ggml : do not use _GNU_SOURCE gratuitously

Co-authored-by: anon <anon@example.org>

.devops/full.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip
+    apt-get install -y build-essential python3 python3-pip git
 
 COPY requirements.txt requirements.txt
 
@@ -16,4 +16,6 @@ COPY . .
 
 RUN make
 
+ENV LC_ALL=C.utf8
+
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/main.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential
+    apt-get install -y build-essential git
 
 WORKDIR /app
 
@@ -15,4 +15,6 @@ FROM ubuntu:$UBUNTU_VERSION as runtime
 
 COPY --from=build /app/main /main
 
+ENV LC_ALL=C.utf8
+
 ENTRYPOINT [ "/main" ]

.devops/tools.sh

@@ -11,7 +11,7 @@ shift
 arg2="$@"
 
 if [[ $arg1 == '--convert' || $arg1 == '-c' ]]; then
-    python3 ./convert-pth-to-ggml.py $arg2
+    python3 ./convert.py $arg2
 elif [[ $arg1 == '--quantize' || $arg1 == '-q' ]]; then
     ./quantize $arg2
 elif [[ $arg1 == '--run' || $arg1 == '-r' ]]; then
@@ -32,7 +32,7 @@ else
     echo "  --run (-r): Run a model previously converted into ggml"
     echo "              ex: -m /models/7B/ggml-model-q4_0.bin -p \"Building a website can be done in 10 simple steps:\" -n 512"
     echo "  --convert (-c): Convert a llama model into ggml"
-    echo "              ex: \"/models/7B/\" 1"
+    echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
     echo "  --all-in-one (-a): Execute --convert & --quantize"

.flake8

@@ -0,0 +1,2 @@
+[flake8]
+max-line-length = 125

.github/workflows/build.yml

@@ -10,10 +10,10 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
 
 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -151,21 +151,21 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
-      CLBLAST_VERSION: 1.5.3
+      CLBLAST_VERSION: 1.6.0
 
     strategy:
       matrix:
         include:
           - build: 'avx2'
-            defines: ''
+            defines: '-DLLAMA_BUILD_SERVER=ON'
           - build: 'avx'
-            defines: '-DLLAMA_AVX2=OFF'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
           - build: 'avx512'
-            defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'clblast'
-            defines: '-DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
-            defines: '-DLLAMA_OPENBLAS=ON -DBLAS_LIBRARIES="/LIBPATH:$env:RUNNER_TEMP/openblas/lib" -DOPENBLAS_INC="$env:RUNNER_TEMP/openblas/include"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
 
     steps:
       - name: Clone
@@ -184,13 +184,13 @@ jobs:
         id: get_clblast
         if: ${{ matrix.build == 'clblast' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/clblast.zip -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-Windows-x64.zip"
+          curl.exe -o $env:RUNNER_TEMP/clblast.7z -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-windows-x64.7z"
           curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/clblast
-          tar.exe -xvf $env:RUNNER_TEMP/clblast.zip -C $env:RUNNER_TEMP/clblast
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/clblast.7z
+          rename-item $env:RUNNER_TEMP/CLBlast-${env:CLBLAST_VERSION}-windows-x64 clblast
           foreach ($f in (gci -Recurse -Path "$env:RUNNER_TEMP/clblast" -Filter '*.cmake')) {
             $txt = Get-Content -Path $f -Raw
-            $txt.Replace('C:/dependencies/opencl/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8
+            $txt.Replace('C:/vcpkg/packages/opencl_x64-windows/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8
           }
 
       - name: Download OpenBLAS
@@ -213,7 +213,6 @@ jobs:
           cd build
           cmake .. ${{ matrix.defines }}
           cmake --build . --config Release
-          cp ../LICENSE ./bin/Release/llama.cpp.txt
 
       - name: Add clblast.dll
         id: add_clblast_dll
@@ -258,6 +257,7 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
+          Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
           7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -292,7 +292,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_CUBLAS=ON
+          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON
           cmake --build . --config Release
 
       - name: Get commit hash

.github/workflows/tidy-post.yml

@@ -1,7 +1,7 @@
 name: clang-tidy review post comments
 
 on:
-  workflow_run:
+  workflow_dispatch:
     workflows: ["clang-tidy-review"]
     types:
       - completed

.gitignore

@@ -7,6 +7,7 @@
 .envrc
 .swiftpm
 .venv
+.clang-tidy
 .vs/
 .vscode/
 
@@ -17,9 +18,11 @@ build-release/
 build-static/
 build-cublas/
 build-opencl/
+build-metal/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
+out/
 
 models/*
 *.bin
@@ -30,13 +33,18 @@ models/*
 /result
 /perplexity
 /embedding
+/train-text-from-scratch
+/simple
 /benchmark-matmult
 /vdot
+/server
 /Pipfile
+/libllama.so
 
 build-info.h
 arm_neon.h
 compile_commands.json
+CMakeSettings.json
 
 __pycache__
 

.pre-commit-config.yaml

@@ -0,0 +1,15 @@
+# See https://pre-commit.com for more information
+# See https://pre-commit.com/hooks.html for more hooks
+exclude: prompts/.*.txt
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+  rev: v3.2.0
+  hooks:
+  - id: trailing-whitespace
+  - id: end-of-file-fixer
+  - id: check-yaml
+  - id: check-added-large-files
+- repo: https://github.com/PyCQA/flake8
+  rev: 6.0.0
+  hooks:
+  -   id: flake8

CMakeLists.txt

@@ -37,40 +37,48 @@ endif()
 #
 
 # general
-option(LLAMA_STATIC                 "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                 "llama: enable -march=native flag"                      OFF)
-option(LLAMA_LTO                    "llama: enable link time optimization"                  OFF)
+option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
+option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
-option(LLAMA_ALL_WARNINGS           "llama: enable all compiler warnings"                   ON)
-option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
-option(LLAMA_GPROF                  "llama: enable gprof"                                   OFF)
+option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"                   ON)
+option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
+option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
 # sanitizers
-option(LLAMA_SANITIZE_THREAD        "llama: enable thread sanitizer"                        OFF)
-option(LLAMA_SANITIZE_ADDRESS       "llama: enable address sanitizer"                       OFF)
-option(LLAMA_SANITIZE_UNDEFINED     "llama: enable undefined sanitizer"                     OFF)
+option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
+option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
+option(LLAMA_SANITIZE_UNDEFINED         "llama: enable undefined sanitizer"                     OFF)
 
 # instruction set specific
-option(LLAMA_AVX                    "llama: enable AVX"                                     ON)
-option(LLAMA_AVX2                   "llama: enable AVX2"                                    ON)
-option(LLAMA_AVX512                 "llama: enable AVX512"                                  OFF)
-option(LLAMA_AVX512_VBMI            "llama: enable AVX512-VBMI"                             OFF)
-option(LLAMA_AVX512_VNNI            "llama: enable AVX512-VNNI"                             OFF)
-option(LLAMA_FMA                    "llama: enable FMA"                                     ON)
+option(LLAMA_AVX                        "llama: enable AVX"                                     ON)
+option(LLAMA_AVX2                       "llama: enable AVX2"                                    ON)
+option(LLAMA_AVX512                     "llama: enable AVX512"                                  OFF)
+option(LLAMA_AVX512_VBMI                "llama: enable AVX512-VBMI"                             OFF)
+option(LLAMA_AVX512_VNNI                "llama: enable AVX512-VNNI"                             OFF)
+option(LLAMA_FMA                        "llama: enable FMA"                                     ON)
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
-    option(LLAMA_F16C               "llama: enable F16C"                                    ON)
+    option(LLAMA_F16C                   "llama: enable F16C"                                    ON)
 endif()
 
 # 3rd party libs
-option(LLAMA_ACCELERATE             "llama: enable Accelerate framework"                    ON)
-option(LLAMA_OPENBLAS               "llama: use OpenBLAS"                                   OFF)
-option(LLAMA_CUBLAS                 "llama: use cuBLAS"                                     OFF)
-option(LLAMA_CLBLAST                "llama: use CLBlast"                                    OFF)
+option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
+option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
+set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
+option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
+set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
+set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
+option(LLAMA_CUDA_DMMV_F16                   "llama: use 16 bit floats for dmmv CUDA kernels"   OFF)
+set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
+option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
+option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 
-option(LLAMA_BUILD_TESTS            "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES         "llama: build examples" ${LLAMA_STANDALONE})
+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"                           OFF)
 
 #
 # Build info header
@@ -145,36 +153,75 @@ if (APPLE AND LLAMA_ACCELERATE)
     endif()
 endif()
 
-if (LLAMA_OPENBLAS)
+if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
     endif()
+    if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
+        set(BLA_SIZEOF_INTEGER 8)
+    endif()
 
-    set(BLA_VENDOR OpenBLAS)
+    set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
     find_package(BLAS)
+
     if (BLAS_FOUND)
-        message(STATUS "OpenBLAS found")
+        message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
+        if ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+            find_package(PkgConfig REQUIRED)
+            if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
+                pkg_check_modules(DepBLAS REQUIRED blas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
+                pkg_check_modules(DepBLAS REQUIRED openblas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
+                pkg_check_modules(DepBLAS REQUIRED blis)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
+                pkg_check_modules(DepBLAS REQUIRED blas-atlas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FlexiBLAS")
+                pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "Intel")
+                # all Intel* libraries share the same include path
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "NVHPC")
+                # this doesn't provide pkg-config
+                # suggest to assign BLAS_INCLUDE_DIRS on your own
+                if ("${NVHPC_VERSION}" STREQUAL "")
+                    message(WARNING "Better to set NVHPC_VERSION")
+                else()
+                    set(DepBLAS_FOUND ON)
+                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+                endif()
+            endif()
+            if (DepBLAS_FOUND)
+                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+            else()
+                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+                " detected by pkgconfig, trying to find cblas.h from possible paths...")
+                find_path(BLAS_INCLUDE_DIRS
+                    NAMES cblas.h
+                    HINTS
+                        /usr/include
+                        /usr/local/include
+                        /usr/include/openblas
+                        /opt/homebrew/opt/openblas/include
+                        /usr/local/opt/openblas/include
+                        /usr/include/x86_64-linux-gnu/openblas/include
+                )
+            endif()
+        endif()
+
+        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+        add_compile_options(${BLAS_LINKER_FLAGS})
         add_compile_definitions(GGML_USE_OPENBLAS)
-        add_link_options(${BLAS_LIBRARIES})
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
-        # find header file
-        set(OPENBLAS_INCLUDE_SEARCH_PATHS
-            /usr/include
-            /usr/include/openblas
-            /usr/include/openblas-base
-            /usr/local/include
-            /usr/local/include/openblas
-            /usr/local/include/openblas-base
-            /opt/OpenBLAS/include
-            $ENV{OpenBLAS_HOME}
-            $ENV{OpenBLAS_HOME}/include
-            )
-        find_path(OPENBLAS_INC NAMES cblas.h PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
-        add_compile_options(-I${OPENBLAS_INC})
     else()
-        message(WARNING "OpenBLAS not found")
+        message(WARNING "BLAS not found, please refer to "
+        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+        " to set correct LLAMA_BLAS_VENDOR")
     endif()
 endif()
 
@@ -187,9 +234,15 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
+        if (LLAMA_CUDA_DMMV_F16)
+            add_compile_definitions(GGML_CUDA_DMMV_F16)
+        endif()
+        add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -197,17 +250,56 @@ if (LLAMA_CUBLAS)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        if (LLAMA_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "61") # needed for f16 CUDA intrinsics
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52") # lowest CUDA 12 standard
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
     else()
         message(WARNING "cuBLAS not found")
     endif()
 endif()
 
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+    find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
+
+    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+    add_compile_definitions(GGML_USE_METAL)
+    add_compile_definitions(GGML_METAL_NDEBUG)
+
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        ${METALPERFORMANCE_FRAMEWORK}
+        )
+endif()
+
+if (LLAMA_K_QUANTS)
+    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
+    add_compile_definitions(GGML_USE_K_QUANTS)
+endif()
+
 if (LLAMA_CLBLAST)
     find_package(CLBlast)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_OPENCL_SOURCES ggml-opencl.c ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
@@ -372,36 +464,42 @@ endif()
 add_library(ggml OBJECT
             ggml.c
             ggml.h
-            ${GGML_CUDA_SOURCES}
-            ${GGML_OPENCL_SOURCES})
+            ${GGML_SOURCES_CUDA}
+            ${GGML_SOURCES_OPENCL}
+            ${GGML_SOURCES_METAL}
+            ${GGML_SOURCES_EXTRA}
+            )
 
-target_include_directories(ggml PUBLIC .)
+target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 
+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>)
+    target_link_libraries(ggml_shared PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 endif()
 
 add_library(llama
             llama.cpp
             llama.h
-            llama-util.h)
+            llama-util.h
+            )
 
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
-target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
+target_link_libraries(llama PRIVATE
+    ggml
+    ${LLAMA_EXTRA_LIBS}
+    )
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
-endif()
-
-if (GGML_CUDA_SOURCES)
-    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF)
-    set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF)
+    if (LLAMA_METAL)
+        set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    endif()
 endif()
 
 

Makefile

@@ -1,5 +1,13 @@
 # Define the default target now so that it is always the first target
-default: main quantize quantize-stats perplexity embedding vdot
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple
+
+ifdef LLAMA_BUILD_SERVER
+	BUILD_TARGETS += server
+	LLAMA_SERVER_VERBOSE ?= 1
+server: private CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
+endif
+
+default: $(BUILD_TARGETS)
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -34,11 +42,26 @@ endif
 #
 
 # keep standard at C11 and C++11
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+# -Ofast tends to produce faster code, but may not be available for some compilers.
+#OPT = -Ofast
+OPT = -O3
+CFLAGS   = -I.              $(OPT) -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples $(OPT) -std=c++11 -fPIC
 LDFLAGS  =
 
-ifndef LLAMA_DEBUG
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+# ref: https://github.com/ggerganov/whisper.cpp/pull/1027
+CFLAGS   += -D_XOPEN_SOURCE=600
+CXXFLAGS += -D_XOPEN_SOURCE=600
+
+ifdef LLAMA_DEBUG
+	CFLAGS   += -O0 -g
+	CXXFLAGS += -O0 -g
+	LDFLAGS  += -g
+else
 	CFLAGS   += -DNDEBUG
 	CXXFLAGS += -DNDEBUG
 endif
@@ -94,7 +117,12 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 	# Usage AVX-only
 	#CFLAGS   += -mfma -mf16c -mavx
 	#CXXFLAGS += -mfma -mf16c -mavx
+
+	# Usage SSSE3-only (Not is SSE3!)
+	#CFLAGS   += -mssse3
+	#CXXFLAGS += -mssse3
 endif
+
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
@@ -106,6 +134,13 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
 	endif
 endif
+
+ifndef LLAMA_NO_K_QUANTS
+	CFLAGS   += -DGGML_USE_K_QUANTS
+	CXXFLAGS += -DGGML_USE_K_QUANTS
+	OBJS     += k_quants.o
+endif
+
 ifndef LLAMA_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework.
 	# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
@@ -113,15 +148,18 @@ ifndef LLAMA_NO_ACCELERATE
 		CFLAGS  += -DGGML_USE_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
-endif
+endif # LLAMA_NO_ACCELERATE
+
 ifdef LLAMA_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
-	ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),)
-		LDFLAGS += -lopenblas -lcblas
-	else
-		LDFLAGS += -lopenblas
-	endif
-endif
+	LDFLAGS += -lopenblas
+endif # LLAMA_OPENBLAS
+
+ifdef LLAMA_BLIS
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
+	LDFLAGS += -lblis -L/usr/local/lib
+endif # LLAMA_BLIS
+
 ifdef LLAMA_CUBLAS
 	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
@@ -129,11 +167,31 @@ ifdef LLAMA_CUBLAS
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
+ifdef LLAMA_CUDA_DMMV_X
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+else
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+endif # LLAMA_CUDA_DMMV_X
+ifdef LLAMA_CUDA_DMMV_Y
+	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=$(LLAMA_CUDA_DMMV_Y)
+else
+	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
+endif # LLAMA_CUDA_DMMV_Y
+ifdef LLAMA_CUDA_DMMV_F16
+	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
+endif # LLAMA_CUDA_DMMV_F16
+ifdef LLAMA_CUDA_KQUANTS_ITER
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+else
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
-endif
+endif # LLAMA_CUBLAS
+
 ifdef LLAMA_CLBLAST
-	CFLAGS  += -DGGML_USE_CLBLAST
+	CFLAGS   += -DGGML_USE_CLBLAST
+	CXXFLAGS += -DGGML_USE_CLBLAST
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
 		LDFLAGS += -lclblast -framework OpenCL
@@ -141,28 +199,48 @@ ifdef LLAMA_CLBLAST
 		LDFLAGS += -lclblast -lOpenCL
 	endif
 	OBJS    += ggml-opencl.o
-ggml-opencl.o: ggml-opencl.c ggml-opencl.h
+
+ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_CLBLAST
+
+ifdef LLAMA_METAL
+	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
+	CXXFLAGS += -DGGML_USE_METAL
+	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+	OBJS     += ggml-metal.o
+
+ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-endif
+endif # LLAMA_METAL
+
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
 	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
+
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, Zero
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
+
 ifneq ($(filter armv7%,$(UNAME_M)),)
 	# Raspberry Pi 2
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
+
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 3, 4, Zero 2 (32-bit)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
+ifdef LLAMA_NO_K_QUANTS
+k_quants.o: k_quants.c k_quants.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_NO_K_QUANTS
+
 #
 # Print build information
 #
@@ -185,7 +263,7 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-cuda.h llama.h llama-util.h
+llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 common.o: examples/common.cpp examples/common.h
@@ -195,33 +273,42 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state build-info.h
+	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot train-text-from-scratch build-info.h
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS)
+simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
+
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
@@ -241,6 +328,6 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-.PHONY: tests
+.PHONY: tests clean
 tests:
 	bash ./tests/run-tests.sh

Package.swift

@@ -11,6 +11,7 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
+            exclude: ["ggml-metal.metal"],
             sources: ["ggml.c", "llama.cpp"],
             publicHeadersPath: "spm-headers",
             cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],

README.md

@@ -5,13 +5,15 @@
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
-- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508)
-- Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405)
-- [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220)
+- New roadmap: https://github.com/users/ggerganov/projects/7
+- Azure CI brainstorming: https://github.com/ggerganov/llama.cpp/discussions/1985
+- p1 : LLM-based code completion engine at the edge : https://github.com/ggml-org/p1/discussions/1
 
 <details>
   <summary>Table of Contents</summary>
@@ -30,6 +32,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
         <li><a href="#instruction-mode-with-alpaca">Instruction mode with Alpaca</a></li>
+        <li><a href="#using-openllama">Using OpenLLaMA</a></li>
         <li><a href="#using-gpt4all">Using GPT4All</a></li>
         <li><a href="#using-pygmalion-7b--metharme-7b">Using Pygmalion 7B & Metharme 7B</a></li>
         <li><a href="#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data">Obtaining the Facebook LLaMA original model and Stanford Alpaca model data</a></li>
@@ -51,12 +54,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quantization on a MacBook
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework
+- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
 - 4-bit, 5-bit and 8-bit integer quantization support
-- Runs on the CPU
-- OpenBLAS support
+- Supports OpenBLAS/Apple BLAS/ARM Performance Lib/ATLAS/BLIS/Intel MKL/NVHPC/ACML/SCSL/SGIMATH and [more](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) in BLAS
 - cuBLAS and CLBlast support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
@@ -236,15 +238,41 @@ In order to build llama.cpp you have three different options.
     zig build -Drelease-fast
     ```
 
+### Metal Build
+
+Using Metal allows the computation to be executed on the GPU for Apple devices:
+
+- Using `make`:
+
+  ```bash
+  LLAMA_METAL=1 make
+  ```
+
+- Using `CMake`:
+
+    ```bash
+    mkdir build-metal
+    cd build-metal
+    cmake -DLLAMA_METAL=ON ..
+    cmake --build . --config Release
+    ```
+
+When built with Metal support, you can enable GPU inference with the `--gpu-layers|-ngl` command-line argument.
+Any value larger than 0 will offload the computation to the GPU. For example:
+
+```bash
+./main -m ./models/7B/ggml-model-q4_0.bin -n 128 -ngl 1
+```
+
 ### BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
 
-- Accelerate Framework:
+- #### Accelerate Framework:
 
   This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
 
-- OpenBLAS:
+- #### OpenBLAS:
 
   This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
 
@@ -274,11 +302,26 @@ Building the program with BLAS support may lead to some performance improvements
       ```bash
       mkdir build
       cd build
-      cmake .. -DLLAMA_OPENBLAS=ON
+      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
       cmake --build . --config Release
       ```
 
-- cuBLAS
+- #### BLIS
+
+  Check [BLIS.md](docs/BLIS.md) for more information.
+
+- #### Intel MKL
+
+  By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. You may also specify it by:
+
+  ```bash
+  mkdir build
+  cd build
+  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+  cmake --build . --config Release
+  ```
+
+- #### cuBLAS
 
   This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
   - Using `make`:
@@ -294,7 +337,88 @@ Building the program with BLAS support may lead to some performance improvements
     cmake --build . --config Release
     ```
 
-Note: Because llama.cpp uses multiple CUDA streams for matrix multiplication results [are not guaranteed to be reproducible](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility). If you need reproducibility, set `GGML_CUDA_MAX_STREAMS` in the file `ggml-cuda.cu` to 1.
+  The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
+
+  | Option                  | Legal values           | Default | Description |
+  |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA dequantization + mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+
+- #### CLBlast
+
+  OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
+
+  You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
+    - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed.
+
+    - <details>
+        <summary>Installing the OpenCL SDK from source</summary>
+
+        ```sh
+        git clone --recurse-submodules https://github.com/KhronosGroup/OpenCL-SDK.git
+        mkdir OpenCL-SDK/build
+        cd OpenCL-SDK/build
+        cmake .. -DBUILD_DOCS=OFF \
+          -DBUILD_EXAMPLES=OFF \
+          -DBUILD_TESTING=OFF \
+          -DOPENCL_SDK_BUILD_SAMPLES=OFF \
+          -DOPENCL_SDK_TEST_SAMPLES=OFF
+        cmake --build . --config Release
+        cmake --install . --prefix /some/path
+        ```
+      </details>
+
+  Installing CLBlast: it may be found in your operating system's packages.
+
+  - <details>
+    <summary>If not, then installing from source:</summary>
+
+      ```sh
+      git clone https://github.com/CNugteren/CLBlast.git
+      mkdir CLBlast/build
+      cd CLBlast/build
+      cmake .. -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF
+      cmake --build . --config Release
+      cmake --install . --prefix /some/path
+      ```
+
+      Where `/some/path` is where the built library will be installed (default is `/usr/local`).
+    </details>
+
+  Building:
+
+  - Build with make:
+    ```sh
+    make LLAMA_CLBLAST=1
+    ```
+  - CMake:
+    ```sh
+    mkdir build
+    cd build
+    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_dir=/some/path
+    cmake --build . --config Release
+    ```
+
+  Running:
+
+  The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
+
+  To select the correct platform (driver) and device (GPU), you can use the environment variables `GGML_OPENCL_PLATFORM` and `GGML_OPENCL_DEVICE`.
+  The selection can be a number (starting from 0) or a text string to search:
+
+  ```sh
+  GGML_OPENCL_PLATFORM=1 ./main ...
+  GGML_OPENCL_DEVICE=2 ./main ...
+  GGML_OPENCL_PLATFORM=Intel ./main ...
+  GGML_OPENCL_PLATFORM=AMD GGML_OPENCL_DEVICE=1 ./main ...
+  ```
+
+  The default behavior is to find the first GPU device, but when it is an integrated GPU on a laptop, for instance, the selectors are useful.
+  Using the variables it is possible to select a CPU-based driver as well, if so desired.
+
+  You can get a list of platforms and devices from the `clinfo -l` command, etc.
 
 ### Prepare Data & Run
 
@@ -376,6 +500,25 @@ Note the use of `--color` to distinguish between user input and generated text.
 
 ![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
 
+### Persistent Interaction
+
+The prompt, user inputs, and model generations can be saved and resumed across calls to `./main` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
+
+```bash
+# Start a new chat
+PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
+
+# Resume that chat
+PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
+
+# Start a different chat with the same prompt/model
+PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/another ./examples/chat-persistent.sh
+
+# Different prompt cache for different prompt/model
+PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
+    CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh
+```
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder
@@ -404,6 +547,13 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 >
 ```
 
+### Using [OpenLLaMA](https://github.com/openlm-research/open_llama)
+
+OpenLLaMA is an openly licensed reproduction of Meta's original LLaMA model. It uses the same architecture and is a drop-in replacement for the original LLaMA weights.
+
+- Download the [3B](https://huggingface.co/openlm-research/open_llama_3b), [7B](https://huggingface.co/openlm-research/open_llama_7b), or [13B](https://huggingface.co/openlm-research/open_llama_13b) model from Hugging Face.
+- Convert the model to ggml FP16 format using `python convert.py <path to OpenLLaMA directory>`
+
 ### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
 
 - Obtain the `tokenizer.model` file from LLaMA model and put it to `models`
@@ -479,8 +629,14 @@ And after 4.45 hours, you will have the final perplexity.
 
 ### Android
 
+#### Building the Project using Android NDK
 You can easily run `llama.cpp` on Android device with [termux](https://termux.dev/).
-First, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
+
+First, install the essential packages for termux:
+```
+pkg install clang wget git cmake
+```
+Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
 ```
 $ mkdir build-android
 $ cd build-android
@@ -493,6 +649,47 @@ Finally, copy the `llama` binary and the model files to your device storage. Her
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
+#### Building the Project using Termux (F-Droid)
+Termux from F-Droid offers an alternative route to execute the project on an Android device. This method empowers you to construct the project right from within the terminal, negating the requirement for a rooted device or SD Card.
+
+Outlined below are the directives for installing the project using OpenBLAS and CLBlast. This combination is specifically designed to deliver peak performance on recent devices that feature a GPU.
+
+If you opt to utilize OpenBLAS, you'll need to install the corresponding package.
+```
+apt install libopenblas
+```
+
+Subsequently, if you decide to incorporate CLBlast, you'll first need to install the requisite OpenCL packages:
+```
+apt install ocl-icd opencl-headers opencl-clhpp clinfo
+```
+
+In order to compile CLBlast, you'll need to first clone the respective Git repository, which can be found at this URL: https://github.com/CNugteren/CLBlast. Alongside this, clone this repository into your home directory. Once this is done, navigate to the CLBlast folder and execute the commands detailed below:
+```
+cmake .
+make
+cp libclblast.so* $PREFIX/lib
+cp ./include/clblast.h ../llama.cpp
+```
+
+Following the previous steps, navigate to the LlamaCpp directory. To compile it with OpenBLAS and CLBlast, execute the command provided below:
+```
+cp /data/data/com.termux/files/usr/include/openblas/cblas.h .
+cp /data/data/com.termux/files/usr/include/openblas/openblas_config.h .
+make LLAMA_CLBLAST=1 //(sometimes you need to run this command twice)
+```
+
+Upon completion of the aforementioned steps, you will have successfully compiled the project. To run it using CLBlast, a slight adjustment is required: a command must be issued to direct the operations towards your device's physical GPU, rather than the virtual one. The necessary command is detailed below:
+```
+GGML_OPENCL_PLATFORM=0
+GGML_OPENCL_DEVICE=0
+export LD_LIBRARY_PATH=/vendor/lib64:$LD_LIBRARY_PATH
+```
+
+For easy and swift re-execution, consider documenting this final part in a .sh script file. This will enable you to rerun the process with minimal hassle.
+
+Place your desired model into the `/llama.cpp/models/` directory and execute the `./main (...)` script.
+
 ### Docker
 
 #### Prerequisites
@@ -548,3 +745,4 @@ docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /mode
 ### Docs
 
 - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [Performance troubleshooting](./docs/token_generation_performance_tips.md)

SHA256SUMS

@@ -1,6 +1,6 @@
 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
 666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_0.bin
+ec2f2d1f0dfb73b72a4cbac7fa121abbe04c37ab327125a38248f930c0f09ddf  models/7B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_1.bin
@@ -8,7 +8,7 @@ ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml
 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
 d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
 2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_0.bin
+fad169e6f0f575402cf75945961cb4a8ecd824ba4da6be2af831f320c4348fa5  models/13B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_1.bin
@@ -18,7 +18,7 @@ e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/con
 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
 7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_0.bin
+d2a441403944819492ec8c2002cc36fa38468149bfb4b7b4c52afc7bd9a7166d  models/30B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_1.bin
@@ -32,7 +32,7 @@ a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/con
 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
 d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
 60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_0.bin
+cde053439fa4910ae454407e2717cc46cc2c2b4995c00c93297a2b52e790fa92  models/65B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_1.bin

build.zig

@@ -1,61 +1,58 @@
 const std = @import("std");
 
+// Zig Version: 0.11.0-dev.3379+629f0d23b
 pub fn build(b: *std.build.Builder) void {
     const target = b.standardTargetOptions(.{});
-    const optimize = b.standardReleaseOptions();
-    const want_lto = b.option(bool, "lto", "Want -fLTO");
-
-    const lib = b.addStaticLibrary("llama", null);
-    lib.want_lto = want_lto;
-    lib.setTarget(target);
-    lib.setBuildMode(optimize);
+    const optimize = b.standardOptimizeOption(.{});
+    const lib = b.addStaticLibrary(.{
+        .name = "llama",
+        .target = target,
+        .optimize = optimize,
+    });
+    lib.linkLibC();
     lib.linkLibCpp();
     lib.addIncludePath(".");
-    lib.addIncludePath("examples");
+    lib.addIncludePath("./examples");
     lib.addCSourceFiles(&.{
         "ggml.c",
     }, &.{"-std=c11"});
     lib.addCSourceFiles(&.{
         "llama.cpp",
     }, &.{"-std=c++11"});
-    lib.install();
+    b.installArtifact(lib);
 
-    const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize, .want_lto = want_lto };
+    const examples = .{
+        "main",
+        "baby-llama",
+        "embedding",
+        // "metal",
+        "perplexity",
+        "quantize",
+        "quantize-stats",
+        "save-load-state",
+        // "server",
+        "simple",
+        "train-text-from-scratch",
+    };
 
-    const exe = build_example("main", build_args);
-    _ = build_example("quantize", build_args);
-    _ = build_example("perplexity", build_args);
-    _ = build_example("embedding", build_args);
-
-    // create "zig build run" command for ./main
-
-    const run_cmd = exe.run();
-    run_cmd.step.dependOn(b.getInstallStep());
-    if (b.args) |args| {
-        run_cmd.addArgs(args);
+    inline for (examples) |example_name| {
+        const exe = b.addExecutable(.{
+            .name = example_name,
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(".");
+        exe.addIncludePath("./examples");
+        exe.addCSourceFiles(&.{
+            std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{example_name, example_name}),
+            "examples/common.cpp",
+        }, &.{"-std=c++11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ example_name, "Run the app");
+        run_step.dependOn(&run_cmd.step);
     }
-
-    const run_step = b.step("run", "Run the app");
-    run_step.dependOn(&run_cmd.step);
-}
-
-fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep {
-    const b = args.b;
-    const lib = args.lib;
-    const want_lto = args.want_lto;
-
-    const exe = b.addExecutable(name, null);
-    exe.want_lto = want_lto;
-    lib.setTarget(args.target);
-    lib.setBuildMode(args.optimize);
-    exe.addIncludePath(".");
-    exe.addIncludePath("examples");
-    exe.addCSourceFiles(&.{
-        std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{name, name}),
-        "examples/common.cpp",
-    }, &.{"-std=c++11"});
-    exe.linkLibrary(lib);
-    exe.install();
-
-    return exe;
 }

convert-pth-to-ggml.py

@@ -4,7 +4,9 @@ import argparse
 
 import convert
 
-parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
+parser = argparse.ArgumentParser(
+    description="""[DEPRECATED - use `convert.py` instead]
+    Convert a LLaMA model checkpoint to a ggml compatible file""")
 parser.add_argument('dir_model',  help='directory containing the model checkpoint')
 parser.add_argument('ftype',      help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
 args = parser.parse_args()

convert.py

@@ -130,6 +130,14 @@ TENSORS_LIST = make_tensors_list()
 TENSORS_SET = set(TENSORS_LIST)
 
 
+def find_n_mult(n_ff: int, n_embd: int) -> int:
+    # hardcoded magic range
+    for n_mult in range(256, 1, -1):
+        calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
+        if calc_ff == n_ff:
+            return n_mult
+    return 1
+
 @dataclass
 class Params:
     n_vocab: int
@@ -137,21 +145,61 @@ class Params:
     n_mult: int
     n_head: int
     n_layer: int
-    file_type: GGMLFileType
 
     @staticmethod
-    def guessed(model: 'LazyModel', file_type: GGMLFileType) -> 'Params':
-        n_vocab, n_embd = model["tok_embeddings.weight"].shape
+    def guessed(model: 'LazyModel') -> 'Params':
+        # try transformer naming first
+        n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
+
+        # try transformer naming first
+        if "model.layers.0.self_attn.q_proj.weight" in model:
+            n_layer=next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
+        else:
+            n_layer=next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
+
+        n_head=n_embd // 128 # guessed
 
         return Params(
             n_vocab=n_vocab,
             n_embd=n_embd,
             n_mult=256,
-            n_head=n_embd // 128,
-            n_layer=next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model),
-            file_type=file_type,
+            n_head=n_head,
+            n_layer=n_layer,
         )
 
+    @staticmethod
+    def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+        config = json.load(open(config_path))
+
+        n_vocab = config["vocab_size"];
+        n_embd = config["hidden_size"];
+        n_head = config["num_attention_heads"];
+        n_layer = config["num_hidden_layers"];
+        n_ff = config["intermediate_size"];
+
+        n_mult = find_n_mult(n_ff, n_embd);
+
+        return Params(
+            n_vocab=n_vocab,
+            n_embd=n_embd,
+            n_mult=n_mult,
+            n_head=n_head,
+            n_layer=n_layer,
+        )
+
+    @staticmethod
+    def load(model_plus: 'ModelPlus') -> 'Params':
+        orig_config_path = model_plus.paths[0].parent / "params.json"
+        hf_transformer_config_path = model_plus.paths[0].parent / "config.json"
+
+        if hf_transformer_config_path.exists():
+            params = Params.loadHFTransformerJson(model_plus.model, hf_transformer_config_path)
+        else:
+            params = Params.guessed(model_plus.model)
+
+        print(f'params: n_vocab:{params.n_vocab} n_embd:{params.n_embd} n_mult:{params.n_mult} n_head:{params.n_head} n_layer:{params.n_layer}')
+        return params
+
 
 class SentencePieceVocab:
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None:
@@ -512,7 +560,11 @@ class LazyTensor:
             if not isinstance(self.data_type, QuantizedDataType):
                 raise Exception(f"Can't turn an unquantized tensor into a quantized type ({data_type})")
             if self.data_type.have_g_idx:
-                sys.stderr.write("Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), which is not yet natively supported by GGML.  For now you can still convert this model by passing `--outtype f16` to dequantize, but that will result in a much larger output file for no quality benefit.\n")
+                sys.stderr.write(
+                    "Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), "
+                    "which is not yet natively supported by GGML. "
+                    "For now you can still convert this model by passing `--outtype f16` to dequantize, "
+                    "but that will result in a much larger output file for no quality benefit.\n")
                 sys.exit(1)
             assert not data_type.have_g_idx and self.data_type.have_addends and data_type.have_addends
 
@@ -591,18 +643,17 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor:
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
 
 
-def convert_transformers_to_orig(model: LazyModel) -> LazyModel:
+def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
     out: LazyModel = {}
     out["tok_embeddings.weight"] = model["model.embed_tokens.weight"]
     out["norm.weight"] = model["model.norm.weight"]
     out["output.weight"] = model["lm_head.weight"]
 
-    n_head = model["model.layers.0.self_attn.q_proj.weight"].shape[1] // 128
     for i in itertools.count():
         if f"model.layers.{i}.self_attn.q_proj.weight" not in model:
             break
-        out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], n_head)
-        out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], n_head)
+        out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
+        out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head)
         out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
         out[f"layers.{i}.attention.wo.weight"] = model[f"model.layers.{i}.self_attn.o_proj.weight"]
 
@@ -694,8 +745,9 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         return LazyStorage(load=load, kind=pid[1], description=description)
 
-   # @staticmethod
-    def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,  # pyright: ignore[reportSelfClsParameterName]
+    # @staticmethod
+    def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
+                               # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
         assert isinstance(storage, LazyStorage)
 
@@ -812,7 +864,7 @@ def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus:
     # Use mmap for the actual data to avoid race conditions with the file offset.
     off = fp.raw.tell()
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
-    fp.raw.seek(off) # needed on Windows
+    fp.raw.seek(off)  # needed on Windows
 
     def read_tensor() -> None:  # this is a function so that variables captured in `load` don't change
         shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12))
@@ -915,7 +967,7 @@ class OutputFile:
     def __init__(self, fname_out: Path) -> None:
         self.fout = open(fname_out, "wb")
 
-    def write_file_header(self, params: Params) -> None:
+    def write_file_header(self, params: Params, file_type: GGMLFileType) -> None:
         self.fout.write(b"ggjt"[::-1])  # magic
         values = [
             1,  # file version
@@ -925,7 +977,7 @@ class OutputFile:
             params.n_head,
             params.n_layer,
             params.n_embd // params.n_head,  # rot (obsolete)
-            params.file_type.value,
+            file_type.value,
         ]
         self.fout.write(struct.pack("i" * len(values), *values))
 
@@ -946,17 +998,17 @@ class OutputFile:
     def write_vocab_only(fname_out: Path, vocab: Vocab) -> None:
         of = OutputFile(fname_out)
         params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0,
-                        n_head=1, n_layer=0, file_type=GGMLFileType.AllF32)
+                        n_head=1, n_layer=0)
         of = OutputFile(fname_out)
-        of.write_file_header(params)
+        of.write_file_header(params, file_type=GGMLFileType.AllF32)
         of.write_vocab(vocab)
         of.fout.close()
 
     @staticmethod
-    def write_all(fname_out: Path, params: Params, model: LazyModel, vocab: Vocab) -> None:
+    def write_all(fname_out: Path, params: Params, file_type: GGMLFileType, model: LazyModel, vocab: Vocab) -> None:
         check_vocab_size(params, vocab)
         of = OutputFile(fname_out)
-        of.write_file_header(params)
+        of.write_file_header(params, file_type)
         print("Writing vocab...")
         of.write_vocab(vocab)
 
@@ -992,11 +1044,11 @@ def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFi
     raise Exception(f"Unexpected combination of types: {name_to_type}")
 
 
-def do_necessary_conversions(model: LazyModel) -> LazyModel:
+def do_necessary_conversions(model: LazyModel, params: Params) -> LazyModel:
     model = handle_quantization(model)
 
     if "lm_head.weight" in model:
-        model = convert_transformers_to_orig(model)
+        model = convert_transformers_to_orig(model, params)
     model = filter_and_sort_tensors(model)
 
     return model
@@ -1054,7 +1106,7 @@ def load_some_model(path: Path) -> ModelPlus:
         files = list(path.glob("model-00001-of-*.safetensors"))
         if not files:
             # Try the PyTorch patterns too, with lower priority
-            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ]
+            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
             files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try GGML too, but with lower priority, since if both a non-GGML
@@ -1094,23 +1146,27 @@ def load_vocab(path: Path) -> SentencePieceVocab:
         elif path3.exists():
             path = path3
         else:
-            raise FileNotFoundError(f"Could not find tokenizer.model in {path} or its parent; if it's in another directory, pass the directory as --vocab-dir")
+            raise FileNotFoundError(
+                f"Could not find tokenizer.model in {path} or its parent; "
+                "if it's in another directory, pass the directory as --vocab-dir")
     added_tokens_path = path.parent / "added_tokens.json"
     print(f"Loading vocab file {path}")
     return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
 
 
-def default_outfile(model_paths: List[Path], params: Params) -> Path:
+def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
     namestr = {
         GGMLFileType.AllF32: "f32",
         GGMLFileType.MostlyF16: "f16",
         GGMLFileType.MostlyQ4_0: "q4_0",
         GGMLFileType.MostlyQ4_1: "q4_1",
         GGMLFileType.PerLayerIsQ4_1: "q4_1",
-    }[params.file_type]
+    }[file_type]
     ret = model_paths[0].parent / f"ggml-model-{namestr}.bin"
     if ret in model_paths:
-        sys.stderr.write(f"Error: Default output path ({ret}) would overwrite the input.  Please explicitly specify a path using --outfile.\n")
+        sys.stderr.write(
+            f"Error: Default output path ({ret}) would overwrite the input. "
+            "Please explicitly specify a path using --outfile.\n")
         sys.exit(1)
     return ret
 
@@ -1131,7 +1187,8 @@ def main(args_in: Optional[List[str]] = None) -> None:
     parser.add_argument("--outtype", choices=["f32", "f16", "q4_1", "q4_0"], help="output format (default: based on input)")
     parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
     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("model", type=Path,
+                        help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     args = parser.parse_args(args_in)
 
     vocab: Vocab
@@ -1154,13 +1211,13 @@ def main(args_in: Optional[List[str]] = None) -> None:
         else:
             vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
             vocab = load_vocab(vocab_dir)
+        params = Params.load(model_plus)
         model = model_plus.model
-        model = do_necessary_conversions(model)
+        model = do_necessary_conversions(model, params)
         output_type = pick_output_type(model, args.outtype)
         model = convert_to_output_type(model, output_type)
-        params = Params.guessed(model, output_type)
-        outfile = args.outfile or default_outfile(model_plus.paths, params)
-        OutputFile.write_all(outfile, params, model, vocab)
+        outfile = args.outfile or default_outfile(model_plus.paths, output_type)
+        OutputFile.write_all(outfile, params, output_type, model, vocab)
         print(f"Wrote {outfile}")
 
 

docs/BLIS.md

@@ -0,0 +1,67 @@
+BLIS Installation Manual
+------------------------
+
+BLIS is a portable software framework for high-performance BLAS-like dense linear algebra libraries. It has received awards and recognition, including the 2023 James H. Wilkinson Prize for Numerical Software and the 2020 SIAM Activity Group on Supercomputing Best Paper Prize. BLIS provides a new BLAS-like API and a compatibility layer for traditional BLAS routine calls. It offers features such as object-based API, typed API, BLAS and CBLAS compatibility layers.
+
+Project URL: https://github.com/flame/blis
+
+### Prepare:
+
+Compile BLIS:
+
+```bash
+git clone https://github.com/flame/blis
+cd blis
+./configure --enable-cblas -t openmp,pthreads auto
+# will install to /usr/local/ by default.
+make -j
+```
+
+Install BLIS:
+
+```bash
+sudo make install
+```
+
+We recommend using openmp since it's easier to modify the cores been used.
+
+### llama.cpp compilation
+
+Makefile:
+
+```bash
+make LLAMA_BLIS=1 -j
+# make LLAMA_BLIS=1 benchmark-matmult
+```
+
+CMake:
+
+```bash
+mkdir build
+cd build
+cmake -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=FLAME ..
+make -j
+```
+
+### llama.cpp execution
+
+According to the BLIS documentation, we could set the following
+environment variables to modify the behavior of openmp:
+
+```
+export GOMP_GPU_AFFINITY="0-19"
+export BLIS_NUM_THREADS=14
+```
+
+And then run the binaries as normal.
+
+
+### Intel specific issue
+
+Some might get the error message saying that `libimf.so` cannot be found.
+Please follow this [stackoverflow page](https://stackoverflow.com/questions/70687930/intel-oneapi-2022-libimf-so-no-such-file-or-directory-during-openmpi-compila).
+
+### Reference:
+
+1. https://github.com/flame/blis#getting-started
+2. https://github.com/flame/blis/blob/master/docs/Multithreading.md

docs/token_generation_performance_tips.md

@@ -0,0 +1,40 @@
+# Token generation performance troubleshooting
+
+## Verifying that the model is running on the GPU with cuBLAS
+Make sure you compiled llama with the correct env variables according to [this guide](../README.md#cublas), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
+```shell
+./main -m "path/to/model.bin" -ngl 200000 -p "Please sir, may I have some "
+```
+
+When running llama, before it starts the inference work, it will output diagnostic information that shows whether cuBLAS is offloading work to the GPU. Look for these lines:
+```shell
+llama_model_load_internal: [cublas] offloading 60 layers to GPU
+llama_model_load_internal: [cublas] offloading output layer to GPU
+llama_model_load_internal: [cublas] total VRAM used: 17223 MB
+... rest of inference
+```
+
+If you see these lines, then the GPU is being used.
+
+## Verifying that the CPU is not oversaturated
+llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physicial CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down.
+
+# Example of runtime flags effect on inference speed benchmark
+These runs were tested on the following machine:
+GPU: A6000 (48GB VRAM)
+CPU: 7 physical cores
+RAM: 32GB
+
+Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.ggmlv3.q4_0.bin` (30B parameters, 4bit quantization, GGML)
+
+Run command: `./main -m "path/to/model.bin" -p "-p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
+
+Result:
+
+| command | tokens/second (higher is better) |
+| - | - |
+| -ngl 2000000 | N/A (less than 0.1) |
+| -t 7 | 1.7 |
+| -t 1 -ngl 2000000 | 5.5 |
+| -t 7 -ngl 2000000 | 8.7 |
+| -t 4 -ngl 2000000 | 9.1 |

examples/CMakeLists.txt

@@ -37,4 +37,12 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
+    add_subdirectory(train-text-from-scratch)
+    add_subdirectory(simple)
+    if (LLAMA_METAL)
+        add_subdirectory(metal)
+    endif()
+    if (LLAMA_BUILD_SERVER)
+        add_subdirectory(server)
+    endif()
 endif()

examples/baby-llama/baby-llama.cpp

@@ -4,6 +4,10 @@
 #include <random>
 #include <cstring>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 float frand() {
     return (float)rand()/(float)RAND_MAX;
 }
@@ -79,34 +83,39 @@ struct ggml_tensor * randomize_tensor_normal(
         int ndims,
         const int64_t ne[],
         struct random_normal_distribution * rnd) {
+    float scale = 1.0; // xavier
     switch (ndims) {
         case 1:
+            scale /= sqrtf(ne[0]);
             for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = frand_normal(rnd);
+                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
             }
             break;
         case 2:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i1 = 0; i1 < ne[1]; i1++) {
                 for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = frand_normal(rnd);
+                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
                 }
             }
             break;
         case 3:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i2 = 0; i2 < ne[2]; i2++) {
                 for (int i1 = 0; i1 < ne[1]; i1++) {
                     for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                     }
                 }
             }
             break;
         case 4:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i3 = 0; i3 < ne[3]; i3++) {
                 for (int i2 = 0; i2 < ne[2]; i2++) {
                     for (int i1 = 0; i1 < ne[1]; i1++) {
                         for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                         }
                     }
                 }
@@ -148,8 +157,8 @@ struct llama_hparams_lora {
     uint32_t n_rot   = 64;
     uint32_t n_lora  = 64;
 
-    bool operator!=(const llama_hparams & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+    bool operator!=(const llama_hparams_lora & other) const {
+        return memcmp(this, &other, sizeof(llama_hparams_lora)) != 0;
     }
 };
 
@@ -1465,7 +1474,7 @@ struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_te
 }
 
 struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-    const float eps = 1e-3;
+    const float eps = 1e-3f;
     return
         ggml_sum(ctx,
             ggml_neg(ctx,

examples/benchmark/benchmark-matmult.cpp

@@ -1,6 +1,7 @@
-#include <locale.h>
 #include "ggml.h"
 #include "build-info.h"
+
+#include <locale.h>
 #include <assert.h>
 #include <math.h>
 #include <cstring>
@@ -15,6 +16,10 @@
 #include <iterator>
 #include <algorithm>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 float tensor_sum_elements(const ggml_tensor * tensor) {
     float sum = 0;
     if (tensor->type==GGML_TYPE_F32) {
@@ -28,9 +33,9 @@ float tensor_sum_elements(const ggml_tensor * tensor) {
 }
 
 void tensor_dump(const ggml_tensor * tensor, const char * name) {
-    printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", name,
+    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
         tensor->type, ggml_type_name(tensor->type),
-        (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
+        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
     float sum = tensor_sum_elements(tensor);
     printf("Sum of tensor %s is %6.2f\n", name, sum);
 }
@@ -119,7 +124,7 @@ int main(int argc, char ** argv)  {
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
 
-    printf("Allocating Memory of size %li bytes, %li MB\n",ctx_size, (ctx_size/1024/1024));
+    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx_size,

examples/chat-persistent.sh

@@ -23,8 +23,8 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
 NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
 NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
 
-SESSION_SIZE_MSG_PATTERN='main: session file matches \d+ / \d+'
-SAMPLE_TIME_MSG_PATTERN='sample time =\s+\d+.\d+ ms /\s+\d+'
+SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
+SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
 SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
 
 CTX_SIZE=2048

examples/chat-vicuna.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+
+set -e
+
+cd "$(dirname "$0")/.." || exit
+
+MODEL="${MODEL:-./models/ggml-vic13b-uncensored-q5_0.bin}"
+PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
+USER_NAME="### Human"
+AI_NAME="### Assistant"
+
+# Adjust to the number of CPU cores you want to use.
+N_THREAD="${N_THREAD:-8}"
+# Number of tokens to predict (made it larger than default because we want a long interaction)
+N_PREDICTS="${N_PREDICTS:-2048}"
+
+# Note: you can also override the generation options by specifying them on the command line:
+# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
+GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
+
+DATE_TIME=$(date +%H:%M)
+DATE_YEAR=$(date +%Y)
+
+PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
+
+sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
+    -e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
+    -e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
+    -e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
+     $PROMPT_TEMPLATE > $PROMPT_FILE
+
+# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
+./bin/main $GEN_OPTIONS \
+  --model "$MODEL" \
+  --threads "$N_THREAD" \
+  --n_predict "$N_PREDICTS" \
+  --color --interactive \
+  --file ${PROMPT_FILE} \
+  --reverse-prompt "### Human:" \
+  --in-prefix ' ' \
+  "$@"

examples/common.cpp

@@ -9,6 +9,7 @@
 #include <algorithm>
 #include <sstream>
 #include <unordered_set>
+#include <regex>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -27,6 +28,10 @@
 #include <wchar.h>
 #endif
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -101,9 +106,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         }
 
         if (arg == "-s" || arg == "--seed") {
-#if defined(GGML_USE_CUBLAS)
-            fprintf(stderr, "WARNING: when using cuBLAS generation results are NOT guaranteed to be reproducible.\n");
-#endif
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -131,6 +133,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.path_prompt_cache = argv[i];
         } else if (arg == "--prompt-cache-all") {
             params.prompt_cache_all = true;
+        } else if (arg == "--prompt-cache-ro") {
+            params.prompt_cache_ro = true;
         } else if (arg == "-f" || arg == "--file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -251,6 +255,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "-a" || arg == "--alias") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_alias = argv[i];
         } else if (arg == "--lora") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -283,11 +293,58 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+        } else if (arg == "--main-gpu" || arg == "-mg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            params.main_gpu = std::stoi(argv[i]);
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
+#endif
+        } else if (arg == "--tensor-split" || arg == "-ts") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            std::string arg_next = argv[i];
+
+            // split string by , and /
+            const std::regex regex{R"([,/]+)"};
+            std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
+            std::vector<std::string> split_arg{it, {}};
+            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+            for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+                if (i < split_arg.size()) {
+                    params.tensor_split[i] = std::stof(split_arg[i]);
+                } else {
+                    params.tensor_split[i] = 0.0f;
+                }
+            }
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--mtest") {
             params.mem_test = true;
+        } else if (arg == "--export") {
+            params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -317,7 +374,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 } else {
                     throw std::exception();
                 }
-            } catch (const std::exception &e) {
+            } catch (const std::exception&) {
                 invalid_param = true;
                 break;
             }
@@ -356,6 +413,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         gpt_print_usage(argc, argv, default_params);
         exit(1);
     }
+
+#ifdef GGML_USE_CUBLAS
+    if (!params.lora_adapter.empty() && params.n_gpu_layers > 0) {
+        fprintf(stderr, "%s: error: the simultaneous use of LoRAs and GPU acceleration is not supported", __func__);
+        exit(1);
+    }
+#endif // GGML_USE_CUBLAS
+
     if (escape_prompt) {
         process_escapes(params.prompt);
     }
@@ -384,6 +449,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
     fprintf(stderr, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
     fprintf(stderr, "                        not supported with --interactive or other interactive options\n");
+    fprintf(stderr, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
     fprintf(stderr, "  --random-prompt       start with a randomized prompt.\n");
     fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
     fprintf(stderr, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
@@ -410,7 +476,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value\n");
+    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
     fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
@@ -421,9 +488,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_mmap_supported()) {
         fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
     fprintf(stderr, "                        number of layers to store in VRAM\n");
+    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
+#endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
+    fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
@@ -462,11 +536,15 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
     return res;
 }
 
-struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params) {
     auto lparams = llama_context_default_params();
 
     lparams.n_ctx        = params.n_ctx;
+    lparams.n_batch      = params.n_batch;
     lparams.n_gpu_layers = params.n_gpu_layers;
+    lparams.main_gpu     = params.main_gpu;
+    memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
     lparams.use_mmap     = params.use_mmap;
@@ -474,25 +552,33 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     lparams.logits_all   = params.perplexity;
     lparams.embedding    = params.embedding;
 
-    llama_context * lctx = llama_init_from_file(params.model.c_str(), lparams);
-
-    if (lctx == NULL) {
+    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
+    if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return NULL;
+        return std::make_tuple(nullptr, nullptr);
+    }
+
+    llama_context * lctx = llama_new_context_with_model(model, lparams);
+    if (lctx == NULL) {
+        fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
+        llama_free_model(model);
+        return std::make_tuple(nullptr, nullptr);
     }
 
     if (!params.lora_adapter.empty()) {
-        int err = llama_apply_lora_from_file(lctx,
+        int err = llama_model_apply_lora_from_file(model,
                                              params.lora_adapter.c_str(),
                                              params.lora_base.empty() ? NULL : params.lora_base.c_str(),
                                              params.n_threads);
         if (err != 0) {
             fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
-            return NULL;
+            llama_free(lctx);
+            llama_free_model(model);
+            return std::make_tuple(nullptr, nullptr);
         }
     }
 
-    return lctx;
+    return std::make_tuple(model, lctx);
 }
 
 void console_init(console_state & con_st) {
@@ -571,6 +657,9 @@ void console_set_color(console_state & con_st, console_color_t color) {
             case CONSOLE_COLOR_USER_INPUT:
                 fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_GREEN);
                 break;
+            case CONSOLE_COLOR_ERROR:
+                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_RED);
+                break;
         }
         con_st.color = color;
         fflush(con_st.out);
@@ -578,6 +667,37 @@ void console_set_color(console_state & con_st, console_color_t color) {
 }
 
 char32_t getchar32() {
+#if defined(_WIN32)
+    HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
+    wchar_t high_surrogate = 0;
+
+    while (true) {
+        INPUT_RECORD record;
+        DWORD count;
+        if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
+            return WEOF;
+        }
+
+        if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
+            wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
+            if (wc == 0) {
+                continue;
+            }
+
+            if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+                high_surrogate = wc;
+                continue;
+            } else if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
+                if (high_surrogate != 0) { // Check if we have a high surrogate
+                    return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
+                }
+            }
+
+            high_surrogate = 0; // Reset the high surrogate
+            return static_cast<char32_t>(wc);
+        }
+    }
+#else
     wchar_t wc = getwchar();
     if (static_cast<wint_t>(wc) == WEOF) {
         return WEOF;
@@ -596,6 +716,7 @@ char32_t getchar32() {
 #endif
 
     return static_cast<char32_t>(wc);
+#endif
 }
 
 void pop_cursor(console_state & con_st) {

examples/common.h

@@ -9,6 +9,7 @@
 #include <random>
 #include <thread>
 #include <unordered_map>
+#include <tuple>
 
 #if !defined (_WIN32)
 #include <stdio.h>
@@ -21,13 +22,16 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed          = -1;  // RNG seed
-    int32_t n_threads     = get_num_physical_cores();
-    int32_t n_predict     = -1;  // new tokens to predict
-    int32_t n_ctx         = 512; // context size
-    int32_t n_batch       = 512; // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep        = 0;   // number of tokens to keep from initial prompt
-    int32_t n_gpu_layers  = 0;   // number of layers to store in VRAM
+    int32_t seed                            = -1;  // RNG seed
+    int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_predict                       = -1;  // new tokens to predict
+    int32_t n_ctx                           = 512; // context size
+    int32_t n_batch                         = 512; // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
+    int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
+    int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
+    float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
+    bool    low_vram                        = 0;   // if true, reduce VRAM usage at the cost of performance
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
@@ -45,6 +49,7 @@ struct gpt_params {
     float   mirostat_eta      = 0.10f; // learning rate
 
     std::string model             = "models/7B/ggml-model.bin"; // model path
+    std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
@@ -59,6 +64,7 @@ struct gpt_params {
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
+    bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
     bool embedding         = false; // get only sentence embedding
     bool interactive_first = false; // wait for user input immediately
@@ -70,6 +76,7 @@ struct gpt_params {
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
+    bool export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
 };
 
@@ -89,7 +96,7 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 // Model utils
 //
 
-struct llama_context * llama_init_from_gpt_params(const gpt_params & params);
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params);
 
 //
 // Console utils
@@ -107,7 +114,8 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params);
 enum console_color_t {
     CONSOLE_COLOR_DEFAULT=0,
     CONSOLE_COLOR_PROMPT,
-    CONSOLE_COLOR_USER_INPUT
+    CONSOLE_COLOR_USER_INPUT,
+    CONSOLE_COLOR_ERROR
 };
 
 struct console_state {

examples/embedding/embedding.cpp

@@ -4,6 +4,10 @@
 
 #include <ctime>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -31,11 +35,14 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    llama_init_backend();
+
+    llama_model * model;
     llama_context * ctx;
 
     // load the model
-    ctx = llama_init_from_gpt_params(params);
-    if (ctx == NULL) {
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
     }
@@ -84,6 +91,7 @@ int main(int argc, char ** argv) {
 
     llama_print_timings(ctx);
     llama_free(ctx);
+    llama_free_model(model);
 
     return 0;
 }

examples/jeopardy/graph.py

@@ -1,5 +1,5 @@
 import matplotlib.pyplot as plt
-import sys, os
+import os
 import csv
 
 labels = []
@@ -8,6 +8,7 @@ numEntries = 1
 
 rows = []
 
+
 def bar_chart(numbers, labels, pos):
     plt.bar(pos, numbers, color='blue')
     plt.xticks(ticks=pos, labels=labels)
@@ -16,6 +17,7 @@ def bar_chart(numbers, labels, pos):
     plt.ylabel("Questions Correct")
     plt.show()
 
+
 def calculatecorrect():
     directory = os.fsencode("./examples/jeopardy/results/")
     csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',')
@@ -38,14 +40,13 @@ def calculatecorrect():
                     print(line)
                 else:
                     print("Correct answer: " + rows[i][2] + "\n")
-                    i+=1
+                    i += 1
                     print("Did the AI get the question right? (y/n)")
                     if input() == "y":
                         totalcorrect += 1
             numbers.append(totalcorrect)
 
 
-
 if __name__ == '__main__':
     calculatecorrect()
     pos = list(range(numEntries))

examples/main/README.md

@@ -69,8 +69,8 @@ In this section, we cover the most commonly used options for running the `main`
 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
--   `-n N, --n_predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
--   `-c N, --ctx_size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 
 ## Input Prompts
 
@@ -136,9 +136,9 @@ During text generation, LLaMA models have a limited context size, which means th
 
 ### Context Size
 
-The `--ctx_size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
+The `--ctx-size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
 
--   `-c N, --ctx_size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
+-   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
 ### Keep Prompt
 
@@ -146,7 +146,7 @@ The `--keep` option allows users to retain the original prompt when the model ru
 
 -   `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
 
-By utilizing context management options like `--ctx_size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
+By utilizing context management options like `--ctx-size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
 
 ## Generation Flags
 
@@ -154,11 +154,11 @@ The following options allow you to control the text generation process and fine-
 
 ### Number of Tokens to Predict
 
--   `-n N, --n_predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
 
-The `--n_predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
+The `--n-predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
 
-It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n_predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
+It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n-predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
 
 ### Temperature
 
@@ -170,33 +170,33 @@ Example usage: `--temp 0.5`
 
 ### Repeat Penalty
 
--   `--repeat_penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
--   `--repeat_last_n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx_size).
+-   `--repeat-penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
+-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
 -   `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
 
-The `repeat_penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
+The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
 
-The `repeat_last_n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx_size`).
+The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
 
 Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases.
 
-Example usage: `--repeat_penalty 1.15 --repeat_last_n 128 --no-penalize-nl`
+Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
 
 ### Top-K Sampling
 
--   `--top_k N`: Limit the next token selection to the K most probable tokens (default: 40).
+-   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
 
-Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
+Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top-k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
 
-Example usage: `--top_k 30`
+Example usage: `--top-k 30`
 
 ### Top-P Sampling
 
--   `--top_p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+-   `--top-p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
 
-Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
+Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top-p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
 
-Example usage: `--top_p 0.95`
+Example usage: `--top-p 0.95`
 
 ### Tail Free Sampling (TFS)
 
@@ -217,16 +217,16 @@ Example usage: `--typical 0.9`
 ### Mirostat Sampling
 
 -   `--mirostat N`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
--   `--mirostat_lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1).
--   `--mirostat_ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0).
+-   `--mirostat-lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1).
+-   `--mirostat-ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0).
 
 Mirostat is an algorithm that actively maintains the quality of generated text within a desired range during text generation. It aims to strike a balance between coherence and diversity, avoiding low-quality output caused by excessive repetition (boredom traps) or incoherence (confusion traps).
 
-The `--mirostat_lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`.
+The `--mirostat-lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`.
 
-The `--mirostat_ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`.
+The `--mirostat-ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`.
 
-Example usage: `--mirostat 2 --mirostat_lr 0.05 --mirostat_ent 3.0`
+Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
 
 ### Logit Bias
 
@@ -264,15 +264,15 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Memory Float 32
 
--   `--memory_f32`: Use 32-bit floats instead of 16-bit floats for memory key+value, allowing higher quality inference at the cost of higher memory usage.
+-   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
 
 ### Batch Size
 
--   `-b N, --batch_size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+-   `-b N, --batch-size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
 
 ### Prompt Caching
 
--   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs.
+-   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
 
 ### Quantization
 
@@ -285,5 +285,9 @@ These options provide extra functionality and customization when running the LLa
 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
+-   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--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.

examples/main/main.cpp

@@ -23,11 +23,17 @@
 #include <unistd.h>
 #elif defined (_WIN32)
 #define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
 #define NOMINMAX
+#endif
 #include <windows.h>
 #include <signal.h>
 #endif
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 static console_state con_st;
 static llama_context ** g_ctx;
 
@@ -81,6 +87,9 @@ int main(int argc, char ** argv) {
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
                 "expect poor results\n", __func__, params.n_ctx);
+    } else if (params.n_ctx < 8) {
+        fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
@@ -96,15 +105,15 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-//    params.prompt = R"(// this function checks if the number n is prime
-//bool is_prime(int n) {)";
+    llama_init_backend();
 
+    llama_model * model;
     llama_context * ctx;
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
-    ctx = llama_init_from_gpt_params(params);
-    if (ctx == NULL) {
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
     }
@@ -131,12 +140,19 @@ int main(int argc, char ** argv) {
 
         llama_print_timings(ctx);
         llama_free(ctx);
+        llama_free_model(model);
 
         return 0;
     }
 
-    // Add a space in front of the first character to match OG llama tokenizer behavior
-    params.prompt.insert(0, 1, ' ');
+    // export the cgraph and exit
+    if (params.export_cgraph) {
+        llama_eval_export(ctx, "llama.ggml");
+        llama_free(ctx);
+        llama_free_model(model);
+
+        return 0;
+    }
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
@@ -156,6 +172,7 @@ int main(int argc, char ** argv) {
                 return 1;
             }
             session_tokens.resize(n_token_count_out);
+            llama_set_rng_seed(ctx, params.seed);
 
             fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
         } else {
@@ -164,7 +181,16 @@ int main(int argc, char ** argv) {
     }
 
     // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> embd_inp;
+
+    if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
+        // Add a space in front of the first character to match OG llama tokenizer behavior
+        params.prompt.insert(0, 1, ' ');
+
+        embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    } else {
+        embd_inp = session_tokens;
+    }
 
     const int n_ctx = llama_n_ctx(ctx);
 
@@ -182,7 +208,9 @@ int main(int argc, char ** argv) {
             }
             n_matching_session_tokens++;
         }
-        if (n_matching_session_tokens >= embd_inp.size()) {
+        if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
+            fprintf(stderr, "%s: using full prompt from session file\n", __func__);
+        } else if (n_matching_session_tokens >= embd_inp.size()) {
             fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__);
         } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
             fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
@@ -193,6 +221,13 @@ int main(int argc, char ** argv) {
         }
     }
 
+    // if we will use the cache for the full prompt without reaching the end of the cache, force
+    // reevaluation of the last token token to recalculate the cached logits
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
+            session_tokens.size() > embd_inp.size()) {
+        session_tokens.resize(embd_inp.size() - 1);
+    }
+
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) {
         params.n_keep = (int)embd_inp.size();
@@ -305,9 +340,29 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd;
 
+    // do one empty run to warm up the model
+    {
+        const std::vector<llama_token> tmp = { llama_token_bos(), };
+        llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
+        llama_reset_timings(ctx);
+    }
+
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (embd.size() > 0) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            auto max_embd_size = n_ctx - 4;
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int)embd.size() > max_embd_size) {
+                auto skipped_tokens = embd.size() - max_embd_size;
+                console_set_color(con_st, CONSOLE_COLOR_ERROR);
+                printf("<<input too long: skipped %zu token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                fflush(stdout);
+                embd.resize(max_embd_size);
+            }
+
             // infinite text generation via context swapping
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)
@@ -394,7 +449,7 @@ int main(int argc, char ** argv) {
             const bool    penalize_nl     = params.penalize_nl;
 
             // optionally save the session on first sample (for faster prompt loading next time)
-            if (!path_session.empty() && need_to_save_session) {
+            if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
                 need_to_save_session = false;
                 llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
             }
@@ -607,13 +662,14 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if (!path_session.empty() && params.prompt_cache_all) {
+    if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
         fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
         llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
     llama_print_timings(ctx);
     llama_free(ctx);
+    llama_free_model(model);
 
     return 0;
 }

examples/metal/CMakeLists.txt

@@ -0,0 +1,3 @@
+set(TEST_TARGET metal)
+add_executable(${TEST_TARGET} metal.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml)

examples/metal/metal.cpp

@@ -0,0 +1,104 @@
+// Evaluate a statically exported ggml computation graph with Metal
+//
+// - First, export a LLaMA graph:
+//
+//  $ ./bin/main -m ../models/7B/ggml-model-q4_0.bin --export
+//
+// - Run this tool to evaluate the exported graph:
+//
+//  $ ./bin/metal llama.ggml
+//
+// The purpose of this tool is mostly for debugging and demonstration purposes.
+// The main limitation of exporting computation graphs is that their sizes are static which often
+// can be a problem for real-world applications.
+//
+
+#include "ggml.h"
+#include "ggml-metal.h"
+
+#include <cstdio>
+#include <cstring>
+#include <cstdlib>
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    if (argc != 2) {
+        fprintf(stderr, "Usage: %s llama.ggml\n", argv[0]);
+        return -1;
+    }
+
+    const char * fname_cgraph = argv[1];
+
+    // load the compute graph
+    struct ggml_context * ctx_data = NULL;
+    struct ggml_context * ctx_eval = NULL;
+
+    struct ggml_cgraph gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
+    gf.n_threads = 1;
+
+    // this allocates all Metal resources and memory buffers
+    auto * ctx_metal = ggml_metal_init();
+
+    const size_t max_size_data = ggml_get_max_tensor_size(ctx_data);
+    const size_t max_size_eval = ggml_get_max_tensor_size(ctx_eval);
+    ggml_metal_add_buffer(ctx_metal, "data", ggml_get_mem_buffer(ctx_data), ggml_get_mem_size(ctx_data), max_size_data);
+    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval), max_size_eval);
+
+    // main
+    {
+        struct ggml_tensor * input = ggml_graph_get_tensor(&gf, "embd");
+        *(int32_t *) input->data = 1; // BOS
+
+        ggml_metal_set_tensor(ctx_metal, input);
+
+        // warmup
+        ggml_metal_graph_compute(ctx_metal, &gf);
+
+        const int n_iter = 16;
+
+        const int64_t t0 = ggml_time_us();
+
+        // the actual inference happens here
+        for (int i = 0; i < n_iter; ++i) {
+            ggml_metal_graph_compute(ctx_metal, &gf);
+        }
+
+        const int64_t t1 = ggml_time_us();
+
+        printf("time: %.2f ms, %.2f ms/tok\n", (t1 - t0) / 1000.0, (t1 - t0) / 1000.0 / n_iter);
+    }
+
+    // debug output
+    {
+        struct ggml_tensor * logits = gf.nodes[gf.n_nodes - 1];
+        ggml_metal_get_tensor(ctx_metal, logits);
+
+        float * ptr = (float *) ggml_get_data(logits);
+
+        printf("logits: ");
+        for (int i = 0; i < 10; i++) {
+            printf("%8.4f ", ptr[i]);
+        }
+        printf("\n");
+        int imax = 0;
+        double sum = 0.0;
+        double vmax = -1e9;
+        for (int i = 0; i < 32000; i++) {
+            sum += (double) ptr[i];
+            if (ptr[i] > vmax) {
+                vmax = ptr[i];
+                imax = i;
+            }
+        }
+        printf("sum: %f, imax = %d, vmax = %f\n", sum, imax, vmax);
+    }
+
+    ggml_metal_free(ctx_metal);
+
+    ggml_free(ctx_data);
+    ggml_free(ctx_eval);
+
+    return 0;
+}
+

examples/perplexity/perplexity.cpp

@@ -5,6 +5,10 @@
 #include <cmath>
 #include <ctime>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];
@@ -143,11 +147,14 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    llama_init_backend();
+
+    llama_model * model;
     llama_context * ctx;
 
     // load the model and apply lora adapter, if any
-    ctx = llama_init_from_gpt_params(params);
-    if (ctx == NULL) {
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
     }
@@ -163,6 +170,7 @@ int main(int argc, char ** argv) {
 
     llama_print_timings(ctx);
     llama_free(ctx);
+    llama_free_model(model);
 
     return 0;
 }

examples/quantize-stats/quantize-stats.cpp

@@ -19,6 +19,10 @@
 #include <thread>
 #include <mutex>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 struct quantize_stats_params {
     std::string model = "models/7B/ggml-model-f16.bin";
     bool verbose = false;
@@ -282,8 +286,9 @@ int main(int argc, char ** argv) {
                 break;
             }
             int j;
-            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) j)) != 0; j++) {
-                // find match
+            for (j = 0; j < GGML_TYPE_COUNT; ++j) {
+               const auto * name = ggml_type_name((ggml_type) j);
+               if (name && strcmp(argv[i], name) == 0) break;
             }
             if (j < GGML_TYPE_COUNT) {
                 params.include_types.push_back((ggml_type) j);
@@ -315,6 +320,7 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "Loading model\n");
 
     const int64_t t_main_start_us = ggml_time_us();
+    llama_model * model;
     llama_context * ctx;
 
     {
@@ -325,10 +331,18 @@ int main(int argc, char ** argv) {
         lparams.f16_kv     = false;
         lparams.use_mlock  = false;
 
-        ctx = llama_init_from_file(params.model.c_str(), lparams);
+        model = llama_load_model_from_file(params.model.c_str(), lparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        ctx = llama_new_context_with_model(model, lparams);
 
         if (ctx == NULL) {
-            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+            fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
+            llama_free_model(model);
             return 1;
         }
     }
@@ -352,6 +366,7 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: error: Quantization should be tested with a float model, "
                 "this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);
             llama_free(ctx);
+            llama_free_model(model);
             return 1;
         }
         included_layers++;
@@ -410,6 +425,7 @@ int main(int argc, char ** argv) {
 
 
     llama_free(ctx);
+    llama_free_model(model);
     // report timing
     {
         const int64_t t_main_end_us = ggml_time_us();

examples/quantize/quantize.cpp

@@ -1,33 +1,138 @@
-#include "ggml.h"
-#include "llama.h"
 #include "build-info.h"
 
+#include "llama.h"
+
 #include <cstdio>
-#include <map>
+#include <cstring>
+#include <vector>
 #include <string>
 
-static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
-  {"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0},
-  {"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1},
-  {"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0},
-  {"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1},
-  {"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0},
+struct quant_option {
+    std::string name;
+    llama_ftype ftype;
+    std::string desc;
 };
 
-bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {
-    auto it = LLAMA_FTYPE_MAP.find(ftype_str);
-    if (it != LLAMA_FTYPE_MAP.end()) {
-        ftype = it->second;
-        ftype_str_out = it->first;
-        return true;
+static const std::vector<struct quant_option> QUANT_OPTIONS = {
+    {
+        "Q4_0",
+        LLAMA_FTYPE_MOSTLY_Q4_0,
+        " 3.50G, +0.2499 ppl @ 7B - small, very high quality loss - legacy, prefer using Q3_K_M",
+    },
+    {
+        "Q4_1",
+        LLAMA_FTYPE_MOSTLY_Q4_1,
+        " 3.90G, +0.1846 ppl @ 7B - small, substantial quality loss - legacy, prefer using Q3_K_L",
+    },
+    {
+        "Q5_0",
+        LLAMA_FTYPE_MOSTLY_Q5_0,
+        " 4.30G, +0.0796 ppl @ 7B - medium, balanced quality - legacy, prefer using Q4_K_M",
+    },
+    {
+        "Q5_1",
+        LLAMA_FTYPE_MOSTLY_Q5_1,
+        " 4.70G, +0.0415 ppl @ 7B - medium, low quality loss - legacy, prefer using Q5_K_M",
+    },
+#ifdef GGML_USE_K_QUANTS
+    {
+        "Q2_K",
+        LLAMA_FTYPE_MOSTLY_Q2_K,
+        " 2.67G, +0.8698 ppl @ 7B - smallest, extreme quality loss - not recommended",
+    },
+    {
+        "Q3_K",
+        LLAMA_FTYPE_MOSTLY_Q3_K_M,
+        "alias for Q3_K_M"
+    },
+    {
+        "Q3_K_S",
+        LLAMA_FTYPE_MOSTLY_Q3_K_S,
+        " 2.75G, +0.5505 ppl @ 7B - very small, very high quality loss",
+    },
+    {
+        "Q3_K_M",
+        LLAMA_FTYPE_MOSTLY_Q3_K_M,
+        " 3.06G, +0.2437 ppl @ 7B - very small, very high quality loss",
+    },
+    {
+        "Q3_K_L",
+        LLAMA_FTYPE_MOSTLY_Q3_K_L,
+        " 3.35G, +0.1803 ppl @ 7B - small, substantial quality loss",
+    },
+    {
+        "Q4_K",
+        LLAMA_FTYPE_MOSTLY_Q4_K_M,
+        "alias for Q4_K_M",
+    },
+    {
+        "Q4_K_S",
+        LLAMA_FTYPE_MOSTLY_Q4_K_S,
+        " 3.56G, +0.1149 ppl @ 7B - small, significant quality loss",
+    },
+    {
+        "Q4_K_M",
+        LLAMA_FTYPE_MOSTLY_Q4_K_M,
+        " 3.80G, +0.0535 ppl @ 7B - medium, balanced quality - *recommended*",
+    },
+    {
+        "Q5_K",
+        LLAMA_FTYPE_MOSTLY_Q5_K_M,
+        "alias for Q5_K_M",
+    },
+    {
+        "Q5_K_S",
+        LLAMA_FTYPE_MOSTLY_Q5_K_S,
+        " 4.33G, +0.0353 ppl @ 7B - large, low quality loss - *recommended*",
+    },
+    {
+        "Q5_K_M",
+        LLAMA_FTYPE_MOSTLY_Q5_K_M,
+        " 4.45G, +0.0142 ppl @ 7B - large, very low quality loss - *recommended*",
+    },
+    {
+        "Q6_K",
+        LLAMA_FTYPE_MOSTLY_Q6_K,
+        " 5.15G, +0.0044 ppl @ 7B - very large, extremely low quality loss",
+    },
+#endif
+    {
+        "Q8_0",
+        LLAMA_FTYPE_MOSTLY_Q8_0,
+        " 6.70G, +0.0004 ppl @ 7B - very large, extremely low quality loss - not recommended",
+    },
+    {
+        "F16",
+        LLAMA_FTYPE_MOSTLY_F16,
+        "13.00G              @ 7B - extremely large, virtually no quality loss - not recommended",
+    },
+    {
+        "F32",
+        LLAMA_FTYPE_ALL_F32,
+        "26.00G              @ 7B - absolutely huge, lossless - not recommended",
+    },
+};
+
+
+bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+    std::string ftype_str;
+
+    for (auto ch : ftype_str_in) {
+        ftype_str.push_back(std::toupper(ch));
+    }
+    for (auto & it : QUANT_OPTIONS) {
+        if (it.name == ftype_str) {
+            ftype = it.ftype;
+            ftype_str_out = it.name;
+            return true;
+        }
     }
-    // try to parse as an integer
     try {
         int ftype_int = std::stoi(ftype_str);
-        for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
-            if (it->second == ftype_int) {
-                ftype = it->second;
-                ftype_str_out = it->first;
+        for (auto & it : QUANT_OPTIONS) {
+            if (it.ftype == ftype_int) {
+                ftype = it.ftype;
+                ftype_str_out = it.name;
                 return true;
             }
         }
@@ -39,36 +144,51 @@ bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::st
 }
 
 // usage:
-//  ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
 //
+void usage(const char * executable) {
+    fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.bin [model-quant.bin] type [nthreads]\n\n", executable);
+    fprintf(stderr, "  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
+    fprintf(stderr, "  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    fprintf(stderr, "\nAllowed quantization types:\n");
+    for (auto & it : QUANT_OPTIONS) {
+        printf("  %2d  or  %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
+    }
+    exit(1);
+}
+
 int main(int argc, char ** argv) {
-    ggml_time_init();
-
     if (argc < 3) {
-        fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]);
-        for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
-            fprintf(stderr, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
-        }
-        return 1;
+        usage(argv[0]);
     }
 
-    // needed to initialize f16 tables
-    {
-        struct ggml_init_params params = { 0, NULL, false };
-        struct ggml_context * ctx = ggml_init(params);
-        ggml_free(ctx);
+    llama_model_quantize_params params = llama_model_quantize_default_params();
+
+    int arg_idx = 1;
+
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
+            params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
+            params.allow_requantize = true;
+        } else {
+            usage(argv[0]);
+        }
     }
 
+    if (argc - arg_idx < 3) {
+        usage(argv[0]);
+    }
+
+    llama_init_backend();
+
     // parse command line arguments
-    const std::string fname_inp = argv[1];
+    const std::string fname_inp = argv[arg_idx];
+    arg_idx++;
     std::string fname_out;
-    int nthread;
-    llama_ftype ftype;
 
-    int arg_idx = 2;
     std::string ftype_str;
-    if (try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
-        // argv[2] is the ftype
+    if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
         std::string fpath;
         const size_t pos = fname_inp.find_last_of('/');
         if (pos != std::string::npos) {
@@ -79,7 +199,6 @@ int main(int argc, char ** argv) {
         arg_idx++;
     }
     else {
-        // argv[2] is the output path
         fname_out = argv[arg_idx];
         arg_idx++;
 
@@ -87,8 +206,7 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: missing ftype\n", __func__);
             return 1;
         }
-        // argv[3] is the ftype
-        if (!try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
+        if (!try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
             fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
             return 1;
         }
@@ -98,43 +216,41 @@ int main(int argc, char ** argv) {
     // parse nthreads
     if (argc > arg_idx) {
         try {
-            nthread = std::stoi(argv[arg_idx]);
+            params.nthread = std::stoi(argv[arg_idx]);
         }
         catch (const std::exception & e) {
             fprintf(stderr, "%s: invalid nthread '%s' (%s)\n", __func__, argv[arg_idx], e.what());
             return 1;
         }
-    } else {
-        nthread = 0;
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
     fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
-    if (nthread > 0) {
-        fprintf(stderr, " using %d threads", nthread);
+    if (params.nthread > 0) {
+        fprintf(stderr, " using %d threads", params.nthread);
     }
     fprintf(stderr, "\n");
 
-    const int64_t t_main_start_us = ggml_time_us();
+    const int64_t t_main_start_us = llama_time_us();
 
     int64_t t_quantize_us = 0;
 
     // load the model
     {
-        const int64_t t_start_us = ggml_time_us();
+        const int64_t t_start_us = llama_time_us();
 
-        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) {
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), &params)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }
 
-        t_quantize_us = ggml_time_us() - t_start_us;
+        t_quantize_us = llama_time_us() - t_start_us;
     }
 
     // report timing
     {
-        const int64_t t_main_end_us = ggml_time_us();
+        const int64_t t_main_end_us = llama_time_us();
 
         printf("\n");
         printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0);

examples/save-load-state/save-load-state.cpp

@@ -35,12 +35,22 @@ int main(int argc, char ** argv) {
     auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
 
     // init
-    auto ctx = llama_init_from_file(params.model.c_str(), lparams);
+    auto model = llama_load_model_from_file(params.model.c_str(), lparams);
+    if (model == nullptr) {
+        return 1;
+    }
+    auto ctx = llama_new_context_with_model(model, lparams);
+    if (ctx == nullptr) {
+        llama_free_model(model);
+        return 1;
+    }
     auto tokens = std::vector<llama_token>(params.n_ctx);
-    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true);
+    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), int(tokens.size()), true);
 
     if (n_prompt_tokens < 1) {
         fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
+        llama_free(ctx);
+        llama_free_model(model);
         return 1;
     }
 
@@ -84,6 +94,8 @@ int main(int argc, char ** argv) {
         printf("%s", next_token_str);
         if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_free(ctx);
+            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -91,23 +103,27 @@ int main(int argc, char ** argv) {
 
     printf("\n\n");
 
-    // free old model
+    // free old context
     llama_free(ctx);
 
-    // load new model
-    auto ctx2 = llama_init_from_file(params.model.c_str(), lparams);
+    // make new context
+    auto ctx2 = llama_new_context_with_model(model, lparams);
 
     // Load state (rng, logits, embedding and kv_cache) from file
     {
         FILE *fp_read = fopen("dump_state.bin", "rb");
         if (state_size != llama_get_state_size(ctx2)) {
             fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
+            llama_free(ctx2);
+            llama_free_model(model);
             return 1;
         }
 
         const size_t ret = fread(state_mem, 1, state_size, fp_read);
         if (ret != state_size) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
+            llama_free(ctx2);
+            llama_free_model(model);
             return 1;
         }
 
@@ -138,6 +154,8 @@ int main(int argc, char ** argv) {
         printf("%s", next_token_str);
         if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_free(ctx2);
+            llama_free_model(model);
             return 1;
         }
         n_past += 1;
@@ -145,5 +163,8 @@ int main(int argc, char ** argv) {
 
     printf("\n\n");
 
+    llama_free(ctx2);
+    llama_free_model(model);
+
     return 0;
 }

examples/server/CMakeLists.txt

@@ -0,0 +1,12 @@
+set(TARGET server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp json.hpp httplib.h)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/server/README.md

@@ -0,0 +1,192 @@
+# llama.cpp/example/server
+
+This example demonstrates a simple HTTP API server to interact with llama.cpp.
+
+Command line options:
+
+-   `--threads N`, `-t N`: Set the number of threads to use during computation.
+-   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
+-   `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
+-   `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
+-   `--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.
+-   `--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.
+-   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
+-   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
+-   `--port`: Set the port to listen. Default: `8080`.
+-   `--embedding`: Enable embedding extraction, Default: disabled.
+
+## Build
+
+Build llama.cpp with server from repository root with either make or CMake.
+
+- Using `make`:
+
+  ```bash
+  LLAMA_BUILD_SERVER=1 make
+  ```
+
+- Using `CMake`:
+
+  ```bash
+  mkdir build-server
+  cd build-server
+  cmake -DLLAMA_BUILD_SERVER=ON ..
+  cmake --build . --config Release
+  ```
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./server -m models/7B/ggml-model.bin -c 2048
+```
+
+### Windows:
+
+```powershell
+server.exe -m models\7B\ggml-model.bin -c 2048
+```
+
+The above command will start a server that by default listens on `127.0.0.1:8080`.
+You can consume the endpoints with Postman or NodeJS with axios library.
+
+## Testing with CURL
+
+Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.
+
+```sh
+curl --request POST \
+    --url http://localhost:8080/completion \
+    --data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
+```
+
+## Node JS Test
+
+You need to have [Node.js](https://nodejs.org/en) installed.
+
+```bash
+mkdir llama-client
+cd llama-client
+npm init
+npm install axios
+```
+
+Create a index.js file and put inside this:
+
+```javascript
+const axios = require("axios");
+
+const prompt = `Building a website can be done in 10 simple steps:`;
+
+async function Test() {
+    let result = await axios.post("http://127.0.0.1:8080/completion", {
+        prompt,
+        n_predict: 512,
+    });
+
+    // the response is received until completion finish
+    console.log(result.data.content);
+}
+
+Test();
+```
+
+And run it:
+
+```bash
+node .
+```
+
+## API Endpoints
+
+-   **POST** `/completion`: Given a prompt, it returns the predicted completion.
+
+    *Options:*
+
+    `temperature`: Adjust the randomness of the generated text (default: 0.8).
+
+    `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
+
+    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+
+    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: 128, -1 = infinity).
+
+    `n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
+    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
+
+    `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
+
+    `prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. A space is inserted in the front like main.cpp does.
+
+    `stop`: Specify a JSON array of stopping strings.
+    These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
+
+    `tfs_z`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
+
+    `typical_p`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
+
+    `repeat_penalty`: Control the repetition of token sequences in the generated text (default: 1.1).
+
+    `repeat_last_n`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+
+    `penalize_nl`: Penalize newline tokens when applying the repeat penalty (default: true).
+
+    `presence_penalty`: Repeat alpha presence penalty (default: 0.0, 0.0 = disabled).
+
+    `frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled);
+
+    `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
+
+    `mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0).
+
+    `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
+
+    `seed`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
+
+    `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: []).
+
+-   **POST** `/tokenize`: Tokenize a given text.
+
+    *Options:*
+
+    `content`: Set the text to tokenize.
+
+    Note that the special `BOS` token is not added in fron of the text and also a space character is not inserted automatically as it is for `/completion`.
+
+-   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+
+    *Options:*
+
+    `content`: Set the text to process.
+
+## More examples
+
+### Interactive mode
+
+Check the sample in [chat.mjs](chat.mjs).
+Run with NodeJS version 16 or later:
+
+```sh
+node chat.mjs
+```
+
+Another sample in [chat.sh](chat.sh).
+Requires [bash](https://www.gnu.org/software/bash/), [curl](https://curl.se) and [jq](https://jqlang.github.io/jq/).
+Run with bash:
+
+```sh
+bash chat.sh
+```

examples/server/chat.mjs

@@ -0,0 +1,89 @@
+import * as readline from 'node:readline'
+import { stdin, stdout } from 'node:process'
+
+const API_URL = 'http://127.0.0.1:8080'
+
+const chat = [
+    {
+        human: "Hello, Assistant.",
+        assistant: "Hello. How may I help you today?"
+    },
+    {
+        human: "Please tell me the largest city in Europe.",
+        assistant: "Sure. The largest city in Europe is Moscow, the capital of Russia."
+    },
+]
+
+const instruction = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.`
+
+function format_prompt(question) {
+    return `${instruction}\n${
+        chat.map(m =>`### Human: ${m.human}\n### Assistant: ${m.assistant}`).join("\n")
+    }\n### Human: ${question}\n### Assistant:`
+}
+
+async function tokenize(content) {
+    const result = await fetch(`${API_URL}/tokenize`, {
+        method: 'POST',
+        body: JSON.stringify({ content })
+    })
+
+    if (!result.ok) {
+        return []
+    }
+
+    return await result.json().tokens
+}
+
+const n_keep = await tokenize(instruction).length
+
+async function chat_completion(question) {
+    const result = await fetch(`${API_URL}/completion`, {
+        method: 'POST',
+        body: JSON.stringify({
+            prompt: format_prompt(question),
+            temperature: 0.2,
+            top_k: 40,
+            top_p: 0.9,
+            n_keep: n_keep,
+            n_predict: 256,
+            stop: ["\n### Human:"], // stop completion after generating this
+            stream: true,
+        })
+    })
+
+    if (!result.ok) {
+        return
+    }
+
+    let answer = ''
+
+    for await (var chunk of result.body) {
+        const t = Buffer.from(chunk).toString('utf8')
+        if (t.startsWith('data: ')) {
+            const message = JSON.parse(t.substring(6))
+            answer += message.content
+            process.stdout.write(message.content)
+            if (message.stop) {
+                if (message.truncated) {
+                    chat.shift()
+                }
+                break
+            }
+        }
+    }
+
+    process.stdout.write('\n')
+    chat.push({ human: question, assistant: answer.trimStart() })
+}
+
+const rl = readline.createInterface({ input: stdin, output: stdout });
+
+const readlineQuestion = (rl, query, options) => new Promise((resolve, reject) => {
+    rl.question(query, options, resolve)
+});
+
+while(true) {
+    const question = await readlineQuestion(rl, '> ')
+    await chat_completion(question)
+}

examples/server/chat.sh

@@ -0,0 +1,77 @@
+#!/bin/bash
+
+API_URL="${API_URL:-http://127.0.0.1:8080}"
+
+CHAT=(
+    "Hello, Assistant."
+    "Hello. How may I help you today?"
+    "Please tell me the largest city in Europe."
+    "Sure. The largest city in Europe is Moscow, the capital of Russia."
+)
+
+INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
+
+trim() {
+    shopt -s extglob
+    set -- "${1##+([[:space:]])}"
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+trim_trailing() {
+    shopt -s extglob
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+format_prompt() {
+    echo -n "${INSTRUCTION}"
+    printf "\n### Human: %s\n### Assistant: %s" "${CHAT[@]}" "$1"
+}
+
+tokenize() {
+    curl \
+        --silent \
+        --request POST \
+        --url "${API_URL}/tokenize" \
+        --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
+    | jq '.tokens[]'
+}
+
+N_KEEP=$(tokenize "${INSTRUCTION}" | wc -l)
+
+chat_completion() {
+    PROMPT="$(trim_trailing "$(format_prompt "$1")")"
+    DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
+        prompt: .,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: $n_keep,
+        n_predict: 256,
+        stop: ["\n### Human:"],
+        stream: true
+    }')"
+
+    ANSWER=''
+
+    while IFS= read -r LINE; do
+        if [[ $LINE = data:* ]]; then
+            CONTENT="$(echo "${LINE:5}" | jq -r '.content')"
+            printf "%s" "${CONTENT}"
+            ANSWER+="${CONTENT}"
+        fi
+    done < <(curl \
+        --silent \
+        --no-buffer \
+        --request POST \
+        --url "${API_URL}/completion" \
+        --data-raw "${DATA}")
+
+    printf "\n"
+
+    CHAT+=("$1" "$(trim "$ANSWER")")
+}
+
+while true; do
+    read -r -e -p "> " QUESTION
+    chat_completion "${QUESTION}"
+done

examples/server/server.cpp

@@ -0,0 +1,975 @@
+#include "common.h"
+#include "llama.h"
+#include "build-info.h"
+
+// single thread
+#define CPPHTTPLIB_THREAD_POOL_COUNT 1
+#ifndef NDEBUG
+// crash the server in debug mode, otherwise send an http 500 error
+#define CPPHTTPLIB_NO_EXCEPTIONS 1
+#endif
+
+#include "httplib.h"
+#include "json.hpp"
+
+#ifndef SERVER_VERBOSE
+#define SERVER_VERBOSE 1
+#endif
+
+using namespace httplib;
+using json = nlohmann::json;
+
+struct server_params {
+    std::string hostname = "127.0.0.1";
+    int32_t port = 8080;
+    int32_t read_timeout = 600;
+    int32_t write_timeout = 600;
+};
+
+static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
+    return i;
+}
+
+enum stop_type {
+    STOP_FULL,
+    STOP_PARTIAL,
+};
+
+static bool ends_with(const std::string & str, const std::string & suffix) {
+    return str.size() >= suffix.size() &&
+        0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
+}
+
+static size_t find_partial_stop_string(const std::string & stop,
+                                       const std::string & text) {
+    if (!text.empty() && !stop.empty()) {
+        const char text_last_char = text.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
+            if (stop[char_index] == text_last_char) {
+                const std::string current_partial = stop.substr(0, char_index + 1);
+                if (ends_with(text, current_partial)) {
+                    return text.size() - char_index - 1;
+                }
+            }
+        }
+    }
+    return std::string::npos;
+}
+
+template<class Iter>
+static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
+    std::string ret;
+    for (; begin != end; ++begin) {
+        ret += llama_token_to_str(ctx, *begin);
+    }
+    return ret;
+}
+
+static void server_log(const char * level, const char * function, int line,
+                       const char * message, const nlohmann::ordered_json & extra) {
+    nlohmann::ordered_json log {
+        { "timestamp", time(nullptr) },
+        { "level", level },
+        { "function", function },
+        { "line", line },
+        { "message", message },
+    };
+
+    if (!extra.empty()) {
+        log.merge_patch(extra);
+    }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    fprintf(stdout, "%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
+}
+
+static bool server_verbose = false;
+
+#if SERVER_VERBOSE != 1
+#  define LOG_VERBOSE(MSG, ...)
+#else
+#  define LOG_VERBOSE(MSG, ...)                                          \
+    do {                                                                 \
+        if (server_verbose) {                                            \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+        }                                                                \
+    } while(0)
+#endif
+
+#define LOG_ERROR(MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
+
+struct llama_server_context {
+    bool stream = false;
+    bool has_next_token = false;
+    std::string generated_text;
+
+    size_t num_tokens_predicted = 0;
+    size_t n_past = 0;
+    size_t n_remain = 0;
+
+    std::vector<llama_token> embd;
+    std::vector<llama_token> last_n_tokens;
+
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    gpt_params params;
+
+    bool truncated = false;
+    bool stopped_eos = false;
+    bool stopped_word = false;
+    bool stopped_limit = false;
+    std::string stopping_word;
+    int32_t multibyte_pending = 0;
+
+    ~llama_server_context() {
+        if (ctx) {
+            llama_free(ctx);
+            ctx = nullptr;
+        }
+        if (model) {
+            llama_free_model(model);
+            model = nullptr;
+        }
+    }
+
+    void rewind() {
+        params.antiprompt.clear();
+        num_tokens_predicted = 0;
+        generated_text = "";
+        generated_text.reserve(params.n_ctx);
+        truncated = false;
+        stopped_eos = false;
+        stopped_word = false;
+        stopped_limit = false;
+        stopping_word = "";
+        multibyte_pending = 0;
+
+        n_remain = 0;
+        n_past = 0;
+    }
+
+    bool loadModel(const gpt_params & params_) {
+        params = params_;
+        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        if (model == nullptr) {
+            LOG_ERROR("unable to load model", { { "model", params_.model } });
+            return false;
+        }
+
+        last_n_tokens.resize(params.n_ctx);
+        std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
+        return true;
+    }
+
+    void loadPrompt() {
+        params.prompt.insert(0, 1, ' '); // always add a first space
+        std::vector<llama_token> prompt_tokens = ::llama_tokenize(ctx, params.prompt, true);
+
+        if (params.n_keep < 0) {
+            params.n_keep = (int)prompt_tokens.size();
+        }
+        params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (prompt_tokens.size() >= (size_t)params.n_ctx) {
+            const int n_left = (params.n_ctx - params.n_keep) / 2;
+            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+            const int erased_blocks = (prompt_tokens.size() - params.n_keep - n_left - 1) / n_left;
+            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
+            std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+
+            LOG_VERBOSE("input truncated", {
+                { "n_ctx", params.n_ctx },
+                { "n_keep", params.n_keep },
+                { "n_left", n_left },
+                { "new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend()) },
+            });
+
+            truncated = true;
+            prompt_tokens = new_tokens;
+        } else {
+            const size_t ps = prompt_tokens.size();
+            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
+            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+        embd = prompt_tokens;
+        if (n_past == prompt_tokens.size()) {
+            // we have to evaluate at least 1 token to generate logits.
+            n_past--;
+        }
+
+        LOG_VERBOSE("prompt ingested", {
+            { "n_past", n_past },
+            { "cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past) },
+            { "to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend()) },
+        });
+
+        has_next_token = true;
+    }
+
+    void beginCompletion() {
+        // number of tokens to keep when resetting context
+        n_remain = params.n_predict;
+        llama_set_rng_seed(ctx, params.seed);
+    }
+
+    llama_token nextToken() {
+        llama_token result = -1;
+
+        if (embd.size() >= (size_t)params.n_ctx) {
+            // Reset context
+            const int n_left = (params.n_ctx - params.n_keep) / 2;
+
+            std::vector<llama_token> new_tokens(embd.begin(), embd.begin() + params.n_keep);
+            new_tokens.insert(new_tokens.end(), embd.end() - n_left, embd.end());
+            embd = new_tokens;
+            n_past = params.n_keep;
+            truncated = true;
+            LOG_VERBOSE("input truncated", {
+                { "n_ctx", params.n_ctx },
+                { "n_keep", params.n_keep },
+                { "n_left", n_left },
+                { "new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend()) },
+            });
+        }
+
+        while (n_past < embd.size()) {
+            int n_eval = (int)embd.size() - n_past;
+            if (n_eval > params.n_batch) {
+                n_eval = params.n_batch;
+            }
+            if (llama_eval(ctx, &embd[n_past], n_eval, n_past, params.n_threads)) {
+                LOG_ERROR("failed to eval", {
+                    { "n_eval", n_eval },
+                    { "n_past", n_past },
+                    { "n_threads", params.n_threads },
+                    { "embd", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend()) },
+                });
+                has_next_token = false;
+                return result;
+            }
+            n_past += n_eval;
+        }
+
+        if (params.n_predict == 0) {
+            has_next_token = false;
+            return llama_token_eos();
+        }
+
+        // out of user input, sample next token
+        const float temp = params.temp;
+        const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
+        const float top_p = params.top_p;
+        const float tfs_z = params.tfs_z;
+        const float typical_p = params.typical_p;
+        const int32_t repeat_last_n = params.repeat_last_n < 0 ? params.n_ctx : params.repeat_last_n;
+        const float repeat_penalty = params.repeat_penalty;
+        const float alpha_presence = params.presence_penalty;
+        const float alpha_frequency = params.frequency_penalty;
+        const int mirostat = params.mirostat;
+        const float mirostat_tau = params.mirostat_tau;
+        const float mirostat_eta = params.mirostat_eta;
+        const bool penalize_nl = params.penalize_nl;
+        llama_token id = 0;
+
+        {
+            auto * logits = llama_get_logits(ctx);
+            auto n_vocab = llama_n_vocab(ctx);
+
+            // Apply params.logit_bias map
+            for (const auto & it : params.logit_bias) {
+                logits[it.first] += it.second;
+            }
+
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            // Apply penalties
+            float nl_logit = logits[llama_token_nl()];
+            auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), params.n_ctx);
+            llama_sample_repetition_penalty(ctx, &candidates_p,
+                last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
+                last_n_repeat, repeat_penalty);
+            llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
+                last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
+                last_n_repeat, alpha_frequency, alpha_presence);
+            if (!penalize_nl) {
+                logits[llama_token_nl()] = nl_logit;
+            }
+
+            if (temp <= 0) {
+                // Greedy sampling
+                id = llama_sample_token_greedy(ctx, &candidates_p);
+            } else {
+                if (mirostat == 1) {
+                    static float mirostat_mu = 2.0f * mirostat_tau;
+                    const int mirostat_m = 100;
+                    llama_sample_temperature(ctx, &candidates_p, temp);
+                    id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+                } else if (mirostat == 2) {
+                    static float mirostat_mu = 2.0f * mirostat_tau;
+                    llama_sample_temperature(ctx, &candidates_p, temp);
+                    id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+                } else {
+                    // Temperature sampling
+                    llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
+                    llama_sample_typical(ctx, &candidates_p, typical_p, 1);
+                    llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+                    llama_sample_temperature(ctx, &candidates_p, temp);
+                    id = llama_sample_token(ctx, &candidates_p);
+                }
+            }
+            last_n_tokens.erase(last_n_tokens.begin());
+            last_n_tokens.push_back(id);
+            num_tokens_predicted++;
+        }
+
+        // add it to the context
+        embd.push_back(id);
+        result = id;
+        // decrement remaining sampling budget
+        --n_remain;
+
+        if (!embd.empty() && embd.back() == llama_token_eos()) {
+            //stopping_word = llama_token_to_str(ctx, embd.back());
+            has_next_token = false;
+            stopped_eos = true;
+            LOG_VERBOSE("eos token found", {});
+            return result;
+        }
+
+        has_next_token = params.n_predict == -1 || n_remain != 0;
+        return result;
+    }
+
+    size_t findStoppingStrings(const std::string & text, const size_t last_token_size,
+                               const stop_type type) {
+        size_t stop_pos = std::string::npos;
+        for (const std::string & word : params.antiprompt) {
+            size_t pos;
+            if (type == STOP_FULL) {
+                const size_t tmp = word.size() + last_token_size;
+                const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0;
+                pos = text.find(word, from_pos);
+            }
+            else {
+                pos = find_partial_stop_string(word, text);
+            }
+            if (pos != std::string::npos &&
+                (stop_pos == std::string::npos || pos < stop_pos)) {
+                if (type == STOP_FULL) {
+                    stopping_word = word;
+                    stopped_word = true;
+                    has_next_token = false;
+                }
+                stop_pos = pos;
+            }
+        }
+        return stop_pos;
+    }
+
+    std::string doCompletion() {
+        const llama_token token = nextToken();
+
+        const std::string token_text = token == -1 ? "" : llama_token_to_str(ctx, token);
+        generated_text += token_text;
+
+        if (multibyte_pending > 0) {
+            multibyte_pending -= token_text.size();
+        } else if (token_text.size() == 1) {
+            const char c = token_text[0];
+            // 2-byte characters: 110xxxxx 10xxxxxx
+            if ((c & 0xE0) == 0xC0) {
+                multibyte_pending = 1;
+            // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
+            } else if ((c & 0xF0) == 0xE0) {
+                multibyte_pending = 2;
+            // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+            } else if ((c & 0xF8) == 0xF0) {
+                multibyte_pending = 3;
+            } else {
+                multibyte_pending = 0;
+            }
+        }
+
+        if (multibyte_pending > 0 && !has_next_token) {
+            has_next_token = true;
+            n_remain++;
+        }
+
+        if (!has_next_token && n_remain == 0) {
+            stopped_limit = true;
+        }
+
+        LOG_VERBOSE("next token", {
+            { "token", token },
+            { "token_text", llama_token_to_str(ctx, token) },
+            { "has_next_token", has_next_token },
+            { "n_remain", n_remain },
+            { "num_tokens_predicted", num_tokens_predicted },
+            { "stopped_eos", stopped_eos },
+            { "stopped_word", stopped_word },
+            { "stopped_limit", stopped_limit },
+            { "stopping_word", stopping_word },
+        });
+
+        return token_text;
+    }
+
+    std::vector<float> getEmbedding() {
+        static const int n_embd = llama_n_embd(ctx);
+        if (!params.embedding) {
+            LOG_WARNING("embedding disabled", {
+                { "params.embedding", params.embedding },
+            });
+            return std::vector<float>(n_embd, 0.0f);
+        }
+        const float * data = llama_get_embeddings(ctx);
+        std::vector<float> embedding(data, data + n_embd);
+        return embedding;
+    }
+};
+
+static void server_print_usage(const char * argv0, const gpt_params & params,
+                               const server_params & sparams) {
+    fprintf(stderr, "usage: %s [options]\n", argv0);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    if (llama_mlock_supported()) {
+        fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+    }
+    if (llama_mmap_supported()) {
+        fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+    }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+    fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
+    fprintf(stderr, "                        number of layers to store in VRAM\n");
+    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    fprintf(stderr, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
+#endif
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -a ALIAS, --alias ALIAS\n");
+    fprintf(stderr, "                        set an alias for the model, will be added as `model` field in completion response\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    fprintf(stderr, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    fprintf(stderr, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    fprintf(stderr, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+    fprintf(stderr, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
+    fprintf(stderr, "\n");
+}
+
+static void server_params_parse(int argc, char ** argv, server_params & sparams,
+                                gpt_params & params) {
+    gpt_params default_params;
+    server_params default_sparams;
+    std::string arg;
+    bool invalid_param = false;
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg == "--port") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.port = std::stoi(argv[i]);
+        } else if (arg == "--host") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.hostname = argv[i];
+        } else if (arg == "--timeout" || arg == "-to") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.read_timeout = std::stoi(argv[i]);
+            sparams.write_timeout = std::stoi(argv[i]);
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model = argv[i];
+        } else if (arg == "-a" || arg == "--alias") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_alias = argv[i];
+        } else if (arg == "-h" || arg == "--help") {
+            server_print_usage(argv[0], default_params, default_sparams);
+            exit(0);
+        } else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--memory-f32" || arg == "--memory_f32") {
+            params.memory_f16 = false;
+        } else if (arg == "--threads" || arg == "-t") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads = std::stoi(argv[i]);
+        } else if (arg == "-b" || arg == "--batch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_batch = std::stoi(argv[i]);
+            params.n_batch = std::min(512, params.n_batch);
+        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers = std::stoi(argv[i]);
+#else
+            LOG_WARNING("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. "
+                        "See main README.md for information on enabling GPU BLAS support", { { "n_gpu_layers", params.n_gpu_layers } });
+#endif
+        }
+        else if (arg == "--tensor-split" || arg == "-ts") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            std::string arg_next = argv[i];
+
+            // split string by , and /
+            const std::regex regex{ R"([,/]+)" };
+            std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
+            std::vector<std::string> split_arg{ it, {} };
+            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+            for (size_t i_device = 0; i_device < LLAMA_MAX_DEVICES; ++i_device) {
+                if (i_device < split_arg.size()) {
+                    params.tensor_split[i_device] = std::stof(split_arg[i_device]);
+                }
+                else {
+                    params.tensor_split[i_device] = 0.0f;
+                }
+            }
+#else
+            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.", {});
+#endif // GGML_USE_CUBLAS
+        }
+        else if (arg == "--low-vram" || arg == "-lv")
+        {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+#endif // GGML_USE_CUBLAS
+        }
+        else if (arg == "--main-gpu" || arg == "-mg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            params.main_gpu = std::stoi(argv[i]);
+#else
+            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {});
+#endif
+        } else if (arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter = argv[i];
+            params.use_mmap = false;
+        } else if (arg == "--lora-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_base = argv[i];
+        } else if (arg == "-v" || arg == "--verbose") {
+#if SERVER_VERBOSE != 1
+            LOG_WARNING("server.cpp is not built with verbose logging.", {});
+#else
+            server_verbose = true;
+#endif
+        } else if (arg == "--mlock") {
+            params.use_mlock = true;
+        } else if (arg == "--no-mmap") {
+            params.use_mmap = false;
+        } else if (arg == "--embedding") {
+            params.embedding = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            server_print_usage(argv[0], default_params, default_sparams);
+            exit(1);
+        }
+    }
+
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        server_print_usage(argv[0], default_params, default_sparams);
+        exit(1);
+    }
+}
+
+static json format_generation_settings(llama_server_context & llama) {
+    const auto eos_bias = llama.params.logit_bias.find(llama_token_eos());
+    const bool ignore_eos = eos_bias != llama.params.logit_bias.end() &&
+        eos_bias->second < 0.0f && std::isinf(eos_bias->second);
+
+    return json {
+        { "seed", llama.params.seed },
+        { "temp", llama.params.temp },
+        { "top_k", llama.params.top_k },
+        { "top_p", llama.params.top_p },
+        { "tfs_z", llama.params.tfs_z },
+        { "typical_p", llama.params.typical_p },
+        { "repeat_last_n", llama.params.repeat_last_n },
+        { "repeat_penalty", llama.params.repeat_penalty },
+        { "presence_penalty", llama.params.presence_penalty },
+        { "frequency_penalty", llama.params.frequency_penalty },
+        { "mirostat", llama.params.mirostat },
+        { "mirostat_tau", llama.params.mirostat_tau },
+        { "mirostat_eta", llama.params.mirostat_eta },
+        { "penalize_nl", llama.params.penalize_nl },
+        { "stop", llama.params.antiprompt },
+        { "n_predict", llama.params.n_predict },
+        { "n_keep", llama.params.n_keep },
+        { "ignore_eos", ignore_eos },
+        { "stream", llama.stream },
+        { "logit_bias", llama.params.logit_bias },
+    };
+}
+
+static json format_embedding_response(llama_server_context & llama) {
+    return json {
+        { "embedding", llama.getEmbedding() },
+    };
+}
+
+static json format_final_response(llama_server_context & llama, const std::string & content) {
+    return json {
+        { "content", content },
+        { "stop", true },
+        { "model", llama.params.model_alias },
+        { "tokens_predicted", llama.num_tokens_predicted },
+        { "generation_settings", format_generation_settings(llama) },
+        { "prompt", llama.params.prompt },
+        { "truncated", llama.truncated },
+        { "stopped_eos", llama.stopped_eos },
+        { "stopped_word", llama.stopped_word },
+        { "stopped_limit", llama.stopped_limit },
+        { "stopping_word", llama.stopping_word },
+    };
+}
+
+static json format_partial_response(const std::string & content) {
+    return json {
+        { "content", content },
+        { "stop", false },
+    };
+}
+
+static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
+    return json {
+        { "tokens", tokens }
+    };
+}
+
+static void parse_options_completion(const json & body, llama_server_context & llama) {
+    gpt_params default_params;
+
+    llama.stream = body.value("stream", false);
+    llama.params.n_predict = body.value("n_predict", default_params.n_predict);
+    llama.params.top_k = body.value("top_k", default_params.top_k);
+    llama.params.top_p = body.value("top_p", default_params.top_p);
+    llama.params.tfs_z = body.value("tfs_z", default_params.tfs_z);
+    llama.params.typical_p = body.value("typical_p", default_params.typical_p);
+    llama.params.repeat_last_n = body.value("repeat_last_n", default_params.repeat_last_n);
+    llama.params.temp = body.value("temperature", default_params.temp);
+    llama.params.repeat_penalty = body.value("repeat_penalty", default_params.repeat_penalty);
+    llama.params.presence_penalty = body.value("presence_penalty", default_params.presence_penalty);
+    llama.params.frequency_penalty = body.value("frequency_penalty", default_params.frequency_penalty);
+    llama.params.mirostat = body.value("mirostat", default_params.mirostat);
+    llama.params.mirostat_tau = body.value("mirostat_tau", default_params.mirostat_tau);
+    llama.params.mirostat_eta = body.value("mirostat_eta", default_params.mirostat_eta);
+    llama.params.penalize_nl = body.value("penalize_nl", default_params.penalize_nl);
+    llama.params.n_keep = body.value("n_keep", default_params.n_keep);
+    llama.params.seed = body.value("seed", default_params.seed);
+    llama.params.prompt = body.value("prompt", default_params.prompt);
+
+    llama.params.logit_bias.clear();
+    if (body.value("ignore_eos", false)) {
+        llama.params.logit_bias[llama_token_eos()] = -INFINITY;
+    }
+
+    const auto & logit_bias = body.find("logit_bias");
+    if (logit_bias != body.end() && logit_bias->is_array()) {
+        const int n_vocab = llama_n_vocab(llama.ctx);
+        for (const auto & el : *logit_bias) {
+            if (el.is_array() && el.size() == 2 && el[0].is_number_integer()) {
+                llama_token tok = el[0].get<llama_token>();
+                if (tok >= 0 && tok < n_vocab) {
+                    if (el[1].is_number()) {
+                        llama.params.logit_bias[tok] = el[1].get<float>();
+                    } else if (el[1].is_boolean() && !el[1].get<bool>()) {
+                        llama.params.logit_bias[tok] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    llama.params.antiprompt.clear();
+    const auto & stop = body.find("stop");
+    if (stop != body.end() && stop->is_array()) {
+        for (const auto & word : *stop) {
+            if (!word.empty()) {
+                llama.params.antiprompt.push_back(word);
+            }
+        }
+    }
+
+    LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
+}
+
+static void log_server_request(const Request & req, const Response & res) {
+    LOG_INFO("request", {
+        { "remote_addr", req.remote_addr },
+        { "remote_port", req.remote_port },
+        { "status", res.status },
+        { "path", req.path },
+        { "request", req.body },
+        { "response", res.body },
+    });
+}
+
+int main(int argc, char ** argv) {
+    // own arguments required by this example
+    gpt_params params;
+    server_params sparams;
+
+    // struct that contains llama context and inference
+    llama_server_context llama;
+
+    server_params_parse(argc, argv, sparams, params);
+
+    if (params.model_alias == "unknown") {
+        params.model_alias = params.model;
+    }
+
+    llama_init_backend();
+
+    LOG_INFO("build info", {
+        { "build", BUILD_NUMBER },
+        { "commit", BUILD_COMMIT }
+    });
+    LOG_INFO("system info", {
+        { "n_threads", params.n_threads },
+        { "total_threads", std::thread::hardware_concurrency() },
+        { "system_info", llama_print_system_info() },
+    });
+
+    // load the model
+    if (!llama.loadModel(params)) {
+        return 1;
+    }
+
+    Server svr;
+
+    svr.set_default_headers({
+        { "Access-Control-Allow-Origin", "*" },
+        { "Access-Control-Allow-Headers", "content-type" }
+    });
+
+    svr.Get("/", [](const Request &, Response & res) {
+        res.set_content("<h1>llama.cpp server works</h1>", "text/html");
+    });
+
+    svr.Post("/completion", [&llama](const Request & req, Response & res) {
+        llama.rewind();
+        llama_reset_timings(llama.ctx);
+
+        parse_options_completion(json::parse(req.body), llama);
+
+        llama.loadPrompt();
+        llama.beginCompletion();
+
+        if (!llama.stream) {
+            size_t stop_pos = std::string::npos;
+
+            while (llama.has_next_token) {
+                const std::string token_text = llama.doCompletion();
+
+                stop_pos = llama.findStoppingStrings(llama.generated_text,
+                    token_text.size(), STOP_FULL);
+            }
+
+            if (stop_pos == std::string::npos) {
+                stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
+            }
+            if (stop_pos != std::string::npos) {
+                llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
+                    llama.generated_text.end());
+            }
+
+            const json data = format_final_response(llama, llama.generated_text);
+
+            llama_print_timings(llama.ctx);
+
+            res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace),
+                            "application/json");
+        } else {
+            const auto chunked_content_provider = [&](size_t, DataSink & sink) {
+                size_t sent_count = 0;
+
+                while (llama.has_next_token) {
+                    const std::string token_text = llama.doCompletion();
+                    if (llama.multibyte_pending > 0) {
+                        continue;
+                    }
+
+                    size_t pos = std::min(sent_count, llama.generated_text.size());
+
+                    const std::string str_test = llama.generated_text.substr(pos);
+                    size_t stop_pos =
+                        llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
+                    if (stop_pos != std::string::npos) {
+                        llama.generated_text.erase(
+                            llama.generated_text.begin() + pos + stop_pos,
+                            llama.generated_text.end());
+                        pos = std::min(sent_count, llama.generated_text.size());
+                    } else {
+                        stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
+                            STOP_PARTIAL);
+                    }
+
+                    const std::string to_send = llama.generated_text.substr(pos, stop_pos);
+                    sent_count += to_send.size();
+
+                    const json data = llama.has_next_token
+                                          ? format_partial_response(to_send)
+                                          // Generation is done, send extra information.
+                                          : format_final_response(llama, to_send);
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+
+                llama_print_timings(llama.ctx);
+                sink.done();
+                return true;
+            };
+            res.set_chunked_content_provider("text/event-stream", chunked_content_provider);
+        }
+    });
+
+    svr.Options(R"(/.*)", [](const Request &, Response & res) {
+        return res.set_content("", "application/json");
+    });
+
+    svr.Post("/tokenize", [&llama](const Request & req, Response & res) {
+        const json body = json::parse(req.body);
+        const std::string content = body.value("content", "");
+        const std::vector<llama_token> tokens = llama_tokenize(llama.ctx, content, false);
+        const json data = format_tokenizer_response(tokens);
+        return res.set_content(data.dump(), "application/json");
+    });
+
+    svr.Post("/embedding", [&llama](const Request & req, Response & res) {
+        const json body = json::parse(req.body);
+
+        llama.rewind();
+        llama_reset_timings(llama.ctx);
+        llama.params.prompt = body.value("content", "");
+        llama.params.n_predict = 0;
+        llama.loadPrompt();
+        llama.beginCompletion();
+        llama.doCompletion();
+
+        const json data = format_embedding_response(llama);
+        return res.set_content(data.dump(), "application/json");
+    });
+
+    svr.set_logger(log_server_request);
+
+    svr.set_exception_handler([](const Request &, Response & res, std::exception_ptr ep) {
+        const auto * fmt = "500 Internal Server Error\n%s";
+        char buf[BUFSIZ];
+        try {
+            std::rethrow_exception(std::move(ep));
+        } catch (std::exception & e) {
+            snprintf(buf, sizeof(buf), fmt, e.what());
+        } catch (...) {
+            snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
+        }
+        res.set_content(buf, "text/plain");
+        res.status = 500;
+    });
+
+    // set timeouts and change hostname and port
+    svr.set_read_timeout(sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port)) {
+        LOG_ERROR("couldn't bind to server socket", {
+            { "hostname", sparams.hostname },
+            { "port", sparams.port },
+        });
+        return 1;
+    }
+
+    LOG_INFO("HTTP server listening", {
+        { "hostname", sparams.hostname },
+        { "port", sparams.port },
+    });
+
+    if (!svr.listen_after_bind()) {
+        return 1;
+    }
+
+    return 0;
+}

examples/simple/CMakeLists.txt

@@ -0,0 +1,7 @@
+set(TARGET simple)
+add_executable(${TARGET} simple.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/simple/simple.cpp

@@ -0,0 +1,179 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
+#include "common.h"
+#include "llama.h"
+#include "build-info.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#include <signal.h>
+#endif
+
+
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc == 1 || argv[1][0] == '-' )
+    {
+        printf( "usage: %s MODEL_PATH [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    if ( argc >= 2 )
+    {
+        params.model = argv[1];
+    }
+
+    if ( argc >= 3 )
+    {
+        params.prompt = argv[2];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "Hello my name is";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_init_backend();
+
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+
+    if ( model == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize( ctx , params.prompt , true );
+
+    const int max_context_size     = llama_n_ctx( ctx );
+    const int max_tokens_list_size = max_context_size - 4 ;
+
+    if ( (int)tokens_list.size() > max_tokens_list_size )
+    {
+        fprintf( stderr , "%s: error: prompt too long (%d tokens, max %d)\n" ,
+             __func__ , (int)tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        printf( "%s" , llama_token_to_str( ctx , id ) );
+    }
+
+    fflush(stdout);
+
+
+    //---------------------------------
+    // Main prediction loop :
+    //---------------------------------
+
+    // The LLM keeps a contextual cache memory of previous token evaluation.
+    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
+    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
+    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+
+    while ( llama_get_kv_cache_token_count( ctx ) < max_context_size )
+    {
+        //---------------------------------
+        // Evaluate the tokens :
+        //---------------------------------
+
+        if ( llama_eval( ctx , tokens_list.data() , tokens_list.size() , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
+        {
+            fprintf( stderr,  "%s : failed to eval\n" , __func__ );
+            return 1;
+        }
+
+        tokens_list.clear();
+
+        //---------------------------------
+        // Select the best prediction :
+        //---------------------------------
+
+        llama_token new_token_id = 0;
+
+        auto logits  = llama_get_logits( ctx );
+        auto n_vocab = llama_n_vocab( ctx ); // the size of the LLM vocabulary (in tokens)
+
+        std::vector<llama_token_data> candidates;
+        candidates.reserve( n_vocab );
+
+        for( llama_token token_id = 0 ; token_id < n_vocab ; token_id++ )
+        {
+            candidates.emplace_back( llama_token_data{ token_id , logits[ token_id ] , 0.0f } );
+        }
+
+        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+        // Select it using the "Greedy sampling" method :
+        new_token_id = llama_sample_token_greedy( ctx , &candidates_p );
+
+
+        // is it an end of stream ?
+        if ( new_token_id == llama_token_eos() )
+        {
+            fprintf(stderr, " [end of text]\n");
+            break;
+        }
+
+        // Print the new token :
+        printf( "%s" , llama_token_to_str( ctx , new_token_id ) );
+        fflush( stdout );
+
+        // Push this new token for next evaluation :
+        tokens_list.push_back( new_token_id );
+
+    } // wend of main loop
+
+    llama_free( ctx );
+    llama_free_model( model );
+
+    return 0;
+}
+
+// EOF

examples/train-text-from-scratch/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET train-text-from-scratch)
+add_executable(${TARGET} train-text-from-scratch.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/train-text-from-scratch/README.md

@@ -0,0 +1,22 @@
+# train-text-from-scratch
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# train
+./bin/train-text-from-scratch \
+        --vocab-model ../models/ggml-vocab.bin \
+        --ctx 64 --embd 256 --head 8 --layer 16 \
+        --checkpoint-in  chk-shakespeare-256x16.bin \
+        --checkpoint-out chk-shakespeare-256x16.bin \
+        --model-out ggml-shakespeare-256x16-f32.bin \
+        --train-data "shakespeare.txt" \
+        -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
+        --print-details-interval 0 --predict 16 --use-flash
+
+# predict
+./bin/main -m ggml-shakespeare-256x16-f32.bin
+```

flake.lock

@@ -1,12 +1,15 @@
 {
   "nodes": {
     "flake-utils": {
+      "inputs": {
+        "systems": "systems"
+      },
       "locked": {
-        "lastModified": 1676283394,
-        "narHash": "sha256-XX2f9c3iySLCw54rJ/CZs+ZK6IQy7GXNY4nSOyu2QG4=",
+        "lastModified": 1685518550,
+        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "3db36a8b464d0c4532ba1c7dda728f4576d6d073",
+        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
         "type": "github"
       },
       "original": {
@@ -17,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1678470307,
-        "narHash": "sha256-OEeMUr3ueLIXyW/OaFUX5jUdimyQwMg/7e+/Q0gC/QE=",
+        "lastModified": 1685931219,
+        "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "0c4800d579af4ed98ecc47d464a5e7b0870c4b1f",
+        "rev": "7409480d5c8584a1a83c422530419efe4afb0d19",
         "type": "github"
       },
       "original": {
@@ -36,6 +39,21 @@
         "flake-utils": "flake-utils",
         "nixpkgs": "nixpkgs"
       }
+    },
+    "systems": {
+      "locked": {
+        "lastModified": 1681028828,
+        "narHash": "sha256-Vy1rq5AaRuLzOxct8nz4T6wlgyUR7zLU309k9mBC768=",
+        "owner": "nix-systems",
+        "repo": "default",
+        "rev": "da67096a3b9bf56a91d16901293e51ba5b49a27e",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-systems",
+        "repo": "default",
+        "type": "github"
+      }
     }
   },
   "root": "root",

flake.nix

@@ -6,29 +6,47 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
-        pkgs = import nixpkgs {
-          inherit system;
-        };
-        llama-python = pkgs.python310.withPackages (ps: with ps; [
-          numpy
-          sentencepiece
-        ]);
-      in
-      {
+        inherit (pkgs.stdenv) isAarch64 isDarwin;
+        inherit (pkgs.lib) optionals;
+        isM1 = isAarch64 && isDarwin;
+        osSpecific = if isM1 then
+          with pkgs.darwin.apple_sdk_11_0.frameworks; [
+            Accelerate
+            MetalKit
+            MetalPerformanceShaders
+            MetalPerformanceShadersGraph
+          ]
+        else if isDarwin then
+          with pkgs.darwin.apple_sdk.frameworks; [
+            Accelerate
+            CoreGraphics
+            CoreVideo
+          ]
+        else
+          [ ];
+        pkgs = import nixpkgs { inherit system; };
+        llama-python =
+          pkgs.python310.withPackages (ps: with ps; [ numpy sentencepiece ]);
+      in {
         packages.default = pkgs.stdenv.mkDerivation {
           name = "llama.cpp";
           src = ./.;
+          postPatch = if isM1 then ''
+            substituteInPlace ./ggml-metal.m \
+              --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+          '' else
+            "";
           nativeBuildInputs = with pkgs; [ cmake ];
-          buildInputs = with pkgs; lib.optionals stdenv.isDarwin [
-            darwin.apple_sdk.frameworks.Accelerate
-          ];
-          cmakeFlags = with pkgs; lib.optionals (system == "aarch64-darwin") [
+          buildInputs = osSpecific;
+          cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" ] ++ (optionals isM1 [
             "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-          ];
+            "-DLLAMA_METAL=ON"
+          ]);
           installPhase = ''
             mkdir -p $out/bin
             mv bin/* $out/bin/
             mv $out/bin/main $out/bin/llama
+            mv $out/bin/server $out/bin/llama-server
 
             echo "#!${llama-python}/bin/python" > $out/bin/convert.py
             cat ${./convert.py} >> $out/bin/convert.py
@@ -36,14 +54,21 @@
           '';
           meta.mainProgram = "llama";
         };
-        devShells.default = pkgs.mkShell {
-          packages = with pkgs; [
-            cmake
-            llama-python
-          ] ++ lib.optionals stdenv.isDarwin [
-            darwin.apple_sdk.frameworks.Accelerate
-          ];
+        apps.llama-server = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/llama-server";
         };
-      }
-    );
+        apps.llama-embedding = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/embedding";
+        };
+        apps.llama = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/llama";
+        };
+        apps.default = self.apps.${system}.llama;
+        devShells.default = pkgs.mkShell {
+          packages = with pkgs; [ cmake llama-python ] ++ osSpecific;
+        };
+      });
 }

ggml-cuda.h

@@ -1,11 +1,21 @@
+#pragma once
+
 #include "ggml.h"
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-void   ggml_init_cublas(void);
+#define GGML_CUDA_MAX_DEVICES       16
 
+struct ggml_tensor_extra_gpu {
+    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
+};
+
+void   ggml_init_cublas(void);
+void   ggml_cuda_set_tensor_split(const float * tensor_split);
+
+void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
@@ -14,7 +24,15 @@ void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
 void * ggml_cuda_host_malloc(size_t size);
 void   ggml_cuda_host_free(void * ptr);
 
-void ggml_cuda_transform_tensor(struct ggml_tensor * tensor);
+void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
+void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
+bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus
 }

ggml-metal.h

@@ -0,0 +1,67 @@
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 16
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_metal_context;
+
+struct ggml_metal_context * ggml_metal_init(void);
+void ggml_metal_free(struct ggml_metal_context * ctx);
+
+// creates a mapping between a host memory buffer and a device memory buffer
+// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+// - the mapping is used during computation to determine the arguments of the compute kernels
+// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+// - max_size specifies the maximum size of a tensor and is used to create shared views such
+//   that it is guaranteed that the tensor will fit in at least one of the views
+//
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                       const char * name,
+                             void * data,
+                           size_t   size,
+                           size_t   max_size);
+
+// set data from host memory into the device
+void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// get data from the device into host memory
+void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// same as ggml_graph_compute but uses Metal
+// creates gf->n_threads command buffers in parallel
+void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+#ifdef __cplusplus
+}
+#endif
+

ggml-metal.m

@@ -0,0 +1,972 @@
+#import "ggml-metal.h"
+
+#import "ggml.h"
+
+#import <Foundation/Foundation.h>
+
+#import <Metal/Metal.h>
+#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
+
+#ifdef GGML_METAL_NDEBUG
+#define metal_printf(...)
+#else
+#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#endif
+
+#define UNUSED(x) (void)(x)
+
+struct ggml_metal_buffer {
+    const char * name;
+
+    void   * data;
+    size_t   size;
+
+    id<MTLBuffer> metal;
+};
+
+struct ggml_metal_context {
+    float * logits;
+
+    id<MTLDevice>       device;
+    id<MTLCommandQueue> queue;
+    id<MTLLibrary>      library;
+
+    int n_buffers;
+    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+
+    // custom kernels
+#define GGML_METAL_DECL_KERNEL(name) \
+    id<MTLFunction>             function_##name; \
+    id<MTLComputePipelineState> pipeline_##name
+
+    GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(mul);
+    GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+    GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(silu);
+    GGML_METAL_DECL_KERNEL(relu);
+    GGML_METAL_DECL_KERNEL(gelu);
+    GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(get_rows_f16);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q2_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q3_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q6_k);
+    GGML_METAL_DECL_KERNEL(rms_norm);
+    GGML_METAL_DECL_KERNEL(norm);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q2_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q3_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q5_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q6_k_f32);
+    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(alibi_f32);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f16);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f32);
+    GGML_METAL_DECL_KERNEL(cpy_f16_f16);
+
+#undef GGML_METAL_DECL_KERNEL
+};
+
+// 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";
+
+// Here to assist with NSBundle Path Hack
+@interface GGMLMetalClass : NSObject
+@end
+@implementation GGMLMetalClass
+@end
+
+struct ggml_metal_context * ggml_metal_init(void) {
+    fprintf(stderr, "%s: allocating\n", __func__);
+
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+
+    ctx->device = MTLCreateSystemDefaultDevice();
+    ctx->queue  = [ctx->device newCommandQueue];
+    ctx->n_buffers = 0;
+
+    // determine if we can use MPS
+    if (MPSSupportsMTLDevice(ctx->device)) {
+        fprintf(stderr, "%s: using MPS\n", __func__);
+    } else {
+        fprintf(stderr, "%s: not using MPS\n", __func__);
+        GGML_ASSERT(false && "MPS not supported");
+    }
+
+#if 0
+    // compile from source string and show compile log
+    {
+        NSError * error = nil;
+
+        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+#else
+    UNUSED(msl_library_source);
+
+    // read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
+    {
+        NSError * error = nil;
+
+        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
+        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
+
+        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+
+        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+#endif
+
+    // load kernels
+    {
+#define GGML_METAL_ADD_KERNEL(name) \
+        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
+        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
+
+        GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(mul);
+        GGML_METAL_ADD_KERNEL(mul_row);
+        GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(silu);
+        GGML_METAL_ADD_KERNEL(relu);
+        GGML_METAL_ADD_KERNEL(gelu);
+        GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(get_rows_f16);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q2_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q3_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q6_k);
+        GGML_METAL_ADD_KERNEL(rms_norm);
+        GGML_METAL_ADD_KERNEL(norm);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q2_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q3_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q5_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q6_k_f32);
+        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(alibi_f32);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f16);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f32);
+        GGML_METAL_ADD_KERNEL(cpy_f16_f16);
+
+#undef GGML_METAL_ADD_KERNEL
+    }
+
+    fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    fprintf(stderr, "%s: hasUnifiedMemory             = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    if (ctx->device.maxTransferRate != 0) {
+        fprintf(stderr, "%s: maxTransferRate              = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+    } else {
+        fprintf(stderr, "%s: maxTransferRate              = built-in GPU\n", __func__);
+    }
+
+    return ctx;
+}
+
+void ggml_metal_free(struct ggml_metal_context * ctx) {
+    fprintf(stderr, "%s: deallocating\n", __func__);
+
+    free(ctx);
+}
+
+// finds the Metal buffer that contains the tensor data on the GPU device
+// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
+// Metal buffer based on the host memory pointer
+//
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
+    //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+
+    const int64_t tsize = ggml_nbytes(t);
+
+    // find the view that contains the tensor fully
+    for (int i = 0; i < ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
+
+        if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
+            *offs = (size_t) ioffs;
+
+            //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+
+            return ctx->buffers[i].metal;
+        }
+    }
+
+    fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+
+    return nil;
+}
+
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                     const char * name,
+                           void * data,
+                         size_t   size,
+                         size_t   max_size) {
+    if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
+        fprintf(stderr, "%s: too many buffers\n", __func__);
+        return false;
+    }
+
+    if (data) {
+        // verify that the buffer does not overlap with any of the existing buffers
+        for (int i = 0; i < ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
+
+            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                return false;
+            }
+        }
+
+        const size_t size_page = getpagesize();
+
+        size_t size_aligned = size;
+        if ((size_aligned % size_page) != 0) {
+            size_aligned += (size_page - (size_aligned % size_page));
+        }
+
+        // the buffer fits into the max buffer size allowed by the device
+        if (size_aligned <= ctx->device.maxBufferLength) {
+            ctx->buffers[ctx->n_buffers].name = name;
+            ctx->buffers[ctx->n_buffers].data = data;
+            ctx->buffers[ctx->n_buffers].size = size;
+
+            ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                return false;
+            }
+
+            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+
+            ++ctx->n_buffers;
+        } else {
+            // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+            // one of the views
+            const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+            const size_t size_step = ctx->device.maxBufferLength - size_ovlp;
+            const size_t size_view = ctx->device.maxBufferLength;
+
+            for (size_t i = 0; i < size; i += size_step) {
+                const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+                ctx->buffers[ctx->n_buffers].name = name;
+                ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
+                ctx->buffers[ctx->n_buffers].size = size_step_aligned;
+
+                ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+                if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                    fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    return false;
+                }
+
+                fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                if (i + size_step < size) {
+                    fprintf(stderr, "\n");
+                }
+
+                ++ctx->n_buffers;
+            }
+        }
+
+        fprintf(stderr, ", (%8.2f / %8.2f)",
+                ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
+                ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+        if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
+            fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
+        } else {
+            fprintf(stderr, "\n");
+        }
+    }
+
+    return true;
+}
+
+void ggml_metal_set_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
+}
+
+void ggml_metal_get_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
+}
+
+void ggml_metal_graph_compute(
+        struct ggml_metal_context * ctx,
+               struct ggml_cgraph * gf) {
+    metal_printf("%s: evaluating graph\n", __func__);
+
+    // create multiple command buffers and enqueue them
+    // then, we encode the graph into the command buffers in parallel
+
+    const int n_cb = gf->n_threads;
+
+    NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
+
+    for (int i = 0; i < n_cb; ++i) {
+        command_buffers[i] = [ctx->queue commandBuffer];
+
+        // enqueue the command buffers in order to specify their execution order
+        [command_buffers[i] enqueue];
+    }
+
+    // TODO: is this the best way to start threads?
+    dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+
+    for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
+        const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
+
+        dispatch_async(queue, ^{
+            size_t offs_src0 = 0;
+            size_t offs_src1 = 0;
+            size_t offs_dst  = 0;
+
+            id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
+
+            id<MTLComputeCommandEncoder> encoder = nil;
+
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+
+            for (int i = node_start; i < node_end; ++i) {
+                metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+
+                struct ggml_tensor * src0 = gf->nodes[i]->src0;
+                struct ggml_tensor * src1 = gf->nodes[i]->src1;
+                struct ggml_tensor * dst  = gf->nodes[i];
+
+                const int64_t  ne00 = src0 ? src0->ne[0] : 0;
+                const int64_t  ne01 = src0 ? src0->ne[1] : 0;
+                const int64_t  ne02 = src0 ? src0->ne[2] : 0;
+                const int64_t  ne03 = src0 ? src0->ne[3] : 0;
+
+                const uint64_t nb00 = src0 ? src0->nb[0] : 0;
+                const uint64_t nb01 = src0 ? src0->nb[1] : 0;
+                const uint64_t nb02 = src0 ? src0->nb[2] : 0;
+                const uint64_t nb03 = src0 ? src0->nb[3] : 0;
+
+                const int64_t  ne10 = src1 ? src1->ne[0] : 0;
+                const int64_t  ne11 = src1 ? src1->ne[1] : 0;
+                const int64_t  ne12 = src1 ? src1->ne[2] : 0;
+                const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+
+                const uint64_t nb10 = src1 ? src1->nb[0] : 0;
+                const uint64_t nb11 = src1 ? src1->nb[1] : 0;
+                const uint64_t nb12 = src1 ? src1->nb[2] : 0;
+                const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+
+                const int64_t  ne0  = dst ? dst->ne[0] : 0;
+                const int64_t  ne1  = dst ? dst->ne[1] : 0;
+                const int64_t  ne2  = dst ? dst->ne[2] : 0;
+                const int64_t  ne3  = dst ? dst->ne[3] : 0;
+
+                const uint64_t nb0  = dst ? dst->nb[0] : 0;
+                const uint64_t nb1  = dst ? dst->nb[1] : 0;
+                const uint64_t nb2  = dst ? dst->nb[2] : 0;
+                const uint64_t nb3  = dst ? dst->nb[3] : 0;
+
+                const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+                const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+                const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
+
+                id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
+                id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
+                id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+
+                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //if (src0) {
+                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //            ggml_is_contiguous(src0), src0->name);
+                //}
+                //if (src1) {
+                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //            ggml_is_contiguous(src1), src1->name);
+                //}
+                //if (dst) {
+                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //            dst->name);
+                //}
+
+                switch (dst->op) {
+                    case GGML_OP_RESHAPE:
+                    case GGML_OP_VIEW:
+                    case GGML_OP_TRANSPOSE:
+                    case GGML_OP_PERMUTE:
+                        {
+                            // noop
+                        } break;
+                    case GGML_OP_ADD:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_mul_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_mul];
+                            }
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SCALE:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float scale = *(const float *) src1->data;
+
+                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SILU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_silu];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_RELU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_relu];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_GELU:
+                    {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_gelu];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                    case GGML_OP_SOFT_MAX:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int nth = 32;
+
+                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_DIAG_MASK_INF:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int n_past = ((int32_t *)(src1->data))[0];
+
+                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL_MAT:
+                        {
+                            // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
+
+                            GGML_ASSERT(ne00 == ne10);
+                            GGML_ASSERT(ne02 == ne12);
+
+                            if (ggml_is_contiguous(src0) &&
+                                ggml_is_contiguous(src1) &&
+                                (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
+
+                                if (encoder != nil) {
+                                    [encoder endEncoding];
+                                    encoder = nil;
+                                }
+
+                                MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                                MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+
+                                // for F32 x F32 we use MPS
+                                MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
+
+                                MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
+
+                                MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
+
+                                MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
+                                    initWithDevice:ctx->device transposeLeft:false transposeRight:true
+                                        resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
+
+                                // we need to do ne02 multiplications
+                                // TODO: is there a way to do this in parallel - currently very slow ..
+                                // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
+                                for (int64_t i02 = 0; i02 < ne02; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                    size_t offs_src1_cur = offs_src1 + i02*nb12;
+                                    size_t offs_dst_cur  = offs_dst  + i02*nb2;
+
+                                    MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
+                                    MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
+                                    MPSMatrix * mat_dst  = [[MPSMatrix alloc] initWithBuffer:id_dst  offset:offs_dst_cur  descriptor:desc ];
+
+                                    [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
+                                }
+                            } else {
+                                if (encoder == nil) {
+                                    encoder = [command_buffer computeCommandEncoder];
+                                }
+
+                                int nth0 = 32;
+                                int nth1 = 1;
+
+                                // use custom matrix x vector kernel
+                                switch (src0t) {
+                                    case GGML_TYPE_F16:
+                                        {
+                                            GGML_ASSERT(ne02 == ne12);
+
+                                            nth0 = 64;
+                                            nth1 = 1;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_1:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q2_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q3_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q6_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
+                                        } break;
+                                    default:
+                                        {
+                                            fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                            GGML_ASSERT(false && "not implemented");
+                                        }
+                                };
+
+                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q2_K ||
+                                         src0t == GGML_TYPE_Q3_K ||
+                                         src0t == GGML_TYPE_Q4_K ||
+                                         src0t == GGML_TYPE_Q5_K ||
+                                         src0t == GGML_TYPE_Q6_K) {
+                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                } else {
+                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                            }
+                        } break;
+                    case GGML_OP_GET_ROWS:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            switch (src0->type) {
+                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
+                                case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
+                                case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
+                                case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
+                                case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
+                                case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
+                                case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+
+                            const int64_t n = ggml_nelements(src1);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_RMS_NORM:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float eps = 1e-6f;
+
+                            const int nth = 256;
+
+                            [encoder setComputePipelineState:ctx->pipeline_rms_norm];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            const int64_t nrows = ggml_nrows(src0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_NORM:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float eps = 1e-5f;
+
+                            const int nth = 256;
+
+                            [encoder setComputePipelineState:ctx->pipeline_norm];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            const int64_t nrows = ggml_nrows(src0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_ALIBI:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            GGML_ASSERT((src0t == GGML_TYPE_F32));
+
+                            const int   n_past   = ((int32_t *) src1->data)[0]; UNUSED(n_past);
+                            const int   n_head   = ((int32_t *) src1->data)[1];
+                            const float max_bias = ((float *)   src1->data)[2];
+
+                            if (__builtin_popcount(n_head) != 1) {
+                                GGML_ASSERT(false && "only power-of-two n_head implemented");
+                            }
+
+                            const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+                            const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+
+                            [encoder setComputePipelineState:ctx->pipeline_alibi_f32];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
+                            const int nth = 32;
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_ROPE:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int n_dims = ((int32_t *) src1->data)[1];
+                            const int mode   = ((int32_t *) src1->data)[2];
+
+                            const int n_past = ((int32_t *)(src1->data))[0];
+
+                            [encoder setComputePipelineState:ctx->pipeline_rope];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00   length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01   length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02   length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03   length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00   length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01   length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02   length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03   length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0    length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1    length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2    length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3    length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0    length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1    length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2    length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3    length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&n_past length:sizeof(     int) atIndex:18];
+                            [encoder setBytes:&n_dims length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&mode   length:sizeof(     int) atIndex:20];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_CPY:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int nth = 32;
+
+                            switch (src0t) {
+                                case GGML_TYPE_F32:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
+                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
+                                case GGML_TYPE_F16:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
+                                            case GGML_TYPE_F32: GGML_ASSERT(false && "cpy_f16_f32 not implemented"); break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    default:
+                        fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                        GGML_ASSERT(false);
+                }
+            }
+
+            if (encoder != nil) {
+                [encoder endEncoding];
+                encoder = nil;
+            }
+
+            [command_buffer commit];
+        });
+    }
+
+    // wait for all threads to finish
+    dispatch_barrier_sync(queue, ^{});
+
+    [command_buffers[n_cb - 1] waitUntilCompleted];
+
+    // check status of command buffers
+    // needed to detect if the device ran out-of-memory for example (#1881)
+    for (int i = 0; i < n_cb; i++) {
+        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status];
+        if (status != MTLCommandBufferStatusCompleted) {
+            fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            GGML_ASSERT(false);
+        }
+    }
+}

ggml-opencl.c

@@ -1,361 +0,0 @@
-#include "ggml-opencl.h"
-
-#define CL_TARGET_OPENCL_VERSION 110
-#include <clblast_c.h>
-
-#include <stdlib.h>
-#include <stdio.h>
-#include <string.h>
-
-#include "ggml.h"
-
-#define MULTILINE_QUOTE(...) #__VA_ARGS__
-const char * clblast_dequant = MULTILINE_QUOTE(
-
-typedef uchar uint8_t;
-typedef int int32_t;
-typedef uint uint32_t;
-
-constant uint QK4_0 = 32;
-struct block_q4_0
-{
-    float d;
-    uint8_t qs[QK4_0 / 2];
-};
-
-constant uint QK4_1 = 32;
-struct block_q4_1
-{
-    float d;
-    float m;
-    uint8_t qs[QK4_1 / 2];
-};
-
-constant uint QK5_0 = 32;
-struct __attribute__ ((packed)) block_q5_0
-{
-    half d;
-    uint32_t qh;
-    uint8_t qs[QK5_0 / 2];
-};
-
-constant uint QK5_1 = 32;
-struct block_q5_1
-{
-    half d;
-    half m;
-    uint32_t qh;
-    uint8_t qs[QK5_1 / 2];
-};
-
-constant uint QK8_0 = 32;
-struct block_q8_0
-{
-    float d;
-    uint8_t qs[QK8_0];
-};
-
-
-__kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float* y) {
-    constant uint qk = QK4_0;
-
-    const uint i = get_global_id(0) / qk;
-    const uint j = get_local_id(0);
-
-    const float d = x[i].d;
-
-    const int x0 = (x[i].qs[j] & 0xf) - 8;
-    const int x1 = (x[i].qs[j] >>  4) - 8;
-
-    y[i*qk + j + 0   ] = x0*d;
-    y[i*qk + j + qk/2] = x1*d;
-}
-
-__kernel void dequantize_row_q4_1(__global struct block_q4_1* x, __global float* y) {
-    constant uint qk = QK4_1;
-
-    const uint i = get_global_id(0) / qk;
-    const uint j = get_local_id(0);
-
-    const float d = x[i].d;
-    const float m = x[i].m;
-
-    const int x0 = (x[i].qs[j] & 0xf);
-    const int x1 = (x[i].qs[j] >>  4);
-
-    y[i*qk + j + 0   ] = x0*d + m;
-    y[i*qk + j + qk/2] = x1*d + m;
-}
-
-__kernel void dequantize_row_q5_0(__global struct block_q5_0* x, __global float* y) {
-    constant uint qk = QK5_0;
-
-    const uint i = get_global_id(0) / qk;
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-
-    uint32_t qh = x[i].qh;
-
-    const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-    const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16;
-    const int32_t x1 = ((x[i].qs[j] >>  4) | xh_1) - 16;
-
-    y[i*qk + j + 0   ] = x0*d;
-    y[i*qk + j + qk/2] = x1*d;
-}
-
-__kernel void dequantize_row_q5_1(__global struct block_q5_1* x, __global float* y) {
-    constant uint qk = QK5_1;
-
-    const uint i = get_global_id(0) / qk;
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-    const float m = vload_half(0, (__global half*) &x[i].m);
-
-    uint32_t qh = x[i].qh;
-
-    const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-    const int x0 = (x[i].qs[j] & 0xf) | xh_0;
-    const int x1 = (x[i].qs[j] >>  4) | xh_1;
-
-    y[i*qk + j + 0   ] = x0*d + m;
-    y[i*qk + j + qk/2] = x1*d + m;
-}
-
-__kernel void dequantize_row_q8_0(__global struct block_q8_0* x, __global float* y) {
-    constant uint qk = QK8_0;
-    const uint i = get_global_id(0) / qk;
-    const uint j = get_local_id(0);
-
-    const float d = x[i].d;
-    y[i*qk + j] = x[i].qs[j]*d;
-}
-
-);
-
-#define CL_CHECK(err, name)                                                                     \
-    do {                                                                                        \
-        cl_int err_ = (err);                                                                    \
-        if (err_ != CL_SUCCESS) {                                                               \
-            fprintf(stderr, "OpenCL %s error %d at %s:%d\n", name, err_, __FILE__, __LINE__);   \
-            exit(1);                                                                            \
-        }                                                                                       \
-    } while (0)
-
-static cl_platform_id platform;
-static cl_device_id device;
-static cl_context context;
-static cl_command_queue queue;
-static cl_program program;
-static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q5_0, kernel_q5_1, kernel_q8_0;
-static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
-static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
-
-static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
-    cl_program p;
-    char *program_log;
-    size_t program_size, log_size;
-    int err;
-
-    program_size = strlen(program_buffer);
-
-    p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
-    if(err < 0) {
-        fprintf(stderr, "OpenCL error creating program");
-        exit(1);
-    }
-
-    err = clBuildProgram(p, 0, NULL, NULL, NULL, NULL);
-    if(err < 0) {
-
-        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
-        program_log = (char*) malloc(log_size + 1);
-        program_log[log_size] = '\0';
-        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
-        printf("%s\n", program_log);
-        free(program_log);
-        exit(1);
-    }
-
-    return p;
-}
-
-void ggml_cl_init(void) {
-    cl_int err = 0;
-    char * GGML_CLBLAST_PLATFORM = getenv("GGML_CLBLAST_PLATFORM");
-    char * GGML_CLBLAST_DEVICE = getenv("GGML_CLBLAST_DEVICE");
-    int plat_num = (GGML_CLBLAST_PLATFORM == NULL ? 0 : atoi(GGML_CLBLAST_PLATFORM));
-    int dev_num = (GGML_CLBLAST_DEVICE == NULL ? 0 : atoi(GGML_CLBLAST_DEVICE));
-    printf("\nInitializing CLBlast (First Run)...");
-    printf("\nAttempting to use: Platform=%d, Device=%d (If invalid, program will crash)\n",plat_num,dev_num);
-    cl_uint num_platforms;
-    clGetPlatformIDs(0, NULL, &num_platforms);
-    cl_platform_id* platforms = (cl_platform_id*)malloc(num_platforms*sizeof(cl_platform_id));
-    clGetPlatformIDs(num_platforms, platforms, NULL);
-    platform = platforms[plat_num];
-    char platform_buffer[1024];
-    clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(platform_buffer), &platform_buffer, NULL);
-    cl_uint num_devices;
-    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 0, NULL, &num_devices);
-    cl_device_id* devices = (cl_device_id*)malloc(num_devices*sizeof(cl_device_id));
-    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, num_devices, devices, NULL);
-    device = devices[dev_num];
-    char device_buffer[1024];
-    clGetDeviceInfo(device, CL_DEVICE_NAME, sizeof(device_buffer), &device_buffer, NULL);
-    printf("Using Platform: %s Device: %s\n", platform_buffer, device_buffer);
-    context = clCreateContext(NULL, 1, &device, NULL, NULL, &err);
-    CL_CHECK(err, "clCreateContext");
-    queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err);
-    CL_CHECK(err, "clCreateCommandQueue");
-
-    free(platforms);
-    free(devices);
-
-    program = build_program_from_source(context, device, clblast_dequant);
-
-    // Prepare dequantize kernels
-    kernel_q4_0 = clCreateKernel(program, "dequantize_row_q4_0", &err);
-    CL_CHECK(err, "clCreateKernel");
-    kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err);
-    CL_CHECK(err, "clCreateKernel");
-    kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err);
-    CL_CHECK(err, "clCreateKernel");
-    kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err);
-    CL_CHECK(err, "clCreateKernel");
-    kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err);
-    CL_CHECK(err, "clCreateKernel");
-}
-
-static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
-    if (req_size <= *cur_size) {
-        return;
-    }
-
-    // Reallocate buffer with enough space
-    if (*cur_size > 0) {
-        clReleaseMemObject(*buf);
-    }
-    cl_int err;
-    *buf = clCreateBuffer(context, flags, req_size, NULL, &err);
-    *cur_size = req_size;
-    CL_CHECK(err, "clCreateBuffer");
-}
-
-void ggml_cl_sgemm_wrapper(
-        const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b,
-        const int m, const int n, const int k,
-        const float alpha, const void *host_a, const int lda,
-        const float *host_b, const int ldb, const float beta,
-        float *host_c, const int ldc, const int btype) {
-    cl_int err = 0;
-
-    cl_kernel kernel;
-    size_t global = n * k, local, size_qb;
-    bool dequant;
-
-    switch (btype) {
-    case GGML_TYPE_F32:
-        dequant = false;
-        break;
-    case GGML_TYPE_Q4_0:
-        dequant = true;
-        kernel = kernel_q4_0;
-        local = 16;
-        size_qb = global * (sizeof(float) + local) / 32;
-        break;
-    case GGML_TYPE_Q4_1:
-        dequant = true;
-        kernel = kernel_q4_1;
-        local = 16;
-        size_qb = global * (sizeof(float) * 2 + local) / 32;
-        break;
-    case GGML_TYPE_Q5_0:
-        dequant = true;
-        kernel = kernel_q5_0;
-        local = 16;
-        size_qb = global * (sizeof(ggml_fp16_t) + sizeof(uint32_t) + local) / 32;
-        break;
-    case GGML_TYPE_Q5_1:
-        dequant = true;
-        kernel = kernel_q5_1;
-        local = 16;
-        size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
-        break;
-    case GGML_TYPE_Q8_0:
-        dequant = true;
-        kernel = kernel_q8_0;
-        local = 32;
-        size_qb = global * (sizeof(float) + local) / 32;
-        break;
-    default:
-        fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
-        abort();
-    }
-
-    const size_t size_a =  m * k * sizeof(float);
-    const size_t size_b =  n * k * sizeof(float);
-    const size_t size_c =  m * n * sizeof(float);
-
-    // Prepare buffers
-    ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a);
-    if (dequant) {
-        ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb);
-    }
-    ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b);
-    ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c);
-
-    cl_event ev_a, ev_qb, ev_b;
-
-    if (dequant) {
-        err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb);
-        err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b);
-        CL_CHECK(err, "clSetKernelArg");
-        err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
-        CL_CHECK(err, "clEnqueueWriteBuffer qb");
-    } else {
-        err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
-        CL_CHECK(err, "clEnqueueWriteBuffer b");
-    }
-
-    err = clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
-    CL_CHECK(err, "clEnqueueWriteBuffer a");
-    if (dequant) {
-        err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b);
-        CL_CHECK(err, "clEnqueueNDRangeKernel");
-        clReleaseEvent(ev_qb);
-    }
-    clWaitForEvents(1, &ev_a);
-    clWaitForEvents(1, &ev_b);
-    clReleaseEvent(ev_a);
-    clReleaseEvent(ev_b);
-
-    cl_event ev_sgemm;
-    CLBlastStatusCode status = CLBlastSgemm((CLBlastLayout)order,
-                                            (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
-                                            m, n, k,
-                                            alpha,
-                                            cl_buffer_a, 0, lda,
-                                            cl_buffer_b, 0, ldb,
-                                            beta,
-                                            cl_buffer_c, 0, ldc,
-                                            &queue, &ev_sgemm);
-
-    if (status != CLBlastSuccess) {
-        fprintf(stderr, "Error: CLBlast SGEMM %d\n", status);
-        abort();
-    }
-
-    cl_event ev_c;
-    clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c);
-
-    // Wait for completion
-    clWaitForEvents(1, &ev_c);
-    clReleaseEvent(ev_sgemm);
-    clReleaseEvent(ev_c);
-}

ggml-opencl.h

@@ -1,23 +1,24 @@
 #pragma once
 
+#include "ggml.h"
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
 void ggml_cl_init(void);
 
-enum ggml_blas_order {
-    GGML_BLAS_ORDER_ROW_MAJOR = 101,
-    GGML_BLAS_ORDER_COLUMN_MAJOR = 102,
-};
+void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
 
-enum ggml_blas_op {
-    GGML_BLAS_OP_N = 111,
-    GGML_BLAS_OP_T = 112,
-    GGML_BLAS_OP_C = 113,
-};
+void * ggml_cl_host_malloc(size_t size);
+void   ggml_cl_host_free(void * ptr);
 
-void ggml_cl_sgemm_wrapper(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype);
+void ggml_cl_free_data(const struct ggml_tensor* tensor);
+
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus
 }

ggml.h

@@ -198,6 +198,7 @@
 #define GGML_MAX_PARAMS        256
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_OPT           4
+#define GGML_MAX_NAME          32
 #define GGML_DEFAULT_N_THREADS 4
 
 #define GGML_ASSERT(x) \
@@ -240,6 +241,13 @@ extern "C" {
         GGML_TYPE_Q5_1 = 7,
         GGML_TYPE_Q8_0 = 8,
         GGML_TYPE_Q8_1 = 9,
+        // k-quantizations
+        GGML_TYPE_Q2_K = 10,
+        GGML_TYPE_Q3_K = 11,
+        GGML_TYPE_Q4_K = 12,
+        GGML_TYPE_Q5_K = 13,
+        GGML_TYPE_Q6_K = 14,
+        GGML_TYPE_Q8_K = 15,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -248,7 +256,8 @@ extern "C" {
 
     enum ggml_backend {
         GGML_BACKEND_CPU = 0,
-        GGML_BACKEND_CUDA = 1,
+        GGML_BACKEND_GPU = 10,
+        GGML_BACKEND_GPU_SPLIT = 20,
     };
 
     // model file types
@@ -262,6 +271,11 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
     };
 
     // available tensor operations:
@@ -282,12 +296,14 @@ extern "C" {
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
+        GGML_OP_REPEAT_BACK,
         GGML_OP_ABS,
         GGML_OP_SGN,
         GGML_OP_NEG,
         GGML_OP_STEP,
         GGML_OP_RELU,
         GGML_OP_GELU,
+        GGML_OP_GELU_QUICK,
         GGML_OP_SILU,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
@@ -295,6 +311,7 @@ extern "C" {
         GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
         GGML_OP_SET,
@@ -310,18 +327,31 @@ extern "C" {
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
+        GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
-        GGML_OP_CONV_1D_1S,
-        GGML_OP_CONV_1D_2S,
+        GGML_OP_CLAMP,
+        GGML_OP_CONV_1D_S1_PH,
+        GGML_OP_CONV_1D_S2_PH,
+        GGML_OP_CONV_2D_SK_P0,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
+        GGML_OP_FLASH_ATTN_BACK,
+        GGML_OP_WIN_PART,
+        GGML_OP_WIN_UNPART,
 
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_MAP_CUSTOM1,
+        GGML_OP_MAP_CUSTOM2,
+        GGML_OP_MAP_CUSTOM3,
+
+        GGML_OP_CROSS_ENTROPY_LOSS,
+        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+
         GGML_OP_COUNT,
     };
 
@@ -370,11 +400,15 @@ extern "C" {
 
         void * data;
 
-        char name[32];
+        char name[GGML_MAX_NAME];
 
-        char padding[16];
+        void * extra; // extra things e.g. for ggml-cuda.cu
+
+        char padding[4];
     };
 
+    static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
+
     // computation graph
     struct ggml_cgraph {
         int n_nodes;
@@ -408,6 +442,25 @@ extern "C" {
         bool   no_alloc;   // don't allocate memory for the tensor data
     };
 
+
+    // compute types
+    enum ggml_task_type {
+        GGML_TASK_INIT = 0,
+        GGML_TASK_COMPUTE,
+        GGML_TASK_FINALIZE,
+    };
+
+    struct ggml_compute_params {
+        enum ggml_task_type type;
+
+        // ith = thread index, nth = number of threads
+        int ith, nth;
+
+        // work buffer for all threads
+        size_t wsize;
+        void * wdata;
+    };
+
     // misc
 
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
@@ -419,14 +472,17 @@ extern "C" {
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
-    GGML_API int64_t ggml_nelements(const struct ggml_tensor * tensor);
-    GGML_API size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nelements   (const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nrows       (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
 
     GGML_API int     ggml_blck_size (enum ggml_type type);
     GGML_API size_t  ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
     GGML_API float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
+    GGML_API const char * ggml_op_name  (enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -435,14 +491,26 @@ extern "C" {
     // TODO: temporary until model loading of ggml examples is refactored
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
+    GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
+
+    // use this to compute the memory overhead of a tensor
+    GGML_API size_t ggml_tensor_overhead(void);
+
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
-    GGML_API void    ggml_free(struct ggml_context * ctx);
+    GGML_API void                  ggml_free(struct ggml_context * ctx);
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
-    GGML_API size_t  ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
+
+    GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_mem_size       (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_max_tensor_size(const struct ggml_context * ctx);
 
     GGML_API struct ggml_tensor * ggml_new_tensor(
             struct ggml_context * ctx,
@@ -482,6 +550,8 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
 
+    GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
+
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
     GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
@@ -495,8 +565,9 @@ extern "C" {
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API const char * ggml_get_name(const struct ggml_tensor * tensor);
-    GGML_API void         ggml_set_name(struct ggml_tensor * tensor, const char * name);
+    GGML_API const char *         ggml_get_name(const struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name);
+    GGML_API struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
     // operations on tensors with backpropagation
@@ -521,6 +592,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -544,24 +620,47 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_sub_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_mul(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_mul_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_div(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_div_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_sqr(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sqr_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_sqrt(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sqrt_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_log(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -592,35 +691,76 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_repeat_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_abs_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_sgn(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sgn_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_neg(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_neg_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_step(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_step_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_relu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_relu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // TODO: double-check this computation is correct
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_gelu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_silu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_silu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // a - x
     // b - dy
     GGML_API struct ggml_tensor * ggml_silu_back(
@@ -634,10 +774,18 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // a - x
     // b - dy
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
@@ -645,14 +793,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // A: m rows, n columns
-    // B: p rows, n columns (i.e. we transpose it internally)
+    // A: n columns, m rows
+    // B: n columns, p rows  (i.e. we transpose it internally)
     // result is m columns, p rows
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // A: m columns, n rows,
+    // B: p columns, n rows,
+    // result is m columns, p rows
+    GGML_API struct ggml_tensor * ggml_out_prod(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     //
     // operations on tensors without backpropagation
     //
@@ -849,7 +1005,7 @@ extern "C" {
             int                   n_past);
 
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * gml_diag_mask_zero_inplace(
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past);
@@ -863,6 +1019,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_soft_max_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
@@ -897,18 +1064,66 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
-            int                   n_head);
+            int                   n_head,
+            float                 bias_max);
 
-    // padding = 1
+    // clamp
+    // in-place, returns view(a)
+    struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
+    // TODO: implement general-purpose convolutions
+    // GGML_API struct ggml_tensor * ggml_conv_1d(
+    //        struct ggml_context * ctx,
+    //        struct ggml_tensor  * a,
+    //        struct ggml_tensor  * b,
+    //        int                   s0
+    //        int                   p0,
+    //        int                   d0);
+    //
+    // GGML_API struct ggml_tensor * ggml_conv_2d(
+    //        struct ggml_context * ctx,
+    //        struct ggml_tensor  * a,
+    //        struct ggml_tensor  * b,
+    //        int                   s0,
+    //        int                   s1,
+    //        int                   p0,
+    //        int                   p1,
+    //        int                   d0,
+    //        int                   d1);
+
+    // padding = half
     // TODO: we don't support extra parameters for now
     //       that's why we are hard-coding the stride, padding, and dilation
     //       not great ..
-    GGML_API struct ggml_tensor * ggml_conv_1d_1s(
+    // example:
+    // a:      3   80  768    1
+    // b:   3000   80    1    1
+    // res: 3000  768    1    1
+    // used in whisper
+    GGML_API struct ggml_tensor * ggml_conv_1d_s1_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    GGML_API struct ggml_tensor * ggml_conv_1d_2s(
+    // used in whisper
+    GGML_API struct ggml_tensor * ggml_conv_1d_s2_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // kernel size is a->ne[0] x a->ne[1]
+    // stride is equal to kernel size
+    // padding is zero
+    // example:
+    // a:     16   16    3  768
+    // b:   1024 1024    3    1
+    // res:   64   64  768    1
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
@@ -920,6 +1135,14 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+    GGML_API struct ggml_tensor * ggml_flash_attn_back(
+           struct ggml_context * ctx,
+           struct ggml_tensor  * q,
+           struct ggml_tensor  * k,
+           struct ggml_tensor  * v,
+           struct ggml_tensor  * d,
+           bool                  masked);
+
     GGML_API struct ggml_tensor * ggml_flash_ff(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -928,21 +1151,106 @@ extern "C" {
             struct ggml_tensor  * c0,
             struct ggml_tensor  * c1);
 
-    // Mapping operations
-    typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
+    // partition into non-overlapping windows with padding if needed
+    // example:
+    // a:   768   64   64    1
+    // w:    14
+    // res: 768   14   14    25
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_win_part(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   w);
+
+    // reverse of ggml_win_part
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_win_unpart(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   w0,
+            int                   h0,
+            int                   w);
+
+    // custom operators
+
+    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
     typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
 
+    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
                    ggml_unary_op_f32_t   fun);
 
+    GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+            struct ggml_context        * ctx,
+            struct ggml_tensor         * a,
+                   ggml_unary_op_f32_t   fun);
+
     GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
                    ggml_binary_op_f32_t   fun);
 
+    GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+                   ggml_binary_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+                   ggml_custom1_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+                   ggml_custom1_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+                   ggml_custom2_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+                   ggml_custom2_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+            struct ggml_tensor           * c,
+                   ggml_custom3_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+            struct ggml_tensor           * c,
+                   ggml_custom3_op_f32_t   fun);
+
+    // loss function
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b);
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+            struct ggml_tensor          * c);
+
     //
     // automatic differentiation
     //
@@ -959,6 +1267,11 @@ extern "C" {
     GGML_API void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph);
     GGML_API void ggml_graph_reset  (struct ggml_cgraph * cgraph);
 
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+
+    GGML_API void               ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
+    GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
+
     // print info and performance information for the graph
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 
@@ -1032,6 +1345,8 @@ extern "C" {
         struct {
             int n_iter;
 
+            float sched; // schedule multiplier (fixed, decay or warmup)
+            float decay; // weight decay for AdamW, use 0.0f to disable
             float alpha; // learning rate
             float beta1;
             float beta2;
@@ -1056,6 +1371,49 @@ extern "C" {
         } lbfgs;
     };
 
+    struct ggml_opt_context {
+        struct ggml_context * ctx;
+        struct ggml_opt_params params;
+
+        int iter;
+        int64_t nx; // number of parameter elements
+
+        bool just_initialized;
+
+        struct {
+            struct ggml_tensor * x;  // view of the parameters
+            struct ggml_tensor * g1; // gradient
+            struct ggml_tensor * g2; // gradient squared
+            struct ggml_tensor * m;  // first moment
+            struct ggml_tensor * v;  // second moment
+            struct ggml_tensor * mh; // first moment hat
+            struct ggml_tensor * vh; // second moment hat
+            struct ggml_tensor * pf; // past function values
+            float fx_best;
+            float fx_prev;
+            int n_no_improvement;
+        } adam;
+
+        struct {
+            struct ggml_tensor * x;    // current parameters
+            struct ggml_tensor * xp;   // previous parameters
+            struct ggml_tensor * g;    // current gradient
+            struct ggml_tensor * gp;   // previous gradient
+            struct ggml_tensor * d;    // search direction
+            struct ggml_tensor * pf;   // past function values
+            struct ggml_tensor * lmal; // the L-BFGS memory alpha
+            struct ggml_tensor * lmys; // the L-BFGS memory ys
+            struct ggml_tensor * lms;  // the L-BFGS memory s
+            struct ggml_tensor * lmy;  // the L-BFGS memory y
+            float fx_best;
+            float step;
+            int j;
+            int k;
+            int end;
+            int n_no_improvement;
+        } lbfgs;
+    };
+
     GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
 
     // optimize the function defined by the tensor f
@@ -1064,6 +1422,27 @@ extern "C" {
             struct ggml_opt_params params,
             struct ggml_tensor * f);
 
+    // initialize optimizer context
+    GGML_API void ggml_opt_init(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_opt_params params,
+            int64_t nx);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume_g(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f,
+            struct ggml_cgraph * gf,
+            struct ggml_cgraph * gb);
+
     //
     // quantization
     //

k_quants.h

@@ -0,0 +1,122 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <stdint.h>
+#include <assert.h>
+#include <stddef.h>
+
+// Super-block size
+#define QK_K 256
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elemenets each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 3.4375 bits per weight
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + 11 * QK_K / 64, "wrong q3_K block size/padding");
+
+// 4-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
+
+// 5-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+// Quantization with histogram collection
+size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+

llama-util.h

@@ -53,6 +53,7 @@ __attribute__((format(gnu_printf, 1, 2)))
 __attribute__((format(printf, 1, 2)))
 #endif
 #endif
+
 static std::string format(const char * fmt, ...) {
     va_list ap, ap2;
     va_start(ap, fmt);
@@ -101,12 +102,12 @@ struct llama_file {
         LLAMA_ASSERT(ret == 0); // same
     }
 
-    void read_raw(void * ptr, size_t size) {
-        if (size == 0) {
+    void read_raw(void * ptr, size_t len) const {
+        if (len == 0) {
             return;
         }
         errno = 0;
-        std::size_t ret = std::fread(ptr, size, 1, fp);
+        std::size_t ret = std::fread(ptr, len, 1, fp);
         if (ferror(fp)) {
             throw std::runtime_error(format("read error: %s", strerror(errno)));
         }
@@ -127,12 +128,12 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
+    void write_raw(const void * ptr, size_t len) const {
+        if (len == 0) {
             return;
         }
         errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
+        size_t ret = std::fwrite(ptr, len, 1, fp);
         if (ret != 1) {
             throw std::runtime_error(format("write error: %s", strerror(errno)));
         }
@@ -172,7 +173,7 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true) {
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
@@ -184,9 +185,9 @@ struct llama_mmap {
             throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
         }
 
-        if (prefetch) {
+        if (prefetch > 0) {
             // Advise the kernel to preload the mapped memory
-            if (madvise(addr, file->size, MADV_WILLNEED)) {
+            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
                 fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
                         strerror(errno));
             }
@@ -267,9 +268,9 @@ struct llama_mlock {
         }
     }
 
-    void init(void * addr) {
-        LLAMA_ASSERT(this->addr == NULL && this->size == 0);
-        this->addr = addr;
+    void init(void * ptr) {
+        LLAMA_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
     }
 
     void grow_to(size_t target_size) {
@@ -340,14 +341,14 @@ struct llama_mlock {
         return (size_t) si.dwPageSize;
     }
 
-    bool raw_lock(void * addr, size_t size) {
+    bool raw_lock(void * ptr, size_t len) {
         for (int tries = 1; ; tries++) {
-            if (VirtualLock(addr, size)) {
+            if (VirtualLock(ptr, len)) {
                 return true;
             }
             if (tries == 2) {
                 fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
-                        size, this->size, llama_format_win_err(GetLastError()).c_str());
+                    len, size, llama_format_win_err(GetLastError()).c_str());
                 return false;
             }
 
@@ -363,7 +364,7 @@ struct llama_mlock {
             // is equal to the number of pages in its minimum working set minus
             // a small overhead."
             // Hopefully a megabyte is enough overhead:
-            size_t increment = size + 1048576;
+            size_t increment = len + 1048576;
             // The minimum must be <= the maximum, so we need to increase both:
             min_ws_size += increment;
             max_ws_size += increment;
@@ -375,8 +376,8 @@ struct llama_mlock {
         }
     }
 
-    void raw_unlock(void * addr, size_t size) {
-        if (!VirtualUnlock(addr, size)) {
+    void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
             fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
                     llama_format_win_err(GetLastError()).c_str());
         }
@@ -388,12 +389,12 @@ struct llama_mlock {
         return (size_t) 65536;
     }
 
-    bool raw_lock(const void * addr, size_t size) {
+    bool raw_lock(const void * addr, size_t len) {
         fprintf(stderr, "warning: mlock not supported on this system\n");
         return false;
     }
 
-    void raw_unlock(const void * addr, size_t size) {}
+    void raw_unlock(const void * addr, size_t len) {}
 #endif
 };
 
@@ -404,14 +405,30 @@ struct llama_buffer {
 
     llama_buffer() = default;
 
-    void resize(size_t size) {
+    void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
         delete[] addr;
-        addr = new uint8_t[size];
-        this->size = size;
+        addr = new uint8_t[len];
+#endif
+        size = len;
     }
 
     ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
         delete[] addr;
+#endif
+        addr = NULL;
     }
 
     // disable copy and move

llama.h

@@ -1,6 +1,13 @@
 #ifndef LLAMA_H
 #define LLAMA_H
 
+#include "ggml.h"
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#else
+#define LLAMA_MAX_DEVICES 1
+#endif // GGML_USE_CUBLAS
 #include <stddef.h>
 #include <stdint.h>
 #include <stdbool.h>
@@ -19,12 +26,31 @@
 #    define LLAMA_API
 #endif
 
+#ifdef __GNUC__
+#    define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define DEPRECATED(func, hint) func
+#endif
+
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
+#define LLAMA_FILE_MAGIC_GGMF        0x67676d66u // 'ggmf'
+#define LLAMA_FILE_MAGIC_GGML        0x67676d6cu // 'ggml'
+#define LLAMA_FILE_MAGIC_GGSN        0x6767736eu // 'ggsn'
+
 #define LLAMA_FILE_VERSION           3
-#define LLAMA_FILE_MAGIC             'ggjt'
-#define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
-#define LLAMA_SESSION_MAGIC          'ggsn'
+#define LLAMA_FILE_MAGIC             LLAMA_FILE_MAGIC_GGJT
+#define LLAMA_FILE_MAGIC_UNVERSIONED LLAMA_FILE_MAGIC_GGML
+#define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+#define LLAMA_SUPPORTS_GPU_OFFLOAD
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -35,14 +61,15 @@ extern "C" {
     // TODO: show sample usage
     //
 
+    struct llama_model;
     struct llama_context;
 
     typedef int llama_token;
 
     typedef struct llama_token_data {
-        llama_token id;  // token id
-        float logit; // log-odds of the token
-        float p;     // probability of the token
+        llama_token id; // token id
+        float logit;    // log-odds of the token
+        float p;        // probability of the token
     } llama_token_data;
 
     typedef struct llama_token_data_array {
@@ -53,61 +80,97 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
-    struct llama_context_params {
-        int n_ctx;        // text context
-        int n_gpu_layers; // number of layers to store in VRAM
-        int seed;         // RNG seed, -1 for random
+   struct llama_context_params {
+        int seed;                              // RNG seed, -1 for random
+        int n_ctx;                             // text context
+        int n_batch;                           // prompt processing batch size
+        int n_gpu_layers;                      // number of layers to store in VRAM
+        int main_gpu;                          // the GPU that is used for scratch and small tensors
+        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
 
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
         bool f16_kv;     // use fp16 for KV cache
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
         bool use_mlock;  // force system to keep model in RAM
         bool embedding;  // embedding mode only
-
-        // called with a progress value between 0 and 1, pass NULL to disable
-        llama_progress_callback progress_callback;
-        // context pointer passed to the progress callback
-        void * progress_callback_user_data;
     };
-
     // model file types
     enum llama_ftype {
-        LLAMA_FTYPE_ALL_F32     = 0,
-        LLAMA_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        LLAMA_FTYPE_ALL_F32              = 0,
+        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
     };
 
+    // model quantization parameters
+    typedef struct llama_model_quantize_params {
+        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype   ftype;    // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+    } llama_model_quantize_params;
+
     LLAMA_API struct llama_context_params llama_context_default_params();
+    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
 
     LLAMA_API bool llama_mmap_supported();
     LLAMA_API bool llama_mlock_supported();
 
+    // TODO: not great API - very likely to change
+    // Initialize the llama + ggml backend
+    // Call once at the start of the program
+    LLAMA_API void llama_init_backend();
+
+    LLAMA_API int64_t llama_time_us();
+
+    LLAMA_API struct llama_model * llama_load_model_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    LLAMA_API void llama_free_model(struct llama_model * model);
+
+    LLAMA_API struct llama_context * llama_new_context_with_model(
+                     struct llama_model * model,
+            struct llama_context_params   params);
+
     // Various functions for loading a ggml llama model.
     // Allocate (almost) all memory needed for the model.
     // Return NULL on failure
-    LLAMA_API struct llama_context * llama_init_from_file(
+    LLAMA_API DEPRECATED(struct llama_context * llama_init_from_file(
                              const char * path_model,
-            struct llama_context_params   params);
+            struct llama_context_params   params),
+            "please use llama_load_model_from_file combined with llama_new_context_with_model instead");
 
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
-    // TODO: not great API - very likely to change
     // Returns 0 on success
-    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-      enum llama_ftype   ftype,
-            int          nthread);
+            const llama_model_quantize_params * params);
 
     // Apply a LoRA adapter to a loaded model
     // path_base_model is the path to a higher quality model to use as a base for
@@ -115,8 +178,15 @@ extern "C" {
     // The model needs to be reloaded before applying a new adapter, otherwise the adapter
     // will be applied on top of the previous one
     // Returns 0 on success
-    LLAMA_API int llama_apply_lora_from_file(
+    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
             struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads),
+            "please use llama_model_apply_lora_from_file instead");
+
+    LLAMA_API int llama_model_apply_lora_from_file(
+            const struct llama_model * model,
                       const char * path_lora,
                       const char * path_base_model,
                              int   n_threads);
@@ -155,6 +225,12 @@ extern "C" {
                              int   n_past,
                              int   n_threads);
 
+    // Export a static computation graph for context of 511 and batch size of 1
+    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
+    //       parameters here to keep things simple
+    // IMPORTANT: do not use for anything else other than debugging and testing!
+    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+
     // Convert the provided text into tokens.
     // The tokens pointer must be large enough to hold the resulting tokens.
     // Returns the number of tokens on success, no more than n_max_tokens
@@ -171,6 +247,14 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
     // Can be mutated in order to change the probabilities of the next token
@@ -186,9 +270,9 @@ extern "C" {
     LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos();
-    LLAMA_API llama_token llama_token_eos();
-    LLAMA_API llama_token llama_token_nl();
+    LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos();  // end-of-sentence
+    LLAMA_API llama_token llama_token_nl();   // next-line
 
     // Sampling functions
 
@@ -253,7 +337,7 @@ extern "C" {
 #include <string>
 struct ggml_tensor;
 
-std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
 
 #endif
 

pocs/vdot/vdot.cpp

@@ -10,6 +10,10 @@
 
 #include <ggml.h>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 constexpr int kVecSize = 1 << 18;
 
 float drawFromGaussianPdf(std::mt19937& rndm) {

scripts/verify-checksum-models.py

@@ -1,9 +1,10 @@
 import os
 import hashlib
 
+
 def sha256sum(file):
     block_size = 16 * 1024 * 1024  # 16 MB block size
-    b  = bytearray(block_size)
+    b = bytearray(block_size)
     file_hash = hashlib.sha256()
     mv = memoryview(b)
     with open(file, 'rb', buffering=0) as f:
@@ -15,6 +16,7 @@ def sha256sum(file):
 
     return file_hash.hexdigest()
 
+
 # Define the path to the llama directory (parent folder of script directory)
 llama_path = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir))
 

spm-headers/ggml.h

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

tests/test-grad0.c

@@ -1,3 +1,4 @@
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
 #include "ggml.h"
 
 #include <math.h>
@@ -5,7 +6,11 @@
 #include <stdlib.h>
 #include <assert.h>
 
-#define MAX_NARGS 2
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define MAX_NARGS 3
 
 #undef MIN
 #undef MAX
@@ -197,8 +202,23 @@ bool check_gradient(
         float max_error_abs,
         float max_error_rel) {
 
+    static int n_threads = -1;
+    if (n_threads < 0) {
+        n_threads = GGML_DEFAULT_N_THREADS;
+
+        const char *env = getenv("GGML_N_THREADS");
+        if (env) {
+            n_threads = atoi(env);
+        }
+
+        printf("GGML_N_THREADS = %d\n", n_threads);
+    }
+
     struct ggml_cgraph gf = ggml_build_forward (f);
+    gf.n_threads = n_threads;
+
     struct ggml_cgraph gb = ggml_build_backward(ctx0, &gf, false);
+    gb.n_threads = n_threads;
 
     ggml_graph_compute(ctx0, &gf);
     ggml_graph_reset  (&gf);
@@ -1090,6 +1110,25 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // cross_entropy_loss
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            for (int ndims = 1; ndims <= 3; ++ndims) {
+                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
+
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
+                // finite differences regularly fails!
+            }
+        }
+
         // rope
         {
             const int nargs = 1;
@@ -1124,6 +1163,45 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // flash_attn
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
         ggml_free(ctx0);
     }
 

tests/test-quantize-fns.cpp

@@ -9,10 +9,15 @@
 #include <string>
 #include <vector>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
 
-const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
-const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
-const float MAX_DOT_PRODUCT_ERROR = 0.02;
+const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
+const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
+const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
+const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
+const float MAX_DOT_PRODUCT_ERROR = 0.02f;
 
 const char* RESULT_STR[] = {"ok", "FAILED"};
 
@@ -122,7 +127,10 @@ int main(int argc, char * argv[]) {
 
         if (qfns.quantize_row_q && qfns.dequantize_row_q) {
             const float total_error = total_quantization_error(qfns, test_size, test_data.data());
-            failed = !(total_error < MAX_QUANTIZATION_TOTAL_ERROR);
+            const float max_quantization_error =
+                type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
+                type == GGML_TYPE_Q3_K ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS : MAX_QUANTIZATION_TOTAL_ERROR;
+            failed = !(total_error < max_quantization_error);
             num_failed += failed;
             if (failed || verbose) {
                 printf("%5s absolute quantization error:    %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);

tests/test-quantize-perf.cpp

@@ -13,6 +13,10 @@
 #include <string>
 #include <vector>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define MAX_ALIGNMENT 64
 #define QK 32
 #define WARMUP 5

tests/test-sampling.cpp

@@ -176,27 +176,28 @@ void test_frequency_presence_penalty(
 int main(void) {
     ggml_time_init();
 
-    test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4}, 1);
-    test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2}, 3);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 1);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 3);
 
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4}, 0);
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3}, 0.7);
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2, 0.1}, 1);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 0);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f}, 0.7f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
 
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3}, 0.25);
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.75);
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.99);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
 
-    test_typical({0.97, 0.01, 0.01, 0.01}, {0.97}, 0.5);
-    test_typical({0.4, 0.2, 0.2, 0.2}, {0.2, 0.2, 0.2}, 0.5);
+    test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
+    test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
 
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0}, {0.25, 0.25, 0.25, 0.25, 0}, 50.0);
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2}, {0.5, 0.5, 0, 0, 0}, 50.0);
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.5, 0.5, 0, 0, 0}, 50.0);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
 
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0},             {0.249997, 0.249997, 0.249997, 0.249997, 0.000011}, 5.0, 5.0);
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2},       {0.499966, 0.499966, 0.000023, 0.000023, 0.000023}, 5.0, 5.0);
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.499977, 0.499977, 0.000023, 0.000023, 0.000000}, 5.0, 5.0);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 5.0f, 5.0f);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 5.0f, 5.0f);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 5.0f, 5.0f);
 
     printf("OK\n");
 }

tests/test-tokenizer-0.cpp

@@ -28,6 +28,7 @@ int main(int argc, char **argv) {
 
     fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
 
+    llama_model * model;
     llama_context * ctx;
 
     // load the vocab
@@ -36,10 +37,18 @@ int main(int argc, char **argv) {
 
         lparams.vocab_only = true;
 
-        ctx = llama_init_from_file(fname.c_str(), lparams);
+        model = llama_load_model_from_file(fname.c_str(), lparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        ctx = llama_new_context_with_model(model, lparams);
 
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
             return 1;
         }
     }
@@ -48,12 +57,14 @@ int main(int argc, char **argv) {
 
     if (n_vocab != 32000) {
         fprintf(stderr, "%s : expected 32000 tokens, got %d\n", __func__, n_vocab);
+        llama_free_model(model);
+        llama_free(ctx);
         return 2;
     }
 
     for (const auto & test_kv : k_tests()) {
         std::vector<llama_token> res(test_kv.first.size());
-        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true);
+        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), int(res.size()), true);
         res.resize(n);
 
         bool correct = res.size() == test_kv.second.size();
@@ -77,10 +88,13 @@ int main(int argc, char **argv) {
             }
             fprintf(stderr, "\n");
 
+            llama_free_model(model);
+            llama_free(ctx);
             return 3;
         }
     }
 
+    llama_free_model(model);
     llama_free(ctx);
 
     return 0;