tokenizer.ggml.unknown_token_id"
-        SEP_ID        = "tokenizer.ggml.seperator_token_id"
-        PAD_ID        = "tokenizer.ggml.padding_token_id"
-        ADD_BOS       = "tokenizer.ggml.add_bos_token"
-        ADD_EOS       = "tokenizer.ggml.add_eos_token"
-        ADD_PREFIX    = "tokenizer.ggml.add_space_prefix"
-        HF_JSON       = "tokenizer.huggingface.json"
-        RWKV          = "tokenizer.rwkv.world"
-        CHAT_TEMPLATE = "tokenizer.chat_template"
+        MODEL            = "tokenizer.ggml.model"
+        LIST             = "tokenizer.ggml.tokens"
+        TOKEN_TYPE       = "tokenizer.ggml.token_type"
+        TOKEN_TYPE_COUNT = "tokenizer.ggml.token_type_count"  # for BERT-style token types
+        SCORES           = "tokenizer.ggml.scores"
+        MERGES           = "tokenizer.ggml.merges"
+        BOS_ID           = "tokenizer.ggml.bos_token_id"
+        EOS_ID           = "tokenizer.ggml.eos_token_id"
+        UNK_ID           = "tokenizer.ggml.unknown_token_id"
+        SEP_ID           = "tokenizer.ggml.seperator_token_id"
+        PAD_ID           = "tokenizer.ggml.padding_token_id"
+        CLS_ID           = "tokenizer.ggml.cls_token_id"
+        MASK_ID          = "tokenizer.ggml.mask_token_id"
+        ADD_BOS          = "tokenizer.ggml.add_bos_token"
+        ADD_EOS          = "tokenizer.ggml.add_eos_token"
+        ADD_PREFIX       = "tokenizer.ggml.add_space_prefix"
+        HF_JSON          = "tokenizer.huggingface.json"
+        RWKV             = "tokenizer.rwkv.world"
+        CHAT_TEMPLATE    = "tokenizer.chat_template"
 
 
 #
@@ -84,27 +89,28 @@ class Keys:
 
 
 class MODEL_ARCH(IntEnum):
-    LLAMA     = auto()
-    FALCON    = auto()
-    BAICHUAN  = auto()
-    GPT2      = auto()
-    GPTJ      = auto()
-    GPTNEOX   = auto()
-    MPT       = auto()
-    STARCODER = auto()
-    PERSIMMON = auto()
-    REFACT    = auto()
-    BERT      = auto()
-    BLOOM     = auto()
-    STABLELM  = auto()
-    QWEN      = auto()
-    QWEN2     = auto()
-    PHI2      = auto()
-    PLAMO     = auto()
-    CODESHELL = auto()
-    ORION     = auto()
+    LLAMA      = auto()
+    FALCON     = auto()
+    BAICHUAN   = auto()
+    GPT2       = auto()
+    GPTJ       = auto()
+    GPTNEOX    = auto()
+    MPT        = auto()
+    STARCODER  = auto()
+    PERSIMMON  = auto()
+    REFACT     = auto()
+    BERT       = auto()
+    NOMIC_BERT = auto()
+    BLOOM      = auto()
+    STABLELM   = auto()
+    QWEN       = auto()
+    QWEN2      = auto()
+    PHI2       = auto()
+    PLAMO      = auto()
+    CODESHELL  = auto()
+    ORION      = auto()
     INTERNLM2  = auto()
-    MINICPM   = auto()
+    MINICPM    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -122,6 +128,7 @@ class MODEL_TENSOR(IntEnum):
     ATTN_OUT        = auto()
     ATTN_NORM       = auto()
     ATTN_NORM_2     = auto()
+    ATTN_OUT_NORM   = auto()
     ATTN_ROT_EMBD   = auto()
     FFN_GATE_INP    = auto()
     FFN_NORM        = auto()
@@ -134,6 +141,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_UP_EXP      = auto()
     ATTN_Q_NORM     = auto()
     ATTN_K_NORM     = auto()
+    LAYER_OUT_NORM  = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -148,6 +156,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PERSIMMON:      "persimmon",
     MODEL_ARCH.REFACT:         "refact",
     MODEL_ARCH.BERT:           "bert",
+    MODEL_ARCH.NOMIC_BERT:     "nomic-bert",
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
@@ -178,6 +187,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ATTN_ROT_EMBD:   "blk.{bid}.attn_rot_embd",
     MODEL_TENSOR.ATTN_Q_NORM:     "blk.{bid}.attn_q_norm",
     MODEL_TENSOR.ATTN_K_NORM:     "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.ATTN_OUT_NORM:   "blk.{bid}.attn_output_norm",
     MODEL_TENSOR.FFN_GATE_INP:    "blk.{bid}.ffn_gate_inp",
     MODEL_TENSOR.FFN_NORM:        "blk.{bid}.ffn_norm",
     MODEL_TENSOR.FFN_GATE:        "blk.{bid}.ffn_gate",
@@ -187,6 +197,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate.{xid}",
     MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down.{xid}",
     MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up.{xid}",
+    MODEL_TENSOR.LAYER_OUT_NORM:  "blk.{bid}.layer_output_norm",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -262,17 +273,32 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     ],
     MODEL_ARCH.BERT: [
         MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
         MODEL_TENSOR.TOKEN_TYPES,
         MODEL_TENSOR.POS_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_K,
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
+    MODEL_ARCH.NOMIC_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
     ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -535,6 +561,12 @@ class RopeScalingType(Enum):
     YARN   = 'yarn'
 
 
+class PoolingType(IntEnum):
+    NONE = 0
+    MEAN = 1
+    CLS  = 2
+
+
 class GGMLQuantizationType(IntEnum):
     F32  = 0
     F16  = 1
@@ -661,5 +693,7 @@ KEY_TOKENIZER_EOS_ID     = Keys.Tokenizer.EOS_ID
 KEY_TOKENIZER_UNK_ID     = Keys.Tokenizer.UNK_ID
 KEY_TOKENIZER_SEP_ID     = Keys.Tokenizer.SEP_ID
 KEY_TOKENIZER_PAD_ID     = Keys.Tokenizer.PAD_ID
+KEY_TOKENIZER_CLS_ID     = Keys.Tokenizer.CLS_ID
+KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
 KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
 KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV

gguf-py/gguf/gguf_writer.py

@@ -19,6 +19,7 @@ from .constants import (
     GGUFValueType,
     Keys,
     RopeScalingType,
+    PoolingType,
     TokenType,
 )
 
@@ -357,6 +358,12 @@ class GGUFWriter:
     def add_layer_norm_rms_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value)
 
+    def add_causal_attention(self, value: bool) -> None:
+        self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
+
+    def add_pooling_type(self, value: PoolingType) -> None:
+        self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value)
+
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
@@ -387,6 +394,9 @@ class GGUFWriter:
     def add_token_types(self, types: Sequence[TokenType] | Sequence[int]) -> None:
         self.add_array(Keys.Tokenizer.TOKEN_TYPE, types)
 
+    def add_token_type_count(self, value: int) -> None:
+        self.add_uint32(Keys.Tokenizer.TOKEN_TYPE_COUNT, value)
+
     def add_token_scores(self, scores: Sequence[float]) -> None:
         self.add_array(Keys.Tokenizer.SCORES, scores)
 
@@ -405,6 +415,12 @@ class GGUFWriter:
     def add_pad_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.PAD_ID, id)
 
+    def add_cls_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.CLS_ID, id)
+
+    def add_mask_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.MASK_ID, id)
+
     def add_add_bos_token(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_BOS, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -15,7 +15,7 @@ class TensorNameMap:
             "word_embeddings",                           # bloom
             "model.embed_tokens",                        # llama-hf
             "tok_embeddings",                            # llama-pth
-            "embeddings.word_embeddings",                # bert
+            "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
             "wte",                                       # gpt2
             "transformer.embd.wte",                      # phi2
@@ -24,12 +24,14 @@ class TensorNameMap:
 
         # Token type embeddings
         MODEL_TENSOR.TOKEN_TYPES: (
-            "embeddings.token_type_embeddings",  # bert
+            "embeddings.token_type_embeddings",  # bert nomic-bert
         ),
 
         # Normalization of token embeddings
         MODEL_TENSOR.TOKEN_EMBD_NORM: (
             "word_embeddings_layernorm",  # bloom
+            "embeddings.LayerNorm",       # bert
+            "emb_ln",                     # nomic-bert
         ),
 
         # Position embeddings
@@ -54,7 +56,6 @@ class TensorNameMap:
             "transformer.ln_f",                        # gpt2 gpt-j falcon
             "model.norm",                              # llama-hf baichuan internlm2
             "norm",                                    # llama-pth
-            "embeddings.LayerNorm",                    # bert
             "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen gpt2
             "language_model.encoder.final_layernorm",  # persimmon
@@ -79,7 +80,6 @@ class TensorNameMap:
             "transformer.h.{bid}.ln_mlp",                           # falcon40b
             "model.layers.{bid}.input_layernorm",                   # llama-hf
             "layers.{bid}.attention_norm",                          # llama-pth
-            "encoder.layer.{bid}.attention.output.LayerNorm",       # bert
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
             "h.{bid}.ln_1",                                         # gpt2
@@ -104,6 +104,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
+            "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
         ),
 
         # Attention query
@@ -153,6 +154,13 @@ class TensorNameMap:
             "transformer.h.{bid}.mixer.out_proj",                        # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
             "model.layers.{bid}.attention.wo",                           # internlm2
+            "encoder.layers.{bid}.attn.out_proj",                        # nomic-bert
+        ),
+
+        # Attention output norm
+        MODEL_TENSOR.ATTN_OUT_NORM: (
+            "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm1",                      # nomic-bert
         ),
 
         # Rotary embeddings
@@ -171,7 +179,6 @@ class TensorNameMap:
             "transformer.blocks.{bid}.norm_2",                               # mpt
             "model.layers.{bid}.post_attention_layernorm",                   # llama-hf
             "layers.{bid}.ffn_norm",                                         # llama-pth
-            "encoder.layer.{bid}.output.LayerNorm",                          # bert
             "language_model.encoder.layers.{bid}.post_attention_layernorm",  # persimmon
             "model.layers.{bid}.ln2",                                        # yi
             "h.{bid}.ln_2",                                                  # gpt2
@@ -202,6 +209,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc1",                             # phi2
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -221,6 +229,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
+            "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -246,6 +255,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc2",                             # phi2
             "model.layers.layers.{bid}.mlp.down_proj",                # plamo
             "model.layers.{bid}.feed_forward.w2",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -266,6 +276,11 @@ class TensorNameMap:
         MODEL_TENSOR.ROPE_FREQS: (
             "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq",  # persimmon
         ),
+
+        MODEL_TENSOR.LAYER_OUT_NORM: (
+            "encoder.layer.{bid}.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm2",            # nomic-bert
+        )
     }
 
     mapping: dict[str, tuple[MODEL_TENSOR, str]]

gguf-py/gguf/vocab.py

@@ -29,7 +29,7 @@ class SpecialVocab:
         if special_token_types is not None:
             self.special_token_types = special_token_types
         else:
-            self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad')
+            self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad', 'cls', 'mask')
         self._load(Path(path))
 
     def __repr__(self) -> str:
@@ -152,10 +152,6 @@ class SpecialVocab:
             add_entry = tokenizer_config.get(f'add_{typ}_token')
             if isinstance(add_entry, bool):
                 self.add_special_token[typ] = add_entry
-            if not added_tokens:
-                # We will need this to get the content for the token, so if it's empty
-                # may as well just give up.
-                continue
             entry = tokenizer_config.get(f'{typ}_token')
             if isinstance(entry, str):
                 tc_content = entry

llama.h

@@ -61,6 +61,7 @@ extern "C" {
     enum llama_vocab_type {
         LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
         LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
     };
 
     enum llama_token_type {
@@ -99,6 +100,7 @@ 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_GUESSED = 1024, // not specified in the model file
     };
@@ -111,6 +113,12 @@ extern "C" {
         LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
     };
 
+    enum llama_pooling_type {
+        LLAMA_POOLING_NONE = 0,
+        LLAMA_POOLING_MEAN = 1,
+        LLAMA_POOLING_CLS  = 2,
+    };
+
     enum llama_split_mode {
         LLAMA_SPLIT_NONE    = 0, // single GPU
         LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
@@ -235,6 +243,7 @@ extern "C" {
         bool logits_all;  // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         bool embedding;   // embedding mode only
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
+        bool do_pooling;  // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
     };
 
     // model quantization parameters
@@ -304,7 +313,10 @@ extern "C" {
     // Initialize the llama + ggml backend
     // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_backend_init(bool numa);
+    LLAMA_API void llama_backend_init(void);
+
+    //optional:
+    LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
 
     // Call once at the end of the program - currently only used for MPI
     LLAMA_API void llama_backend_free(void);
@@ -627,6 +639,10 @@ extern "C" {
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
+    // Get the embeddings for the ith sequence
+    // llama_get_embeddings(ctx) + i*n_embd
+    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
+
     //
     // Vocab
     //

pocs/vdot/q8dot.cpp

@@ -156,8 +156,8 @@ int main(int argc, char** argv) {
 
         t1 = std::chrono::high_resolution_clock::now();
         float fs;
-        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, x40.data(), y.data());
-        else funcs.vec_dot(kVecSize * QK4_1, &fs, x41.data(), y.data());
+        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
+        else funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
         t2 = std::chrono::high_resolution_clock::now();
         t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
         if (iloop > 3) ggml.addResult(fs, t);

pocs/vdot/vdot.cpp

@@ -284,8 +284,8 @@ int main(int argc, char** argv) {
         else {
             auto vdot = ggml_internal_get_type_traits(funcs.vec_dot_type);
             vdot.from_float(y1.data(), q8.data(), kVecSize);
-            if (useQ4_1) funcs.vec_dot(kVecSize, &result, q41.data(), q8.data());
-            else funcs.vec_dot(kVecSize, &result, q40.data(), q8.data());
+            if (useQ4_1) funcs.vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
+            else funcs.vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
         }
         sumq += result;
         t2 = std::chrono::high_resolution_clock::now();

scripts/compare-commits.sh

@@ -0,0 +1,37 @@
+#!/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/hf.sh

@@ -0,0 +1,107 @@
+#!/bin/bash
+#
+# Shortcut for downloading HF models
+#
+# Usage:
+#   ./main -m $(./examples/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./examples/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./examples/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#
+
+# all logs go to stderr
+function log {
+    echo "$@" 1>&2
+}
+
+function usage {
+    log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [-h|--help]"
+    exit 1
+}
+
+# check for curl or wget
+function has_cmd {
+    if ! [ -x "$(command -v $1)" ]; then
+        return 1
+    fi
+}
+
+if has_cmd wget; then
+    cmd="wget -q --show-progress -c -O %s %s"
+elif has_cmd curl; then
+    cmd="curl -C - -f -o %s -L %s"
+else
+    log "[E] curl or wget not found"
+    exit 1
+fi
+
+url=""
+repo=""
+file=""
+
+# parse args
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --url)
+            url="$2"
+            shift 2
+            ;;
+        --repo)
+            repo="$2"
+            shift 2
+            ;;
+        --file)
+            file="$2"
+            shift 2
+            ;;
+        -h|--help)
+            usage
+            ;;
+        *)
+            url="$1"
+            shift
+            ;;
+    esac
+done
+
+if [ -n "$repo" ] && [ -n "$file" ]; then
+    url="https://huggingface.co/$repo/resolve/main/$file"
+fi
+
+if [ -z "$url" ]; then
+    log "[E] missing --url"
+    usage
+fi
+
+# check if the URL is a HuggingFace model, and if so, try to download it
+is_url=false
+
+if [[ ${#url} -gt 22 ]]; then
+    if [[ ${url:0:22} == "https://huggingface.co" ]]; then
+        is_url=true
+    fi
+fi
+
+if [ "$is_url" = false ]; then
+    log "[E] invalid URL, must start with https://huggingface.co"
+    exit 0
+fi
+
+# replace "blob/main" with "resolve/main"
+url=${url/blob\/main/resolve\/main}
+
+basename=$(basename $url)
+
+log "[+] attempting to download $basename"
+
+if [ -n "$cmd" ]; then
+    cmd=$(printf "$cmd" "$basename" "$url")
+    log "[+] $cmd"
+    if $cmd; then
+        echo $basename
+        exit 0
+    fi
+fi
+
+log "[-] failed to download"
+
+exit 1

scripts/sync-ggml-am.sh

@@ -97,6 +97,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-cuda.cu            -> ggml-cuda.cu
     # src/ggml-cuda.h             -> ggml-cuda.h
     # src/ggml-impl.h             -> ggml-impl.h
+    # src/ggml-kompute.cpp        -> ggml-kompute.cpp
+    # src/ggml-kompute.h          -> ggml-kompute.h
     # src/ggml-metal.h            -> ggml-metal.h
     # src/ggml-metal.m            -> ggml-metal.m
     # src/ggml-mpi.h              -> ggml-mpi.h
@@ -105,6 +107,10 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-opencl.h           -> ggml-opencl.h
     # src/ggml-quants.c           -> ggml-quants.c
     # src/ggml-quants.h           -> ggml-quants.h
+    # src/ggml-sycl.cpp           -> ggml-sycl.cpp
+    # src/ggml-sycl.h             -> ggml-sycl.h
+    # src/ggml-vulkan.cpp         -> ggml-vulkan.cpp
+    # src/ggml-vulkan.h           -> ggml-vulkan.h
     # include/ggml/ggml.h         -> ggml.h
     # include/ggml/ggml-alloc.h   -> ggml-alloc.h
     # include/ggml/ggml-backend.h -> ggml-backend.h
@@ -123,6 +129,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/src\/ggml-cuda\.cu/ggml-cuda.cu/g' \
         -e 's/src\/ggml-cuda\.h/ggml-cuda.h/g' \
         -e 's/src\/ggml-impl\.h/ggml-impl.h/g' \
+        -e 's/src\/ggml-kompute\.cpp/ggml-kompute.cpp/g' \
+        -e 's/src\/ggml-kompute\.h/ggml-kompute.h/g' \
         -e 's/src\/ggml-metal\.h/ggml-metal.h/g' \
         -e 's/src\/ggml-metal\.m/ggml-metal.m/g' \
         -e 's/src\/ggml-mpi\.h/ggml-mpi.h/g' \
@@ -131,6 +139,10 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \
         -e 's/src\/ggml-quants\.c/ggml-quants.c/g' \
         -e 's/src\/ggml-quants\.h/ggml-quants.h/g' \
+        -e 's/src\/ggml-sycl\.cpp/ggml-sycl.cpp/g' \
+        -e 's/src\/ggml-sycl\.h/ggml-sycl.h/g' \
+        -e 's/src\/ggml-vulkan\.cpp/ggml-vulkan.cpp/g' \
+        -e 's/src\/ggml-vulkan\.h/ggml-vulkan.h/g' \
         -e 's/include\/ggml\/ggml\.h/ggml.h/g' \
         -e 's/include\/ggml\/ggml-alloc\.h/ggml-alloc.h/g' \
         -e 's/include\/ggml\/ggml-backend\.h/ggml-backend.h/g' \

scripts/sync-ggml.last

@@ -1 +1 @@
-475cbad5c1c834e31e26a2283bc1413181644360
+5070f078a67c18c11736e78316ab715ca9afde16

scripts/sync-ggml.sh

@@ -7,6 +7,8 @@ cp -rpv ../ggml/src/ggml-backend.c          ./ggml-backend.c
 cp -rpv ../ggml/src/ggml-cuda.cu            ./ggml-cuda.cu
 cp -rpv ../ggml/src/ggml-cuda.h             ./ggml-cuda.h
 cp -rpv ../ggml/src/ggml-impl.h             ./ggml-impl.h
+cp -rpv ../ggml/src/ggml-kompute.cpp        ./ggml-kompute.cpp
+cp -rpv ../ggml/src/ggml-kompute.h          ./ggml-kompute.h
 cp -rpv ../ggml/src/ggml-metal.h            ./ggml-metal.h
 cp -rpv ../ggml/src/ggml-metal.m            ./ggml-metal.m
 cp -rpv ../ggml/src/ggml-metal.metal        ./ggml-metal.metal
@@ -16,6 +18,10 @@ cp -rpv ../ggml/src/ggml-opencl.cpp         ./ggml-opencl.cpp
 cp -rpv ../ggml/src/ggml-opencl.h           ./ggml-opencl.h
 cp -rpv ../ggml/src/ggml-quants.c           ./ggml-quants.c
 cp -rpv ../ggml/src/ggml-quants.h           ./ggml-quants.h
+cp -rpv ../ggml/src/ggml-sycl.cpp           ./ggml-sycl.cpp
+cp -rpv ../ggml/src/ggml-sycl.h             ./ggml-sycl.h
+cp -rpv ../ggml/src/ggml-vulkan.cpp         ./ggml-vulkan.cpp
+cp -rpv ../ggml/src/ggml-vulkan.h           ./ggml-vulkan.h
 cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
 cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h
 cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h

spm-headers/ggml-alloc.h

@@ -0,0 +1 @@
+../ggml-alloc.h

spm-headers/ggml-backend.h

@@ -0,0 +1 @@
+../ggml-backend.h

spm-headers/ggml.h

@@ -0,0 +1 @@
+../ggml.h

tests/test-autorelease.cpp

@@ -12,7 +12,7 @@ int main(int argc, char ** argv) {
     auto * model_path = get_model_or_exit(argc, argv);
 
     std::thread([&model_path]() {
-        llama_backend_init(false);
+        llama_backend_init();
         auto * model = llama_load_model_from_file(model_path, llama_model_default_params());
         auto * ctx = llama_new_context_with_model(model, llama_context_default_params());
         llama_free(ctx);

tests/test-backend-ops.cpp

@@ -1085,24 +1085,32 @@ 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 float scale;
     const bool mask;
+    const float scale;
+    const float max_bias;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, scale, mask);
+        return VARS_TO_STR5(type, ne, mask, scale, max_bias);
     }
 
     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,
-            bool mask = false)
-        : type(type), ne(ne), scale(scale), mask(mask) {}
+            float max_bias = 0.0f)
+        : type(type), ne(ne), mask(mask), scale(scale), max_bias(max_bias) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        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);
+        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);
         return out;
     }
 };
@@ -1147,30 +1155,6 @@ 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;
@@ -1488,7 +1472,7 @@ 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, 1.0f/sqrtf(n_embd));
+        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, nullptr, 1.0f/sqrtf(n_embd), 0.0f);
 
         // select experts
         ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok);
@@ -1617,7 +1601,6 @@ 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,
@@ -1636,7 +1619,7 @@ public:
 
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
 
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale);
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, nullptr, kq_scale, 0.0f);
 
         // split cached v into n_head heads
         struct ggml_tensor * v =
@@ -1934,7 +1917,7 @@ 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_IQ3_XXS, GGML_TYPE_IQ1_S,
     };
 
     // unary ops
@@ -2083,6 +2066,7 @@ 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)) {
@@ -2091,14 +2075,29 @@ 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}));
+            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));
         }
 
         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}, 0.1f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, 0.1f, true));
+    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));
 
     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
@@ -2113,7 +2112,6 @@ 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));
 
@@ -2129,14 +2127,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_leaky_relu());
 
+    // these tests are disabled to save execution time, but they can be handy for debugging
+#if 0
 #if !defined(__SANITIZE_THREAD__)
     // FIXME: these tests use too much memory with thread sanitizer
     test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
     //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
 #endif
-
-    // these tests are disabled to save execution time, but they can be handy for debugging
-#if 0
     test_cases.emplace_back(new test_llama(1));
     test_cases.emplace_back(new test_llama(2));
     test_cases.emplace_back(new test_falcon(1));

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, %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: %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 != 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: %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, "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, "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, %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: %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 != 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: %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, "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, "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, %d\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %d\n", element->type, element->value);
+                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 != actual_element\n");
             }
 

tests/test-model-load-cancel.cpp

@@ -14,7 +14,7 @@ int main(int argc, char *argv[] ) {
     fprintf(stderr, "using '%s'\n", model_path);
     fclose(file);
 
-    llama_backend_init(false);
+    llama_backend_init();
     auto params = llama_model_params{};
     params.use_mmap = false;
     params.progress_callback = [](float progress, void * ctx){

tests/test-quantize-fns.cpp

@@ -87,7 +87,7 @@ static float dot_product_error(
     vdot.from_float(test_data2, tmp_q2.data(), test_size);
 
     float result = INFINITY;
-    qfns.vec_dot(test_size, &result, tmp_q1.data(), tmp_q2.data());
+    qfns.vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
 
     const float dot_ref = dot_product(test_data1, test_data2, test_size);
 

tests/test-quantize-perf.cpp

@@ -346,7 +346,7 @@ int main(int argc, char * argv[]) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
                         float result;
-                        qfns.vec_dot(size, &result, test_q1, test_q2);
+                        qfns.vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
                         return result;
                     };
                     size_t quantized_size = ggml_row_size(type, size);

tests/test-tokenizer-0-falcon.cpp

@@ -61,7 +61,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {

tests/test-tokenizer-0-llama.cpp

@@ -60,7 +60,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {

tests/test-tokenizer-1-bpe.cpp

@@ -4,13 +4,13 @@
 #include "console.h"
 
 #include <cassert>
+#include <codecvt>
 #include <cstdio>
 #include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
 #include <locale>
+#include <string>
+#include <thread>
+#include <vector>
 
 int main(int argc, char **argv) {
     if (argc < 2) {
@@ -25,7 +25,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {
@@ -74,45 +74,46 @@ int main(int argc, char **argv) {
             }
         }
         catch (const std::invalid_argument &) {
-            fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
+            //fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
         }
     }
 
-    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
-        // NOTE: these exceptions seem to be necessary, because the GPT2 tokenizer doesn't want to interfere with some ASCII control characters
-        if ((cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 && (cp < 0x13 || cp > 0x17) && cp != 0x19 && (cp < 0x1c || cp > 0x1e) && (cp < 0xd800 || cp > 0xdfff)) {
-            std::string str = " " + codepoint_to_utf8(cp);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-            std::string check = llama_detokenize_bpe(ctx, tokens);
-            if (str != check) {
-                fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-                return 3;
-            }
-        }
-    }
-    // Restrict to assigned unicode planes
-    // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-    for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_bpe(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
-        }
-    }
-    for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_bpe(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
+    // unicode
+    {
+        const int nthread = std::thread::hardware_concurrency();
+
+        std::vector<std::thread> threads(nthread);
+
+        for (int i = 0; i < nthread; ++i) {
+            threads[i] = std::thread([i, nthread, ctx]() {
+                for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
+                    if (!( // NOLINT
+                                (cp < 0x03       || cp >  0x05)   && cp != 0x0b && cp != 0x11 &&
+                                (cp < 0x13       || cp >  0x17)   && cp != 0x19 &&
+                                (cp < 0x1c       || cp >  0x1e)   &&
+                                (cp < 0xd800     || cp >  0xdfff) &&
+                                (cp < 0x00040000 || cp >= 0x000e0000)
+                        )) {
+                        continue;
+                    }
+
+                    std::string str = codepoint_to_utf8(cp);
+                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+                    std::string check = llama_detokenize_bpe(ctx, tokens);
+                    if (cp != 9601 && str != check) {
+                        fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                                cp, check.c_str(), check.length(), str.c_str(), str.length());
+                        std::exit(3);
+                    }
+                }
+            });
+        }
+
+        for (auto & t : threads) {
+            t.join();
         }
     }
+
     llama_free_model(model);
     llama_free(ctx);
 

tests/test-tokenizer-1-llama.cpp

@@ -4,13 +4,13 @@
 #include "console.h"
 
 #include <cassert>
+#include <codecvt>
 #include <cstdio>
 #include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
 #include <locale>
+#include <string>
+#include <thread>
+#include <vector>
 
 int main(int argc, char **argv) {
     if (argc < 2) {
@@ -25,7 +25,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {
@@ -72,26 +72,33 @@ int main(int argc, char **argv) {
         }
     }
 
-    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
-        if (cp < 0xd800 || cp > 0xdfff) {
-            std::string str = codepoint_to_utf8(cp);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-            std::string check = llama_detokenize_spm(ctx, tokens);
-            if (cp != 9601 && str != check) {
-                fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-                return 3;
-            }
+    // unicode
+    {
+        const int nthread = std::thread::hardware_concurrency();
+
+        std::vector<std::thread> threads(nthread);
+
+        for (int i = 0; i < nthread; ++i) {
+            threads[i] = std::thread([i, nthread, ctx]() {
+                for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
+                    if (cp >= 0xd800 && cp <= 0xdfff) {
+                        continue;
+                    }
+
+                    std::string str = codepoint_to_utf8(cp);
+                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+                    std::string check = llama_detokenize_spm(ctx, tokens);
+                    if (cp != 9601 && str != check) {
+                        fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                                cp, check.c_str(), check.length(), str.c_str(), str.length());
+                        std::exit(3);
+                    }
+                }
+            });
         }
-    }
-    for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_spm(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
+
+        for (auto & t : threads) {
+            t.join();
         }
     }