CUDA: Faster Mixtral prompt processing (#4538)

* CUDA: make MoE tensors contiguous for batch size>1

* Update ggml-cuda.cu

Co-authored-by: slaren <slarengh@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

.devops/tools.sh

@@ -13,6 +13,8 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     ./main "$@"
+elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
+    ./finetune "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -34,6 +36,8 @@ else
     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 "  --finetune (-f): Run finetune command to create a lora finetune of the model"
+    echo "              See documentation for finetune for command-line parameters"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
     echo "  --server (-s): Run a model on the server"

.editorconfig

@@ -15,8 +15,14 @@ indent_size = 4
 [Makefile]
 indent_style = tab
 
+[scripts/*.mk]
+indent_style = tab
+
 [prompts/*.txt]
 insert_final_newline = unset
 
 [examples/server/public/*]
 indent_size = 2
+
+[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
+indent_style = tab

.github/ISSUE_TEMPLATE/bug.md

@@ -1,8 +1,7 @@
 ---
-name: Issue and enhancement template
-about: Used to report issues and request enhancements for llama.cpp
-title: "[User] Insert summary of your issue or enhancement.."
-labels: ''
+name: Bug template
+about: Used to report bugs in llama.cpp
+labels: ["bug-unconfirmed"]
 assignees: ''
 
 ---
@@ -46,7 +45,7 @@ $ g++ --version
 
 # Failure Information (for bugs)
 
-Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template.
+Please help provide information about the failure / bug.
 
 # Steps to Reproduce
 

.github/ISSUE_TEMPLATE/enhancement.md

@@ -0,0 +1,28 @@
+---
+name: Enhancement template
+about: Used to request enhancements for llama.cpp
+labels: ["enhancement"]
+assignees: ''
+
+---
+
+# Prerequisites
+
+Please answer the following questions for yourself before submitting an issue.
+
+- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
+- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
+- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
+
+# Feature Description
+
+Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+
+# Motivation
+
+Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+
+# Possible Implementation
+
+If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

.github/workflows/build.yml

@@ -143,6 +143,9 @@ jobs:
           cd build
           ctest --verbose
 
+  # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
   macOS-latest-make:
     runs-on: macos-latest
 
@@ -160,14 +163,18 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          make -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: make_test
         run: |
-          make tests -j $(sysctl -n hw.logicalcpu)
-          make test -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make test  -j $(sysctl -n hw.logicalcpu)
 
+  # TODO: build with LLAMA_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
+  #       would be great if we fix these
   macOS-latest-cmake:
     runs-on: macos-latest
 
@@ -188,7 +195,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake ..
+          cmake -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -288,6 +295,7 @@ jobs:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
+      SDE_VERSION: 9.21.1-2023-04-24
 
     strategy:
       matrix:
@@ -383,11 +391,23 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
+        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
         run: |
           cd build
           ctest -C Release --verbose --timeout 900
 
+      - name: Test (Intel SDE)
+        id: cmake_test_sde
+        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz"
+          # for some weird reason windows tar doesn't like sde tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
+          $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
+          cd build
+          & $sde -future -- ctest -C Release --verbose --timeout 900
+
       - name: Determine tag name
         id: tag
         shell: bash
@@ -485,6 +505,17 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
+  ios-xcode-build:
+    runs-on: macos-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      - name: Build Xcode project
+        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
+
+
 #  freeBSD-latest:
 #    runs-on: macos-12
 #    steps:

.github/workflows/python-lint.yml

@@ -0,0 +1,20 @@
+name: flake8 Lint
+
+on: [push, pull_request]
+
+jobs:
+  flake8-lint:
+    runs-on: ubuntu-latest
+    name: Lint
+    steps:
+      - name: Check out source repository
+        uses: actions/checkout@v3
+      - name: Set up Python environment
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.11"
+      - name: flake8 Lint
+        uses: py-actions/flake8@v2
+        with:
+            ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704"
+            exclude: "examples/*,examples/*/**,*/**/__init__.py"

.gitignore

@@ -10,10 +10,12 @@
 *.gcno
 *.gcda
 *.dot
+*.bat
 *.metallib
 .DS_Store
 .build/
 .cache/
+.ccls-cache/
 .direnv/
 .envrc
 .swiftpm
@@ -44,7 +46,8 @@ models-mnt
 /infill
 /libllama.so
 /llama-bench
-/llava
+/llava-cli
+/lookahead
 /main
 /metal
 /perplexity
@@ -62,8 +65,9 @@ models-mnt
 /speculative
 /parallel
 /train-text-from-scratch
+/tokenize
 /vdot
-build-info.h
+/common/build-info.cpp
 arm_neon.h
 compile_commands.json
 CMakeSettings.json
@@ -84,15 +88,17 @@ poetry.lock
 poetry.toml
 
 # Test binaries
-tests/test-grammar-parser
-tests/test-llama-grammar
-tests/test-double-float
-tests/test-grad0
-tests/test-opt
-tests/test-quantize-fns
-tests/test-quantize-perf
-tests/test-sampling
-tests/test-tokenizer-0-llama
-tests/test-tokenizer-0-falcon
-tests/test-tokenizer-1-llama
-tests/test-tokenizer-1-bpe
+/tests/test-grammar-parser
+/tests/test-llama-grammar
+/tests/test-double-float
+/tests/test-grad0
+/tests/test-opt
+/tests/test-quantize-fns
+/tests/test-quantize-perf
+/tests/test-sampling
+/tests/test-tokenizer-0-llama
+/tests/test-tokenizer-0-falcon
+/tests/test-tokenizer-1-llama
+/tests/test-tokenizer-1-bpe
+/tests/test-rope
+/tests/test-backend-ops

CMakeLists.txt

@@ -10,7 +10,7 @@ endif()
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
     # configure project version
@@ -43,6 +43,7 @@ else()
 endif()
 
 # general
+option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
 option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
@@ -82,6 +83,7 @@ set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
 #option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
+option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
@@ -93,46 +95,15 @@ option(LLAMA_CLBLAST                         "llama: use CLBlast"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
-option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
-option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)
+option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
-#
-# Build info header
-#
-
-# Generate initial build-info.h
+# Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-
-    # Is git submodule
-    if(NOT IS_DIRECTORY "${GIT_DIR}")
-        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
-        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
-        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
-    endif()
-
-    # Add a custom target for build-info.h
-    add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
-
-    # Add a custom command to rebuild build-info.h when .git/index changes
-    add_custom_command(
-        OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
-        COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        DEPENDS "${GIT_DIR}/index"
-        VERBATIM
-    )
-else()
-    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-endif()
-
 #
 # Compile flags
 #
@@ -145,6 +116,11 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+# enable libstdc++ assertions for debug builds
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
+endif()
+
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
@@ -194,7 +170,7 @@ if (LLAMA_METAL)
     #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)
+    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
@@ -277,13 +253,8 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_K_QUANTS)
-    set(GGML_HEADERS_EXTRA k_quants.h)
-    set(GGML_SOURCES_EXTRA k_quants.c)
-    add_compile_definitions(GGML_USE_K_QUANTS)
-    if (LLAMA_QKK_64)
-        add_compile_definitions(GGML_QKK_64)
-    endif()
+if (LLAMA_QKK_64)
+    add_compile_definitions(GGML_QKK_64)
 endif()
 
 if (LLAMA_CUBLAS)
@@ -305,6 +276,9 @@ if (LLAMA_CUBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+        endif()
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         if (DEFINED LLAMA_CUDA_DMMV_Y)
@@ -317,7 +291,12 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
 
         if (LLAMA_STATIC)
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            if (WIN32)
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            endif()
         else()
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
@@ -331,6 +310,7 @@ if (LLAMA_CUBLAS)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -404,6 +384,9 @@ if (LLAMA_HIPBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
         endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
+        endif()
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
@@ -419,57 +402,102 @@ if (LLAMA_HIPBLAS)
     endif()
 endif()
 
-if (LLAMA_ALL_WARNINGS)
-    if (NOT MSVC)
-        set(warning_flags -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration)
-        set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
-        set(host_cxx_flags "")
+function(get_flags CCID CCVER)
+    set(C_FLAGS "")
+    set(CXX_FLAGS "")
 
-        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
-            set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
-            set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
+    if (CCID MATCHES "Clang")
+        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
+        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
 
-            if (
-                (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
-                (CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
-            )
-                set(c_flags ${c_flags} -Wdouble-promotion)
-            endif()
-        elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
-            set(c_flags ${c_flags} -Wdouble-promotion)
-            set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
-
-            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
-                set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
-            endif()
-            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
-                set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
-            endif()
+        if (
+            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
+            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
+        )
+            set(C_FLAGS ${C_FLAGS} -Wdouble-promotion)
+        endif()
+    elseif (CCID STREQUAL "GNU")
+        set(C_FLAGS   -Wdouble-promotion)
+        set(CXX_FLAGS -Wno-array-bounds)
+
+        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
+            set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation)
+        endif()
+        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
+            set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi)
         endif()
-    else()
-        # todo : msvc
     endif()
 
-    set(c_flags   ${c_flags}   ${warning_flags})
-    set(cxx_flags ${cxx_flags} ${warning_flags})
-    add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
-                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
-                        "$<$<COMPILE_LANGUAGE:CXX>:${host_cxx_flags}>")
+    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
+    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
+endfunction()
 
+if (LLAMA_ALL_WARNINGS)
+    if (NOT MSVC)
+        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                          -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+
+        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+
+        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+
+        add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
+                            "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+    else()
+        # todo : msvc
+        set(C_FLAGS   "")
+        set(CXX_FLAGS "")
+    endif()
 endif()
 
-if (NOT MSVC)
-    set(cuda_flags -Wno-pedantic)
-endif()
-set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+if (LLAMA_CUBLAS)
+    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
+    if (NOT MSVC)
+        set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
+    endif()
 
-list(JOIN host_cxx_flags " " cuda_host_flags)  # pass host compiler flags as a single argument
-if (NOT cuda_host_flags STREQUAL "")
-    set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
-endif()
+    if (LLAMA_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
 
-add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS)  # pass host compiler flags as a single argument
+        if (NOT CUDA_CXX_FLAGS STREQUAL "")
+            set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS})
+        endif()
+    endif()
+
+    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+endif()
 
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
@@ -489,6 +517,16 @@ if (LLAMA_LTO)
     endif()
 endif()
 
+# this version of Apple ld64 is buggy
+execute_process(
+    COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
+    ERROR_VARIABLE output
+    OUTPUT_QUIET
+)
+if (output MATCHES "dyld-1015\.7")
+    add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
+endif()
+
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -541,6 +579,10 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
     message(STATUS "x86 detected")
     if (MSVC)
+        # instruction set detection for MSVC only
+        if (LLAMA_NATIVE)
+            include(cmake/FindSIMD.cmake)
+        endif ()
         if (LLAMA_AVX512)
             add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
             add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
@@ -592,12 +634,21 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
-    add_compile_options(-mcpu=native -mtune=native)
-    #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+        add_compile_options(-mcpu=powerpc64le)
+    else()
+        add_compile_options(-mcpu=native -mtune=native)
+        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    endif()
 else()
     message(STATUS "Unknown architecture")
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=0x602)
+endif()
+
 #
 # POSIX conformance
 #
@@ -665,11 +716,13 @@ add_library(ggml OBJECT
             ggml-alloc.h
             ggml-backend.c
             ggml-backend.h
-            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
+            ggml-quants.c
+            ggml-quants.h
+            ${GGML_SOURCES_CUDA}   ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_METAL}  ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}    ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}  ${GGML_HEADERS_EXTRA}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

Makefile

@@ -1,14 +1,15 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server embd-input-test gguf llama-bench llava baby-llama beam-search  \
-	speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
+	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
-	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
+	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
+	tests/test-backend-ops
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -25,20 +26,6 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
-ifeq '' '$(findstring clang,$(shell $(CC) --version))'
-	CC_IS_GCC=1
-	CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
-else
-	CC_IS_CLANG=1
-	ifeq '' '$(findstring Apple LLVM,$(shell $(CC) --version))'
-		CC_IS_LLVM_CLANG=1
-	else
-		CC_IS_APPLE_CLANG=1
-	endif
-	CC_VER := $(shell $(CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
-				| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
-endif
-
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
@@ -120,12 +107,12 @@ MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-MK_CFLAGS        += -Ofast
-MK_HOST_CXXFLAGS += -Ofast
-MK_CUDA_CXXFLAGS += -O3
+MK_CFLAGS     += -Ofast
+HOST_CXXFLAGS += -Ofast
+MK_NVCCFLAGS  += -O3
 else
-MK_CFLAGS        += -O3
-MK_CXXFLAGS      += -O3
+MK_CFLAGS     += -O3
+MK_CXXFLAGS   += -O3
 endif
 
 # clock_gettime came in POSIX.1b (1993)
@@ -174,6 +161,10 @@ ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
 	MK_LDFLAGS  += -g
+
+	ifeq ($(UNAME_S),Linux)
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
+	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
 endif
@@ -215,28 +206,9 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
-ifeq ($(CC_IS_CLANG), 1)
-	# clang options
-	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
-	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
-
-	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
-		MK_CFLAGS += -Wdouble-promotion
-	endif
-	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
-		MK_CFLAGS += -Wdouble-promotion
-	endif
-else
-	# gcc options
-	MK_CFLAGS        += -Wdouble-promotion
-	MK_HOST_CXXFLAGS += -Wno-array-bounds
-
-	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
-		MK_HOST_CXXFLAGS += -Wno-format-truncation
-	endif
-	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
-		MK_HOST_CXXFLAGS += -Wextra-semi
-	endif
+# this version of Apple ld64 is buggy
+ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
+	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
 endif
 
 # OS specific
@@ -284,8 +256,8 @@ ifndef RISCV
 
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
-	MK_CFLAGS   += -march=native -mtune=native
-	MK_HOST_CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS     += -march=native -mtune=native
+	HOST_CXXFLAGS += -march=native -mtune=native
 
 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
@@ -296,12 +268,15 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	#MK_CXXFLAGS += -mssse3
 endif
 
-# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
-# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-# https://github.com/ggerganov/llama.cpp/issues/2922
 ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
+	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+	# https://github.com/ggerganov/llama.cpp/issues/2922
 	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
 	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
+
+	# Target Windows 8 for PrefetchVirtualMemory
+	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
@@ -337,18 +312,20 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 	endif
 endif
 
+ifneq ($(filter ppc64le%,$(UNAME_M)),)
+	MK_CFLAGS   += -mcpu=powerpc64le
+	MK_CXXFLAGS += -mcpu=powerpc64le
+	CUDA_POWER_ARCH = 1
+endif
+
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifndef LLAMA_NO_K_QUANTS
-	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
-	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
 	MK_CPPFLAGS += -DGGML_QKK_64
 endif
-endif
 
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
@@ -365,7 +342,7 @@ ifdef LLAMA_MPI
 	MK_CPPFLAGS += -DGGML_USE_MPI
 	MK_CFLAGS   += -Wno-cast-qual
 	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS     += ggml-mpi.o
+	OBJS        += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
@@ -382,57 +359,65 @@ endif # LLAMA_BLIS
 ifdef LLAMA_CUBLAS
 	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
-	OBJS      += ggml-cuda.o
-	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+	OBJS         += ggml-cuda.o
+	MK_NVCCFLAGS  = --forward-unknown-to-host-compiler -use_fast_math
+
+ifdef LLAMA_DEBUG
+	MK_NVCCFLAGS += -lineinfo
+endif
+
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
 	NVCC = nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
-	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else
-	NVCCFLAGS += -arch=native
+	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else ifndef CUDA_POWER_ARCH
+	MK_NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
-	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_FORCE_MMQ
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
-	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
-	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
 ifdef LLAMA_CUDA_MMV_Y
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 else ifdef LLAMA_CUDA_DMMV_Y
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_F16
-	NVCCFLAGS += -DGGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	NVCCFLAGS += -DGGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
-	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
-	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
 ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
 else
-	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
-#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
-	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) -c $< -o $@
+	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -454,9 +439,15 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 endif # LLAMA_CLBLAST
 
 ifdef LLAMA_HIPBLAS
-	ROCM_PATH	?= /opt/rocm
-	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
-	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+
+	ifeq ($(wildcard /opt/rocm),)
+		ROCM_PATH	?= /usr
+		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
+	else
+		ROCM_PATH	?= /opt/rocm
+		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	endif
+	HIPCC                   ?= $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
@@ -494,21 +485,22 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifndef LLAMA_NO_K_QUANTS
-k_quants.o: k_quants.c k_quants.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_NO_K_QUANTS
+GF_CC := $(CC)
+include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
-override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
-override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
-override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
-override LDFLAGS       := $(MK_LDFLAGS) $(LDFLAGS)
+override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
+override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
+override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
-# save CXXFLAGS before we add host-only options
-NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
-override CXXFLAGS += $(HOST_CXXFLAGS)
+# identify CUDA host compiler
+ifdef LLAMA_CUBLAS
+GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
+include scripts/get-flags.mk
+CUDA_CXXFLAGS := $(GF_CXXFLAGS)
+endif
 
 #
 # Print build information
@@ -539,13 +531,16 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
 
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h
-COMMON_DEPS   = $(COMMON_H_DEPS) common.o sampling.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -566,54 +561,50 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tokenize: examples/tokenize/tokenize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.h ggml.o llama.o common.o $(OBJS)
+batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o 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.o 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_DEPS) $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o 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_DEPS) $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(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_DEPS) $(OBJS)
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
+save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
-
-
-embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -624,28 +615,34 @@ train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratc
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
+
+llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ifdef LLAMA_METAL
@@ -658,7 +655,7 @@ swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
 endif
 
-build-info.h: $(wildcard .git/index) scripts/build-info.sh
+common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
@@ -666,13 +663,16 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 		rm $@.tmp; \
 	fi
 
+build-info.o: common/build-info.cpp
+	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
+
 #
 # Tests
 #
 
 tests: $(TEST_TARGETS)
 
-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 run-benchmark-matmult: benchmark-matmult
@@ -686,41 +686,47 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-opt: tests/test-opt.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-c.o: tests/test-c.c llama.h
 	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
+
+tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

Package.swift

@@ -2,33 +2,14 @@
 
 import PackageDescription
 
-#if arch(arm) || arch(arm64)
-let platforms: [SupportedPlatform]? = [
-    .macOS(.v12),
-    .iOS(.v14),
-    .watchOS(.v4),
-    .tvOS(.v14)
-]
-let exclude: [String] = []
-let resources: [Resource] = [
-    .process("ggml-metal.metal")
-]
-let additionalSources: [String] = ["ggml-metal.m"]
-let additionalSettings: [CSetting] = [
-    .unsafeFlags(["-fno-objc-arc"]),
-    .define("GGML_USE_METAL")
-]
-#else
-let platforms: [SupportedPlatform]? = nil
-let exclude: [String] = ["ggml-metal.metal"]
-let resources: [Resource] = []
-let additionalSources: [String] = []
-let additionalSettings: [CSetting] = []
-#endif
-
 let package = Package(
     name: "llama",
-    platforms: platforms,
+    platforms: [
+        .macOS(.v12),
+        .iOS(.v14),
+        .watchOS(.v4),
+        .tvOS(.v14)
+    ],
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -36,26 +17,30 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: exclude,
+            exclude: [],
             sources: [
                 "ggml.c",
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
-                "k_quants.c",
-            ] + additionalSources,
-            resources: resources,
+                "ggml-quants.c",
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
+            ],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_K_QUANTS"),
-                .define("GGML_USE_ACCELERATE")
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL"),
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14
                 // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
                 // .define("ACCELERATE_NEW_LAPACK"),
                 // .define("ACCELERATE_LAPACK_ILP64")
-            ] + additionalSettings,
+            ],
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]

README.md

@@ -2,7 +2,6 @@
 
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
-[![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) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
@@ -10,13 +9,12 @@
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
-- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252)
-- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
-- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
-  **Devs should become familiar with the new API**
-- Local Falcon 180B inference on Mac Studio
 
-  https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
+- Collecting Apple Silicon performance stats:
+  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
+  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
+- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406
+- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
 
@@ -99,13 +97,25 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
+- [X] [StableLM-3b-4e1t](https://github.com/ggerganov/llama.cpp/pull/3586)
+- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
+- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
+- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
+
+**Multimodal models:**
+
+- [x] [Llava 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
+- [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
+- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
+- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
 
 
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
@@ -119,6 +129,8 @@ as the main playground for developing new features for the [ggml](https://github
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
 - [withcatai/catai](https://github.com/withcatai/catai)
+- [semperai/amica](https://github.com/semperai/amica)
+- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
 
 ---
 
@@ -325,7 +337,7 @@ mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
 
 ### 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:
+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). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
 
 - #### Accelerate Framework:
 
@@ -415,22 +427,31 @@ Building the program with BLAS support may lead to some performance improvements
   This provides BLAS acceleration on HIP-supported AMD GPUs.
   Make sure to have ROCm installed.
   You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
-  Windows support is coming soon...
 
   - Using `make`:
     ```bash
     make LLAMA_HIPBLAS=1
     ```
-  - Using `CMake`:
+  - Using `CMake` for Linux:
     ```bash
     mkdir build
     cd build
     CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
     cmake --build .
     ```
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS):
+    ```bash
+    set PATH=%HIP_PATH%\bin;%PATH%
+    mkdir build
+    cd build
+    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ..
+    cmake --build .
+    ```
+    Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+
 
   The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
-  If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
+  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
   The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
 
   | Option                  | Legal values           | Default | Description |
@@ -888,7 +909,7 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
-The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
 #### Usage
 
@@ -966,7 +987,6 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 
 - [main](./examples/main/README.md)
 - [server](./examples/server/README.md)
-- [embd-input](./examples/embd-input/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [BLIS](./docs/BLIS.md)
 - [Performance troubleshooting](./docs/token_generation_performance_tips.md)

build.zig

@@ -10,7 +10,6 @@ const Maker = struct {
     builder: *std.build.Builder,
     target: CrossTarget,
     optimize: Mode,
-    config_header: *ConfigHeader,
     enable_lto: bool,
 
     include_dirs: ArrayList([]const u8),
@@ -41,26 +40,24 @@ const Maker = struct {
         const commit_hash = try std.ChildProcess.exec(
             .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
         );
-        const config_header = builder.addConfigHeader(
-            .{ .style = .blank, .include_path = "build-info.h" },
-            .{
-                .BUILD_NUMBER = 0,
-                .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline
-                .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}),
-                .BUILD_TARGET = try target.allocDescription(builder.allocator),
-            },
-        );
+        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
+            \\int LLAMA_BUILD_NUMBER = {};
+            \\char const *LLAMA_COMMIT = "{s}";
+            \\char const *LLAMA_COMPILER = "Zig {s}";
+            \\char const *LLAMA_BUILD_TARGET = "{s}";
+            \\
+        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
         var m = Maker{
             .builder = builder,
             .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
-            .config_header = config_header,
             .enable_lto = false,
             .include_dirs = ArrayList([]const u8).init(builder.allocator),
             .cflags = ArrayList([]const u8).init(builder.allocator),
             .cxxflags = ArrayList([]const u8).init(builder.allocator),
             .objs = ArrayList(*Compile).init(builder.allocator),
         };
+
         try m.addCFlag("-std=c11");
         try m.addCxxFlag("-std=c++11");
         try m.addProjectInclude(&.{});
@@ -72,7 +69,7 @@ const Maker = struct {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
         if (o.target.getAbi() != .msvc)
             o.defineCMacro("_GNU_SOURCE", null);
-        o.addConfigHeader(m.config_header);
+
         if (std.mem.endsWith(u8, src, ".c")) {
             o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
@@ -85,7 +82,6 @@ const Maker = struct {
                 o.linkLibCpp();
             }
         }
-        o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
         o.want_lto = m.enable_lto;
         return o;
@@ -105,7 +101,6 @@ const Maker = struct {
             // linkLibCpp already add (libc++ + libunwind + libc)
             e.linkLibCpp();
         }
-        e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
         e.want_lto = m.enable_lto;
         return e;
@@ -116,30 +111,27 @@ pub fn build(b: *std.build.Builder) !void {
     var make = try Maker.init(b);
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
-    if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
-        try make.addFlag("-DGGML_USE_K_QUANTS");
-        const k_quants = make.obj("k_quants", "k_quants.c");
-        try make.objs.append(k_quants);
-    }
-
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
+    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const llama = make.obj("llama", "llama.cpp");
+    const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
+    const clip = make.obj("clip", "examples/llava/clip.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

cmake/FindSIMD.cmake

@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(LLAMA_AVX OFF)
+else()
+    set(LLAMA_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(LLAMA_AVX2 OFF)
+else()
+    set(LLAMA_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(LLAMA_AVX512 OFF)
+else()
+    set(LLAMA_AVX512 ON)
+endif()

common/CMakeLists.txt

@@ -1,8 +1,52 @@
 # common
 
+
+# Build info header
+#
+
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+
+    # Is git submodule
+    if(NOT IS_DIRECTORY "${GIT_DIR}")
+        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
+        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
+        string(FIND "${REAL_GIT_DIR}" "/" SLASH_POS)
+        if (SLASH_POS EQUAL 0)
+            set(GIT_DIR "${REAL_GIT_DIR}")
+        else()
+            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+        endif()
+    endif()
+
+    set(GIT_INDEX "${GIT_DIR}/index")
+else()
+    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
+endif()
+
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/gen-build-info-cpp.cmake"
+    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
+set(TARGET build_info)
+add_library(${TARGET} OBJECT build-info.cpp)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+
 set(TARGET common)
 
-add_library(${TARGET} OBJECT
+add_library(${TARGET} STATIC
+    base64.hpp
     common.h
     common.cpp
     sampling.h
@@ -21,4 +65,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama)
+target_link_libraries(${TARGET} PRIVATE llama build_info)

common/base64.hpp

@@ -0,0 +1,392 @@
+/*
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or
+distribute this software, either in source code form or as a compiled
+binary, for any purpose, commercial or non-commercial, and by any
+means.
+
+In jurisdictions that recognize copyright laws, the author or authors
+of this software dedicate any and all copyright interest in the
+software to the public domain. We make this dedication for the benefit
+of the public at large and to the detriment of our heirs and
+successors. We intend this dedication to be an overt act of
+relinquishment in perpetuity of all present and future rights to this
+software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <http://unlicense.org>
+*/
+
+#ifndef PUBLIC_DOMAIN_BASE64_HPP_
+#define PUBLIC_DOMAIN_BASE64_HPP_
+
+#include <cstdint>
+#include <iterator>
+#include <stdexcept>
+#include <string>
+
+class base64_error : public std::runtime_error
+{
+public:
+    using std::runtime_error::runtime_error;
+};
+
+class base64
+{
+public:
+    enum class alphabet
+    {
+        /** the alphabet is detected automatically */
+        auto_,
+        /** the standard base64 alphabet is used */
+        standard,
+        /** like `standard` except that the characters `+` and `/` are replaced by `-` and `_` respectively*/
+        url_filename_safe
+    };
+
+    enum class decoding_behavior
+    {
+        /** if the input is not padded, the remaining bits are ignored */
+        moderate,
+        /** if a padding character is encounter decoding is finished */
+        loose
+    };
+
+    /**
+     Encodes all the elements from `in_begin` to `in_end` to `out`.
+
+     @warning The source and destination cannot overlap. The destination must be able to hold at least
+     `required_encode_size(std::distance(in_begin, in_end))`, otherwise the behavior depends on the output iterator.
+
+     @tparam Input_iterator the source; the returned elements are cast to `std::uint8_t` and should not be greater than
+     8 bits
+     @tparam Output_iterator the destination; the elements written to it are from the type `char`
+     @param in_begin the beginning of the source
+     @param in_end the ending of the source
+     @param out the destination iterator
+     @param alphabet which alphabet should be used
+     @returns the iterator to the next element past the last element copied
+     @throws see `Input_iterator` and `Output_iterator`
+    */
+    template<typename Input_iterator, typename Output_iterator>
+    static Output_iterator encode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
+                                  alphabet alphabet = alphabet::standard)
+    {
+        constexpr auto pad = '=';
+        const char* alpha  = alphabet == alphabet::url_filename_safe
+                                ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
+                                : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+        while (in_begin != in_end) {
+            std::uint8_t i0 = 0, i1 = 0, i2 = 0;
+
+            // first character
+            i0 = static_cast<std::uint8_t>(*in_begin);
+            ++in_begin;
+
+            *out = alpha[i0 >> 2 & 0x3f];
+            ++out;
+
+            // part of first character and second
+            if (in_begin != in_end) {
+                i1 = static_cast<std::uint8_t>(*in_begin);
+                ++in_begin;
+
+                *out = alpha[((i0 & 0x3) << 4) | (i1 >> 4 & 0x0f)];
+                ++out;
+            } else {
+                *out = alpha[(i0 & 0x3) << 4];
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                break;
+            }
+
+            // part of second character and third
+            if (in_begin != in_end) {
+                i2 = static_cast<std::uint8_t>(*in_begin);
+                ++in_begin;
+
+                *out = alpha[((i1 & 0xf) << 2) | (i2 >> 6 & 0x03)];
+                ++out;
+            } else {
+                *out = alpha[(i1 & 0xf) << 2];
+                ++out;
+
+                // last padding
+                *out = pad;
+                ++out;
+
+                break;
+            }
+
+            // rest of third
+            *out = alpha[i2 & 0x3f];
+            ++out;
+        }
+
+        return out;
+    }
+    /**
+     Encodes a string.
+
+     @param str the string that should be encoded
+     @param alphabet which alphabet should be used
+     @returns the encoded base64 string
+     @throws see base64::encode()
+    */
+    static std::string encode(const std::string& str, alphabet alphabet = alphabet::standard)
+    {
+        std::string result;
+
+        result.reserve(required_encode_size(str.length()) + 1);
+
+        encode(str.begin(), str.end(), std::back_inserter(result), alphabet);
+
+        return result;
+    }
+    /**
+     Encodes a char array.
+
+     @param buffer the char array
+     @param size the size of the array
+     @param alphabet which alphabet should be used
+     @returns the encoded string
+    */
+    static std::string encode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::standard)
+    {
+        std::string result;
+
+        result.reserve(required_encode_size(size) + 1);
+
+        encode(buffer, buffer + size, std::back_inserter(result), alphabet);
+
+        return result;
+    }
+    /**
+     Decodes all the elements from `in_begin` to `in_end` to `out`. `in_begin` may point to the same location as `out`,
+     in other words: inplace decoding is possible.
+
+     @warning The destination must be able to hold at least `required_decode_size(std::distance(in_begin, in_end))`,
+     otherwise the behavior depends on the output iterator.
+
+     @tparam Input_iterator the source; the returned elements are cast to `char`
+     @tparam Output_iterator the destination; the elements written to it are from the type `std::uint8_t`
+     @param in_begin the beginning of the source
+     @param in_end the ending of the source
+     @param out the destination iterator
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the iterator to the next element past the last element copied
+     @throws base64_error depending on the set behavior
+     @throws see `Input_iterator` and `Output_iterator`
+    */
+    template<typename Input_iterator, typename Output_iterator>
+    static Output_iterator decode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
+                                  alphabet alphabet          = alphabet::auto_,
+                                  decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        //constexpr auto pad = '=';
+        std::uint8_t last  = 0;
+        auto bits          = 0;
+
+        while (in_begin != in_end) {
+            auto c = *in_begin;
+            ++in_begin;
+
+            if (c == '=') {
+                break;
+            }
+
+            auto part = _base64_value(alphabet, c);
+
+            // enough bits for one byte
+            if (bits + 6 >= 8) {
+                *out = (last << (8 - bits)) | (part >> (bits - 2));
+                ++out;
+
+                bits -= 2;
+            } else {
+                bits += 6;
+            }
+
+            last = part;
+        }
+
+        // check padding
+        if (behavior != decoding_behavior::loose) {
+            while (in_begin != in_end) {
+                auto c = *in_begin;
+                ++in_begin;
+
+                if (c != '=') {
+                    throw base64_error("invalid base64 character.");
+                }
+            }
+        }
+
+        return out;
+    }
+    /**
+     Decodes a string.
+
+     @param str the base64 encoded string
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the decoded string
+     @throws see base64::decode()
+    */
+    static std::string decode(const std::string& str, alphabet alphabet = alphabet::auto_,
+                              decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        std::string result;
+
+        result.reserve(max_decode_size(str.length()));
+
+        decode(str.begin(), str.end(), std::back_inserter(result), alphabet, behavior);
+
+        return result;
+    }
+    /**
+     Decodes a string.
+
+     @param buffer the base64 encoded buffer
+     @param size the size of the buffer
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the decoded string
+     @throws see base64::decode()
+    */
+    static std::string decode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::auto_,
+                              decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        std::string result;
+
+        result.reserve(max_decode_size(size));
+
+        decode(buffer, buffer + size, std::back_inserter(result), alphabet, behavior);
+
+        return result;
+    }
+    /**
+     Decodes a string inplace.
+
+     @param[in,out] str the base64 encoded string
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @throws base64::decode_inplace()
+    */
+    static void decode_inplace(std::string& str, alphabet alphabet = alphabet::auto_,
+                               decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        str.resize(decode(str.begin(), str.end(), str.begin(), alphabet, behavior) - str.begin());
+    }
+    /**
+     Decodes a char array inplace.
+
+     @param[in,out] str the string array
+     @param size the length of the array
+     @param alphabet which alphabet should be used
+     @param behavior the behavior when an error was detected
+     @returns the pointer to the next element past the last element decoded
+     @throws base64::decode_inplace()
+    */
+    static char* decode_inplace(char* str, std::size_t size, alphabet alphabet = alphabet::auto_,
+                                decoding_behavior behavior = decoding_behavior::moderate)
+    {
+        return decode(str, str + size, str, alphabet, behavior);
+    }
+    /**
+     Returns the required decoding size for a given size. The value is calculated with the following formula:
+
+     $$
+     \lceil \frac{size}{4} \rceil \cdot 3
+     $$
+
+     @param size the size of the encoded input
+     @returns the size of the resulting decoded buffer; this the absolute maximum
+    */
+    static std::size_t max_decode_size(std::size_t size) noexcept
+    {
+        return (size / 4 + (size % 4 ? 1 : 0)) * 3;
+    }
+    /**
+     Returns the required encoding size for a given size. The value is calculated with the following formula:
+
+     $$
+     \lceil \frac{size}{3} \rceil \cdot 4
+     $$
+
+     @param size the size of the decoded input
+     @returns the size of the resulting encoded buffer
+    */
+    static std::size_t required_encode_size(std::size_t size) noexcept
+    {
+        return (size / 3 + (size % 3 ? 1 : 0)) * 4;
+    }
+
+private:
+    static std::uint8_t _base64_value(alphabet& alphabet, char c)
+    {
+        if (c >= 'A' && c <= 'Z') {
+            return c - 'A';
+        } else if (c >= 'a' && c <= 'z') {
+            return c - 'a' + 26;
+        } else if (c >= '0' && c <= '9') {
+            return c - '0' + 52;
+        }
+
+        // comes down to alphabet
+        if (alphabet == alphabet::standard) {
+            if (c == '+') {
+                return 62;
+            } else if (c == '/') {
+                return 63;
+            }
+        } else if (alphabet == alphabet::url_filename_safe) {
+            if (c == '-') {
+                return 62;
+            } else if (c == '_') {
+                return 63;
+            }
+        } // auto detect
+        else {
+            if (c == '+') {
+                alphabet = alphabet::standard;
+
+                return 62;
+            } else if (c == '/') {
+                alphabet = alphabet::standard;
+
+                return 63;
+            } else if (c == '-') {
+                alphabet = alphabet::url_filename_safe;
+
+                return 62;
+            } else if (c == '_') {
+                alphabet = alphabet::url_filename_safe;
+
+                return 63;
+            }
+        }
+
+        throw base64_error("invalid base64 character.");
+    }
+};
+
+#endif // !PUBLIC_DOMAIN_BASE64_HPP_

common/build-info.cpp.in

@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/common.cpp

@@ -1,5 +1,4 @@
 #include "common.h"
-#include "build-info.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -13,6 +12,7 @@
 #include <regex>
 #include <sstream>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 #include <cinttypes>
@@ -91,6 +91,19 @@ void process_escapes(std::string& input) {
                 case '\'': input[output_idx++] = '\''; break;
                 case '\"': input[output_idx++] = '\"'; break;
                 case '\\': input[output_idx++] = '\\'; break;
+                case 'x':
+                    // Handle \x12, etc
+                    if (input_idx + 2 < input_len) {
+                        const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 };
+                        char *err_p = nullptr;
+                        const long val = std::strtol(x, &err_p, 16);
+                        if (err_p == x + 2) {
+                            input_idx += 2;
+                            input[output_idx++] = char(val);
+                            break;
+                        }
+                    }
+                    // fall through
                 default:   input[output_idx++] = '\\';
                            input[output_idx++] = input[input_idx]; break;
             }
@@ -103,11 +116,26 @@ void process_escapes(std::string& input) {
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    bool result = true;
+    try {
+        if (!gpt_params_parse_ex(argc, argv, params)) {
+            gpt_print_usage(argc, argv, gpt_params());
+            exit(0);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        gpt_print_usage(argc, argv, gpt_params());
+        exit(1);
+    }
+    return result;
+}
+
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
-    gpt_params default_params;
     const std::string arg_prefix = "--";
-    llama_sampling_params & sparams = params.sampling_params;
+    llama_sampling_params & sparams = params.sparams;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -204,26 +232,83 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scaling") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
+            else { invalid_param = true; break; }
         } else if (arg == "--rope-scale") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rope_freq_scale = 1.0f/std::stof(argv[i]);
-        } else if (arg == "--memory-f32") {
-            params.memory_f16 = false;
+        } else if (arg == "--yarn-orig-ctx") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_orig_ctx = std::stoi(argv[i]);
+        } else if (arg == "--yarn-ext-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_ext_factor = std::stof(argv[i]);
+        } else if (arg == "--yarn-attn-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_attn_factor = std::stof(argv[i]);
+        } else if (arg == "--yarn-beta-fast") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_fast = std::stof(argv[i]);
+        } else if (arg == "--yarn-beta-slow") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_slow = std::stof(argv[i]);
+        } else if (arg == "--samplers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.samplers_sequence = parse_samplers_input(argv[i]);
+        } else if (arg == "--sampling-seq") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.samplers_sequence = argv[i];
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             sparams.top_p = std::stof(argv[i]);
+        } else if (arg == "--min-p") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.min_p = std::stof(argv[i]);
         } else if (arg == "--temp") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             sparams.temp = std::stof(argv[i]);
+            sparams.temp = std::max(sparams.temp, 0.0f);
         } else if (arg == "--tfs") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -241,25 +326,26 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.repeat_last_n = std::stoi(argv[i]);
+            sparams.penalty_last_n = std::stoi(argv[i]);
+            sparams.n_prev = std::max(sparams.n_prev, sparams.penalty_last_n);
         } else if (arg == "--repeat-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.repeat_penalty = std::stof(argv[i]);
+            sparams.penalty_repeat = std::stof(argv[i]);
         } else if (arg == "--frequency-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.frequency_penalty = std::stof(argv[i]);
+            sparams.penalty_freq = std::stof(argv[i]);
         } else if (arg == "--presence-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.presence_penalty = std::stof(argv[i]);
+            sparams.penalty_present = std::stof(argv[i]);
         } else if (arg == "--mirostat") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -341,6 +427,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_sequences = std::stoi(argv[i]);
+        } else if (arg == "--p-accept" || arg == "-pa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.p_accept = std::stof(argv[i]);
+        } else if (arg == "--p-split" || arg == "-ps") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.p_split = std::stof(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -404,8 +502,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
+        } else if (arg == "-cml" || arg == "--chatml") {
+            params.chatml = true;
         } else if (arg == "--infill") {
             params.infill = true;
+        } else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
+            params.dump_kv_cache = true;
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            params.no_kv_offload = true;
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
+            params.cache_type_k = argv[++i];
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            params.cache_type_v = argv[++i];
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -546,10 +654,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
         } else if (arg == "-h" || arg == "--help") {
-            gpt_print_usage(argc, argv, default_params);
-#ifndef LOG_DISABLE_LOGS
-            log_print_usage();
-#endif // LOG_DISABLE_LOGS
+            return false;
+
+        } else if (arg == "--version") {
+            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -572,7 +681,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.grammar = argv[i];
+            sparams.grammar = argv[i];
         } else if (arg == "--grammar-file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -587,8 +696,49 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             std::copy(
                 std::istreambuf_iterator<char>(file),
                 std::istreambuf_iterator<char>(),
-                std::back_inserter(params.grammar)
+                std::back_inserter(sparams.grammar)
             );
+        } else if (arg == "--override-kv") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            char * sep = strchr(argv[i], '=');
+            if (sep == nullptr || sep - argv[i] >= 128) {
+                fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            struct llama_model_kv_override kvo;
+            std::strncpy(kvo.key, argv[i], sep - argv[i]);
+            kvo.key[sep - argv[i]] = 0;
+            sep++;
+            if (strncmp(sep, "int:", 4) == 0) {
+                sep += 4;
+                kvo.tag = LLAMA_KV_OVERRIDE_INT;
+                kvo.int_value = std::atol(sep);
+            } else if (strncmp(sep, "float:", 6) == 0) {
+                sep += 6;
+                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
+                kvo.float_value = std::atof(sep);
+            } else if (strncmp(sep, "bool:", 5) == 0) {
+                sep += 5;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
+                if (std::strcmp(sep, "true") == 0) {
+                    kvo.bool_value = true;
+                } else if (std::strcmp(sep, "false") == 0) {
+                    kvo.bool_value = false;
+                } else {
+                    fprintf(stderr, "error: Invalid boolean value for KV override: %s\n", argv[i]);
+                    invalid_param = true;
+                    break;
+                }
+            } else {
+                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            params.kv_overrides.push_back(kvo);
 #ifndef LOG_DISABLE_LOGS
         // Parse args for logging parameters
         } else if ( log_param_single_parse( argv[i] ) ) {
@@ -609,46 +759,50 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         // End of Parse args for logging parameters
 #endif // LOG_DISABLE_LOGS
         } else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            gpt_print_usage(argc, argv, default_params);
-            exit(1);
+            throw std::invalid_argument("error: unknown argument: " + arg);
         }
     }
     if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        gpt_print_usage(argc, argv, default_params);
-        exit(1);
+        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
     }
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
              params.instruct)) {
-        fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
-        gpt_print_usage(argc, argv, default_params);
-        exit(1);
+
+        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
     if (params.escape) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
         process_escapes(params.input_suffix);
+        process_escapes(sparams.cfg_negative_prompt);
         for (auto & antiprompt : params.antiprompt) {
             process_escapes(antiprompt);
         }
     }
 
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back(llama_model_kv_override());
+        params.kv_overrides.back().key[0] = 0;
+    }
+
     return true;
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    const llama_sampling_params & sparams = params.sampling_params;
+    const llama_sampling_params & sparams = params.sparams;
 
+    printf("\n");
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
+    printf("      --version         show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    printf("  -cml, --chatml        run in chatml mode (use with ChatML-compatible models)\n");
     printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
     printf("  -r PROMPT, --reverse-prompt PROMPT\n");
     printf("                        halt generation at PROMPT, return control in interactive mode\n");
@@ -674,14 +828,17 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
+    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
+    printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
     printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
     printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
-    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.repeat_last_n);
-    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.repeat_penalty);
-    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.presence_penalty);
-    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.frequency_penalty);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.penalty_repeat);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_present);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_freq);
     printf("  --mirostat N          use Mirostat sampling.\n");
     printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
     printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", sparams.mirostat);
@@ -698,13 +855,18 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --cfg-negative-prompt-file FNAME\n");
     printf("                        negative prompt file to use for guidance. (default: empty)\n");
     printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale);
-    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-scaling {none,linear,yarn}\n");
+    printf("                        RoPE frequency scaling method, defaults to linear unless specified by the model\n");
+    printf("  --rope-scale N        RoPE context scaling factor, expands context by a factor of N\n");
     printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor, expands context by a factor of 1/N\n");
+    printf("  --yarn-orig-ctx N     YaRN: original context size of model (default: 0 = model training context size)\n");
+    printf("  --yarn-ext-factor N   YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
+    printf("  --yarn-attn-factor N  YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
+    printf("  --yarn-beta-slow N    YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
+    printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
-    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
@@ -714,6 +876,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
     printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -pa N, --p-accept N   speculative decoding accept probability (default: %.1f)\n", (double)params.p_accept);
+    printf("  -ps N, --p-split N    speculative decoding split probability (default: %.1f)\n", (double)params.p_split);
     printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
     printf("  --image IMAGE_FILE    path to an image file. use with multimodal models\n");
@@ -741,7 +905,15 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #endif // GGML_USE_CUBLAS
 #endif
     printf("  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
+    printf("  -dkvc, --dump-kv-cache\n");
+    printf("                        verbose print of the KV cache\n");
+    printf("  -nkvo, --no-kv-offload\n");
+    printf("                        disable KV offload\n");
+    printf("  -ctk TYPE, --cache-type-k TYPE\n");
+    printf("                        KV cache data type for K (default: %s)\n", params.cache_type_k.c_str());
+    printf("  -ctv TYPE, --cache-type-v TYPE\n");
+    printf("                        KV cache data type for V (default: %s)\n", params.cache_type_v.c_str());
+    printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
     printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
@@ -751,7 +923,13 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
+    printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("\n");
+#ifndef LOG_DISABLE_LOGS
+    log_print_usage();
+#endif // LOG_DISABLE_LOGS
 }
 
 std::string get_system_info(const gpt_params & params) {
@@ -784,6 +962,48 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
     GGML_UNREACHABLE();
 }
 
+//
+// String parsing
+//
+
+std::string parse_samplers_input(std::string input) {
+    std::string output = "";
+    // since samplers names are written multiple ways
+    // make it ready for both system names and input names
+    std::unordered_map<std::string, char> samplers_symbols {
+        {"top_k",      'k'},
+        {"top-k",      'k'},
+        {"top_p",      'p'},
+        {"top-p",      'p'},
+        {"nucleus",    'p'},
+        {"typical_p",  'y'},
+        {"typical-p",  'y'},
+        {"typical",    'y'},
+        {"min_p",      'm'},
+        {"min-p",      'm'},
+        {"tfs_z",      'f'},
+        {"tfs-z",      'f'},
+        {"tfs",        'f'},
+        {"temp",       't'},
+        {"temperature",'t'}
+    };
+    // expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
+    size_t separator = input.find(';');
+    while (separator != input.npos) {
+        std::string name = input.substr(0,separator);
+        input = input.substr(separator+1);
+        separator = input.find(';');
+
+        if (samplers_symbols.find(name) != samplers_symbols.end()) {
+            output += samplers_symbols[name];
+        }
+    }
+    if (samplers_symbols.find(input) != samplers_symbols.end()) {
+        output += samplers_symbols[input];
+    }
+    return output;
+}
+
 //
 // Model utils
 //
@@ -798,28 +1018,87 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
+    if (params.kv_overrides.empty()) {
+        mparams.kv_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
+        mparams.kv_overrides = params.kv_overrides.data();
+    }
 
     return mparams;
 }
 
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    throw std::runtime_error("Invalid cache type: " + s);
+}
+
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto cparams = llama_context_default_params();
 
-    cparams.n_ctx           = params.n_ctx;
-    cparams.n_batch         = params.n_batch;
-    cparams.n_threads       = params.n_threads;
-    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
-    cparams.mul_mat_q       = params.mul_mat_q;
-    cparams.seed            = params.seed;
-    cparams.f16_kv          = params.memory_f16;
-    cparams.logits_all      = params.logits_all;
-    cparams.embedding       = params.embedding;
-    cparams.rope_freq_base  = params.rope_freq_base;
-    cparams.rope_freq_scale = params.rope_freq_scale;
+    cparams.n_ctx             = params.n_ctx;
+    cparams.n_batch           = params.n_batch;
+    cparams.n_threads         = params.n_threads;
+    cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q         = params.mul_mat_q;
+    cparams.seed              = params.seed;
+    cparams.logits_all        = params.logits_all;
+    cparams.embedding         = params.embedding;
+    cparams.rope_scaling_type = params.rope_scaling_type;
+    cparams.rope_freq_base    = params.rope_freq_base;
+    cparams.rope_freq_scale   = params.rope_freq_scale;
+    cparams.yarn_ext_factor   = params.yarn_ext_factor;
+    cparams.yarn_attn_factor  = params.yarn_attn_factor;
+    cparams.yarn_beta_fast    = params.yarn_beta_fast;
+    cparams.yarn_beta_slow    = params.yarn_beta_slow;
+    cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
+    cparams.offload_kqv       = !params.no_kv_offload;
+
+    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
+    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
 
     return cparams;
 }
 
+void llama_batch_clear(struct llama_batch & batch) {
+    batch.n_tokens = 0;
+}
+
+void llama_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits) {
+    batch.token   [batch.n_tokens] = id;
+    batch.pos     [batch.n_tokens] = pos;
+    batch.n_seq_id[batch.n_tokens] = seq_ids.size();
+    for (size_t i = 0; i < seq_ids.size(); ++i) {
+        batch.seq_id[batch.n_tokens][i] = seq_ids[i];
+    }
+    batch.logits  [batch.n_tokens] = logits;
+
+    batch.n_tokens++;
+}
+
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
     auto mparams = llama_model_params_from_gpt_params(params);
 
@@ -857,15 +1136,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     if (params.ignore_eos) {
-        params.sampling_params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
+        params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
     {
         LOG("warming up the model with an empty run\n");
 
-        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_kv_cache_clear(lctx);
         llama_reset_timings(lctx);
     }
 
@@ -918,7 +1197,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
 }
 
 std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    const llama_token bos_id = llama_token_bos(ctx);
+    const llama_token bos_id = llama_token_bos(llama_get_model(ctx));
 
     std::string piece;
     std::string result;
@@ -951,6 +1230,12 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
     return result;
 }
 
+bool llama_should_add_bos_token(const llama_model * model) {
+    const int add_bos = llama_add_bos_token(model);
+
+    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
+}
+
 //
 // YAML utils
 //
@@ -1067,6 +1352,7 @@ void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const cha
     if (!data_str.empty() && (std::isspace(data_str[0]) || std::isspace(data_str.back()))) {
         data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
         data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
+        data_str = std::regex_replace(data_str, std::regex(R"(\\[^n"])"), R"(\$&)");
         data_str = "\"" + data_str + "\"";
         fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
         return;
@@ -1102,28 +1388,28 @@ std::string get_sortable_timestamp() {
 
 void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
-    const llama_sampling_params & sparams = params.sampling_params;
+    const llama_sampling_params & sparams = params.sparams;
 
-    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
-    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
-    fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
-    fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false");
-    fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false");
-    fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
+    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
+    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
+    fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
+    fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
-    fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false");
-    fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false");
-    fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
-    fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
-    fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
-    fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
-    fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
-    fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
-    fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
-    fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
+    fprintf(stream, "cpu_has_cublas: %s\n",      ggml_cpu_has_cublas()      ? "true" : "false");
+    fprintf(stream, "cpu_has_clblast: %s\n",     ggml_cpu_has_clblast()     ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
 
 #ifdef NDEBUG
     fprintf(stream, "debug: false\n");
@@ -1157,13 +1443,13 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
-    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.frequency_penalty);
-    dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
+    dump_string_yaml_multiline(stream, "grammar", sparams.grammar.c_str());
     fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
     fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
-    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(lctx));
+    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx)));
     const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
     fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
 
@@ -1200,7 +1486,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
-    fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
@@ -1217,14 +1502,14 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
-    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.presence_penalty);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
     dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
     fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
     fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
     fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
     dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
     fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
-    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.repeat_penalty);
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
 
     fprintf(stream, "reverse_prompt:\n");
     for (std::string ap : params.antiprompt) {
@@ -1251,6 +1536,81 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
+    fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }
+
+//
+// KV cache utils
+//
+
+void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
+        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        int seq_count = 0;
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] >= 0) { seq_count++; }
+        }
+        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
+void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
+        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    std::unordered_map<llama_seq_id, size_t> seqs;
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] < 0) { continue; }
+            if (seqs.find(cs_curr[j]) == seqs.end()) {
+                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+                seqs[cs_curr[j]] = seqs.size();
+            }
+        }
+        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+    }
+
+    printf("=== Sequence legend: ");
+    for (const auto & it : seqs) {
+        printf("%zu=%d, ", it.second, it.first);
+    }
+    printf("'+'=other sequence ids");
+
+    c_curr = view.cells;
+    cs_curr = view.cells_sequences;
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] >= 0) {
+                const auto & it = seqs.find(cs_curr[j]);
+                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
+            } else {
+                putchar('.');
+            }
+        }
+        putchar(' ');
+    }
+
+    printf("\n=== Done dumping\n");
+}

common/common.h

@@ -9,6 +9,7 @@
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
+#include <cmath>
 #include <string>
 #include <vector>
 #include <random>
@@ -25,38 +26,54 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
-#define print_build_info() do {                                                             \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+#define print_build_info() do {                                                                     \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
+// build info
+extern int LLAMA_BUILD_NUMBER;
+extern char const *LLAMA_COMMIT;
+extern char const *LLAMA_COMPILER;
+extern char const *LLAMA_BUILD_TARGET;
+
 //
 // CLI argument parsing
 //
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                           = -1;   // RNG seed
+    uint32_t seed                           = -1;    // RNG seed
+
     int32_t n_threads                       = get_num_physical_cores();
-    int32_t n_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
-    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_draft                         = 16;   // number of tokens to draft during speculative decoding
-    int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
-    int32_t n_parallel                      = 1;    // number of parallel sequences to decode
-    int32_t n_sequences                     = 1;    // number of sequences to decode
-    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
-    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
-    int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
-    float   rope_freq_base                  = 0.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
+    int32_t n_threads_batch                 = -1;    // number of threads to use for batch processing (-1 = use n_threads)
+    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_draft                         = 16;    // number of tokens to draft during speculative decoding
+    int32_t n_chunks                        = -1;    // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel                      = 1;     // number of parallel sequences to decode
+    int32_t n_sequences                     = 1;     // number of sequences to decode
+    float   p_accept                        = 0.5f;  // speculative decoding accept probability
+    float   p_split                         = 0.1f;  // speculative decoding split probability
+    int32_t n_gpu_layers                    = -1;    // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft              = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    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
+    int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
+    float   rope_freq_base                  = 0.0f;  // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
+    float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
+    float   yarn_attn_factor                = 1.0f;  // YaRN magnitude scaling factor
+    float   yarn_beta_fast                  = 32.0f; // YaRN low correction dim
+    float   yarn_beta_slow                  = 1.0f;  // YaRN high correction dim
+    int32_t yarn_orig_ctx                   = 0;     // YaRN original context length
+    int8_t  rope_scaling_type               = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
+                                                                              //       pinging @cebtenzzre
 
     // // sampling parameters
-    struct llama_sampling_params sampling_params;
+    struct llama_sampling_params sparams;
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
     std::string model_draft       = "";                              // draft model for speculative decoding
@@ -66,10 +83,12 @@ struct gpt_params {
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
-    std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
     std::string logdir            = "";  // directory in which to save YAML log files
 
+    std::vector<llama_model_kv_override> kv_overrides;
+
+    // TODO: avoid tuple, use struct
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
 
@@ -77,14 +96,14 @@ struct gpt_params {
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
                                     //                                       (which is more convenient to use for plotting)
                                     //
-    bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    bool   hellaswag       = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
-    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
+    bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     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
 
@@ -104,12 +123,19 @@ struct gpt_params {
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool infill            = false; // use infill mode
+    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
+    bool no_kv_offload     = false; // disable KV offloading
+
+    std::string cache_type_k = "f16"; // KV cache data type for the K
+    std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
     std::string mmproj = ""; // path to multimodal projector
-    std::string image = ""; // path to an image file
+    std::string image  = ""; // path to an image file
 };
 
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
+
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
@@ -120,14 +146,33 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 
 void process_escapes(std::string& input);
 
+//
+// String parsing
+//
+
+std::string parse_samplers_input(std::string input);
+
 //
 // Model utils
 //
 
+// TODO: avoid tuplue, use struct
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
-struct llama_model_params   llama_model_params_from_gpt_params(const gpt_params & params);
+
+struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
+// Batch utils
+
+void llama_batch_clear(struct llama_batch & batch);
+
+void llama_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits);
+
 //
 // Vocab utils
 //
@@ -168,6 +213,10 @@ std::string llama_detokenize_bpe(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens);
 
+// Uses the value from the model metadata if possible, otherwise
+// defaults to true when model type is SPM, otherwise false.
+bool llama_should_add_bos_token(const llama_model * model);
+
 //
 // YAML utils
 //
@@ -181,3 +230,13 @@ std::string get_sortable_timestamp();
 void dump_non_result_info_yaml(
     FILE * stream, const gpt_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
+
+//
+// KV cache utils
+//
+
+// Dump the KV cache view with the number of sequences per cell.
+void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
+
+// Dump the KV cache view showing individual sequences in each cell (long output).
+void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);

common/grammar-parser.cpp

@@ -190,7 +190,7 @@ namespace grammar_parser {
                 pos = parse_space(pos + 1, is_nested);
             } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
                 if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
                 }
 
                 // apply transformation to previous symbol (last_sym_start to end) according to
@@ -399,7 +399,7 @@ namespace grammar_parser {
     void print_grammar(FILE * file, const parse_state & state) {
         try {
             std::map<uint32_t, std::string> symbol_id_names;
-            for (auto kv : state.symbol_ids) {
+            for (const auto & kv : state.symbol_ids) {
                 symbol_id_names[kv.second] = kv.first;
             }
             for (size_t i = 0, end = state.rules.size(); i < end; i++) {

common/log.h

@@ -61,13 +61,13 @@
 //  #define LOG_TARGET stderr
 //  #include "log.h"
 //
-//  The log target can also be redirected to a diffrent function
+//  The log target can also be redirected to a different function
 //  like so:
 //
-//  #define LOG_TARGET log_handler_diffrent()
+//  #define LOG_TARGET log_handler_different()
 //  #include "log.h"
 //
-//  FILE* log_handler_diffrent()
+//  FILE* log_handler_different()
 //  {
 //      return stderr;
 //  }
@@ -97,37 +97,56 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
 // Utility to obtain "pid" like unique process id and use it when creating log files.
 inline std::string log_get_pid()
 {
-    static std::string pid;
-    if (pid.empty())
-    {
-        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
-        //  it's not the same as "pid" but is unique enough to solve multiple instances
-        //  trying to write to the same log.
-        std::stringstream ss;
-        ss << std::this_thread::get_id();
-        pid = ss.str();
-    }
+   static std::string pid;
+   if (pid.empty())
+   {
+       // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+       //  it's not the same as "pid" but is unique enough to solve multiple instances
+       //  trying to write to the same log.
+       std::stringstream ss;
+       ss << std::this_thread::get_id();
+       pid = ss.str();
+   }
 
-    return pid;
+   return pid;
 }
 
 // Utility function for generating log file names with unique id based on thread id.
 //  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
 //  where the number is a runtime id of the current thread.
 
-#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
 
 // INTERNAL, DO NOT USE
-inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
 {
+    static bool _multilog = false;
+
+    if (multilog != LogTriStateSame)
+    {
+        _multilog = multilog == LogTriStateTrue;
+    }
+
     std::stringstream buf;
 
     buf << log_file_basename;
-    buf << ".";
-    buf << log_get_pid();
+    if (_multilog)
+    {
+        buf << ".";
+        buf << log_get_pid();
+    }
     buf << ".";
     buf << log_file_extension;
 
@@ -212,15 +231,6 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
     #define LOG_TEE_FLF_VAL ,""
 #endif
 
-// Utility for synchronizing log configuration state
-//  since std::optional was introduced only in c++17
-enum LogTriState
-{
-    LogTriStateSame,
-    LogTriStateFalse,
-    LogTriStateTrue
-};
-
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
@@ -314,16 +324,23 @@ enum LogTriState
 #endif
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
 {
-    static bool _initialized{false};
-    static bool _disabled{(filename.empty() && target == nullptr)};
+    static bool _initialized = false;
+    static bool _append = false;
+    static bool _disabled = filename.empty() && target == nullptr;
     static std::string log_current_filename{filename};
     static FILE *log_current_target{target};
     static FILE *logfile = nullptr;
 
     if (change)
     {
+        if (append != LogTriStateSame)
+        {
+            _append = append == LogTriStateTrue;
+            return logfile;
+        }
+
         if (disable == LogTriStateTrue)
         {
             // Disable primary target
@@ -376,7 +393,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
             }
         }
 
-        logfile = fopen(filename.c_str(), "w");
+        logfile = fopen(filename.c_str(), _append ? "a" : "w");
     }
 
     if (!logfile)
@@ -397,20 +414,20 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
 }
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
 {
-    return log_handler1_impl(change, disable, filename, target);
+    return log_handler1_impl(change, append, disable, filename, target);
 }
 
 // Disables logs entirely at runtime.
 //  Makes LOG() and LOG_TEE() produce no output,
-//  untill enabled back.
+//  until enabled back.
 #define log_disable() log_disable_impl()
 
 // INTERNAL, DO NOT USE
 inline FILE *log_disable_impl()
 {
-    return log_handler1_impl(true, LogTriStateTrue);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
 }
 
 // Enables logs at runtime.
@@ -419,19 +436,31 @@ inline FILE *log_disable_impl()
 // INTERNAL, DO NOT USE
 inline FILE *log_enable_impl()
 {
-    return log_handler1_impl(true, LogTriStateFalse);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
 }
 
 // Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
 #define log_set_target(target) log_set_target_impl(target)
 
 // INTERNAL, DO NOT USE
-inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
-inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
 
 // INTERNAL, DO NOT USE
 inline FILE *log_handler() { return log_handler1_impl(); }
 
+// Enable or disable creating separate log files for each run.
+//  can ONLY be invoked BEFORE first log use.
+#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
+// Enable or disable append mode for log file.
+//  can ONLY be invoked BEFORE first log use.
+#define log_append(enable) log_append_impl(enable)
+// INTERNAL, DO NOT USE
+inline FILE *log_append_impl(bool enable)
+{
+    return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
+}
+
 inline void log_test()
 {
     log_disable();
@@ -493,6 +522,18 @@ inline bool log_param_single_parse(const std::string & param)
         return true;
     }
 
+    if (param == "--log-new")
+    {
+        log_multilog(true);
+        return true;
+    }
+
+    if (param == "--log-append")
+    {
+        log_append(true);
+        return true;
+    }
+
     return false;
 }
 
@@ -522,7 +563,9 @@ inline void log_print_usage()
     printf("  --log-disable         Disable trace logs\n");
     printf("  --log-enable          Enable trace logs\n");
     printf("  --log-file            Specify a log filename (without extension)\n");
-    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+    printf("  --log-new             Create a separate new log file on start. "
+                                   "Each log file will have unique name: \"<name>.<ID>.log\"\n");
+    printf("  --log-append          Don't truncate the old log file.\n");
 }
 
 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
@@ -579,38 +622,75 @@ inline std::string log_var_to_string_impl(const std::vector<int> & var)
     return buf.str();
 }
 
-#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
-    [&tokens, &ctx]()                                                        \
-    {                                                                        \
-        std::stringstream buf;                                               \
-        buf << "[ ";                                                         \
-                                                                             \
-        bool first = true;                                                   \
-        for (const auto &token : tokens)                                     \
-        {                                                                    \
-            if (!first)                                                      \
-                buf << ", ";                                                 \
-            else                                                             \
-                first = false;                                               \
-                                                                             \
-            auto detokenized = llama_token_to_piece(ctx, token);             \
-                                                                             \
-            detokenized.erase(                                               \
-                std::remove_if(                                              \
-                    detokenized.begin(),                                     \
-                    detokenized.end(),                                       \
-                    [](const unsigned char c) { return !std::isprint(c); }), \
-                detokenized.end());                                          \
-                                                                             \
-            buf                                                              \
-                << "'" << detokenized << "'"                                 \
-                << ":" << std::to_string(token);                             \
-        }                                                                    \
-        buf << " ]";                                                         \
-                                                                             \
-        return buf.str();                                                    \
-    }()                                                                      \
-        .c_str()
+template <typename C, typename T>
+inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (const auto &token : tokens)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, token);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "'" << detokenized << "'"
+            << ":" << std::to_string(token);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+template <typename C, typename B>
+inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (int i = 0; i < batch.n_tokens; ++i)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "\n" << std::to_string(i)
+            << ":token '" << detokenized << "'"
+            << ":pos " << std::to_string(batch.pos[i])
+            << ":n_seq_id  " << std::to_string(batch.n_seq_id[i])
+            << ":seq_id " << std::to_string(batch.seq_id[i][0])
+            << ":logits " << std::to_string(batch.logits[i]);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
 
 #ifdef LOG_DISABLE_LOGS
 

common/sampling.cpp

@@ -1,113 +1,208 @@
 #include "sampling.h"
 
-llama_sampling_context::~llama_sampling_context() {
-    for (auto & it : sequence_contexts) {
-        if (it.second.grammar != NULL) {
-            llama_grammar_free(it.second.grammar);
-            it.second.grammar = NULL;
+struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
+    struct llama_sampling_context * result = new llama_sampling_context();
+
+    result->params  = params;
+    result->grammar = nullptr;
+
+    // if there is a grammar, parse it
+    if (!params.grammar.empty()) {
+        result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+
+        // will be empty (default) if there are parse errors
+        if (result->parsed_grammar.rules.empty()) {
+            fprintf(stderr, "%s: failed to parse grammar\n", __func__);
+            return nullptr;
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
+
+        result->grammar = llama_grammar_init(
+                grammar_rules.data(),
+                grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
+    }
+
+    result->prev.resize(params.n_prev);
+
+    return result;
+}
+
+void llama_sampling_free(struct llama_sampling_context * ctx) {
+    if (ctx->grammar != NULL) {
+        llama_grammar_free(ctx->grammar);
+    }
+
+    delete ctx;
+}
+
+void llama_sampling_reset(llama_sampling_context * ctx) {
+    if (ctx->grammar != NULL) {
+        llama_grammar_free(ctx->grammar);
+        ctx->grammar = NULL;
+    }
+
+    if (!ctx->parsed_grammar.rules.empty()) {
+        std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
+
+        ctx->grammar = llama_grammar_init(
+                grammar_rules.data(),
+                grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
+    }
+
+    std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
+    ctx->cur.clear();
+}
+
+void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
+    if (dst->grammar) {
+        llama_grammar_free(dst->grammar);
+        dst->grammar = nullptr;
+    }
+
+    if (src->grammar) {
+        dst->grammar = llama_grammar_copy(src->grammar);
+    }
+
+    dst->prev = src->prev;
+}
+
+llama_token llama_sampling_last(llama_sampling_context * ctx) {
+    return ctx->prev.back();
+}
+
+std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
+    const int size = ctx_sampling->prev.size();
+
+    n = std::min(n, size);
+
+    std::string result;
+
+    for (int i = size - n; i < size; i++) {
+        result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
+    }
+
+    return result;
+}
+
+std::string llama_sampling_print(const llama_sampling_params & params) {
+    char result[1024];
+
+    snprintf(result, sizeof(result),
+            "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
+            params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
+            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
+            params.mirostat, params.mirostat_eta, params.mirostat_tau);
+
+    return std::string(result);
+}
+
+std::string llama_sampling_order_print(const llama_sampling_params & params) {
+    std::string result = "CFG -> Penalties ";
+    if (params.mirostat == 0) {
+        for (auto s : params.samplers_sequence) {
+            switch (s) {
+                case 'k': result += "-> top_k "; break;
+                case 'f': result += "-> tfs_z "; break;
+                case 'y': result += "-> typical_p "; break;
+                case 'p': result += "-> top_p "; break;
+                case 'm': result += "-> min_p "; break;
+                case 't': result += "-> temp "; break;
+                default : break;
+            }
+        }
+    } else {
+        result += "-> mirostat ";
+    }
+
+    return result;
+}
+
+// no reasons to expose this function in header
+static void sampler_queue(
+                   struct llama_context * ctx_main,
+            const llama_sampling_params & params,
+                 llama_token_data_array & cur_p,
+                                 size_t & min_keep) {
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+
+    const float         temp              = params.temp;
+    const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float         top_p             = params.top_p;
+    const float         min_p             = params.min_p;
+    const float         tfs_z             = params.tfs_z;
+    const float         typical_p         = params.typical_p;
+    const std::string & samplers_sequence = params.samplers_sequence;
+
+    for (auto s : samplers_sequence) {
+        switch (s){
+            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case 't': llama_sample_temp     (ctx_main, &cur_p, temp); break;
+            default : break;
         }
     }
 }
 
-llama_sampling_context llama_sampling_context_init(
-        const struct gpt_params & params,
-                  llama_grammar * grammar) {
-  llama_sampling_context result;
-
-  result.params = params.sampling_params;
-  result.grammar = grammar;
-  return result;
-}
-
-// Note: Creates the context if it doesn't exist, so this always return something.
-llama_sampler_sequence_context & llama_sampling_get_sequence_context(
-              llama_sampling_context & ctx_sampling,
-        const llama_seq_id             seq) {
-    const auto it = ctx_sampling.sequence_contexts.find(seq);
-    if (it != ctx_sampling.sequence_contexts.end()) {
-        return it->second;
-    }
-    llama_sampler_sequence_context new_ctx = {
-        2.0f * ctx_sampling.params.mirostat_tau,
-        ctx_sampling.grammar != NULL ? llama_grammar_copy(ctx_sampling.grammar) : NULL,
-    };
-    return ctx_sampling.sequence_contexts.insert({seq, new_ctx}).first->second;
-}
-
-bool llama_sampling_context_reset(
-              llama_sampling_context & ctx_sampling,
-        const llama_seq_id             seq) {
-    const auto it = ctx_sampling.sequence_contexts.find(seq);
-    if (it == ctx_sampling.sequence_contexts.end()) return false;
-    if (it->second.grammar != NULL) {
-        llama_grammar_free(it->second.grammar);
-        it->second.grammar = NULL;
-    }
-    ctx_sampling.sequence_contexts.erase(it);
-    return true;
-}
-
 llama_token llama_sampling_sample(
-                  struct llama_context * ctx,
-                  struct llama_context * ctx_guidance,
-                  struct llama_sampling_context & ctx_sampling,
-        const std::vector<llama_token> & last_tokens,
-         std::vector<llama_token_data> & candidates,
-        const                      int   idx,
-                          llama_seq_id   seq) {
-    const int n_ctx   = llama_n_ctx(ctx);
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx) {
+    const llama_sampling_params & params = ctx_sampling->params;
+
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
-    const llama_sampling_params & params = ctx_sampling.params;
     const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? n_vocab : 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 ? 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 int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
+    const float   penalty_repeat  = params.penalty_repeat;
+    const float   penalty_freq    = params.penalty_freq;
+    const float   penalty_present = params.penalty_present;
     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;
 
+    auto & prev = ctx_sampling->prev;
+    auto & cur  = ctx_sampling->cur;
+
     llama_token id = 0;
 
-    float * logits = llama_get_logits_ith(ctx, idx);
+    float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    // Apply params.logit_bias map
+    // apply params.logit_bias map
     for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
         logits[it->first] += it->second;
     }
 
-    candidates.clear();
+    cur.clear();
+
     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});
+        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
     }
 
-    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
 
-    if (ctx_guidance) {
-        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+    if (ctx_cfg) {
+        llama_sample_classifier_free_guidance(ctx_main, &cur_p, ctx_cfg, params.cfg_scale);
     }
 
     // apply penalties
-    if (!last_tokens.empty()) {
-        const float nl_logit = logits[llama_token_nl(ctx)];
-        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
+    if (!prev.empty()) {
+        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
 
-        llama_sample_repetition_penalty(ctx, &cur_p,
-                last_tokens.data() + last_tokens.size() - last_n_repeat,
-                last_n_repeat, repeat_penalty);
-        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
-                last_tokens.data() + last_tokens.size() - last_n_repeat,
-                last_n_repeat, alpha_frequency, alpha_presence);
+        llama_sample_repetition_penalties(ctx_main, &cur_p,
+                prev.data() + prev.size() - penalty_last_n,
+                penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
                     cur_p.data[idx].logit = nl_logit;
                     break;
                 }
@@ -115,52 +210,60 @@ llama_token llama_sampling_sample(
         }
     }
 
-    llama_sampler_sequence_context & ctx_seq = llama_sampling_get_sequence_context(ctx_sampling, seq);
-
-    if (ctx_seq.grammar != NULL) {
-        llama_sample_grammar(ctx, &cur_p, ctx_seq.grammar);
+    if (ctx_sampling->grammar != NULL) {
+        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
-    if (temp <= 0) {
-        // Greedy sampling
-        id = llama_sample_token_greedy(ctx, &cur_p);
+    if (temp < 0.0) {
+        // greedy sampling, with probs
+        llama_sample_softmax(ctx_main, &cur_p);
+        id = cur_p.data[0].id;
+    } else if (temp == 0.0) {
+        // greedy sampling, no probs
+        id = llama_sample_token_greedy(ctx_main, &cur_p);
     } else {
         if (mirostat == 1) {
             const int mirostat_m = 100;
-            llama_sample_temp(ctx, &cur_p, temp);
-            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_seq.mirostat_mu);
+            llama_sample_temp(ctx_main, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
         } else if (mirostat == 2) {
-            llama_sample_temp(ctx, &cur_p, temp);
-            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &ctx_seq.mirostat_mu);
+            llama_sample_temp(ctx_main, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
-            // Temperature sampling
+            // temperature sampling
             size_t min_keep = std::max(1, params.n_probs);
-            llama_sample_top_k      (ctx, &cur_p, top_k, min_keep);
-            llama_sample_tail_free  (ctx, &cur_p, tfs_z, min_keep);
-            llama_sample_typical    (ctx, &cur_p, typical_p, min_keep);
-            llama_sample_top_p      (ctx, &cur_p, top_p, min_keep);
-            llama_sample_temp(ctx, &cur_p, temp);
 
-            {
-                const int n_top = 10;
-                LOG("top %d candidates:\n", n_top);
+            sampler_queue(ctx_main, params, cur_p, min_keep);
 
-                for (int i = 0; i < n_top; i++) {
-                    const llama_token id = cur_p.data[i].id;
-                    (void)id; // To avoid a warning that id is unused when logging is disabled.
-                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
-                }
-            }
+            id = llama_sample_token(ctx_main, &cur_p);
 
-            id = llama_sample_token(ctx, &cur_p);
+            //{
+            //    const int n_top = 10;
+            //    LOG("top %d candidates:\n", n_top);
 
-            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+            //    for (int i = 0; i < n_top; i++) {
+            //        const llama_token id = cur_p.data[i].id;
+            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
+            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
+            //    }
+            //}
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
         }
     }
 
-    if (ctx_seq.grammar != NULL) {
-        llama_grammar_accept_token(ctx, ctx_seq.grammar, id);
-    }
-
     return id;
 }
+
+void llama_sampling_accept(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        llama_token id,
+        bool apply_grammar) {
+    ctx_sampling->prev.erase(ctx_sampling->prev.begin());
+    ctx_sampling->prev.push_back(id);
+
+    if (ctx_sampling->grammar != NULL && apply_grammar) {
+        llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
+    }
+}

common/sampling.h

@@ -2,107 +2,113 @@
 
 #include "llama.h"
 
+#include "grammar-parser.h"
+
 #include <string>
 #include <vector>
 #include <unordered_map>
 
 // sampling parameters
 typedef struct llama_sampling_params {
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typical_p         = 1.00f; // 1.0 = disabled
-    float   temp              = 0.80f; // 1.0 = disabled
-    float   repeat_penalty    = 1.10f; // 1.0 = disabled
-    int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   frequency_penalty = 0.00f; // 0.0 = disabled
-    float   presence_penalty  = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
+    int32_t     n_prev                = 64;       // number of previous tokens to remember
+    int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     top_k                 = 40;       // <= 0 to use vocab size
+    float       top_p                 = 0.95f;    // 1.0 = disabled
+    float       min_p                 = 0.05f;    // 0.0 = disabled
+    float       tfs_z                 = 1.00f;    // 1.0 = disabled
+    float       typical_p             = 1.00f;    // 1.0 = disabled
+    float       temp                  = 0.80f;    // 1.0 = disabled
+    int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
+    float       penalty_freq          = 0.00f;    // 0.0 = disabled
+    float       penalty_present       = 0.00f;    // 0.0 = disabled
+    int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float       mirostat_tau          = 5.00f;    // target entropy
+    float       mirostat_eta          = 0.10f;    // learning rate
+    bool        penalize_nl           = true;     // consider newlines as a repeatable token
+    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
 
-    bool    penalize_nl       = true;  // consider newlines as a repeatable token
-
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    std::string grammar;  // optional BNF-like grammar to constrain sampling
 
     // Classifier-Free Guidance
     // https://arxiv.org/abs/2306.17806
-    std::string cfg_negative_prompt;   // string to help guidance
-    float       cfg_scale     = 1.f;   // How strong is guidance
+    std::string cfg_negative_prompt; // string to help guidance
+    float       cfg_scale     = 1.f; // how strong is guidance
 
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
-
 } llama_sampling_params;
 
-// per-sequence sampler context
-typedef struct llama_sampler_sequence_context {
-    float mirostat_mu; // mirostat sampler state
-    llama_grammar * grammar;
-} llama_sampler_sequence_context;
-
 // general sampler context
-typedef struct llama_sampling_context {
-    ~llama_sampling_context();
-
-    // parameters that will be used for sampling and when creating
-    // new llama_sampler_sequence_context instances
+// TODO: move to llama.h
+struct llama_sampling_context {
+    // parameters that will be used for sampling
     llama_sampling_params params;
 
-    // map of sequence ids to sampler contexts
-    std::unordered_map<llama_seq_id, llama_sampler_sequence_context> sequence_contexts;
+    // mirostat sampler state
+    float mirostat_mu;
 
-    // when non-NULL, new instances of llama_sampler_sequence_context
-    // will get a copy of the grammar here
-    // note: only the pointer is stored here, it is not a copy of
-    //       the grammar and shouldn't be freed
     llama_grammar * grammar;
-} llama_sampling_context;
+
+    // internal
+    grammar_parser::parse_state parsed_grammar;
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token>      prev;
+    std::vector<llama_token_data> cur;
+};
 
 #include "common.h"
 
 // Create a new sampling context instance.
-llama_sampling_context llama_sampling_context_init(
-        const struct gpt_params & params,
-                  llama_grammar * grammar = NULL);
+struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
 
-// Fetches the sampler context for the specified sequence id (defaults to 0).
-// If the context for that sequence id doesn't already exist, it will be created with
-// default values based on the parameters in the ctx_sampling argument.
-llama_sampler_sequence_context & llama_sampling_get_sequence_context(
-              llama_sampling_context & ctx_sampling,
-        const llama_seq_id             seq = 0);
+void llama_sampling_free(struct llama_sampling_context * ctx);
 
-// Reset the sampler context for the supplied sequence id (defaults to 0).
-// This is necessary to reuse a sequence id or free memory used by sequences
-// that are no longer required.
-bool llama_sampling_context_reset(
-              llama_sampling_context & ctx_sampling,
-        const llama_seq_id             seq = 0);
+// Reset the sampler context
+// - clear prev tokens
+// - reset grammar
+void llama_sampling_reset(llama_sampling_context * ctx);
+
+// Copy the sampler context
+void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
+
+// Get the last sampled token
+llama_token llama_sampling_last(llama_sampling_context * ctx);
+
+// Get a string representation of the last sampled tokens
+std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
+
+// Print sampling parameters into a string
+std::string llama_sampling_print(const llama_sampling_params & params);
+
+// Print sampling order into a string
+std::string llama_sampling_order_print(const llama_sampling_params & params);
 
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call
-//       llama_sampling_context_reset when a sequence ends
+//       llama_sampling_reset when a sequence ends
 //
 // required:
-//  - ctx:          context to use for sampling
+//  - ctx_main:     context to use for sampling
 //  - ctx_sampling: sampling-specific context
 //
 // optional:
-//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
-//  - last_tokens:   needed for repetition penalty, ignore if empty
-//  - idx:           sample from llama_get_logits_ith(ctx, idx)
-//  - seq:           sequence id to associate sampler state with
+//  - ctx_cfg:      context to use for classifier-free guidance
+//  - idx:          sample from llama_get_logits_ith(ctx, idx)
 //
 // returns:
 //  - token:      sampled token
 //  - candidates: vector of candidate tokens
 //
 llama_token llama_sampling_sample(
-                  struct llama_context * ctx,
-                  struct llama_context * ctx_guidance,
-                  struct llama_sampling_context & ctx_sampling,
-        const std::vector<llama_token> & last_tokens,
-         std::vector<llama_token_data> & candidates,
-        const                      int   idx = 0,
-                          llama_seq_id   seq = 0);
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        struct llama_context * ctx_cfg,
+        int idx = 0);
+
+void llama_sampling_accept(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        llama_token id,
+        bool apply_grammar);

common/train.cpp

@@ -32,6 +32,7 @@ struct train_state  * init_train_state() {
     state->opt = new struct ggml_opt_context;
     state->opt->ctx = NULL;
     state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
     state->opt->loss_after = 0.0f;
 
     return state;
@@ -70,7 +71,7 @@ void free_random_uniform_distribution(struct random_uniform_distribution * rnd)
 
 struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
     float scale = 1.0f; // xavier
-    switch (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
         case 1:
             scale /= sqrtf((float) tensor->ne[0]);
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
@@ -118,7 +119,7 @@ struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct
 }
 
 struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
-    switch (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
         case 1:
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
                 float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@@ -182,25 +183,27 @@ float fclamp(const float v, const float min, const float max) {
 }
 
 void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
     GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == 1);
+    GGML_ASSERT(tensor->ne[2] == 1);
+    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
+    GGML_ASSERT(tensor->ne[2] == 1);
+    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
+    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
@@ -224,8 +227,8 @@ int64_t get_example_targets_batch(
     bool                   sample_random_offsets
 ) {
     GGML_ASSERT(samples_count > 0);
-    GGML_ASSERT(tokens_input->n_dims  == 2);
-    GGML_ASSERT(target_probs->n_dims  == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(target_probs));
     int64_t n_vocab  = target_probs->ne[0];
     int64_t n_tokens = tokens_input->ne[0];
     int64_t n_batch  = tokens_input->ne[1];
@@ -236,8 +239,8 @@ int64_t get_example_targets_batch(
     int64_t used_samples = 0;
 
     ggml_set_f32(target_probs, 0.0f);
-    llama_token bos = llama_token_bos(lctx);
-    llama_token eos = llama_token_eos(lctx);
+    llama_token bos = llama_token_bos(llama_get_model(lctx));
+    llama_token eos = llama_token_eos(llama_get_model(lctx));
     // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
     for (int k=0; k<n_batch; ++k) {
         // printf("%s: batch %d\n", __func__, k);
@@ -924,7 +927,7 @@ size_t tokenize_file(
         for (llama_token token=0; token < n_vocab; ++token) {
             max_token_text_size = std::max(
                 max_token_text_size,
-                strlen(llama_token_get_text(lctx, token)));
+                strlen(llama_token_get_text(llama_get_model(lctx), token)));
         }
 
         // upper bound of context byte length.
@@ -1045,6 +1048,7 @@ struct train_params_common get_default_train_params_common() {
     params.n_batch    =    8;
     params.n_gradient_accumulation = 1;
     params.n_epochs   = -1;
+    params.n_gpu_layers = 0;
 
     params.custom_n_ctx = false;
 
@@ -1080,6 +1084,7 @@ struct train_params_common get_default_train_params_common() {
     params.adam_beta2          = 0.999f;
     params.adam_gclip          = 1.0f;
     params.adam_eps_f          = 0.0f;
+
     return params;
 }
 
@@ -1133,6 +1138,7 @@ void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train
     fprintf(stderr, "  --adam-beta2 N             AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2);
     fprintf(stderr, "  --adam-gclip N             AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip);
     fprintf(stderr, "  --adam-epsf N              AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f);
+    fprintf(stderr, "  -ngl N, --n-gpu-layers N   Number of model layers to offload to GPU (default %d)", params->n_gpu_layers);
     fprintf(stderr, "\n");
 }
 
@@ -1352,6 +1358,17 @@ bool consume_common_train_arg(
             return true;
         }
         params->adam_gclip = std::stof(argv[i]);
+    } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                *invalid_param = true;
+                return true;
+            }
+#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 == "-h" || arg == "--help") {
         params->print_usage = true;
         return true;
@@ -1425,7 +1442,7 @@ void train_opt_callback(void * vdata, int accum_step, float * sched, bool * canc
 
         int impr_plot = -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
         if (impr_plot > 0) impr_plot = 0;
-        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_before)) impr_plot = 0;
+        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_after)) impr_plot = 0;
         printf("%s: iter=%6d sample=%zu/%zu sched=%f loss=%f",
             __func__, opt->iter, std::min(1+train->shuffle_next_sample, train->shuffle_sample_count), train->shuffle_sample_count,
             *sched, opt->loss_after);

common/train.h

@@ -9,6 +9,8 @@
 #include "ggml.h"
 #include "llama.h"
 
+#define LLAMA_TRAIN_MAX_NODES 16384
+
 typedef std::string mt19937_state;
 
 struct train_state {
@@ -44,6 +46,7 @@ struct train_params_common {
     int n_batch;
     int n_gradient_accumulation;
     int n_epochs;
+    int n_gpu_layers;
 
     bool custom_n_ctx;
 

convert-baichuan-hf-to-gguf.py

@@ -1,310 +0,0 @@
-#!/usr/bin/env python3
-# HF baichuan --> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import TYPE_CHECKING, Any
-import itertools
-import numpy as np
-import torch
-from sentencepiece import SentencePieceProcessor  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-if TYPE_CHECKING:
-    from typing import TypeAlias
-
-NDArray: TypeAlias = 'np.ndarray[Any, Any]'
-
-# reverse HF permute back to original pth layout
-
-
-def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray:
-    if n_kv_head is not None and n_head != n_kv_head:
-        n_head //= n_kv_head
-
-    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-            .swapaxes(1, 2)
-            .reshape(weights.shape))
-
-def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray:
-        r = weights.shape[0] // 3
-        return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
-
-def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray:
-        r = weights.shape[0] // 3
-        return weights[r * n_part : r * n_part + r, ...]
-
-def count_model_parts(dir_model: str) -> int:
-    num_parts = 0
-
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-
-    return num_parts
-
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-print("hello print: ",hparams["architectures"][0])
-if hparams["architectures"][0] != "BaichuanForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-print(f"num_parts:{num_parts}\n")
-ARCH=gguf.MODEL_ARCH.BAICHUAN
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-head_count = hparams["num_attention_heads"]
-
-if "num_key_value_heads" in hparams:
-    head_count_kv = hparams["num_key_value_heads"]
-else:
-    head_count_kv = head_count
-
-if "_name_or_path" in hparams:
-    hf_repo = hparams["_name_or_path"]
-else:
-    hf_repo = ""
-
-if "max_sequence_length" in hparams:
-    ctx_length = hparams["max_sequence_length"]
-elif "max_position_embeddings" in hparams:
-    ctx_length = hparams["max_position_embeddings"]
-elif "model_max_length" in hparams:
-    ctx_length = hparams["model_max_length"]
-else:
-    print("gguf: can not find ctx length parameter.")
-
-    sys.exit()
-
-
-gguf_writer.add_name(dir_model.name)
-gguf_writer.add_source_hf_repo(hf_repo)
-gguf_writer.add_tensor_data_layout("Meta AI original pth")
-gguf_writer.add_context_length(ctx_length)
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
-gguf_writer.add_head_count(head_count)
-gguf_writer.add_head_count_kv(head_count_kv)
-gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
-
-if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
-    if "type" in hparams["rope_scaling"]:
-        if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
-
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytes] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-tokenizer_model_file = dir_model / 'tokenizer.model'
-if not tokenizer_model_file.is_file():
-    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
-    sys.exit(1)
-
-# vocab type sentencepiece
-print("gguf: get sentencepiece tokenizer vocab, scores and token types")
-
-tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
-vocab_size = hparams.get('vocab_size')
-if vocab_size is None:
-    vocab_size = tokenizer.vocab_size()
-
-for i in range(vocab_size):
-    text: bytes
-    score: float
-
-    piece = tokenizer.id_to_piece(i)
-    text = piece.encode("utf-8")
-    score = tokenizer.get_score(i)
-
-    toktype = 1  # defualt to normal token type
-    if tokenizer.is_unknown(i):
-        toktype = 2
-    if tokenizer.is_control(i):
-        toktype = 3
-
-    # toktype = 4 is user-defined = tokens from added_tokens.json
-
-    if tokenizer.is_unused(i):
-        toktype = 5
-    if tokenizer.is_byte(i):
-        toktype = 6
-
-    tokens.append(text)
-    scores.append(score)
-    toktypes.append(toktype)
-
-added_tokens_file = dir_model / 'added_tokens.json'
-if added_tokens_file.is_file():
-    with open(added_tokens_file, "r", encoding="utf-8") as f:
-        addtokens_json = json.load(f)
-
-        print("gguf: get added tokens")
-
-        for key in addtokens_json:
-            tokens.append( key.encode("utf-8") )
-            scores.append(-1000.0)
-            toktypes.append(4) # user-defined token type
-
-
-gguf_writer.add_tokenizer_model("llama")
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    tmp=model_part
-    for i in range(block_count):
-        if f"model.layers.{i}.self_attn.W_pack.weight" in model_part:
-            print(f"Unpacking and permuting layer {i}")
-            tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count)
-            tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv)
-            tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2)
-            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
-
-    for name in model_part.keys():
-        data = model_part[name]
-        # we don't need these
-        if name.endswith(".rotary_emb.inv_freq"):
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name + " -> " +  new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-bloom-hf-to-gguf.py

@@ -1,238 +0,0 @@
-#!/usr/bin/env python3
-# HF bloom --> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import re
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-# Supported Models:
-#   https://huggingface.co/bigscience/bloom-1b7
-#   https://huggingface.co/bigscience/bloom-3b
-#   https://huggingface.co/bigscience/bloom-7b1
-#   https://huggingface.co/Langboat/bloom-1b4-zh
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a Bloom model to a GGML compatible file")
-    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    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 (*.bin)")
-    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "BloomForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.BLOOM
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["n_layer"]
-
-gguf_writer.add_name("Bloom")
-n_embed = hparams.get("hidden_size", hparams.get("n_embed"))
-n_head = hparams.get("n_head", hparams.get("num_attention_heads"))
-gguf_writer.add_context_length(hparams.get("seq_length", n_embed))
-gguf_writer.add_embedding_length(n_embed)
-gguf_writer.add_feed_forward_length(4 * n_embed)
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(n_head)
-gguf_writer.add_head_count_kv(n_head)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0)  # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
-
-# params for qkv transform
-n_head_kv = hparams.get("n_head_kv", n_head)
-head_dim = n_embed // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(dir_model / part_name, map_location="cpu")
-
-    has_lm_head = True
-    if "lm_head.weight" not in model_part.keys() and "output.weight" not in model_part.keys():
-        has_lm_head = False
-
-    for original_name in model_part.keys():
-        data = model_part[original_name]
-        name = re.sub(r'transformer\.', '', original_name)
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
-            # Map bloom-style qkv_linear to gpt-style qkv_linear
-            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
-            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
-            qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
-            data = np.concatenate(
-                (qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
-                 qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
-                 qkv_weights[:, 2, :, :].reshape((-1, n_embed))),
-                axis=0
-            )
-            print("re-format attention.linear_qkv.weight")
-        elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
-            qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
-            data = np.concatenate(
-                (qkv_bias[:, 0, :].reshape((n_embed,)),
-                 qkv_bias[:, 1, :].reshape((n_embed,)),
-                 qkv_bias[:, 2, :].reshape((n_embed,))),
-                axis=0
-            )
-            print("re-format attention.linear_qkv.bias")
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-        if not has_lm_head and name == "word_embeddings.weight":
-            gguf_writer.add_tensor("output.weight", data)
-            print(name, "=>", "output.weight" + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))  # noqa
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-falcon-hf-to-gguf.py

@@ -1,250 +0,0 @@
-#!/usr/bin/env python3
-# HF falcon--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import contextlib
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path, prefix: str) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith(prefix):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "FalconForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model, "model-00")
-if num_parts:
-    is_safetensors = True
-    from safetensors import safe_open
-else:
-    is_safetensors = False
-    num_parts = count_model_parts(dir_model, "pytorch_model-")
-
-ARCH=gguf.MODEL_ARCH.FALCON
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-
-gguf_writer.add_name("Falcon")
-gguf_writer.add_context_length(2048) # not in config.json
-gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["num_attention_heads"])
-if "num_kv_heads" in hparams:
-    gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
-else:
-    gguf_writer.add_head_count_kv(1)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i])
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# params for qkv transform
-n_head    = hparams["num_attention_heads"]
-n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
-
-head_dim = hparams["hidden_size"] // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-elif is_safetensors:
-    part_names = (
-        f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1)
-    )
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    if is_safetensors:
-        ctx = safe_open(dir_model / part_name, framework="pt", device="cpu")
-    else:
-        ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu"))
-
-    with ctx as model_part:
-        for name in model_part.keys():
-            data = model_part.get_tensor(name) if is_safetensors else model_part[name]
-
-            old_dtype = data.dtype
-
-            # convert any unsupported data types to float32
-            if data.dtype != torch.float16 and data.dtype != torch.float32:
-                data = data.to(torch.float32)
-
-            # QKV tensor transform
-            # The original query_key_value tensor contains n_head_kv "kv groups",
-            # each consisting of n_head/n_head_kv query weights followed by one key
-            # and one value weight (shared by all query heads in the kv group).
-            # This layout makes it a big pain to work with in GGML.
-            # So we rearrange them here,, so that we have n_head query weights
-            # followed by n_head_kv key weights followed by n_head_kv value weights,
-            # in contiguous fashion.
-            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
-
-            if "query_key_value" in name:
-                qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
-                q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
-                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
-                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
-                data = torch.cat((q,k,v)).reshape_as(data)
-
-            data = data.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-            if new_name is None:
-                print("Can not map tensor '" + name + "'")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-            gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-gptneox-hf-to-gguf.py

@@ -1,212 +0,0 @@
-#!/usr/bin/env python3
-# HF gptneox--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "GPTNeoXForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.GPTNEOX
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-
-gguf_writer.add_name(dir_model.name)
-gguf_writer.add_context_length(hparams["max_position_embeddings"])
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"])))
-gguf_writer.add_head_count(hparams["num_attention_heads"])
-gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name.endswith(".attention.masked_bias") or name.endswith(".attention.bias") or name.endswith(".attention.rotary_emb.inv_freq"):
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-llama-ggml-to-gguf.py

@@ -2,7 +2,6 @@
 from __future__ import annotations
 
 import argparse
-import math
 import struct
 import sys
 from enum import IntEnum
@@ -12,34 +11,16 @@ import numpy as np
 
 import os
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
-# Note: Does not support GGML_QKK_64
-QK_K = 256
-# Items here are (block size, type size)
-GGML_QUANT_SIZES = {
-    gguf.GGMLQuantizationType.F32  : (1, 4),
-    gguf.GGMLQuantizationType.F16  : (1, 2),
-    gguf.GGMLQuantizationType.Q4_0 : (32, 2 + 16),
-    gguf.GGMLQuantizationType.Q4_1 : (32, 2 + 2 + 16),
-    gguf.GGMLQuantizationType.Q5_0 : (32, 2 + 4 + 16),
-    gguf.GGMLQuantizationType.Q5_1 : (32, 2 + 2 + 4 + 16),
-    gguf.GGMLQuantizationType.Q8_0 : (32, 2 + 32),
-    gguf.GGMLQuantizationType.Q8_1 : (32, 4 + 4 + 32),
-    gguf.GGMLQuantizationType.Q2_K : (256, 2 + 2 + QK_K // 16 + QK_K // 4),
-    gguf.GGMLQuantizationType.Q3_K : (256, 2 + QK_K // 4 + QK_K // 8 + 12),
-    gguf.GGMLQuantizationType.Q4_K : (256, 2 + 2 + QK_K // 2 + 12),
-    gguf.GGMLQuantizationType.Q5_K : (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12),
-    gguf.GGMLQuantizationType.Q6_K : (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16),
-    gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8),
-}
 
 class GGMLFormat(IntEnum):
     GGML = 0
     GGMF = 1
     GGJT = 2
 
+
 class GGMLFType(IntEnum):
     ALL_F32              = 0
     MOSTLY_F16           = 1
@@ -59,6 +40,7 @@ class GGMLFType(IntEnum):
     MOSTLY_Q5_K_M        = 17
     MOSTLY_Q6_K          = 18
 
+
 class Hyperparameters:
     def __init__(self):
         self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
@@ -90,6 +72,7 @@ class Hyperparameters:
     def __str__(self):
         return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'
 
+
 class Vocab:
     def __init__(self, load_scores = True):
         self.items = []
@@ -111,6 +94,7 @@ class Vocab:
             self.items.append((item_text, item_score))
         return offset - orig_offset
 
+
 class Tensor:
     def __init__(self, use_padding = True):
         self.name = None
@@ -125,7 +109,7 @@ class Tensor:
         (n_dims, name_len, dtype) = struct.unpack('<3I', data[offset:offset + 12])
         assert n_dims >= 0 and n_dims <= 4, f'Invalid tensor dimensions {n_dims}'
         assert name_len < 4096, 'Absurd tensor name length'
-        quant = GGML_QUANT_SIZES.get(dtype)
+        quant = gguf.GGML_QUANT_SIZES.get(dtype)
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
@@ -144,6 +128,7 @@ class Tensor:
         # print(n_dims, name_len, dtype, self.dims, self.name, pad)
         return offset - orig_offset
 
+
 class GGMLModel:
     def __init__(self):
         self.hyperparameters = None
@@ -180,8 +165,8 @@ class GGMLModel:
             if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
                 err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
         elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
-            if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
-                          GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+            if ftype in (GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                         GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
                 err = 'Q4 and Q8 quantizations changed in GGJTv3.'
         if len(err) > 0:
             raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
@@ -208,6 +193,7 @@ class GGMLModel:
         hp.set_n_ff(self)
         return offset
 
+
 class GGMLToGGUF:
     def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
         hp = ggml_model.hyperparameters
@@ -238,7 +224,7 @@ class GGMLToGGUF:
         gguf_writer = gguf.GGUFWriter(
             self.cfg.output,
             gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
-            use_temp_file = False )
+            use_temp_file = False)
         self.add_params(gguf_writer)
         self.add_vocab(gguf_writer)
         if self.special_vocab is not None:
@@ -362,7 +348,8 @@ class GGMLToGGUF:
                 mapped_name,
                 data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
                 raw_shape = tempdims,
-                raw_dtype = tensor.dtype )
+                raw_dtype = tensor.dtype)
+
 
 def handle_metadata(cfg, hp):
     import convert
@@ -386,36 +373,40 @@ def handle_metadata(cfg, hp):
         raise ValueError('Unable to load metadata')
     vocab = convert.load_vocab(
         cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
-        cfg.vocabtype )
+        cfg.vocabtype)
     # FIXME: Respect cfg.vocab_dir?
-    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
+                               load_merges = cfg.vocabtype == 'bpe',
+                               n_vocab = vocab.vocab_size)
     convert.check_vocab_size(params, vocab)
     return (params, vocab, svocab)
 
+
 def handle_args():
     parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
     parser.add_argument('--input', '-i', type = Path, required = True,
-        help = 'Input GGMLv3 filename')
+                        help = 'Input GGMLv3 filename')
     parser.add_argument('--output', '-o', type = Path, required = True,
-        help ='Output GGUF filename')
+                        help ='Output GGUF filename')
     parser.add_argument('--name',
-        help = 'Set model name')
+                        help = 'Set model name')
     parser.add_argument('--desc',
-        help = 'Set model description')
+                        help = 'Set model description')
     parser.add_argument('--gqa', type = int, default = 1,
-        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+                        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
     parser.add_argument('--eps', default = '5.0e-06',
-        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+                        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
     parser.add_argument('--context-length', '-c', type=int, default = 2048,
-        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+                        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
     parser.add_argument('--model-metadata-dir', '-m', type = Path,
-        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+                        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
     parser.add_argument("--vocab-dir", type=Path,
-        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
+                        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
     parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
-        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
     return parser.parse_args()
 
+
 def main():
     cfg = handle_args()
     print(f'* Using config: {cfg}')
@@ -425,7 +416,7 @@ def main():
     data = np.memmap(cfg.input, mode = 'r')
     model = GGMLModel()
     print('* Scanning GGML input file')
-    offset = model.load(data, 0)
+    offset = model.load(data, 0)  # noqa
     print(f'* GGML model hyperparameters: {model.hyperparameters}')
     vocab_override = None
     params_override = None
@@ -440,12 +431,15 @@ def main():
         print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
         if model.file_format == GGMLFormat.GGML:
             print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
-    converter = GGMLToGGUF(model, data, cfg,
+    converter = GGMLToGGUF(
+        model, data, cfg,
         params_override = params_override,
         vocab_override = vocab_override,
-        special_vocab = special_vocab )
+        special_vocab = special_vocab
+    )
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 
+
 if __name__ == '__main__':
     main()

convert-lora-to-ggml.py

@@ -3,7 +3,6 @@ from __future__ import annotations
 
 import json
 import os
-import re
 import struct
 import sys
 from typing import Any, BinaryIO, Sequence
@@ -11,43 +10,15 @@ from typing import Any, BinaryIO, Sequence
 import numpy as np
 import torch
 
+from pathlib import Path
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
 NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
 
 
-HF_SUBLAYER_TO_GGML = {
-    "self_attn.q_proj": "attn_q",
-    "self_attn.k_proj": "attn_k",
-    "self_attn.v_proj": "attn_v",
-    "self_attn.o_proj": "attn_output",
-    "mlp.gate_proj": "ffn_gate",
-    "mlp.down_proj": "ffn_down",
-    "mlp.up_proj": "ffn_up",
-    "input_layernorm": "attn_norm",
-    "post_attention_layernorm": "ffn_norm",
-}
-
-
-def translate_tensor_name(t: str) -> str:
-    match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t)
-    if match:
-        nn = match.group(1)
-        sub_layer = match.group(2)
-        lora_type = match.group(3)
-
-        sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer)
-        if sub_layer_renamed is None:
-            print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}")
-            sys.exit(1)
-
-        output_string = (
-            f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
-        )
-        return output_string
-    else:
-        print(f"Error: unrecognized tensor {t}")
-        sys.exit(1)
-
-
 def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
@@ -61,9 +32,7 @@ def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(struct.pack("i", int(params["lora_alpha"])))
 
 
-def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
-) -> None:
+def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
     sname = name.encode("utf-8")
     fout.write(
         struct.pack(
@@ -78,11 +47,12 @@ def write_tensor_header(
     fout.seek((fout.tell() + 31) & -32)
 
 
-if len(sys.argv) != 2:
-    print(f"Usage: python {sys.argv[0]} <path>")
+if len(sys.argv) < 2:
+    print(f"Usage: python {sys.argv[0]} <path> [arch]")
     print(
         "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
     )
+    print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
     sys.exit(1)
 
 input_json = os.path.join(sys.argv[1], "adapter_config.json")
@@ -90,6 +60,14 @@ input_model = os.path.join(sys.argv[1], "adapter_model.bin")
 output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
 
 model = torch.load(input_model, map_location="cpu")
+arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
+
+if arch_name not in gguf.MODEL_ARCH_NAMES.values():
+    print(f"Error: unsupported architecture {arch_name}")
+    sys.exit(1)
+
+arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
+name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
 
 with open(input_json, "r") as f:
     params = json.load(f)
@@ -117,6 +95,7 @@ with open(output_path, "wb") as fout:
 
     write_file_header(fout, params)
     for k, v in model.items():
+        orig_k = k
         if k.endswith(".default.weight"):
             k = k.replace(".default.weight", ".weight")
         if k in ["llama_proj.weight", "llama_proj.bias"]:
@@ -129,7 +108,32 @@ with open(output_path, "wb") as fout:
             v = v.float()
 
         t = v.detach().numpy()
-        tname = translate_tensor_name(k)
+
+        prefix = "base_model.model."
+        if k.startswith(prefix):
+            k = k[len(prefix) :]
+
+        lora_suffixes = (".lora_A.weight", ".lora_B.weight")
+        if k.endswith(lora_suffixes):
+            suffix = k[-len(lora_suffixes[0]):]
+            k = k[: -len(lora_suffixes[0])]
+        else:
+            print(f"Error: unrecognized tensor name {orig_k}")
+            sys.exit(1)
+
+        tname = name_map.get_name(k)
+        if tname is None:
+            print(f"Error: could not map tensor name {orig_k}")
+            print(" Note: the arch parameter must be specified if the model is not llama")
+            sys.exit(1)
+
+        if suffix == ".lora_A.weight":
+            tname += ".weight.loraA"
+        elif suffix == ".lora_B.weight":
+            tname += ".weight.loraB"
+        else:
+            assert False
+
         print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
         write_tensor_header(fout, tname, t.shape, t.dtype)
         t.tofile(fout)

convert-mpt-hf-to-gguf.py

@@ -1,218 +0,0 @@
-#!/usr/bin/env python3
-# HF mpt--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert an MPT model to a GGML compatible file")
-    parser.add_argument(
-        "--vocab-only", action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "MPTForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.MPT
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["n_layers"]
-
-gguf_writer.add_name(dir_model.name)
-gguf_writer.add_context_length(hparams["max_seq_len"])
-gguf_writer.add_embedding_length(hparams["d_model"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
-gguf_writer.add_head_count(hparams["n_heads"])
-if kv_n_heads := hparams["attn_config"].get("kv_n_heads"):
-    gguf_writer.add_head_count_kv(kv_n_heads)
-gguf_writer.add_layer_norm_eps(1e-05)
-if hparams["attn_config"]["clip_qkv"] is not None:
-    gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
-gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# MPT token embedding tensors have dimension 50432 (hparams["vocab_size"]), but
-# there are only 50254 (len(tokenizer.vocab)) tokens in the vocab, presumably to
-# accomodate some "reserved" tokens; this is causing problems down the line in
-# llama.cpp, so we pad the vocab with dummy tokens:
-
-vocab_size = hparams["vocab_size"]
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Cannot map tensor '" + name + "'")
-            continue # for the sake of compatibility with some old published models, don't quit
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-        # note: MPT output is tied to (same as) wte in original model;
-        # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
-        if new_name == "token_embd.weight":
-            gguf_writer.add_tensor("output.weight", data)
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-persimmon-to-gguf.py

@@ -6,9 +6,10 @@ import argparse
 from pathlib import Path
 from sentencepiece import SentencePieceProcessor
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
+
 def _flatten_dict(dct, tensors, prefix=None):
     assert isinstance(dct, dict)
     for key in dct.keys():
@@ -21,6 +22,7 @@ def _flatten_dict(dct, tensors, prefix=None):
             raise ValueError(type(dct[key]))
     return None
 
+
 def _get_sentencepiece_tokenizer_info(dir_model: Path):
     tokenizer_path = dir_model / 'adept_vocab.model'
     print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
@@ -54,6 +56,7 @@ def _get_sentencepiece_tokenizer_info(dir_model: Path):
         pass
     return tokens, scores, toktypes
 
+
 def main():
     parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
     parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
@@ -125,6 +128,5 @@ def main():
     print("")
 
 
-
 if __name__ == '__main__':
     main()

convert-refact-hf-to-gguf.py

@@ -1,263 +0,0 @@
-#!/usr/bin/env python3
-# HF refact--> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import sys
-from pathlib import Path
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if "NO_LOCAL_GGUF" not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf"))
-import gguf
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Convert a Refact model to a GGML compatible file"
-    )
-    parser.add_argument(
-        "--vocab-only",
-        action="store_true",
-        help="extract only the vocab",
-    )
-    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 (*.bin)",
-    )
-    parser.add_argument(
-        "ftype",
-        type=int,
-        choices=[0, 1],
-        default=1,
-        nargs="?",
-        help="output format - use 0 for float32, 1 for float16",
-    )
-    return parser.parse_args()
-
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f"Error: {args.model} is not a directory", file=sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf"
-
-print("gguf: loading model " + dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "GPTRefactForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH = gguf.MODEL_ARCH.REFACT
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-# Get refact feed forward dimension
-hidden_dim = hparams["n_embd"]
-inner_dim = 4 * hidden_dim
-hidden_dim = int(2 * inner_dim / 3)
-multiple_of = 256
-ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
-
-block_count = hparams["n_layer"]
-
-gguf_writer.add_name("Refact")
-# refact uses Alibi. So this is from config.json which might be used by training.
-gguf_writer.add_context_length(hparams["n_positions"])
-gguf_writer.add_embedding_length(hparams["n_embd"])
-
-gguf_writer.add_feed_forward_length(ff_dim)
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
-gguf_writer.add_head_count_kv(1)
-gguf_writer.add_layer_norm_rms_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
-
-# params for qkv transform
-n_head = hparams["n_head"]
-n_head_kv = 1
-
-head_dim = hparams["n_embd"] // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(dir_model / part_name, map_location="cpu")
-
-    for i in range(block_count):
-        if f"transformer.h.{i}.attn.kv.weight" in model_part:
-            data = model_part[f"transformer.h.{i}.attn.kv.weight"]
-            model_part[f"model.layers.{i}.self_attn.k_proj.weight"] = data[
-                : n_head_kv * head_dim
-            ]
-            model_part[f"model.layers.{i}.self_attn.v_proj.weight"] = data[
-                n_head_kv * head_dim :
-            ]
-            del model_part[f"transformer.h.{i}.attn.kv.weight"]
-        if f"transformer.h.{i}.attn.q.weight" in model_part:
-            model_part[f"model.layers.{i}.self_attn.q_proj.weight"] = model_part[
-                f"transformer.h.{i}.attn.q.weight"
-            ]
-            del model_part[f"transformer.h.{i}.attn.q.weight"]
-        if f"transformer.h.{i}.mlp.gate_up_proj.weight" in model_part:
-            data = model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
-            model_part[f"model.layers.{i}.mlp.gate_proj.weight"] = data[:ff_dim]
-            model_part[f"model.layers.{i}.mlp.up_proj.weight"] = data[ff_dim:]
-            del model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if (
-            ftype == 1
-            and data_dtype == np.float32
-            and name.endswith(".weight")
-            and n_dims == 2
-        ):
-            data = data.astype(np.float16)
-
-        print(
-            new_name
-            + ", n_dims = "
-            + str(n_dims)
-            + ", "
-            + str(old_dtype)
-            + " --> "
-            + str(data.dtype)
-        )
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-starcoder-hf-to-gguf.py

@@ -1,202 +0,0 @@
-#!/usr/bin/env python3
-# HF starcoder --> gguf conversion
-
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-from transformers import AutoTokenizer  # type: ignore[import]
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
-def count_model_parts(dir_model: Path) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file")
-    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
-    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 (*.bin)")
-    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
-    return parser.parse_args()
-
-args = parse_args()
-
-dir_model = args.model
-ftype = args.ftype
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file = sys.stderr)
-    sys.exit(1)
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
-
-print("gguf: loading model "+dir_model.name)
-
-with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "GPTBigCodeForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit(1)
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.STARCODER
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["n_layer"]
-
-gguf_writer.add_name("StarCoder")
-gguf_writer.add_context_length(hparams["n_positions"])
-gguf_writer.add_embedding_length(hparams["n_embd"])
-gguf_writer.add_feed_forward_length(4 * hparams["n_embd"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
-gguf_writer.add_head_count_kv(1)
-gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
-gguf_writer.add_file_type(ftype)
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-# gpt2 tokenizer
-gguf_writer.add_tokenizer_model("gpt2")
-
-print("gguf: get gpt2 tokenizer vocab")
-
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-# The number of tokens in tokenizer.json can differ from the expected vocab size.
-# This causes downstream issues with mismatched tensor sizes when running the inference
-vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
-assert max(tokenizer.vocab.values()) < vocab_size
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-
-for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-    scores.append(0.0) # dummy
-    toktypes.append(gguf.TokenType.NORMAL)
-
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
-special_vocab.add_to_gguf(gguf_writer)
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# params for qkv transform
-n_head    = hparams["n_head"]
-n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
-
-head_dim = hparams["n_embd"] // n_head
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = iter(("pytorch_model.bin",))
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(dir_model / part_name, map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert.py

@@ -3,15 +3,14 @@ from __future__ import annotations
 
 import argparse
 import concurrent.futures
-import copy
 import enum
 import faulthandler
 import functools
-import io
 import itertools
 import json
 import math
 import mmap
+import os
 import pickle
 import re
 import signal
@@ -20,17 +19,17 @@ import sys
 import time
 import zipfile
 from abc import ABCMeta, abstractmethod
+from collections import OrderedDict
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar
+from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, Optional, TypeVar, cast
 
 import numpy as np
-from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+from sentencepiece import SentencePieceProcessor
 
-import os
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
 if TYPE_CHECKING:
@@ -44,10 +43,12 @@ NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
+
 #
 # data types
 #
 
+
 @dataclass(frozen=True)
 class DataType:
     name: str
@@ -57,14 +58,17 @@ class DataType:
     def elements_to_bytes(self, n_elements: int) -> int:
         return n_elements * self.dtype.itemsize
 
+
 @dataclass(frozen=True)
 class UnquantizedDataType(DataType):
     pass
 
-DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
-DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
+
+DT_F16  = UnquantizedDataType('F16',  dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32',  dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32',  dtype = np.dtype(np.int16),   valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16),  valid_conversions = ['F32', 'F16', 'Q8_0'])
+
 
 @dataclass(frozen=True)
 class QuantizedDataType(DataType):
@@ -79,6 +83,7 @@ class QuantizedDataType(DataType):
         assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
         return self.quantized_dtype.itemsize * (n_elements // self.block_size)
 
+
 @dataclass(frozen=True)
 class Q8_0QuantizedDataType(QuantizedDataType):
     # Mini Q8_0 quantization in Python!
@@ -88,6 +93,7 @@ class Q8_0QuantizedDataType(QuantizedDataType):
         n_blocks = arr.size // self.block_size
         blocks = arr.reshape((n_blocks, self.block_size))
         # Much faster implementation of block quantization contributed by @Cebtenzzre
+
         def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
             d = abs(blocks).max(axis = 1) / np.float32(127)
             with np.errstate(divide = 'ignore'):
@@ -96,10 +102,11 @@ class Q8_0QuantizedDataType(QuantizedDataType):
             yield from zip(d, qs)
         return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
 
+
 DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
-    dtype = np.dtype(np.float32), valid_conversions = [],
-    ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
-    quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
+                                dtype = np.dtype(np.float32), valid_conversions = [],
+                                ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+                                quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
 
 # Quantized types skipped here because they may also map to np.float32
 NUMPY_TYPE_TO_DATA_TYPE: dict[np.dtype[Any], DataType] = {}
@@ -118,6 +125,8 @@ SAFETENSORS_DATA_TYPES: dict[str, DataType] = {
 # TODO: match this with `llama_ftype`
 # TODO: rename to LLAMAFileType
 # TODO: move to `gguf.py`
+
+
 class GGMLFileType(enum.IntEnum):
     AllF32     = 0
     MostlyF16  = 1  # except 1d tensors
@@ -130,6 +139,7 @@ class GGMLFileType(enum.IntEnum):
         # 1D tensors are always F32.
         return dt if len(tensor.shape) > 1 else DT_F32
 
+
 GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
     GGMLFileType.AllF32    : DT_F32,
     GGMLFileType.MostlyF16 : DT_F16,
@@ -140,19 +150,25 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
 # hparams loading
 #
 
+
 @dataclass
 class Params:
-    n_vocab:    int
-    n_embd:     int
-    n_layer:    int
-    n_ctx:      int
-    n_ff:       int
-    n_head:     int
-    n_head_kv:  int
-    f_norm_eps: float
+    n_vocab:        int
+    n_embd:         int
+    n_layer:        int
+    n_ctx:          int
+    n_ff:           int
+    n_head:         int
+    n_head_kv:      int
+    n_experts:      int | None = None
+    n_experts_used: int | None = None
+    f_norm_eps:     float | None = None
 
+    rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
     f_rope_scale: float | None = None
+    n_orig_ctx: int | None = None
+    rope_finetuned: bool | None = None
 
     ftype: GGMLFileType | None = None
 
@@ -166,11 +182,11 @@ class Params:
 
         # 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)
+            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
         elif "model.layers.0.self_attn.W_pack.weight" in model:   # next: try baichuan naming
-            n_layer=next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model)
+            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.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_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
 
         if n_layer < 1:
             raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n"
@@ -198,20 +214,20 @@ class Params:
     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_layer          = config["num_hidden_layers"]
-        n_ff             = config["intermediate_size"]
-        n_head           = config["num_attention_heads"]
-        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
-        f_norm_eps       = config["rms_norm_eps"]
-        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
-
+        rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
         rope_scaling = config.get("rope_scaling")
-        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
-            f_rope_scale = config["rope_scaling"].get("factor")
-        else:
-            f_rope_scale = None
+
+        if rope_scaling is not None and (typ := rope_scaling.get("type")):
+            rope_factor = rope_scaling.get("factor")
+            f_rope_scale = rope_factor
+            if typ == "linear":
+                rope_scaling_type = gguf.RopeScalingType.LINEAR
+            elif typ == "yarn":
+                rope_scaling_type = gguf.RopeScalingType.YARN
+                n_orig_ctx = rope_scaling['original_max_position_embeddings']
+                rope_finetuned = rope_scaling['finetuned']
+            else:
+                raise NotImplementedError(f'Unknown rope scaling type: {typ}')
 
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
@@ -221,17 +237,29 @@ class Params:
             raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
+        n_experts      = None
+        n_experts_used = None
+
+        if "num_local_experts" in config:
+            n_experts = config["num_local_experts"]
+            n_experts_used = config["num_experts_per_tok"]
+
         return Params(
-            n_vocab          = n_vocab,
-            n_embd           = n_embd,
-            n_layer          = n_layer,
-            n_ctx            = n_ctx,
-            n_ff             = n_ff,
-            n_head           = n_head,
-            n_head_kv        = n_head_kv,
-            f_norm_eps       = f_norm_eps,
-            f_rope_freq_base = f_rope_freq_base,
-            f_rope_scale     = f_rope_scale,
+            n_vocab           = config["vocab_size"],
+            n_embd            = config["hidden_size"],
+            n_layer           = config["num_hidden_layers"],
+            n_ctx             = n_ctx,
+            n_ff              = config["intermediate_size"],
+            n_head            = (n_head := config["num_attention_heads"]),
+            n_head_kv         = config.get("num_key_value_heads", n_head),
+            n_experts         = n_experts,
+            n_experts_used    = n_experts_used,
+            f_norm_eps        = config["rms_norm_eps"],
+            f_rope_freq_base  = config.get("rope_theta"),
+            rope_scaling_type = rope_scaling_type,
+            f_rope_scale      = f_rope_scale,
+            n_orig_ctx        = n_orig_ctx,
+            rope_finetuned    = rope_finetuned,
         )
 
     # LLaMA v2 70B params.json
@@ -240,17 +268,15 @@ class Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
-        n_embd           = config["dim"]
-        n_layer          = config["n_layers"]
-        n_ff             = -1
-        n_head           = config["n_heads"]
-        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
-        f_norm_eps       = config["norm_eps"]
-        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
+        n_experts      = None
+        n_experts_used = None
+        f_rope_freq_base = None
 
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if f_rope_freq_base == 1000000:
+        if config.get("moe"):
+            # Mixtral
+            n_ctx = 32768
+        elif config.get("rope_theta") == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -260,22 +286,27 @@ class Params:
             # LLaMA v1
             n_ctx = 2048
 
-        if n_vocab == -1:
-            n_vocab = model["tok_embeddings.weight"].shape[0]
-
-        if n_ff == -1:
+        if "layers.0.feed_forward.w1.weight" in model:
             n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
 
+        if config.get("moe"):
+            n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
+            n_experts      = config["moe"]["num_experts"]
+            n_experts_used = config["moe"]["num_experts_per_tok"]
+            f_rope_freq_base = 1e6
+
         return Params(
-            n_vocab          = n_vocab,
-            n_embd           = n_embd,
-            n_layer          = n_layer,
+            n_vocab          = model["tok_embeddings.weight"].shape[0],
+            n_embd           = config["dim"],
+            n_layer          = config["n_layers"],
             n_ctx            = n_ctx,
             n_ff             = n_ff,
-            n_head           = n_head,
-            n_head_kv        = n_head_kv,
-            f_norm_eps       = f_norm_eps,
-            f_rope_freq_base = f_rope_freq_base,
+            n_head           = (n_head := config["n_heads"]),
+            n_head_kv        = config.get("n_kv_heads", n_head),
+            n_experts        = n_experts,
+            n_experts_used   = n_experts_used,
+            f_norm_eps       = config["norm_eps"],
+            f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
         )
 
     @staticmethod
@@ -297,137 +328,152 @@ class Params:
         return params
 
 
-#
-# vocab
-#
+class VocabLoader:
+    def __init__(self, params: Params, fname_tokenizer: Path) -> None:
+        try:
+            from transformers import AutoTokenizer
+        except ImportError as e:
+            raise ImportError(
+                "To use VocabLoader, please install the `transformers` package. "
+                "You can install it with `pip install transformers`."
+            ) from e
 
-class BpeVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        added_tokens: dict[str, int]
-        if fname_added_tokens is not None:
-            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
-            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
+        try:
+            self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), trust_remote_code=True)
+        except ValueError:
+            self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), use_fast=False, trust_remote_code=True)
+
+        self.added_tokens_dict: OrderedDict[str, int] = OrderedDict()
+
+        for tok, tokidx in sorted(self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]):
+            if tokidx >= params.n_vocab or tokidx < self.tokenizer.vocab_size:
+                continue
+
+            self.added_tokens_dict[tok] = tokidx
+
+        self.unk_token_id: int = self.tokenizer.unk_token_id
+        self.specials: dict[str, int] = {
+            tok: self.tokenizer.get_vocab()[tok]
+            for tok in self.tokenizer.all_special_tokens
+        }
+        self.special_ids: set[int] = set(self.tokenizer.all_special_ids)
+        self.vocab_size_base: int = self.tokenizer.vocab_size
+        self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_dict)
+        self.fname_tokenizer: Path = fname_tokenizer
+
+        vocab_file = "tokenizer.model"
+        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
+        if path_candidate is not None:
+            self.spm = SentencePieceProcessor(str(path_candidate))
+            print(self.spm.vocab_size(), self.vocab_size_base)
         else:
-            # Fall back to trying to find the added tokens in tokenizer.json
-            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
-            if not tokenizer_json_file.is_file():
-                added_tokens = {}
-            else:
-                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
-                added_tokens = dict(
-                    (item['content'], item['id'])
-                    for item in tokenizer_json.get('added_tokens', [])
-                    # Added tokens here can be duplicates of the main vocabulary.
-                    if item['content'] not in self.bpe_tokenizer )
+            self.spm = None
 
-        vocab_size: int = len(self.bpe_tokenizer)
-        expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids      = sorted(added_tokens.values())
-        if expected_ids != actual_ids:
-            expected_end_id = vocab_size + len(actual_ids) - 1
-            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
+    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        tokenizer = self.tokenizer
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.get_vocab().items()}
+        added_tokens_ids = set(self.added_tokens_dict.values())
 
-        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
-        self.added_tokens_list    = [text for (text, idx) in items]
-        self.vocab_size_base: int = vocab_size
-        self.vocab_size: int      = self.vocab_size_base + len(self.added_tokens_list)
-        self.fname_tokenizer      = fname_tokenizer
-        self.fname_added_tokens   = fname_added_tokens
+        for i in range(self.vocab_size_base):
+            if i in added_tokens_ids:
+                continue
 
-    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        tokenizer = self.bpe_tokenizer
-        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
-        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
+            text = reverse_vocab[i].encode("utf-8")
+            yield text, self.get_token_score(i), self.get_token_type(i)
 
-        for i, _ in enumerate(tokenizer):
-            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
+    def get_token_type(self, token_id: int) -> gguf.TokenType:
+        toktype = gguf.TokenType.NORMAL
 
-    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        for text in self.added_tokens_list:
-            score = -1000.0
-            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
-
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.bpe_tokens()
-        yield from self.added_tokens()
-
-    def __repr__(self) -> str:
-        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-class SentencePieceVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        added_tokens: dict[str, int]
-        if fname_added_tokens is not None:
-            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
-        else:
-            added_tokens = {}
-
-        vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
-        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids   = sorted(added_tokens.values())
-        if expected_ids != actual_ids:
-            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
-
-        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
-        self.added_tokens_list = [text for (text, idx) in items]
-        self.vocab_size_base: int = vocab_size
-        self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
-        self.fname_tokenizer = fname_tokenizer
-        self.fname_added_tokens = fname_added_tokens
-
-    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        tokenizer = self.sentencepiece_tokenizer
-        for i in range(tokenizer.vocab_size()):
-            piece = tokenizer.id_to_piece(i)
-            text: bytes = piece.encode("utf-8")
-            score: float = tokenizer.get_score(i)
-
-            toktype = gguf.TokenType.NORMAL
-            if tokenizer.is_unknown(i):
+        if self.spm is not None and token_id < self.spm.vocab_size():
+            if self.spm.is_unknown(token_id):
                 toktype = gguf.TokenType.UNKNOWN
-            if tokenizer.is_control(i):
+            if self.spm.is_control(token_id):
+                toktype = gguf.TokenType.CONTROL
+            if self.spm.is_unused(token_id):
+                toktype = gguf.TokenType.UNUSED
+            if self.spm.is_byte(token_id):
+                toktype = gguf.TokenType.BYTE
+        else:
+            if token_id == self.unk_token_id:
+                toktype = gguf.TokenType.UNKNOWN
+            if token_id in self.special_ids:
                 toktype = gguf.TokenType.CONTROL
 
-            # NOTE: I think added_tokens are user defined.
-            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
-            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
+        return toktype
 
-            if tokenizer.is_unused(i):
-                toktype = gguf.TokenType.UNUSED
-            if tokenizer.is_byte(i):
-                toktype = gguf.TokenType.BYTE
-
-            yield text, score, toktype
+    def get_token_score(self, token_id: int) -> float:
+        if self.spm is not None and token_id < self.spm.vocab_size():
+            return cast(float, self.spm.get_score(token_id))
+        return 0.0
 
     def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        for text in self.added_tokens_list:
-            score = -1000.0
-            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+
+        for text in self.added_tokens_dict:
+            if text in self.specials:
+
+                toktype = self.get_token_type(self.specials[text])
+                score = self.get_token_score(self.specials[text])
+
+            else:
+                toktype = gguf.TokenType.USER_DEFINED
+                score = -1000.0
+
+            yield text.encode("utf-8"), score, toktype
+
+    def has_newline_token(self) -> bool:
+        return '<0x0A>' in self.tokenizer.vocab or '\n' in self.tokenizer.vocab
 
     def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.sentencepiece_tokens()
+        yield from self.hf_tokens()
         yield from self.added_tokens()
 
-    def __repr__(self) -> str:
-        return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+    def get_vocab_type(self) -> str:
+        path_candidates = []
+        vocab_file = "tokenizer.model"
+        path_candidates.append(vocab_file)
+        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
+        if path_candidate is not None:
+            return "llama"
+
+        vocab_file = "vocab.json"
+        path_candidates.append(vocab_file)
+        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
+        if path_candidate is not None:
+            return "gpt2"
+
+        vocab_file = "tokenizer.json"
+        path_candidates.append(vocab_file)
+        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
+        if path_candidate:
+            if not self.has_newline_token():
+                return "gpt2"
+            return "llama"
+
+        raise FileNotFoundError(
+            f"Could not find {path_candidates} in {self.fname_tokenizer} or its parent; "
+            "if it's in another directory, pass the directory as --vocab-dir"
+        )
+
+    def __repr__(self) -> str:
+        return f"<VocabLoader with {self.vocab_size_base} base tokens and {len(self.added_tokens_dict)} added tokens>"
+
+
+Vocab: TypeAlias = 'VocabLoader'
 
-Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 
 #
 # data loading
 # TODO: reuse (probably move to gguf.py?)
 #
 
+
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
-    #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
+    # print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
     if n_head_kv is not None and n_head != n_head_kv:
         n_head = n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                .swapaxes(1, 2)
-                .reshape(weights.shape))
+            .swapaxes(1, 2)
+            .reshape(weights.shape))
 
 
 class Tensor(metaclass=ABCMeta):
@@ -508,7 +554,7 @@ class LazyTensor:
         ret = self._load()
         # Should be okay if it maps to the same numpy type?
         assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
-                (self.data_type, ret.data_type, self.description)
+            (self.data_type, ret.data_type, self.description)
         return ret
 
     def astype(self, data_type: DataType) -> LazyTensor:
@@ -581,7 +627,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
 
     if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
         # Transformers models put different tensors in different files, but
-        # don't split indivdual tensors between files.
+        # don't split individual tensors between files.
         model: LazyModel = {}
         for mp in models_plus:
             model.update(mp.model)
@@ -596,6 +642,7 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe
         return lazy_tensor.load().permute(n_head, n_head_kv)
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
+
 def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
         return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
@@ -603,6 +650,7 @@ def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_
     s[0] = s[0] // 3
     return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
+
 def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     def load() -> Tensor:
         return lazy_tensor.load().part(n_part)
@@ -672,7 +720,7 @@ class LazyUnpickler(pickle.Unpickler):
         return func(*args)
 
     CLASSES: dict[tuple[str, str], Any] = {
-        # getattr used here as a workaround for mypy not being smart enough to detrmine
+        # getattr used here as a workaround for mypy not being smart enough to determine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
         ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
@@ -698,6 +746,7 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
                               data_base_path=pickle_paths[0][:-4],
                               zip_file=zf)
     model = unpickler.load()
+    if 'model' in model: model = model['model']
     as_dict = dict(model.items())
     return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None)
 
@@ -751,6 +800,7 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
+
 def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
@@ -785,26 +835,34 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
                     break
             yield result
 
-def check_vocab_size(params: Params, vocab: Vocab) -> None:
+
+def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
     if params.n_vocab != vocab.vocab_size:
-        assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
-        if params.n_vocab == vocab.vocab_size_base:
+        if params.n_vocab == vocab.vocab_size:
             print("Ignoring added_tokens.json since model matches vocab size without it.")
-            vocab.added_tokens_list = []
-            vocab.vocab_size = vocab.vocab_size_base
+            vocab.added_tokens_dict = OrderedDict()
+            vocab.vocab_size = vocab.vocab_size
+            return
+
+        if pad_vocab and params.n_vocab > vocab.vocab_size:
+            pad_count = params.n_vocab - vocab.vocab_size
+            print(f'Padding vocab with {pad_count} token(s) - <dummy00001> through <dummy{pad_count:05}>')
+            for i in range(1, (params.n_vocab - vocab.vocab_size) + 1):
+                vocab.added_tokens_dict[f'<dummy{i:05}>'] = -1
+            vocab.vocab_size = params.n_vocab
             return
         msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer}"
-        if vocab.fname_added_tokens is not None:
-            msg += f" combined with {vocab.fname_added_tokens}"
         msg += f" has {vocab.vocab_size})."
-        if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20 and vocab.fname_added_tokens is None:
+        if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
             msg += f"  Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
+        if vocab.vocab_size < params.n_vocab:
+            msg += " Possibly try using the --padvocab option."
         raise Exception(msg)
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path) -> None:
-        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
+        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
     def add_meta_arch(self, params: Params) -> None:
         name = "LLaMA"
@@ -823,13 +881,31 @@ class OutputFile:
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
-        self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
+
+        if params.n_experts:
+            self.gguf.add_expert_count(params.n_experts)
+
+        if params.n_experts_used:
+            self.gguf.add_expert_used_count(params.n_experts_used)
+
+        if params.f_norm_eps:
+            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
+        else:
+            raise ValueError('f_norm_eps is None')
 
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
-        if params.f_rope_scale is not None:
-            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+        if params.rope_scaling_type:
+            assert params.f_rope_scale is not None
+            self.gguf.add_rope_scaling_type(params.rope_scaling_type)
+            self.gguf.add_rope_scaling_factor(params.f_rope_scale)
+
+        if params.n_orig_ctx is not None:
+            self.gguf.add_rope_scaling_orig_ctx_len(params.n_orig_ctx)
+
+        if params.rope_finetuned is not None:
+            self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
 
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
@@ -844,12 +920,8 @@ class OutputFile:
             scores.append(score)
             toktypes.append(toktype)
 
-        if isinstance(vocab, SentencePieceVocab):
-            self.gguf.add_tokenizer_model("llama")
-        elif isinstance(vocab, BpeVocab):
-            self.gguf.add_tokenizer_model("gpt2")
-        else:
-            raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab')
+        vocab_type = vocab.get_vocab_type()
+        self.gguf.add_tokenizer_model(vocab_type)
         self.gguf.add_token_list(tokens)
         self.gguf.add_token_scores(scores)
         self.gguf.add_token_types(toktypes)
@@ -875,10 +947,14 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
-        check_vocab_size(params, vocab)
+    def write_vocab_only(
+        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
+        pad_vocab: bool = False,
+    ) -> None:
+        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
 
-        of = OutputFile(fname_out)
+        of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
         of.add_meta_arch(params)
@@ -903,10 +979,15 @@ class OutputFile:
         return dt.quantize(arr)
 
     @staticmethod
-    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
-        check_vocab_size(params, vocab)
+    def write_all(
+        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
+        concurrency: int = DEFAULT_CONCURRENCY,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
+        pad_vocab: bool = False,
+    ) -> None:
+        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
 
-        of = OutputFile(fname_out)
+        of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
         of.add_meta_arch(params)
@@ -937,8 +1018,9 @@ class OutputFile:
 
         of.close()
 
+
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
-    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
     if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32
@@ -951,10 +1033,12 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
 
     raise Exception(f"Unexpected combination of types: {name_to_type}")
 
+
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
     return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
             for (name, tensor) in model.items()}
 
+
 def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
     should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
@@ -967,7 +1051,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
             print(f"Permuting layer {i}")
             tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head)
             tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv)
-           #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
+            # tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             print(f"Unpacking and permuting layer {i}")
             tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
@@ -992,6 +1076,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
 
     return out
 
+
 def nth_multifile_path(path: Path, n: int) -> Path | None:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the nth path in the model.
@@ -1036,7 +1121,8 @@ def load_some_model(path: Path) -> ModelPlus:
     # Be extra-friendly and accept either a file or a directory:
     if path.is_dir():
         # Check if it's a set of safetensors files first
-        files = list(path.glob("model-00001-of-*.safetensors"))
+        globs = ["model-00001-of-*.safetensors", "model.safetensors"]
+        files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try the PyTorch patterns too, with lower priority
             globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
@@ -1057,35 +1143,17 @@ def load_some_model(path: Path) -> ModelPlus:
     return model_plus
 
 
-def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
-    # Be extra-friendly and accept either a file or a directory.  Also, if it's
-    # a directory, it might be the model directory, and tokenizer.model might
-    # be in the parent of that.
-    if path.is_dir():
-        vocab_file = "tokenizer.model"
-        if vocabtype == 'bpe':
-            vocab_file = "vocab.json"
-        path2 = path / vocab_file
-        # Use `.parent` instead of /.. to handle the symlink case better.
-        path3 = path.parent / vocab_file
-        if path2.exists():
-            path = path2
-        elif path3.exists():
-            path = path3
-        else:
-            raise FileNotFoundError(
-                f"Could not find {vocab_file} in {path} or its parent; "
-                "if it's in another directory, pass the directory as --vocab-dir")
+def find_vocab_file_path(path: Path, vocab_file: str) -> Optional[Path]:
+    path2 = path / vocab_file
+    # Use `.parent` instead of /.. to handle the symlink case better.
+    path3 = path.parent / vocab_file
 
-    print(f"Loading vocab file '{path}', type '{vocabtype}'")
+    if path2.exists():
+        return path2
+    if path3.exists():
+        return path3
 
-    added_tokens_path = path.parent / "added_tokens.json"
-    if vocabtype == "bpe":
-        return BpeVocab(path, added_tokens_path if added_tokens_path.exists() else None)
-    elif vocabtype == "spm":
-        return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
-    else:
-        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
+    return None
 
 
 def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
@@ -1112,19 +1180,24 @@ def do_dump_model(model_plus: ModelPlus) -> None:
 
 
 def main(args_in: list[str] | None = None) -> None:
+    output_choices = ["f32", "f16"]
+    if np.uint32(1) == np.uint32(1).newbyteorder("<"):
+        # We currently only support Q8_0 output on little endian systems.
+        output_choices.append("q8_0")
     parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
     parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
+    parser.add_argument("--outtype",     choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--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("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
     parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
-    args = parser.parse_args(args_in)
+    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--padvocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
 
+    args = parser.parse_args(args_in)
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
@@ -1138,6 +1211,9 @@ def main(args_in: list[str] | None = None) -> None:
     if args.dump:
         do_dump_model(model_plus)
         return
+    endianess = gguf.GGUFEndian.LITTLE
+    if args.bigendian:
+        endianess = gguf.GGUFEndian.BIG
 
     params = Params.load(model_plus)
     if params.n_ctx == -1:
@@ -1159,12 +1235,16 @@ def main(args_in: list[str] | None = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        assert args.outfile, "need --outfile if using --vocab-only"
+        if not args.outfile:
+            raise ValueError("need --outfile if using --vocab-only")
         # FIXME: Try to respect vocab_dir somehow?
-        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
-        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+        vocab = VocabLoader(params, args.vocab_dir or args.model)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+                                          load_merges = True,
+                                          n_vocab = vocab.vocab_size)
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
+                                    endianess = endianess, pad_vocab = args.padvocab)
         print(f"Wrote {outfile}")
         return
 
@@ -1172,10 +1252,15 @@ def main(args_in: list[str] | None = None) -> None:
         vocab = model_plus.vocab
     else:
         vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-        vocab = load_vocab(vocab_dir, args.vocabtype)
-    # FIXME: Try to respect vocab_dir somehow?
-    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+        vocab = VocabLoader(params, vocab_dir)
 
+    # FIXME: Try to respect vocab_dir somehow?
+    print(f"Vocab info: {vocab}")
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+                                      load_merges = True,
+                                      n_vocab = vocab.vocab_size)
+
+    print(f"Special vocab info: {special_vocab}")
     model   = model_plus.model
     model   = convert_model_names(model, params)
     ftype   = pick_output_type(model, args.outtype)
@@ -1185,7 +1270,8 @@ def main(args_in: list[str] | None = None) -> None:
     params.ftype = ftype
     print(f"Writing {outfile}, format {ftype}")
 
-    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
+                         concurrency = args.concurrency, endianess = endianess, pad_vocab = args.padvocab)
     print(f"Wrote {outfile}")
 
 

docs/token_generation_performance_tips.md

@@ -17,7 +17,7 @@ llama_model_load_internal: [cublas] total VRAM used: 17223 MB
 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.
+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 physical 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:

examples/CMakeLists.txt

@@ -12,26 +12,28 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(main)
-    add_subdirectory(quantize)
-    add_subdirectory(quantize-stats)
-    add_subdirectory(perplexity)
-    add_subdirectory(embedding)
-    add_subdirectory(save-load-state)
-    add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
-    add_subdirectory(train-text-from-scratch)
-    add_subdirectory(finetune)
-    add_subdirectory(convert-llama2c-to-ggml)
-    add_subdirectory(simple)
     add_subdirectory(batched)
     add_subdirectory(batched-bench)
-    add_subdirectory(speculative)
-    add_subdirectory(parallel)
-    add_subdirectory(embd-input)
-    add_subdirectory(llava)
-    add_subdirectory(llama-bench)
     add_subdirectory(beam-search)
+    add_subdirectory(benchmark)
+    add_subdirectory(convert-llama2c-to-ggml)
+    add_subdirectory(embedding)
+    add_subdirectory(finetune)
+    add_subdirectory(infill)
+    add_subdirectory(llama-bench)
+    add_subdirectory(llava)
+    add_subdirectory(main)
+    add_subdirectory(tokenize)
+    add_subdirectory(parallel)
+    add_subdirectory(perplexity)
+    add_subdirectory(quantize)
+    add_subdirectory(quantize-stats)
+    add_subdirectory(save-load-state)
+    add_subdirectory(simple)
+    add_subdirectory(speculative)
+    add_subdirectory(lookahead)
+    add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -1258,9 +1258,9 @@ static struct ggml_tensor * forward_lora(
 }
 
 static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
-    assert(logits->n_dims == 2);
-    assert(probs->n_dims == 2);
-    assert(best_samples->n_dims == 1);
+    assert(ggml_is_matrix(logits));
+    assert(ggml_is_matrix(probs));
+    assert(ggml_is_vector(best_samples));
     assert(logits->ne[1] == best_samples->ne[0]);
     assert(logits->ne[0] == probs->ne[0]);
     assert(logits->ne[1] == probs->ne[1]);
@@ -1292,9 +1292,9 @@ static void sample_softmax_batch(
     struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
     struct ggml_tensor * best_samples
 ) {
-    GGML_ASSERT(best_samples->n_dims == 2);
-    GGML_ASSERT(logits->n_dims == 3);
-    GGML_ASSERT(probs->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(best_samples));
+    GGML_ASSERT(ggml_is_3d(logits));
+    GGML_ASSERT(ggml_is_3d(probs));
     int n_tokens = best_samples->ne[0];
     int n_batch  = best_samples->ne[1];
     int n_vocab  = logits->ne[0];
@@ -1334,7 +1334,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
@@ -1386,8 +1386,8 @@ static void get_example_targets(int example_id, struct ggml_tensor * tokens_inpu
 static void get_example_targets_batch(
     struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
 ) {
-    GGML_ASSERT(tokens_input->n_dims == 2);
-    GGML_ASSERT(     targets->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(targets));
     int n_tokens = tokens_input->ne[0];
     int n_batch  = tokens_input->ne[1];
     GGML_ASSERT(n_tokens == targets->ne[1]);

examples/batched-bench/batched-bench.cpp

@@ -114,7 +114,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_batch batch = llama_batch_init(n_kv_max, 0);
+    llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
 
     // decode in batches of ctx_params.n_batch tokens
     auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
@@ -123,11 +123,12 @@ int main(int argc, char ** argv) {
 
             llama_batch batch_view = {
                 n_tokens,
-                batch.token  + i,
+                batch.token    + i,
                 nullptr,
-                batch.pos    + i,
-                batch.seq_id + i,
-                batch.logits + i,
+                batch.pos      + i,
+                batch.n_seq_id + i,
+                batch.seq_id   + i,
+                batch.logits   + i,
                 0, 0, 0, // unused
             };
 
@@ -143,13 +144,8 @@ int main(int argc, char ** argv) {
 
     // warm up
     {
-        batch.n_tokens = 16;
-
-        for (int i = 0; i < batch.n_tokens; ++i) {
-            batch.token[i]  = 0;
-            batch.pos[i]    = i;
-            batch.seq_id[i] = 0;
-            batch.logits[i] = false;
+        for (int i = 0; i < 16; ++i) {
+            llama_batch_add(batch, 0, i, { 0 }, false);
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -158,6 +154,10 @@ int main(int argc, char ** argv) {
         }
     }
 
+    LOG_TEE("\n");
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("\n");
+
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
     LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
 
@@ -174,19 +174,18 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                batch.n_tokens = is_pp_shared ? pp : pl*pp;
+                llama_batch_clear(batch);
 
-                for (int i = 0; i < batch.n_tokens; ++i) {
-                    batch.token[i]  = 0;
-                    batch.pos[i]    = i;
-                    batch.seq_id[i] = 0;
-                    batch.logits[i] = false;
+                const int n_tokens = is_pp_shared ? pp : pl*pp;
+
+                for (int i = 0; i < n_tokens; ++i) {
+                    llama_batch_add(batch, 0, i, { 0 }, false);
                 }
                 batch.logits[batch.n_tokens - 1] = true;
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_tokens_rm(ctx, -1, -1);
+                llama_kv_cache_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_TEE("%s: llama_decode() failed\n", __func__);
@@ -204,13 +203,10 @@ int main(int argc, char ** argv) {
                 const auto t_tg_start = ggml_time_us();
 
                 for (int i = 0; i < tg; ++i) {
-                    batch.n_tokens = pl;
+                    llama_batch_clear(batch);
 
                     for (int j = 0; j < pl; ++j) {
-                        batch.token[j]  = 0;
-                        batch.pos[j]    = pp + i;
-                        batch.seq_id[j] = j;
-                        batch.logits[j] = true;
+                        llama_batch_add(batch, 0, pp + i, { j }, true);
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {

examples/batched.swift/README.md

@@ -1,4 +1,4 @@
 This is a swift clone of `examples/batched`.
 
 $ `make`
-$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`
+$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched.swift/Sources/main.swift

@@ -69,7 +69,7 @@ for id: llama_token in tokens {
 
 print("\n")
 
-var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0)
+var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0, 1)
 defer {
     llama_batch_free(batch)
 }
@@ -80,7 +80,12 @@ batch.n_tokens = Int32(tokens.count)
 for (i, token) in tokens.enumerated() {
     batch.token[i] = token
     batch.pos[i] = Int32(i)
-    batch.seq_id[i] = 0
+    batch.n_seq_id[i] = 1
+    // batch.seq_id[i][0] = 0
+    // TODO: is this the proper way to do this?
+    if let seq_id = batch.seq_id[i] {
+        seq_id[0] = 0
+    }
     batch.logits[i] = 0
 }
 
@@ -148,7 +153,7 @@ while n_cur <= n_len {
         // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
         // is it an end of stream? -> mark the stream as finished
-        if new_token_id == llama_token_eos(context) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {
@@ -169,7 +174,10 @@ while n_cur <= n_len {
         // push this new token for next evaluation
         batch.token[Int(batch.n_tokens)] = new_token_id
         batch.pos[Int(batch.n_tokens)] = n_cur
-        batch.seq_id[Int(batch.n_tokens)] = Int32(i)
+        batch.n_seq_id[Int(batch.n_tokens)] = 1
+        if let seq_id = batch.seq_id[Int(batch.n_tokens)] {
+            seq_id[0] = Int32(i)
+        }
         batch.logits[Int(batch.n_tokens)] = 1
 
         i_batch[i] = batch.n_tokens
@@ -207,9 +215,10 @@ print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end
 llama_print_timings(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
-    let n_tokens = text.count + (add_bos ? 1 : 0)
+    let utf8Count = text.utf8.count
+    let n_tokens = utf8Count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
+    let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])
@@ -222,18 +231,15 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
     var result = [CChar](repeating: 0, count: 8)
     let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
     if nTokens < 0 {
-        if result.count >= -Int(nTokens) {
-            result.removeLast(-Int(nTokens))
-        } else {
-            result.removeAll()
-        }
+        let actualTokensCount = -Int(nTokens)
+        result = .init(repeating: 0, count: actualTokensCount)
         let check = llama_token_to_piece(
             model,
             token,
             &result,
             Int32(result.count)
         )
-        assert(check == nTokens)
+        assert(check == actualTokensCount)
     } else {
         result.removeLast(result.count - Int(nTokens))
     }
@@ -251,5 +257,4 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
         buffer = []
         return bufferString
     }
-    return nil
 }

examples/batched/batched.cpp

@@ -11,12 +11,19 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
         return 1 ;
     }
 
+    // number of parallel batches
     int n_parallel = 1;
 
+    // total length of the sequences including the prompt
+    int n_len = 32;
+
+    // number of layers to offload to the GPU
+    int n_gpu_layers = 0;
+
     if (argc >= 2) {
         params.model = argv[1];
     }
@@ -29,13 +36,18 @@ int main(int argc, char ** argv) {
         n_parallel = std::atoi(argv[3]);
     }
 
+    if (argc >= 5) {
+        n_len = std::atoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        n_gpu_layers = std::atoi(argv[5]);
+    }
+
     if (params.prompt.empty()) {
         params.prompt = "Hello my name is";
     }
 
-    // total length of the sequences including the prompt
-    const int n_len = 32;
-
     // init LLM
 
     llama_backend_init(params.numa);
@@ -44,7 +56,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+    model_params.n_gpu_layers = n_gpu_layers;
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -97,20 +109,15 @@ int main(int argc, char ** argv) {
 
     fflush(stderr);
 
-    // create a llama_batch with size 512
+    // create a llama_batch
     // we use this object to submit token data for decoding
-
-    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0);
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
 
     // evaluate the initial prompt
-    batch.n_tokens = tokens_list.size();
-
-    for (int32_t i = 0; i < batch.n_tokens; i++) {
-        batch.token[i]  = tokens_list[i];
-        batch.pos[i]    = i;
-        batch.seq_id[i] = 0;
-        batch.logits[i] = false;
+    for (size_t i = 0; i < tokens_list.size(); ++i) {
+        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
     }
+    GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
     // llama_decode will output logits only for the last token of the prompt
     batch.logits[batch.n_tokens - 1] = true;
@@ -146,7 +153,7 @@ int main(int argc, char ** argv) {
 
     while (n_cur <= n_len) {
         // prepare the next batch
-        batch.n_tokens = 0;
+        llama_batch_clear(batch);
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -180,7 +187,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream? -> mark the stream as finished
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {
@@ -198,15 +205,10 @@ int main(int argc, char ** argv) {
 
             streams[i] += llama_token_to_piece(ctx, new_token_id);
 
-            // push this new token for next evaluation
-            batch.token [batch.n_tokens] = new_token_id;
-            batch.pos   [batch.n_tokens] = n_cur;
-            batch.seq_id[batch.n_tokens] = i;
-            batch.logits[batch.n_tokens] = true;
-
             i_batch[i] = batch.n_tokens;
 
-            batch.n_tokens += 1;
+            // push this new token for next evaluation
+            llama_batch_add(batch, new_token_id, n_cur, { i }, true);
 
             n_decode += 1;
         }

examples/beam-search/beam-search.cpp

@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
 }
 
 // Function matching type llama_beam_search_callback_fn_t.

examples/benchmark/CMakeLists.txt

@@ -1,9 +1,6 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/benchmark/benchmark-matmult.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 
@@ -130,13 +129,13 @@ int main(int argc, char ** argv)  {
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
     size_t ctx_size = 0;
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
-    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
+    ctx_size += ggml_row_size(qtype,         sizex*sizey);
+    ctx_size += ggml_row_size(qtype,         sizex*sizey);
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
     ctx_size += 1024*1024*16;
 
     printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
@@ -172,7 +171,8 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
 
     // printf("Creating compute graph\n");
-    struct ggml_cgraph gf = ggml_build_forward(m11xm2);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand(gf, m11xm2);
 
     printf("n_threads=%i\n", benchmark_params.n_threads);
 
@@ -181,9 +181,9 @@ int main(int argc, char ** argv)  {
 
     std::vector<uint8_t> work_buffer;
 
-    ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads);
+    ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
 
-    TENSOR_DUMP(gf.nodes[0]);
+    TENSOR_DUMP(gf->nodes[0]);
 
     printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
@@ -201,7 +201,8 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
 
     // printf("Creating compute graph\n");
-    struct ggml_cgraph gf31 = ggml_build_forward(q31);
+    struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
+    ggml_build_forward_expand(gf31, q31);
 
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
@@ -212,7 +213,8 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
 
     //printf("Creating compute graph\n");
-    struct ggml_cgraph gf32 = ggml_build_forward(q32);
+    struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
+    ggml_build_forward_expand(gf32, q32);
     printf("n_threads=%i\n", benchmark_params.n_threads);
 
     const int dimx = sizex;
@@ -224,7 +226,7 @@ int main(int argc, char ** argv)  {
 
 
     // Let's use the F32 result from above as a reference for the quantized multiplication
-    float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
+    float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
     printf("=====================================================================================\n");
@@ -234,7 +236,7 @@ int main(int argc, char ** argv)  {
 
         long long int start = ggml_time_us();
         //printf("Running ggml_graph_compute\n");
-        ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads);
+        ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads);
 
         long long int stop = ggml_time_us();
         long long int usec = stop-start;
@@ -252,7 +254,7 @@ int main(int argc, char ** argv)  {
 
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
-        float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
+        float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
         float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 
@@ -267,7 +269,7 @@ int main(int argc, char ** argv)  {
         }
 
         // Running a different graph computation to make sure we override the CPU cache lines
-        ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads);
+        ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads);
     }
     printf("\n");
     printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));

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

@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
@@ -536,7 +536,7 @@ static bool is_ggml_file(const char * filename) {
     if (file.size < 4) {
         return false;
     }
-    uint32_t magic = file.read_u32();
+    std::string magic = file.read_string(4);
     return magic == GGUF_MAGIC;
 }
 
@@ -639,7 +639,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
     int ct;
-    switch (gg_weights->n_dims){
+    switch (ggml_n_dims(gg_weights)) {
         case 1:
             ct = 0;
             for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){

examples/embd-input/.gitignore

@@ -1,4 +0,0 @@
-PandaGPT
-MiniGPT-4
-*.pth
-

examples/embd-input/CMakeLists.txt

@@ -1,17 +0,0 @@
-set(TARGET embdinput)
-add_library(${TARGET} embd-input-lib.cpp embd-input.h)
-install(TARGETS ${TARGET} LIBRARY)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
-
-set(TARGET embd-input-test)
-add_executable(${TARGET} embd-input-test.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama embdinput ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/embd-input/README.md

@@ -1,63 +0,0 @@
-### Examples for input embedding directly
-
-## Requirement
-build  `libembdinput.so`
-run the following comman in main dir (../../).
-```
-make
-```
-
-## [LLaVA](https://github.com/haotian-liu/LLaVA/) example  (llava.py)
-
-1. Obtian LLaVA model (following https://github.com/haotian-liu/LLaVA/ , use https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/).
-2. Convert it to ggml format.
-3. `llava_projection.pth` is [pytorch_model-00003-of-00003.bin](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin).
-
-```
-import torch
-
-bin_path = "../LLaVA-13b-delta-v1-1/pytorch_model-00003-of-00003.bin"
-pth_path = "./examples/embd-input/llava_projection.pth"
-
-dic = torch.load(bin_path)
-used_key = ["model.mm_projector.weight","model.mm_projector.bias"]
-torch.save({k: dic[k] for k in used_key}, pth_path)
-```
-4. Check the path of LLaVA model and `llava_projection.pth` in `llava.py`.
-
-
-## [PandaGPT](https://github.com/yxuansu/PandaGPT) example (panda_gpt.py)
-
-1. Obtian PandaGPT lora model from https://github.com/yxuansu/PandaGPT. Rename the file to `adapter_model.bin`. Use [convert-lora-to-ggml.py](../../convert-lora-to-ggml.py) to convert it to ggml format.
-The `adapter_config.json` is
-```
-{
-  "peft_type": "LORA",
-  "fan_in_fan_out": false,
-  "bias": null,
-  "modules_to_save": null,
-  "r": 32,
-  "lora_alpha": 32,
-  "lora_dropout": 0.1,
-  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"]
-}
-```
-2. Papare the `vicuna` v0 model.
-3. Obtain the [ImageBind](https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth) model.
-4. Clone the PandaGPT source.
-```
-git clone https://github.com/yxuansu/PandaGPT
-```
-5. Install the requirement of PandaGPT.
-6. Check the path of PandaGPT source, ImageBind model, lora model and vicuna model in panda_gpt.py.
-
-## [MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4/) example (minigpt4.py)
-
-1. Obtain MiniGPT-4 model from https://github.com/Vision-CAIR/MiniGPT-4/ and put it in `embd-input`.
-2. Clone the MiniGPT-4 source.
-```
-git clone https://github.com/Vision-CAIR/MiniGPT-4/
-```
-3. Install the requirement of PandaGPT.
-4. Papare the `vicuna` v0 model.
-5. Check the path of MiniGPT-4 source, MiniGPT-4 model and vicuna model in `minigpt4.py`.

examples/embd-input/embd-input-lib.cpp

@@ -1,221 +0,0 @@
-#include "build-info.h"
-#include "common.h"
-#include "embd-input.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <ctime>
-#include <fstream>
-#include <iostream>
-#include <string>
-#include <vector>
-
-static llama_context ** g_ctx;
-
-extern "C" {
-
-struct MyModel* create_mymodel(int argc, char ** argv) {
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        return nullptr;
-    }
-
-    print_build_info();
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = uint32_t(time(NULL));
-    }
-    fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
-
-    llama_backend_init(params.numa);
-
-    llama_model * model;
-    llama_context * ctx;
-
-    g_ctx = &ctx;
-
-    // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
-        return nullptr;
-    }
-
-    // print system information
-    {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
-    }
-    struct MyModel * ret = new MyModel();
-    ret->ctx = ctx;
-    ret->params = params;
-    ret->n_past = 0;
-    // printf("ctx: %d\n", ret->ctx);
-    return ret;
-}
-
-void free_mymodel(struct MyModel * mymodel) {
-    llama_context * ctx = mymodel->ctx;
-    llama_print_timings(ctx);
-    llama_free(ctx);
-    delete mymodel;
-}
-
-
-bool eval_float(void * model, float * input, int N){
-    MyModel * mymodel = (MyModel*)model;
-    llama_context * ctx = mymodel->ctx;
-    gpt_params params = mymodel->params;
-    int n_emb = llama_n_embd(llama_get_model(ctx));
-    int n_past = mymodel->n_past;
-    int n_batch = N; // params.n_batch;
-
-    for (int i = 0; i < (int) N; i += n_batch) {
-        int n_eval = (int) N - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        llama_batch batch = {  int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
-        if (llama_decode(ctx, batch)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        n_past += n_eval;
-    }
-    mymodel->n_past = n_past;
-    return true;
-}
-
-bool eval_tokens(void * model, std::vector<llama_token> tokens) {
-    MyModel * mymodel = (MyModel* )model;
-    llama_context * ctx;
-    ctx = mymodel->ctx;
-    gpt_params params = mymodel->params;
-    int n_past = mymodel->n_past;
-    for (int i = 0; i < (int) tokens.size(); i += params.n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > params.n_batch) {
-            n_eval = params.n_batch;
-        }
-        if (llama_decode(ctx, llama_batch_get_one(&tokens[i], n_eval, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        n_past += n_eval;
-    }
-    mymodel->n_past = n_past;
-    return true;
-}
-
-bool eval_id(struct MyModel* mymodel, int id) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(mymodel, tokens);
-}
-
-bool eval_string(struct MyModel * mymodel,const char* str){
-    llama_context * ctx = mymodel->ctx;
-    std::string str2 = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx, str2, true);
-    eval_tokens(mymodel, embd_inp);
-    return true;
-}
-
-llama_token sampling_id(struct MyModel* mymodel) {
-    llama_context* ctx = mymodel->ctx;
-    gpt_params params = mymodel->params;
-    llama_sampling_params & sparams = params.sampling_params;
-    // int n_ctx = llama_n_ctx(ctx);
-
-    // out of user input, sample next token
-    const float   temp            = sparams.temp;
-    const int32_t top_k           = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : sparams.top_k;
-    const float   top_p           = sparams.top_p;
-    const float   tfs_z           = sparams.tfs_z;
-    const float   typical_p       = sparams.typical_p;
-    // const int32_t repeat_last_n   = params.repeat_last_n < 0 ? 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        = sparams.mirostat;
-    const float   mirostat_tau    = sparams.mirostat_tau;
-    const float   mirostat_eta    = sparams.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(llama_get_model(ctx));
-
-        // Apply params.logit_bias map
-        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
-            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 };
-
-        // TODO: Apply penalties
-        // float nl_logit = logits[llama_token_nl(ctx)];
-        // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), 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(ctx)] = 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_temp(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_temp(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_temp(ctx, &candidates_p, temp);
-                id = llama_sample_token(ctx, &candidates_p);
-            }
-        }
-    }
-
-    return id;
-}
-
-const char * sampling(struct MyModel * mymodel) {
-    llama_context * ctx = mymodel->ctx;
-    int id = sampling_id(mymodel);
-    static std::string ret;
-    if (id == llama_token_eos(ctx)) {
-        ret = "</s>";
-    } else {
-        ret = llama_token_to_piece(ctx, id);
-    }
-    eval_id(mymodel, id);
-    return ret.c_str();
-}
-
-}

examples/embd-input/embd-input-test.cpp

@@ -1,35 +0,0 @@
-#include "embd-input.h"
-#include <stdlib.h>
-#include <random>
-#include <string.h>
-
-int main(int argc, char** argv) {
-
-    auto mymodel = create_mymodel(argc, argv);
-    int N = 10;
-    int max_tgt_len = 500;
-    int n_embd = llama_n_embd(llama_get_model(mymodel->ctx));
-
-    // add random float embd to test evaluation
-    float * data = new float[N*n_embd];
-    std::default_random_engine e;
-    std::uniform_real_distribution<float>  u(0,1);
-    for (int i=0;i<N*n_embd;i++) {
-        data[i] = u(e);
-    }
-
-    eval_string(mymodel, "user: what is the color of the flag of UN?");
-    eval_float(mymodel, data, N);
-    eval_string(mymodel, "assistant:");
-    eval_string(mymodel, mymodel->params.prompt.c_str());
-    const char* tmp;
-    for (int i=0; i<max_tgt_len; i++) {
-        tmp = sampling(mymodel);
-        if (strcmp(tmp, "</s>")==0) break;
-        printf("%s", tmp);
-        fflush(stdout);
-    }
-    printf("\n");
-    free_mymodel(mymodel);
-    return 0;
-}

examples/embd-input/embd-input.h

@@ -1,27 +0,0 @@
-#ifndef _EMBD_INPUT_H_
-#define _EMBD_INPUT_H_ 1
-
-#include "common.h"
-#include "llama.h"
-
-extern "C" {
-
-typedef struct MyModel {
-    llama_context* ctx;
-    gpt_params params;
-    int n_past = 0;
-} MyModel;
-
-struct MyModel* create_mymodel(int argc, char ** argv);
-
-bool eval_float(void* model, float* input, int N);
-bool eval_tokens(void* model, std::vector<llama_token> tokens);
-bool eval_id(struct MyModel* mymodel, int id);
-bool eval_string(struct MyModel* mymodel, const char* str);
-const char * sampling(struct MyModel* mymodel);
-llama_token sampling_id(struct MyModel* mymodel);
-void free_mymodel(struct MyModel* mymodel);
-
-}
-
-#endif

examples/embd-input/embd_input.py

@@ -1,72 +0,0 @@
-#!/usr/bin/env python3
-import ctypes
-from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
-import numpy as np
-import os
-
-libc = cdll.LoadLibrary("./libembdinput.so")
-libc.sampling.restype=c_char_p
-libc.create_mymodel.restype=c_void_p
-libc.eval_string.argtypes=[c_void_p, c_char_p]
-libc.sampling.argtypes=[c_void_p]
-libc.eval_float.argtypes=[c_void_p, POINTER(c_float), c_int]
-
-
-class MyModel:
-    def __init__(self, args):
-        argc = len(args)
-        c_str = [c_char_p(i.encode()) for i in args]
-        args_c = (c_char_p * argc)(*c_str)
-        self.model = c_void_p(libc.create_mymodel(argc, args_c))
-        self.max_tgt_len = 512
-        self.print_string_eval = True
-
-    def __del__(self):
-        libc.free_mymodel(self.model)
-
-    def eval_float(self, x):
-        libc.eval_float(self.model, x.astype(np.float32).ctypes.data_as(POINTER(c_float)), x.shape[1])
-
-    def eval_string(self, x):
-        libc.eval_string(self.model, x.encode()) # c_char_p(x.encode()))
-        if self.print_string_eval:
-            print(x)
-
-    def eval_token(self, x):
-        libc.eval_id(self.model, x)
-
-    def sampling(self):
-        s = libc.sampling(self.model)
-        return s
-
-    def stream_generate(self, end="</s>"):
-        ret = b""
-        end = end.encode()
-        for _ in range(self.max_tgt_len):
-            tmp = self.sampling()
-            ret += tmp
-            yield tmp
-            if ret.endswith(end):
-                break
-
-    def generate_with_print(self, end="</s>"):
-        ret = b""
-        for i in self.stream_generate(end=end):
-            ret += i
-            print(i.decode(errors="replace"), end="", flush=True)
-        print("")
-        return ret.decode(errors="replace")
-
-
-    def generate(self, end="</s>"):
-        text = b"".join(self.stream_generate(end=end))
-        return text.decode(errors="replace")
-
-if __name__ == "__main__":
-    model = MyModel(["main", "--model", "../llama.cpp/models/ggml-vic13b-q4_1.bin", "-c", "2048"])
-    model.eval_string("""user: what is the color of the flag of UN?""")
-    x = np.random.random((5120,10))# , dtype=np.float32)
-    model.eval_float(x)
-    model.eval_string("""assistant:""")
-    for i in model.generate():
-        print(i.decode(errors="replace"), end="", flush=True)

examples/embd-input/llava.py

@@ -1,71 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import os
-sys.path.insert(0, os.path.dirname(__file__))
-from embd_input import MyModel
-import numpy as np
-from torch import nn
-import torch
-from transformers import CLIPVisionModel,  CLIPImageProcessor
-from PIL import Image
-
-# model parameters from 'liuhaotian/LLaVA-13b-delta-v1-1'
-vision_tower = "openai/clip-vit-large-patch14"
-select_hidden_state_layer = -2
-# (vision_config.image_size // vision_config.patch_size) ** 2
-image_token_len = (224//14)**2
-
-class Llava:
-    def __init__(self, args):
-        self.image_processor = CLIPImageProcessor.from_pretrained(vision_tower)
-        self.vision_tower = CLIPVisionModel.from_pretrained(vision_tower)
-        self.mm_projector = nn.Linear(1024, 5120)
-        self.model = MyModel(["main", *args])
-
-    def load_projection(self, path):
-        state = torch.load(path)
-        self.mm_projector.load_state_dict({
-            "weight": state["model.mm_projector.weight"],
-            "bias": state["model.mm_projector.bias"]})
-
-    def chat(self, question):
-        self.model.eval_string("user: ")
-        self.model.eval_string(question)
-        self.model.eval_string("\nassistant: ")
-        return self.model.generate_with_print()
-
-    def chat_with_image(self, image, question):
-        with torch.no_grad():
-            embd_image = self.image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
-            image_forward_out = self.vision_tower(embd_image.unsqueeze(0), output_hidden_states=True)
-            select_hidden_state = image_forward_out.hidden_states[select_hidden_state_layer]
-            image_feature = select_hidden_state[:, 1:]
-            embd_image = self.mm_projector(image_feature)
-            embd_image = embd_image.cpu().numpy()[0]
-        self.model.eval_string("user: ")
-        self.model.eval_token(32003-2) # im_start
-        self.model.eval_float(embd_image.T)
-        for i in range(image_token_len-embd_image.shape[0]):
-            self.model.eval_token(32003-3) # im_patch
-        self.model.eval_token(32003-1) # im_end
-        self.model.eval_string(question)
-        self.model.eval_string("\nassistant: ")
-        return self.model.generate_with_print()
-
-
-if __name__=="__main__":
-    # model form liuhaotian/LLaVA-13b-delta-v1-1
-    a = Llava(["--model", "./models/ggml-llava-13b-v1.1.bin", "-c", "2048"])
-    # Extract from https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin.
-    # Also here can use pytorch_model-00003-of-00003.bin directly.
-    a.load_projection(os.path.join(
-        os.path.dirname(__file__) ,
-        "llava_projection.pth"))
-    respose = a.chat_with_image(
-        Image.open("./media/llama1-logo.png").convert('RGB'),
-        "what is the text in the picture?")
-    respose
-    a.chat("what is the color of it?")
-
-
-

examples/embd-input/minigpt4.py

@@ -1,129 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import os
-sys.path.insert(0, os.path.dirname(__file__))
-from embd_input import MyModel
-import numpy as np
-from torch import nn
-import torch
-from PIL import Image
-
-minigpt4_path = os.path.join(os.path.dirname(__file__), "MiniGPT-4")
-sys.path.insert(0, minigpt4_path)
-from minigpt4.models.blip2 import Blip2Base
-from minigpt4.processors.blip_processors import Blip2ImageEvalProcessor
-
-
-class MiniGPT4(Blip2Base):
-    """
-    MiniGPT4 model from https://github.com/Vision-CAIR/MiniGPT-4
-    """
-    def __init__(self,
-        args,
-        vit_model="eva_clip_g",
-        q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth",
-        img_size=224,
-        drop_path_rate=0,
-        use_grad_checkpoint=False,
-        vit_precision="fp32",
-        freeze_vit=True,
-        freeze_qformer=True,
-        num_query_token=32,
-        llama_model="",
-        prompt_path="",
-        prompt_template="",
-        max_txt_len=32,
-        end_sym='\n',
-        low_resource=False,  # use 8 bit and put vit in cpu
-        device_8bit=0
-    ):
-        super().__init__()
-        self.img_size = img_size
-        self.low_resource = low_resource
-        self.preprocessor = Blip2ImageEvalProcessor(img_size)
-
-        print('Loading VIT')
-        self.visual_encoder, self.ln_vision = self.init_vision_encoder(
-            vit_model, img_size, drop_path_rate, use_grad_checkpoint, vit_precision
-        )
-        print('Loading VIT Done')
-        print('Loading Q-Former')
-        self.Qformer, self.query_tokens = self.init_Qformer(
-            num_query_token, self.visual_encoder.num_features
-        )
-        self.Qformer.cls = None
-        self.Qformer.bert.embeddings.word_embeddings = None
-        self.Qformer.bert.embeddings.position_embeddings = None
-        for layer in self.Qformer.bert.encoder.layer:
-            layer.output = None
-            layer.intermediate = None
-        self.load_from_pretrained(url_or_filename=q_former_model)
-        print('Loading Q-Former Done')
-        self.llama_proj = nn.Linear(
-            self.Qformer.config.hidden_size, 5120 # self.llama_model.config.hidden_size
-        )
-        self.max_txt_len = max_txt_len
-        self.end_sym = end_sym
-        self.model = MyModel(["main", *args])
-        # system prompt
-        self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
-           "You will be able to see the image once I provide it to you. Please answer my questions."
-           "###")
-
-    def encode_img(self, image):
-        image = self.preprocessor(image)
-        image = image.unsqueeze(0)
-        device = image.device
-        if self.low_resource:
-            self.vit_to_cpu()
-            image = image.to("cpu")
-
-        with self.maybe_autocast():
-            image_embeds = self.ln_vision(self.visual_encoder(image)).to(device)
-            image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(device)
-
-            query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
-            query_output = self.Qformer.bert(
-                query_embeds=query_tokens,
-                encoder_hidden_states=image_embeds,
-                encoder_attention_mask=image_atts,
-                return_dict=True,
-            )
-
-            inputs_llama = self.llama_proj(query_output.last_hidden_state)
-            # atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
-        return inputs_llama
-
-    def load_projection(self, path):
-        state = torch.load(path)["model"]
-        self.llama_proj.load_state_dict({
-            "weight": state["llama_proj.weight"],
-            "bias": state["llama_proj.bias"]})
-
-    def chat(self, question):
-        self.model.eval_string("Human: ")
-        self.model.eval_string(question)
-        self.model.eval_string("\n### Assistant:")
-        return self.model.generate_with_print(end="###")
-
-    def chat_with_image(self, image, question):
-        with torch.no_grad():
-            embd_image = self.encode_img(image)
-        embd_image = embd_image.cpu().numpy()[0]
-        self.model.eval_string("Human: <Img>")
-        self.model.eval_float(embd_image.T)
-        self.model.eval_string("</Img> ")
-        self.model.eval_string(question)
-        self.model.eval_string("\n### Assistant:")
-        return self.model.generate_with_print(end="###")
-
-
-if __name__=="__main__":
-    a = MiniGPT4(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048"])
-    a.load_projection(os.path.join(
-        os.path.dirname(__file__) ,
-        "pretrained_minigpt4.pth"))
-    respose = a.chat_with_image(
-        Image.open("./media/llama1-logo.png").convert('RGB'),
-        "what is the text in the picture?")
-    a.chat("what is the color of it?")

examples/embd-input/panda_gpt.py

@@ -1,99 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import os
-sys.path.insert(0, os.path.dirname(__file__))
-from embd_input import MyModel
-import numpy as np
-from torch import nn
-import torch
-
-# use PandaGPT path
-panda_gpt_path = os.path.join(os.path.dirname(__file__), "PandaGPT")
-imagebind_ckpt_path = "./models/panda_gpt/"
-
-sys.path.insert(0, os.path.join(panda_gpt_path,"code","model"))
-from ImageBind.models import imagebind_model
-from ImageBind import data
-
-ModalityType = imagebind_model.ModalityType
-max_tgt_len = 400
-
-class PandaGPT:
-    def __init__(self, args):
-        self.visual_encoder,_ = imagebind_model.imagebind_huge(pretrained=True, store_path=imagebind_ckpt_path)
-        self.visual_encoder.eval()
-        self.llama_proj = nn.Linear(1024, 5120) # self.visual_hidden_size, 5120)
-        self.max_tgt_len = max_tgt_len
-        self.model = MyModel(["main", *args])
-        self.generated_text = ""
-        self.device = "cpu"
-
-    def load_projection(self, path):
-        state = torch.load(path, map_location="cpu")
-        self.llama_proj.load_state_dict({
-            "weight": state["llama_proj.weight"],
-            "bias": state["llama_proj.bias"]})
-
-    def eval_inputs(self, inputs):
-        self.model.eval_string("<Img>")
-        embds = self.extract_multimoal_feature(inputs)
-        for i in embds:
-            self.model.eval_float(i.T)
-        self.model.eval_string("</Img> ")
-
-    def chat(self, question):
-        return self.chat_with_image(None, question)
-
-    def chat_with_image(self, inputs, question):
-        if self.generated_text == "":
-            self.model.eval_string("###")
-        self.model.eval_string(" Human: ")
-        if inputs:
-            self.eval_inputs(inputs)
-        self.model.eval_string(question)
-        self.model.eval_string("\n### Assistant:")
-        ret = self.model.generate_with_print(end="###")
-        self.generated_text += ret
-        return ret
-
-    def extract_multimoal_feature(self, inputs):
-        features = []
-        for key in ["image", "audio", "video", "thermal"]:
-            if key + "_paths" in inputs:
-                embeds = self.encode_data(key, inputs[key+"_paths"])
-                features.append(embeds)
-        return features
-
-    def encode_data(self, data_type, data_paths):
-
-        type_map = {
-            "image": ModalityType.VISION,
-            "audio": ModalityType.AUDIO,
-            "video": ModalityType.VISION,
-            "thermal": ModalityType.THERMAL,
-        }
-        load_map = {
-            "image": data.load_and_transform_vision_data,
-            "audio": data.load_and_transform_audio_data,
-            "video": data.load_and_transform_video_data,
-            "thermal": data.load_and_transform_thermal_data
-        }
-
-        load_function = load_map[data_type]
-        key = type_map[data_type]
-
-        inputs = {key: load_function(data_paths, self.device)}
-        with torch.no_grad():
-            embeddings = self.visual_encoder(inputs)
-            embeds = embeddings[key]
-            embeds = self.llama_proj(embeds).cpu().numpy()
-        return embeds
-
-
-if __name__=="__main__":
-    a = PandaGPT(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048", "--lora", "./models/panda_gpt/ggml-adapter-model.bin","--temp", "0"])
-    a.load_projection("./models/panda_gpt/adapter_model.bin")
-    a.chat_with_image(
-        {"image_paths": ["./media/llama1-logo.png"]},
-        "what is the text in the picture? 'llama' or 'lambda'?")
-    a.chat("what is the color of it?")

examples/embedding/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${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/embedding/embedding.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 

examples/export-lora/export-lora.cpp

@@ -240,7 +240,7 @@ static struct lora_data * load_lora(struct lora_info * info) {
     }
 
     struct ggml_init_params params_ggml;
-    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE;
     params_ggml.mem_buffer = NULL;
     params_ggml.no_alloc   = true;
     result->ctx = ggml_init(params_ggml);
@@ -334,7 +334,7 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
     float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
 
     struct ggml_init_params params;
-    params.mem_size   = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_size   = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
     params.mem_buffer = NULL;
     params.no_alloc   = true;
     struct ggml_context * ctx = NULL;

examples/finetune/README.md

@@ -21,7 +21,7 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
 ./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
 ```
 
-Finetune output files will be saved every N iterations (config with `--save-every N`).
+**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
 The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
 So in above example after 10 iterations these files will be written:
 - chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf

examples/finetune/convert-finetune-checkpoint-to-gguf.py

@@ -3,9 +3,7 @@
 
 import argparse
 import gguf
-import os
 import struct
-import sys
 import numpy as np
 from pathlib import Path
 

examples/finetune/finetune.cpp

@@ -548,35 +548,35 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
     struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
 
     randomize_tensor_normal(lora->tok_embeddings_a, rnd);
-    randomize_tensor_normal(lora->tok_embeddings_b, rnd);
+    ggml_set_zero(lora->tok_embeddings_b);
     randomize_tensor_normal(lora->norm_a,           rnd);
-    randomize_tensor_normal(lora->norm_b,           rnd);
+    ggml_set_zero(lora->norm_b);
     randomize_tensor_normal(lora->output_a,         rnd);
-    randomize_tensor_normal(lora->output_b,         rnd);
+    ggml_set_zero(lora->output_b);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = lora->layers[i];
         randomize_tensor_normal(layer.attention_norm_a, rnd);
-        randomize_tensor_normal(layer.attention_norm_b, rnd);
+        ggml_set_zero(layer.attention_norm_b);
 
         randomize_tensor_normal(layer.wq_a, rnd);
-        randomize_tensor_normal(layer.wq_b, rnd);
+        ggml_set_zero(layer.wq_b);
         randomize_tensor_normal(layer.wk_a, rnd);
-        randomize_tensor_normal(layer.wk_b, rnd);
+        ggml_set_zero(layer.wk_b);
         randomize_tensor_normal(layer.wv_a, rnd);
-        randomize_tensor_normal(layer.wv_b, rnd);
+        ggml_set_zero(layer.wv_b);
         randomize_tensor_normal(layer.wo_a, rnd);
-        randomize_tensor_normal(layer.wo_b, rnd);
+        ggml_set_zero(layer.wo_b);
 
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
-        randomize_tensor_normal(layer.ffn_norm_b, rnd);
+        ggml_set_zero(layer.ffn_norm_b);
 
         randomize_tensor_normal(layer.w1_a, rnd);
-        randomize_tensor_normal(layer.w1_b, rnd);
+        ggml_set_zero(layer.w1_b);
         randomize_tensor_normal(layer.w2_a, rnd);
-        randomize_tensor_normal(layer.w2_b, rnd);
+        ggml_set_zero(layer.w2_b);
         randomize_tensor_normal(layer.w3_a, rnd);
-        randomize_tensor_normal(layer.w3_b, rnd);
+        ggml_set_zero(layer.w3_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -642,8 +642,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const int rope_mode = 0;
 
         return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx,
-            rope_freq_base, rope_freq_scale);
+            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+            rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+        );
     };
 
     set_name(tokens_input, "tokens_input");
@@ -652,7 +653,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
 
     auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-        if (ggml_is_quantized(a->type)) {
+        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) {
             return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
         } else if (a->type == GGML_TYPE_F32) {
             return ggml_add(ctx, a, b);
@@ -771,7 +772,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     if (enable_checkpointing) {
         ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size());
     } else {
-        *gb = *gf;
+        ggml_graph_cpy(gf, gb);
         ggml_build_backward_expand(ctx, gf, gb, true);
     }
 
@@ -1109,7 +1110,7 @@ static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor,
         name = ggml_get_name(tensor);
     }
     uint32_t name_len = strlen(name);
-    uint32_t nd = tensor->n_dims;
+    uint32_t nd = ggml_n_dims(tensor);
     uint32_t ne[4] = { (uint32_t)tensor->ne[0],
                        (uint32_t)tensor->ne[1],
                        (uint32_t)tensor->ne[2],
@@ -1545,6 +1546,7 @@ int main(int argc, char ** argv) {
     srand(params.common.seed);
 
     struct llama_model_params llama_mparams = llama_model_default_params();
+    llama_mparams.n_gpu_layers = params.common.n_gpu_layers;
     llama_mparams.vocab_only = false;
 
     printf("%s: model base = '%s'\n", __func__, params.fn_model_base);
@@ -1602,6 +1604,7 @@ int main(int argc, char ** argv) {
     opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
+    opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;
     opt->params.n_threads               = params.common.n_threads;
     opt->params.past                    = params.common.opt_past;
     opt->params.delta                   = params.common.opt_delta;
@@ -1617,8 +1620,6 @@ int main(int argc, char ** argv) {
     opt->params.adam.gclip              = params.common.adam_gclip;
     opt->params.adam.eps_f              = params.common.adam_eps_f;
 
-    ggml_allocr * alloc = NULL;
-
     printf("%s: init model\n", __func__);
     bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
 
@@ -1722,17 +1723,14 @@ int main(int argc, char ** argv) {
 
     // allocate input tensors
     mem_input_data.resize(max_input_size);
-    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc, tokens_input);
-    ggml_allocr_alloc(alloc, target_probs);
-    ggml_allocr_free(alloc);
+    ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
+    ggml_allocr_alloc(alloc_inps, tokens_input);
+    ggml_allocr_alloc(alloc_inps, target_probs);
 
     // context for compute tensors without their data
-    size_t estimated_compute_size_wo_data = (
-        ggml_tensor_overhead()*GGML_MAX_NODES*2
-      + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*(
-            params.common.use_checkpointing ? 3 : 2
-        )
+    const size_t estimated_compute_size_wo_data = (
+            2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() +
+            (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true))
     );
     struct ggml_init_params ctx_compute_params = {
         estimated_compute_size_wo_data, // mem_size
@@ -1754,12 +1752,12 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        alloc = ggml_allocr_new_measure(tensor_alignment);
-        gf = ggml_new_graph(ctx_compute);
+        ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment);
+        gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
-        gb = ggml_new_graph(ctx_compute);
+        gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gb_tmp = params.common.use_checkpointing
-            ? ggml_new_graph(ctx_compute)
+            ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true)
             : NULL;
         loss = llama_build_lora_finetune_graphs(
             &model, &lora, alloc, ctx_compute,
@@ -1787,12 +1785,12 @@ int main(int argc, char ** argv) {
     // allocate compute tensors
     mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
-    gf = ggml_new_graph(ctx_compute);
+    ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
-    gb = ggml_new_graph(ctx_compute);
+    gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gb_tmp = params.common.use_checkpointing
-        ? ggml_new_graph(ctx_compute)
+        ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true)
         : NULL;
     loss = llama_build_lora_finetune_graphs(
         &model, &lora, alloc, ctx_compute,
@@ -1803,6 +1801,8 @@ int main(int argc, char ** argv) {
         params.common.use_checkpointing
     );
     ggml_allocr_free(alloc);
+    ggml_allocr_free(alloc_inps);
+
 
     // tokenize data
     std::vector<llama_token> train_tokens;

examples/finetune/finetune.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+cd `dirname $0`
+cd ../..
+
+EXE="./finetune"
+
+if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi
+if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi
+
+# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses.
+MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing.
+
+while getopts "dg" opt; do
+  case $opt in
+    d)
+      DEBUGGER="gdb --args"
+      ;;
+    g)
+      EXE="./build/bin/Release/finetune"
+      GPUARG="--gpu-layers 25"
+      ;;
+  esac
+done
+
+$DEBUGGER $EXE \
+        --model-base $MODEL \
+        $GPUARG \
+        --checkpoint-in  chk-ol3b-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \
+        --lora-out lora-ol3b-shakespeare-ITERATION.bin \
+        --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \
+        --save-every 10 \
+        --threads 10 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing

examples/gguf/gguf.cpp

@@ -195,7 +195,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;

examples/infill/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} infill.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${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/infill/infill.cpp

@@ -2,7 +2,6 @@
 
 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"
 
 #include <cassert>
@@ -39,8 +38,8 @@ static gpt_params               * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
-static bool is_interacting = false;
 
+static bool is_interacting = false;
 
 static void write_logfile(
     const llama_context * ctx, const gpt_params & params, const llama_model * model,
@@ -104,7 +103,7 @@ static void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    llama_sampling_params & sparams = params.sampling_params;
+    llama_sampling_params & sparams = params.sparams;
     g_params = &params;
 
     if (!gpt_params_parse(argc, argv, params)) {
@@ -147,6 +146,13 @@ int main(int argc, char ** argv) {
 
         return 0;
     }
+    if (params.chatml) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for chatml mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
     if (!params.antiprompt.empty()) {
         printf("\n************\n");
         printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
@@ -184,8 +190,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -231,7 +237,7 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
         LOG_TEE("%s\n", get_system_info(params).c_str());
     }
-    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = llama_should_add_bos_token(model);
     LOG("add_bos: %d\n", add_bos);
 
     bool suff_rm_leading_spc = params.escape;
@@ -246,23 +252,23 @@ int main(int argc, char ** argv) {
     if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
         inp_sfx.erase(inp_sfx.begin());
     }
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
     if (add_bos) {
-        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
     }
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(ctx));
+    embd_inp.push_back(llama_token_middle(model));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+        embd_inp.push_back(llama_token_bos(model));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
     // Tokenize negative prompt
@@ -273,10 +279,10 @@ int main(int argc, char ** argv) {
         LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
         guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
 
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
 
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@@ -294,8 +300,8 @@ int main(int argc, char ** argv) {
         params.n_keep = (int)embd_inp.size();
     }
 
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
 
 
     // enable interactive mode if interactive start is specified
@@ -358,39 +364,10 @@ int main(int argc, char ** argv) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
-            sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
+    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
-    struct llama_grammar * grammar = NULL;
-    grammar_parser::parse_state parsed_grammar;
-
-    if (!params.grammar.empty()) {
-        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
-        // will be empty (default) if there are parse errors
-        if (parsed_grammar.rules.empty()) {
-            return 1;
-        }
-        LOG_TEE("%s: grammar:\n", __func__);
-        grammar_parser::print_grammar(stderr, parsed_grammar);
-        LOG_TEE("\n");
-
-        {
-            auto it = sparams.logit_bias.find(llama_token_eos(ctx));
-            if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
-                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
-            }
-        }
-
-        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
-        grammar = llama_grammar_init(
-            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
-    }
-
-    // TODO: replace with ring-buffer
-    std::vector<llama_token> last_tokens(n_ctx);
-    std::fill(last_tokens.begin(), last_tokens.end(), 0);
     LOG_TEE("\n#####  Infill mode  #####\n\n");
     if (params.infill) {
         printf("\n************\n");
@@ -433,11 +410,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    const int n_vocab = llama_n_vocab(model);
-
-    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
 
     while (n_remain != 0 || params.interactive) {
         // predict
@@ -484,7 +457,7 @@ int main(int argc, char ** argv) {
 
                 LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
 
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
             }
 
@@ -512,7 +485,7 @@ int main(int argc, char ** argv) {
                     input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
 
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
                 } else {
                     input_buf  = embd.data();
                     input_size = embd.size();
@@ -535,7 +508,7 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_TEE("%s : failed to eval\n", __func__);
@@ -554,12 +527,11 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
 
-            const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
+            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
 
-            last_tokens.erase(last_tokens.begin());
-            last_tokens.push_back(id);
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
 
             embd.push_back(id);
 
@@ -575,8 +547,11 @@ int main(int argc, char ** argv) {
             LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-                last_tokens.erase(last_tokens.begin());
-                last_tokens.push_back(embd_inp[n_consumed]);
+
+                // push the prompt in the sampling context in order to apply repetition penalties later
+                // for the prompt, we don't apply grammar rules
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -608,10 +583,10 @@ int main(int argc, char ** argv) {
         if ((int) embd_inp.size() <= n_consumed) {
 
             // deal with eot token in infill mode
-            if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
+            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if(is_interacting && !params.interactive_first) {
                     // print an eot token
-                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
                 fflush(stdout);
                 printf("\n");
@@ -625,7 +600,7 @@ int main(int argc, char ** argv) {
                     buffer += line;
                 } while (another_line);
                 // check if we got an empty line, if so we use the old input
-                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
                     params.input_prefix = buffer;
                 }
                 buffer.clear();
@@ -635,7 +610,7 @@ int main(int argc, char ** argv) {
                     buffer += line;
                 } while (another_line);
                 // check if we got an empty line
-                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
                     params.input_suffix = buffer;
                 }
                 buffer.clear();
@@ -648,7 +623,7 @@ int main(int argc, char ** argv) {
                     process_escapes(params.input_suffix);
                 }
                 suff_rm_leading_spc = params.escape;
-                if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+                if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
                     params.input_suffix.erase(0, 1);
                     suff_rm_leading_spc = false;
                 }
@@ -658,14 +633,14 @@ int main(int argc, char ** argv) {
                 if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
                     inp_sfx.erase(inp_sfx.begin());
                 }
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
                 if (add_bos) {
-                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
                 }
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(ctx));
+                embd_inp.push_back(llama_token_middle(model));
                 embd.clear();
                 embd_guidance.clear();
                 n_remain = params.n_predict;
@@ -675,7 +650,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of text token in interactive mode
-            else if (last_tokens.back() == llama_token_eos(ctx)) {
+            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -692,7 +667,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -727,7 +702,7 @@ int main(int argc, char ** argv) {
                     const size_t original_size = embd_inp.size();
 
                     const auto line_inp = ::llama_tokenize(ctx, buffer, false);
-                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
 
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
@@ -748,22 +723,14 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    // reset grammar state if we're restarting generation
-                    if (grammar != NULL) {
-                        llama_grammar_free(grammar);
-
-                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
-                        grammar = llama_grammar_init(
-                            grammar_rules.data(), grammar_rules.size(),
-                            parsed_grammar.symbol_ids.at("root"));
-                    }
+                    llama_sampling_reset(ctx_sampling);
                 }
                 is_interacting = false;
             }
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
             break;
         }
 
@@ -775,7 +742,7 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
         fflush(stdout);
     }
 
@@ -786,9 +753,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
-    if (grammar != NULL) {
-        llama_grammar_free(grammar);
-    }
+    llama_sampling_free(ctx_sampling);
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/llama-bench/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} llama-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${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/llama-bench/llama-bench.cpp

@@ -19,7 +19,6 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
-#include "build-info.h"
 #include "ggml-cuda.h"
 
 // utils
@@ -54,6 +53,13 @@ static std::vector<T> split(const std::string & str, char delim) {
     return values;
 }
 
+template<typename T, typename F>
+static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
+    std::vector<std::string> str_values;
+    std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
+    return str_values;
+}
+
 template<typename T>
 static T avg(const std::vector<T> & v) {
     if (v.empty()) {
@@ -127,7 +133,8 @@ struct cmd_params {
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
-    std::vector<bool> f32_kv;
+    std::vector<ggml_type> type_k;
+    std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
     std::vector<int> main_gpu;
@@ -143,7 +150,8 @@ static const cmd_params cmd_params_defaults = {
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
     /* n_batch       */ {512},
-    /* f32_kv        */ {false},
+    /* type_k        */ {GGML_TYPE_F16},
+    /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
     /* main_gpu      */ {0},
@@ -163,7 +171,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -ctk <t>, --cache-type-k <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv <t>, --cache-type-v <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
     printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
@@ -174,9 +183,32 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
-
 }
 
+static ggml_type ggml_type_from_name(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    return GGML_TYPE_COUNT;
+}
+
+
 static cmd_params parse_cmd_params(int argc, char ** argv) {
     cmd_params params;
     std::string arg;
@@ -225,13 +257,38 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
-        } else if (arg == "--memory-f32") {
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split<int>(argv[i], split_delim);
-            params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -322,7 +379,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
-    if (params.f32_kv.empty())       { params.f32_kv = cmd_params_defaults.f32_kv; }
+    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
+    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
@@ -337,7 +395,8 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
-    bool f32_kv;
+    ggml_type type_k;
+    ggml_type type_v;
     int n_threads;
     int n_gpu_layers;
     int main_gpu;
@@ -366,7 +425,8 @@ struct cmd_params_instance {
 
         cparams.n_ctx = n_prompt + n_gen;
         cparams.n_batch = n_batch;
-        cparams.f16_kv = !f32_kv;
+        cparams.type_k = type_k;
+        cparams.type_v = type_v;
         cparams.mul_mat_q = mul_mat_q;
 
         return cparams;
@@ -381,7 +441,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         cmd_params_instance instance = {
@@ -389,7 +450,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
             /* .n_prompt     = */ n_prompt,
             /* .n_gen        = */ n_gen,
             /* .n_batch      = */ nb,
-            /* .f32_kv       = */ fk,
+            /* .type_k       = */ tk,
+            /* .type_v       = */ tv,
             /* .n_threads    = */ nt,
             /* .n_gpu_layers = */ nl,
             /* .main_gpu     = */ mg,
@@ -411,7 +473,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -423,7 +486,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
-                /* .f32_kv       = */ fk,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
                 /* .main_gpu     = */ mg,
@@ -442,7 +506,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
-                /* .f32_kv       = */ fk,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
                 /* .main_gpu     = */ mg,
@@ -490,7 +555,8 @@ struct test {
     uint64_t model_n_params;
     int n_batch;
     int n_threads;
-    bool f32_kv;
+    ggml_type type_k;
+    ggml_type type_v;
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
@@ -509,7 +575,8 @@ struct test {
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
-        f32_kv = inst.f32_kv;
+        type_k = inst.type_k;
+        type_v = inst.type_v;
         n_gpu_layers = inst.n_gpu_layers;
         main_gpu = inst.main_gpu;
         mul_mat_q = inst.mul_mat_q;
@@ -572,7 +639,7 @@ struct test {
             "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_threads", "f16_kv",
+            "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
@@ -622,7 +689,7 @@ struct test {
             std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
+            std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
@@ -641,8 +708,8 @@ struct test {
     }
 };
 
-const std::string test::build_commit = BUILD_COMMIT;
-const int         test::build_number = BUILD_NUMBER;
+const std::string test::build_commit = LLAMA_COMMIT;
+const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::metal        = !!ggml_cpu_has_metal();
@@ -806,8 +873,11 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.push_back("n_batch");
         }
-        if (params.f32_kv.size() > 1 || params.f32_kv != cmd_params_defaults.f32_kv) {
-            fields.push_back("f16_kv");
+        if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
+            fields.push_back("type_k");
+        }
+        if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
+            fields.push_back("type_v");
         }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.push_back("main_gpu");
@@ -933,7 +1003,7 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
+    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
     int n_processed = 0;
 
     llama_set_n_threads(ctx, n_threads, n_threads);
@@ -946,7 +1016,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(ctx);
+    llama_token token = llama_token_bos(llama_get_model(ctx));
 
     llama_set_n_threads(ctx, n_threads, n_threads);
 
@@ -1037,7 +1107,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         // warmup run
         if (t.n_prompt > 0) {
@@ -1048,7 +1118,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_tokens_rm(ctx, -1, -1);
+            llama_kv_cache_clear(ctx);
 
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {

examples/llama.swiftui/.gitignore

@@ -0,0 +1,2 @@
+xcuserdata
+xcshareddata

examples/llama.swiftui/README.md

@@ -0,0 +1,7 @@
+# llama.swiftui
+
+Local inference of llama.cpp on an iPhone.
+So far I only tested with starcoder 1B model, but it can most likely handle 7B models as well.
+
+https://github.com/bachittle/llama.cpp/assets/39804642/e290827a-4edb-4093-9642-2a5e399ec545
+

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

@@ -0,0 +1,340 @@
+import Foundation
+
+// import llama
+
+enum LlamaError: Error {
+    case couldNotInitializeContext
+}
+
+func llama_batch_clear(_ batch: inout llama_batch) {
+    batch.n_tokens = 0
+}
+
+func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama_pos, _ seq_ids: [llama_seq_id], _ logits: Bool) {
+    batch.token   [Int(batch.n_tokens)] = id
+    batch.pos     [Int(batch.n_tokens)] = pos
+    batch.n_seq_id[Int(batch.n_tokens)] = Int32(seq_ids.count)
+    for i in 0..<seq_ids.count {
+        batch.seq_id[Int(batch.n_tokens)]![Int(i)] = seq_ids[i]
+    }
+    batch.logits  [Int(batch.n_tokens)] = logits ? 1 : 0
+
+    batch.n_tokens += 1
+}
+
+actor LlamaContext {
+    private var model: OpaquePointer
+    private var context: OpaquePointer
+    private var batch: llama_batch
+    private var tokens_list: [llama_token]
+
+    /// This variable is used to store temporarily invalid cchars
+    private var temporary_invalid_cchars: [CChar]
+
+    var n_len: Int32 = 64
+    var n_cur: Int32 = 0
+
+    var n_decode: Int32 = 0
+
+    init(model: OpaquePointer, context: OpaquePointer) {
+        self.model = model
+        self.context = context
+        self.tokens_list = []
+        self.batch = llama_batch_init(512, 0, 1)
+        self.temporary_invalid_cchars = []
+    }
+
+    deinit {
+        llama_batch_free(batch)
+        llama_free(context)
+        llama_free_model(model)
+        llama_backend_free()
+    }
+
+    static func create_context(path: String) throws -> LlamaContext {
+        llama_backend_init(false)
+        var model_params = llama_model_default_params()
+
+#if targetEnvironment(simulator)
+        model_params.n_gpu_layers = 0
+        print("Running on simulator, force use n_gpu_layers = 0")
+#endif
+        let model = llama_load_model_from_file(path, model_params)
+        guard let model else {
+            print("Could not load model at \(path)")
+            throw LlamaError.couldNotInitializeContext
+        }
+
+        let n_threads = max(1, min(8, ProcessInfo.processInfo.processorCount - 2))
+        print("Using \(n_threads) threads")
+
+        var ctx_params = llama_context_default_params()
+        ctx_params.seed  = 1234
+        ctx_params.n_ctx = 2048
+        ctx_params.n_threads       = UInt32(n_threads)
+        ctx_params.n_threads_batch = UInt32(n_threads)
+
+        let context = llama_new_context_with_model(model, ctx_params)
+        guard let context else {
+            print("Could not load context!")
+            throw LlamaError.couldNotInitializeContext
+        }
+
+        return LlamaContext(model: model, context: context)
+    }
+
+    func model_info() -> String {
+        let result = UnsafeMutablePointer<Int8>.allocate(capacity: 256)
+        result.initialize(repeating: Int8(0), count: 256)
+        defer {
+            result.deallocate()
+        }
+
+        // TODO: this is probably very stupid way to get the string from C
+
+        let nChars = llama_model_desc(model, result, 256)
+        let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nChars))
+
+        var SwiftString = ""
+        for char in bufferPointer {
+            SwiftString.append(Character(UnicodeScalar(UInt8(char))))
+        }
+
+        return SwiftString
+    }
+
+    func get_n_tokens() -> Int32 {
+        return batch.n_tokens;
+    }
+
+    func completion_init(text: String) {
+        print("attempting to complete \"\(text)\"")
+
+        tokens_list = tokenize(text: text, add_bos: true)
+        temporary_invalid_cchars = []
+
+        let n_ctx = llama_n_ctx(context)
+        let n_kv_req = tokens_list.count + (Int(n_len) - tokens_list.count)
+
+        print("\n n_len = \(n_len), n_ctx = \(n_ctx), n_kv_req = \(n_kv_req)")
+
+        if n_kv_req > n_ctx {
+            print("error: n_kv_req > n_ctx, the required KV cache size is not big enough")
+        }
+
+        for id in tokens_list {
+            print(String(cString: token_to_piece(token: id) + [0]))
+        }
+
+        llama_batch_clear(&batch)
+
+        for i1 in 0..<tokens_list.count {
+            let i = Int(i1)
+            llama_batch_add(&batch, tokens_list[i], Int32(i), [0], false)
+        }
+        batch.logits[Int(batch.n_tokens) - 1] = 1 // true
+
+        if llama_decode(context, batch) != 0 {
+            print("llama_decode() failed")
+        }
+
+        n_cur = batch.n_tokens
+    }
+
+    func completion_loop() -> String {
+        var new_token_id: llama_token = 0
+
+        let n_vocab = llama_n_vocab(model)
+        let logits = llama_get_logits_ith(context, batch.n_tokens - 1)
+
+        var candidates = Array<llama_token_data>()
+        candidates.reserveCapacity(Int(n_vocab))
+
+        for token_id in 0..<n_vocab {
+            candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
+        }
+        candidates.withUnsafeMutableBufferPointer() { buffer in
+            var candidates_p = llama_token_data_array(data: buffer.baseAddress, size: buffer.count, sorted: false)
+
+            new_token_id = llama_sample_token_greedy(context, &candidates_p)
+        }
+
+        if new_token_id == llama_token_eos(context) || n_cur == n_len {
+            print("\n")
+            let new_token_str = String(cString: temporary_invalid_cchars + [0])
+            temporary_invalid_cchars.removeAll()
+            return new_token_str
+        }
+
+        let new_token_cchars = token_to_piece(token: new_token_id)
+        temporary_invalid_cchars.append(contentsOf: new_token_cchars)
+        let new_token_str: String
+        if let string = String(validatingUTF8: temporary_invalid_cchars + [0]) {
+            temporary_invalid_cchars.removeAll()
+            new_token_str = string
+        } else if (0 ..< temporary_invalid_cchars.count).contains(where: {$0 != 0 && String(validatingUTF8: Array(temporary_invalid_cchars.suffix($0)) + [0]) != nil}) {
+            // in this case, at least the suffix of the temporary_invalid_cchars can be interpreted as UTF8 string
+            let string = String(cString: temporary_invalid_cchars + [0])
+            temporary_invalid_cchars.removeAll()
+            new_token_str = string
+        } else {
+            new_token_str = ""
+        }
+        print(new_token_str)
+        // tokens_list.append(new_token_id)
+
+        llama_batch_clear(&batch)
+        llama_batch_add(&batch, new_token_id, n_cur, [0], true)
+
+        n_decode += 1
+        n_cur    += 1
+
+        if llama_decode(context, batch) != 0 {
+            print("failed to evaluate llama!")
+        }
+
+        return new_token_str
+    }
+
+    func bench(pp: Int, tg: Int, pl: Int, nr: Int = 1) -> String {
+        var pp_avg: Double = 0
+        var tg_avg: Double = 0
+
+        var pp_std: Double = 0
+        var tg_std: Double = 0
+
+        for _ in 0..<nr {
+            // bench prompt processing
+
+            llama_batch_clear(&batch)
+
+            let n_tokens = pp
+
+            for i in 0..<n_tokens {
+                llama_batch_add(&batch, 0, Int32(i), [0], false)
+            }
+            batch.logits[Int(batch.n_tokens) - 1] = 1 // true
+
+            llama_kv_cache_clear(context)
+
+            let t_pp_start = ggml_time_us()
+
+            if llama_decode(context, batch) != 0 {
+                print("llama_decode() failed during prompt")
+            }
+
+            let t_pp_end = ggml_time_us()
+
+            // bench text generation
+
+            llama_kv_cache_clear(context)
+
+            let t_tg_start = ggml_time_us()
+
+            for i in 0..<tg {
+                llama_batch_clear(&batch)
+
+                for j in 0..<pl {
+                    llama_batch_add(&batch, 0, Int32(i), [Int32(j)], true)
+                }
+
+                if llama_decode(context, batch) != 0 {
+                    print("llama_decode() failed during text generation")
+                }
+            }
+
+            let t_tg_end = ggml_time_us()
+
+            llama_kv_cache_clear(context)
+
+            let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
+            let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
+
+            let speed_pp = Double(pp)    / t_pp
+            let speed_tg = Double(pl*tg) / t_tg
+
+            pp_avg += speed_pp
+            tg_avg += speed_tg
+
+            pp_std += speed_pp * speed_pp
+            tg_std += speed_tg * speed_tg
+
+            print("pp \(speed_pp) t/s, tg \(speed_tg) t/s")
+        }
+
+        pp_avg /= Double(nr)
+        tg_avg /= Double(nr)
+
+        if nr > 1 {
+            pp_std = sqrt(pp_std / Double(nr - 1) - pp_avg * pp_avg * Double(nr) / Double(nr - 1))
+            tg_std = sqrt(tg_std / Double(nr - 1) - tg_avg * tg_avg * Double(nr) / Double(nr - 1))
+        } else {
+            pp_std = 0
+            tg_std = 0
+        }
+
+        let model_desc     = model_info();
+        let model_size     = String(format: "%.2f GiB", Double(llama_model_size(model)) / 1024.0 / 1024.0 / 1024.0);
+        let model_n_params = String(format: "%.2f B", Double(llama_model_n_params(model)) / 1e9);
+        let backend        = "Metal";
+        let pp_avg_str     = String(format: "%.2f", pp_avg);
+        let tg_avg_str     = String(format: "%.2f", tg_avg);
+        let pp_std_str     = String(format: "%.2f", pp_std);
+        let tg_std_str     = String(format: "%.2f", tg_std);
+
+        var result = ""
+
+        result += String("| model | size | params | backend | test | t/s |\n")
+        result += String("| --- | --- | --- | --- | --- | --- |\n")
+        result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | pp \(pp) | \(pp_avg_str) ± \(pp_std_str) |\n")
+        result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | tg \(tg) | \(tg_avg_str) ± \(tg_std_str) |\n")
+
+        return result;
+    }
+
+    func clear() {
+        tokens_list.removeAll()
+        temporary_invalid_cchars.removeAll()
+        llama_kv_cache_clear(context)
+    }
+
+    private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
+        let utf8Count = text.utf8.count
+        let n_tokens = utf8Count + (add_bos ? 1 : 0) + 1
+        let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
+        let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
+
+        var swiftTokens: [llama_token] = []
+        for i in 0..<tokenCount {
+            swiftTokens.append(tokens[Int(i)])
+        }
+
+        tokens.deallocate()
+
+        return swiftTokens
+    }
+
+    /// - note: The result does not contain null-terminator
+    private func token_to_piece(token: llama_token) -> [CChar] {
+        let result = UnsafeMutablePointer<Int8>.allocate(capacity: 8)
+        result.initialize(repeating: Int8(0), count: 8)
+        defer {
+            result.deallocate()
+        }
+        let nTokens = llama_token_to_piece(model, token, result, 8)
+
+        if nTokens < 0 {
+            let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
+            newResult.initialize(repeating: Int8(0), count: Int(-nTokens))
+            defer {
+                newResult.deallocate()
+            }
+            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens)
+            let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
+            return Array(bufferPointer)
+        } else {
+            let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nTokens))
+            return Array(bufferPointer)
+        }
+    }
+}

examples/llama.swiftui/llama.cpp.swift/bridging-header.h

@@ -0,0 +1,5 @@
+//
+//  Use this file to import your target's public headers that you would like to expose to Swift.
+//
+
+#import "llama.h"

examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj

@@ -0,0 +1,483 @@
+// !$*UTF8*$!
+{
+	archiveVersion = 1;
+	classes = {
+	};
+	objectVersion = 56;
+	objects = {
+
+/* Begin PBXBuildFile section */
+		542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
+		5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
+		542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
+		542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
+		542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
+		542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
+		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
+		549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
+		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
+		8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
+		8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
+		8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
+		8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
+		8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
+		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
+		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
+		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
+/* End PBXBuildFile section */
+
+/* Begin PBXFileReference section */
+		542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
+		542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
+		542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
+		5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
+		542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
+		542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
+		542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
+		542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
+		542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
+		542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
+		549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
+		549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
+		549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
+		549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
+		7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
+		8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
+		8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
+		8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
+		8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
+		8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
+		8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
+		8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
+		8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
+		8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
+		8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
+/* End PBXFileReference section */
+
+/* Begin PBXFrameworksBuildPhase section */
+		8A1C83702AC328BD0096AF73 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
+				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXFrameworksBuildPhase section */
+
+/* Begin PBXGroup section */
+		8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
+			isa = PBXGroup;
+			children = (
+				5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
+				542376092B0D9C40008E6A1C /* ggml-backend.h */,
+				542376062B0D9BEA008E6A1C /* ggml-quants.h */,
+				542376072B0D9BFB008E6A1C /* ggml-quants.c */,
+				549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
+				549479C62AC9E0F200E0F78B /* ggml-metal.h */,
+				549479C52AC9E0F200E0F78B /* ggml-metal.m */,
+				542EA09B2AC8723900A8AEE9 /* ggml.c */,
+				542EA09C2AC8723900A8AEE9 /* ggml.h */,
+				542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
+				542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
+				542EA0A12AC8729100A8AEE9 /* llama.cpp */,
+				542EA0A22AC8729100A8AEE9 /* llama.h */,
+			);
+			name = llama.cpp;
+			sourceTree = "<group>";
+		};
+		8A1C836A2AC328BD0096AF73 = {
+			isa = PBXGroup;
+			children = (
+				8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
+				8A907F312AC7134E006146EA /* llama.cpp.swift */,
+				8A3F84232AC4C891005E2EE8 /* models */,
+				8A1C83752AC328BD0096AF73 /* llama.swiftui */,
+				8A1C83742AC328BD0096AF73 /* Products */,
+				8A39BE082AC7601000BFEB40 /* Frameworks */,
+			);
+			sourceTree = "<group>";
+		};
+		8A1C83742AC328BD0096AF73 /* Products */ = {
+			isa = PBXGroup;
+			children = (
+				8A1C83732AC328BD0096AF73 /* llama.swiftui.app */,
+			);
+			name = Products;
+			sourceTree = "<group>";
+		};
+		8A1C83752AC328BD0096AF73 /* llama.swiftui */ = {
+			isa = PBXGroup;
+			children = (
+				8A3F84102AC4BD85005E2EE8 /* Resources */,
+				8A9F7C4B2AC332DC008AE1EA /* Models */,
+				8A9F7C4A2AC332BF008AE1EA /* UI */,
+				8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
+				8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
+				8A1C837C2AC328BE0096AF73 /* Preview Content */,
+			);
+			path = llama.swiftui;
+			sourceTree = "<group>";
+		};
+		8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
+			isa = PBXGroup;
+			children = (
+				8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
+			);
+			path = "Preview Content";
+			sourceTree = "<group>";
+		};
+		8A39BE082AC7601000BFEB40 /* Frameworks */ = {
+			isa = PBXGroup;
+			children = (
+				549479CA2AC9E16000E0F78B /* Metal.framework */,
+				8A39BE092AC7601000BFEB40 /* Accelerate.framework */,
+			);
+			name = Frameworks;
+			sourceTree = "<group>";
+		};
+		8A3F84102AC4BD85005E2EE8 /* Resources */ = {
+			isa = PBXGroup;
+			children = (
+				8A3F84112AC4BD8C005E2EE8 /* models */,
+			);
+			path = Resources;
+			sourceTree = "<group>";
+		};
+		8A3F84112AC4BD8C005E2EE8 /* models */ = {
+			isa = PBXGroup;
+			children = (
+			);
+			path = models;
+			sourceTree = "<group>";
+		};
+		8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
+			isa = PBXGroup;
+			children = (
+				8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
+				8A907F322AC7134E006146EA /* LibLlama.swift */,
+			);
+			path = llama.cpp.swift;
+			sourceTree = "<group>";
+		};
+		8A9F7C4A2AC332BF008AE1EA /* UI */ = {
+			isa = PBXGroup;
+			children = (
+				7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */,
+				8A1C83782AC328BD0096AF73 /* ContentView.swift */,
+			);
+			path = UI;
+			sourceTree = "<group>";
+		};
+		8A9F7C4B2AC332DC008AE1EA /* Models */ = {
+			isa = PBXGroup;
+			children = (
+				8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */,
+			);
+			path = Models;
+			sourceTree = "<group>";
+		};
+/* End PBXGroup section */
+
+/* Begin PBXNativeTarget section */
+		8A1C83722AC328BD0096AF73 /* llama.swiftui */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = 8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */;
+			buildPhases = (
+				8A1C836F2AC328BD0096AF73 /* Sources */,
+				8A1C83702AC328BD0096AF73 /* Frameworks */,
+				8A1C83712AC328BD0096AF73 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+			);
+			name = llama.swiftui;
+			packageProductDependencies = (
+			);
+			productName = llama.swiftui;
+			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
+			productType = "com.apple.product-type.application";
+		};
+/* End PBXNativeTarget section */
+
+/* Begin PBXProject section */
+		8A1C836B2AC328BD0096AF73 /* Project object */ = {
+			isa = PBXProject;
+			attributes = {
+				BuildIndependentTargetsInParallel = 1;
+				LastSwiftUpdateCheck = 1500;
+				LastUpgradeCheck = 1500;
+				TargetAttributes = {
+					8A1C83722AC328BD0096AF73 = {
+						CreatedOnToolsVersion = 15.0;
+						LastSwiftMigration = 1500;
+					};
+				};
+			};
+			buildConfigurationList = 8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */;
+			compatibilityVersion = "Xcode 14.0";
+			developmentRegion = en;
+			hasScannedForEncodings = 0;
+			knownRegions = (
+				en,
+				Base,
+			);
+			mainGroup = 8A1C836A2AC328BD0096AF73;
+			packageReferences = (
+			);
+			productRefGroup = 8A1C83742AC328BD0096AF73 /* Products */;
+			projectDirPath = "";
+			projectRoot = "";
+			targets = (
+				8A1C83722AC328BD0096AF73 /* llama.swiftui */,
+			);
+		};
+/* End PBXProject section */
+
+/* Begin PBXResourcesBuildPhase section */
+		8A1C83712AC328BD0096AF73 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */,
+				8A3F84242AC4C891005E2EE8 /* models in Resources */,
+				8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
+				8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXResourcesBuildPhase section */
+
+/* Begin PBXSourcesBuildPhase section */
+		8A1C836F2AC328BD0096AF73 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
+				549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
+				542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
+				8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
+				542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
+				8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
+				8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
+				8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
+				7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */,
+				542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
+				5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXSourcesBuildPhase section */
+
+/* Begin XCBuildConfiguration section */
+		8A1C837F2AC328BE0096AF73 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = dwarf;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_TESTABILITY = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu17;
+				GCC_DYNAMIC_NO_PIC = NO;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_OPTIMIZATION_LEVEL = 0;
+				GCC_PREPROCESSOR_DEFINITIONS = (
+					"DEBUG=1",
+					"$(inherited)",
+				);
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 17.0;
+				LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
+				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
+				MTL_FAST_MATH = YES;
+				ONLY_ACTIVE_ARCH = YES;
+				SDKROOT = iphoneos;
+				SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+			};
+			name = Debug;
+		};
+		8A1C83802AC328BE0096AF73 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
+				ENABLE_NS_ASSERTIONS = NO;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu17;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 17.0;
+				LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
+				MTL_ENABLE_DEBUG_INFO = NO;
+				MTL_FAST_MATH = YES;
+				SDKROOT = iphoneos;
+				SWIFT_COMPILATION_MODE = wholemodule;
+				VALIDATE_PRODUCT = YES;
+			};
+			name = Release;
+		};
+		8A1C83822AC328BE0096AF73 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CLANG_ENABLE_MODULES = YES;
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
+				DEVELOPMENT_TEAM = STLSG3FG8Q;
+				ENABLE_PREVIEWS = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
+				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
+				INFOPLIST_KEY_UILaunchScreen_Generation = YES;
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/Frameworks",
+				);
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Debug;
+		};
+		8A1C83832AC328BE0096AF73 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CLANG_ENABLE_MODULES = YES;
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
+				DEVELOPMENT_TEAM = STLSG3FG8Q;
+				ENABLE_PREVIEWS = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
+				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
+				INFOPLIST_KEY_UILaunchScreen_Generation = YES;
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/Frameworks",
+				);
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Release;
+		};
+/* End XCBuildConfiguration section */
+
+/* Begin XCConfigurationList section */
+		8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				8A1C837F2AC328BE0096AF73 /* Debug */,
+				8A1C83802AC328BE0096AF73 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				8A1C83822AC328BE0096AF73 /* Debug */,
+				8A1C83832AC328BE0096AF73 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+/* End XCConfigurationList section */
+	};
+	rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
+}

examples/llama.swiftui/llama.swiftui.xcodeproj/project.xcworkspace/contents.xcworkspacedata

@@ -0,0 +1,7 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<Workspace
+   version = "1.0">
+   <FileRef
+      location = "self:">
+   </FileRef>
+</Workspace>

examples/llama.swiftui/llama.swiftui.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>IDEDidComputeMac32BitWarning</key>
+    <true/>
+</dict>
+</plist>

examples/llama.swiftui/llama.swiftui/Assets.xcassets/AccentColor.colorset/Contents.json

@@ -0,0 +1,11 @@
+{
+  "colors" : [
+    {
+      "idiom" : "universal"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/llama.swiftui/llama.swiftui/Assets.xcassets/AppIcon.appiconset/Contents.json

@@ -0,0 +1,13 @@
+{
+  "images" : [
+    {
+      "idiom" : "universal",
+      "platform" : "ios",
+      "size" : "1024x1024"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/llama.swiftui/llama.swiftui/Assets.xcassets/Contents.json

@@ -0,0 +1,6 @@
+{
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift

@@ -0,0 +1,85 @@
+import Foundation
+
+@MainActor
+class LlamaState: ObservableObject {
+    @Published var messageLog = ""
+    @Published var cacheCleared = false
+
+    private var llamaContext: LlamaContext?
+    private var defaultModelUrl: URL? {
+        Bundle.main.url(forResource: "ggml-model", withExtension: "gguf", subdirectory: "models")
+        // Bundle.main.url(forResource: "llama-2-7b-chat", withExtension: "Q2_K.gguf", subdirectory: "models")
+    }
+
+    init() {
+        do {
+            try loadModel(modelUrl: defaultModelUrl)
+        } catch {
+            messageLog += "Error!\n"
+        }
+    }
+
+    func loadModel(modelUrl: URL?) throws {
+        messageLog += "Loading model...\n"
+        if let modelUrl {
+            llamaContext = try LlamaContext.create_context(path: modelUrl.path())
+            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
+        } else {
+            messageLog += "Could not locate model\n"
+        }
+    }
+
+    func complete(text: String) async {
+        guard let llamaContext else {
+            return
+        }
+
+        await llamaContext.completion_init(text: text)
+        messageLog += "\(text)"
+
+        while await llamaContext.n_cur <= llamaContext.n_len {
+            let result = await llamaContext.completion_loop()
+            messageLog += "\(result)"
+        }
+        await llamaContext.clear()
+        messageLog += "\n\ndone\n"
+    }
+
+    func bench() async {
+        guard let llamaContext else {
+            return
+        }
+
+        messageLog += "\n"
+        messageLog += "Running benchmark...\n"
+        messageLog += "Model info: "
+        messageLog += await llamaContext.model_info() + "\n"
+
+        let t_start = DispatchTime.now().uptimeNanoseconds
+        await llamaContext.bench(pp: 8, tg: 4, pl: 1) // heat up
+        let t_end = DispatchTime.now().uptimeNanoseconds
+
+        let t_heat = Double(t_end - t_start) / 1_000_000_000.0
+        messageLog += "Heat up time: \(t_heat) seconds, please wait...\n"
+
+        // if more than 5 seconds, then we're probably running on a slow device
+        if t_heat > 5.0 {
+            messageLog += "Heat up time is too long, aborting benchmark\n"
+            return
+        }
+
+        let result = await llamaContext.bench(pp: 512, tg: 128, pl: 1, nr: 3)
+
+        messageLog += "\(result)"
+        messageLog += "\n"
+    }
+
+    func clear() async {
+        guard let llamaContext else {
+            return
+        }
+
+        await llamaContext.clear()
+        messageLog = ""
+    }
+}

examples/llama.swiftui/llama.swiftui/Preview Content/Preview Assets.xcassets/Contents.json

@@ -0,0 +1,6 @@
+{
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

examples/llama.swiftui/llama.swiftui/UI/ContentView.swift

@@ -0,0 +1,161 @@
+import SwiftUI
+
+struct ContentView: View {
+    @StateObject var llamaState = LlamaState()
+
+    @State private var multiLineText = ""
+
+    private static func cleanupModelCaches() {
+        // Delete all models (*.gguf)
+        let fileManager = FileManager.default
+        let documentsUrl =  FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
+        do {
+            let fileURLs = try fileManager.contentsOfDirectory(at: documentsUrl, includingPropertiesForKeys: nil)
+            for fileURL in fileURLs {
+                if fileURL.pathExtension == "gguf" {
+                    try fileManager.removeItem(at: fileURL)
+                }
+            }
+        } catch {
+            print("Error while enumerating files \(documentsUrl.path): \(error.localizedDescription)")
+        }
+    }
+
+    var body: some View {
+        VStack {
+            ScrollView(.vertical, showsIndicators: true) {
+                Text(llamaState.messageLog)
+                .font(.system(size: 12))
+                .frame(maxWidth: .infinity, alignment: .leading)
+                .padding()
+                .onTapGesture {
+                    UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
+                }
+            }
+
+            TextEditor(text: $multiLineText)
+                .frame(height: 80)
+                .padding()
+                .border(Color.gray, width: 0.5)
+
+            HStack {
+                Button("Send") {
+                    sendText()
+                }
+                .padding(8)
+                .background(Color.blue)
+                .foregroundColor(.white)
+                .cornerRadius(8)
+
+                Button("Bench") {
+                    bench()
+                }
+                .padding(8)
+                .background(Color.blue)
+                .foregroundColor(.white)
+                .cornerRadius(8)
+
+                Button("Clear") {
+                    clear()
+                }
+                .padding(8)
+                .background(Color.blue)
+                .foregroundColor(.white)
+                .cornerRadius(8)
+
+                Button("Copy") {
+                    UIPasteboard.general.string = llamaState.messageLog
+                }
+                .padding(8)
+                .background(Color.blue)
+                .foregroundColor(.white)
+                .cornerRadius(8)
+            }
+
+            VStack {
+                DownloadButton(
+                    llamaState: llamaState,
+                    modelName: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",
+                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
+                    filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
+                )
+                .font(.system(size: 12))
+                .padding(.top, 4)
+                .frame(maxWidth: .infinity, alignment: .leading)
+
+                DownloadButton(
+                    llamaState: llamaState,
+                    modelName: "TinyLlama-1.1B (Q8_0, 1.1 GiB)",
+                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q8_0.gguf?download=true",
+                    filename: "tinyllama-1.1b-1t-openorca.Q8_0.gguf"
+                )
+                .font(.system(size: 12))
+
+                DownloadButton(
+                    llamaState: llamaState,
+                    modelName: "TinyLlama-1.1B (F16, 2.2 GiB)",
+                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
+                    filename: "tinyllama-1.1b-f16.gguf"
+                )
+                .font(.system(size: 12))
+                .frame(maxWidth: .infinity, alignment: .leading)
+
+                DownloadButton(
+                    llamaState: llamaState,
+                    modelName: "Phi-2.7B (Q4_0, 1.6 GiB)",
+                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
+                    filename: "phi-2-q4_0.gguf"
+                )
+                .font(.system(size: 12))
+
+                DownloadButton(
+                    llamaState: llamaState,
+                    modelName: "Phi-2.7B (Q8_0, 2.8 GiB)",
+                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
+                    filename: "phi-2-q8_0.gguf"
+                )
+                .font(.system(size: 12))
+                .frame(maxWidth: .infinity, alignment: .leading)
+
+                DownloadButton(
+                    llamaState: llamaState,
+                    modelName: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
+                    modelUrl: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
+                    filename: "mistral-7b-v0.1.Q4_0.gguf"
+                )
+                .font(.system(size: 12))
+
+                Button("Clear downloaded models") {
+                    ContentView.cleanupModelCaches()
+                    llamaState.cacheCleared = true
+                }
+                .padding(8)
+                .font(.system(size: 12))
+            }
+        }
+        .padding()
+    }
+
+    func sendText() {
+        Task {
+            await llamaState.complete(text: multiLineText)
+            multiLineText = ""
+        }
+    }
+
+    func bench() {
+        Task {
+            await llamaState.bench()
+        }
+    }
+
+    func clear() {
+        Task {
+            await llamaState.clear()
+        }
+    }
+}
+
+//#Preview {
+//    ContentView()
+//}

examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift

@@ -0,0 +1,122 @@
+import SwiftUI
+
+struct DownloadButton: View {
+    @ObservedObject private var llamaState: LlamaState
+    private var modelName: String
+    private var modelUrl: String
+    private var filename: String
+
+    @State private var status: String
+
+    @State private var downloadTask: URLSessionDownloadTask?
+    @State private var progress = 0.0
+    @State private var observation: NSKeyValueObservation?
+
+    private static func getFileURL(filename: String) -> URL {
+        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
+    }
+
+    private func checkFileExistenceAndUpdateStatus() {
+    }
+
+    init(llamaState: LlamaState, modelName: String, modelUrl: String, filename: String) {
+        self.llamaState = llamaState
+        self.modelName = modelName
+        self.modelUrl = modelUrl
+        self.filename = filename
+
+        let fileURL = DownloadButton.getFileURL(filename: filename)
+        status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
+    }
+
+    private func download() {
+        status = "downloading"
+        print("Downloading model \(modelName) from \(modelUrl)")
+        guard let url = URL(string: modelUrl) else { return }
+        let fileURL = DownloadButton.getFileURL(filename: filename)
+
+        downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
+            if let error = error {
+                print("Error: \(error.localizedDescription)")
+                return
+            }
+
+            guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
+                print("Server error!")
+                return
+            }
+
+            do {
+                if let temporaryURL = temporaryURL {
+                    try FileManager.default.copyItem(at: temporaryURL, to: fileURL)
+                    print("Writing to \(filename) completed")
+
+                    llamaState.cacheCleared = false
+
+                    status = "downloaded"
+                }
+            } catch let err {
+                print("Error: \(err.localizedDescription)")
+            }
+        }
+
+        observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
+            self.progress = progress.fractionCompleted
+        }
+
+        downloadTask?.resume()
+    }
+
+    var body: some View {
+        VStack {
+            if status == "download" {
+                Button(action: download) {
+                    Text("Download " + modelName)
+                }
+            } else if status == "downloading" {
+                Button(action: {
+                    downloadTask?.cancel()
+                    status = "download"
+                }) {
+                    Text("\(modelName) (Downloading \(Int(progress * 100))%)")
+                }
+            } else if status == "downloaded" {
+                Button(action: {
+                    let fileURL = DownloadButton.getFileURL(filename: filename)
+                    if !FileManager.default.fileExists(atPath: fileURL.path) {
+                        download()
+                        return
+                    }
+                    do {
+                        try llamaState.loadModel(modelUrl: fileURL)
+                    } catch let err {
+                        print("Error: \(err.localizedDescription)")
+                    }
+                }) {
+                    Text("\(modelName) (Downloaded)")
+                }
+            } else {
+                Text("Unknown status")
+            }
+        }
+        .onDisappear() {
+            downloadTask?.cancel()
+        }
+        .onChange(of: llamaState.cacheCleared) { newValue in
+            if newValue {
+                downloadTask?.cancel()
+                let fileURL = DownloadButton.getFileURL(filename: filename)
+                status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
+            }
+        }
+    }
+}
+
+// #Preview {
+//    DownloadButton(
+//        llamaState: LlamaState(),
+//        modelName: "TheBloke / TinyLlama-1.1B-1T-OpenOrca-GGUF (Q4_0)",
+//        modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
+//        filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
+//    )
+// }

examples/llama.swiftui/llama.swiftui/llama_swiftuiApp.swift

@@ -0,0 +1,10 @@
+import SwiftUI
+
+@main
+struct llama_swiftuiApp: App {
+    var body: some Scene {
+        WindowGroup {
+            ContentView()
+        }
+    }
+}

examples/llava/CMakeLists.txt

@@ -1,20 +1,36 @@
-set(TARGET clip)
-add_library(${TARGET} clip.cpp clip.h)
-install(TARGETS ${TARGET} LIBRARY)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if (NOT MSVC)
-    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
-    endif()
-if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
+add_library(llava OBJECT
+            llava.cpp
+            llava.h
+            clip.cpp
+            clip.h
+            )
+
+target_link_libraries(llava PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+
+target_include_directories(llava PUBLIC .)
+target_include_directories(llava PUBLIC ../..)
+target_include_directories(llava PUBLIC ../../common)
+
+target_compile_features(llava PRIVATE cxx_std_11)
+
+add_library(llava_static STATIC $<TARGET_OBJECTS:llava>)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(llava PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(llava PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    add_library(llava_shared SHARED $<TARGET_OBJECTS:llava>)
+    target_link_libraries(llava_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+    install(TARGETS llava_shared LIBRARY)
 endif()
 
-set(TARGET llava)
-add_executable(${TARGET} llava.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if (NOT MSVC)
+    target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
+    endif()
 if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
+    add_dependencies(llava BUILD_INFO)
 endif()
+
+set(TARGET llava-cli)
+add_executable(llava-cli llava-cli.cpp)
+install(TARGETS llava-cli RUNTIME)
+target_link_libraries(llava-cli PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(llava PRIVATE cxx_std_11)

examples/llava/README.md

@@ -9,12 +9,12 @@ models are available.
 After API is confirmed, more models will be supported / uploaded.
 
 ## Usage
-Build with cmake or run `make llava` to build it.
+Build with cmake or run `make llava-cli` to build it.
 
-After building, run: `./llava` to see the usage. For example:
+After building, run: `./llava-cli` to see the usage. For example:
 
 ```sh
-./llava -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
+./llava-cli -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
@@ -51,7 +51,6 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 
 ## TODO
 
-- [ ] Support server mode.
 - [ ] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.cpp

@@ -112,8 +112,7 @@ static float get_f32(const gguf_context * ctx, const std::string & key) {
 static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::string & name) {
     struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
     if (!cur) {
-        printf("unable to find tensor %s\n", name.c_str());
-        throw std::runtime_error(format("unable to find tensor %s\n", name.c_str()));
+        throw std::runtime_error(format("%s: unable to find tensor %s\n", __func__, name.c_str()));
     }
 
     return cur;
@@ -136,7 +135,7 @@ static std::string get_ftype(int ftype) {
     case 8:
         return "q8_0";
     default:
-        throw std::runtime_error(format("Unrecognized file type: %d\n", ftype));
+        throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype));
     }
 }
 
@@ -462,6 +461,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     };
 
     struct gguf_context * ctx = gguf_init_from_file(fname, params);
+    if (!ctx) {
+        throw std::runtime_error(format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname));
+    }
 
     if (verbosity >= 1) {
         const int n_tensors = gguf_get_n_tensors(ctx);
@@ -512,7 +514,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             ctx_size += padded_size;
             if (verbosity >= 3) {
                 printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i,
-                       cur->n_dims, cur->name, tensor_size, padded_size, offset);
+                       ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
             }
         }
     }
@@ -608,8 +610,8 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
         int idx_std = get_key_idx(ctx, KEY_IMAGE_STD);
         for (int i = 0; i < 3; ++i) {
-            new_clip->image_mean[i] = *((float *)gguf_get_arr_data(ctx, idx_mean));
-            new_clip->image_std[i] = *((float *)gguf_get_arr_data(ctx, idx_std));
+            new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean));
+            new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std));
         }
 
         if (verbosity >= 2) {
@@ -662,7 +664,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 // measure mem requirement and allocate
     {
         static const size_t tensor_alignment = 32;
-        new_clip->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
+        new_clip->buf_compute.resize(ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead());
         new_clip->alloc = ggml_allocr_new_measure(tensor_alignment);
         clip_image_f32_batch batch;
         batch.size = 1;
@@ -678,26 +680,44 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     return new_clip;
 }
 
-clip_image_u8 * make_clip_image_u8() { return new clip_image_u8(); }
-
+clip_image_u8 * make_clip_image_u8() {
+    auto img = new clip_image_u8();
+    return img;
+}
 clip_image_f32 * make_clip_image_f32() { return new clip_image_f32(); }
 
-bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
-    int nx, ny, nc;
-    auto data = stbi_load(fname, &nx, &ny, &nc, 3);
-    if (!data) {
-        fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname);
-        return false;
-    }
+void clip_image_u8_free(clip_image_u8 * img) { if (img->data) { delete[] img->data; } delete img; }
+void clip_image_f32_free(clip_image_f32 * img) { if (img->data) { delete[] img->data; } delete img; }
 
+static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
     img->nx = nx;
     img->ny = ny;
     img->size = nx * ny * 3;
     img->data = new uint8_t[img->size]();
     memcpy(img->data, data, img->size);
+}
 
+bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto data = stbi_load(fname, &nx, &ny, &nc, 3);
+    if (!data) {
+        fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
     stbi_image_free(data);
+    return true;
+}
 
+bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) {
+    int nx, ny, nc;
+    auto data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
+    if (!data) {
+        fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
+        return false;
+    }
+    build_clip_img_from_data(data, nx, ny, img);
+    stbi_image_free(data);
     return true;
 }
 
@@ -712,39 +732,40 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
     // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
 
-    clip_image_u8 temp; // we will keep the input image data here temporarily
+    clip_image_u8 * temp = make_clip_image_u8(); // we will keep the input image data here temporarily
     if (pad2square && img->nx != img->ny) {
         int longer_side = std::max(img->nx, img->ny);
-        temp.nx = longer_side;
-        temp.ny = longer_side;
-        temp.size = 3 * longer_side * longer_side;
-        temp.data = new uint8_t[temp.size]();
-        uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA
+        temp->nx = longer_side;
+        temp->ny = longer_side;
+        temp->size = 3 * longer_side * longer_side;
+        temp->data = new uint8_t[temp->size]();
+        uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
 
         // fill with background color
-        for (size_t i = 0; i < temp.size; i++) {
-            temp.data[i] = bc[i % 3];
+        for (size_t i = 0; i < temp->size; i++) {
+            temp->data[i] = bc[i % 3];
         }
 
         // copy from the input image
         for (int y = 0; y < img->ny; y++) {
             for (int x = 0; x < img->nx; x++) {
                 const int i = 3 * (y * img->nx + x);
-                const int j = 3 * (y * temp.nx + x);
-                temp.data[j] = img->data[i];
-                temp.data[j+1] = img->data[i+1];
-                temp.data[j+2] = img->data[i+2];
+                const int j = 3 * (y * temp->nx + x);
+                temp->data[j] = img->data[i];
+                temp->data[j+1] = img->data[i+1];
+                temp->data[j+2] = img->data[i+2];
             }
         }
     } else {
-        temp.nx   = img->nx;
-        temp.ny   = img->ny;
-        temp.size = img->size;
-        temp.data = img->data;
+        temp->nx   = img->nx;
+        temp->ny   = img->ny;
+        temp->size = img->size;
+        temp->data = new uint8_t[temp->size]();
+        memcpy(&temp->data[0], &img->data[0], temp->size); // copy
     }
 
-    const int nx = temp.nx;
-    const int ny = temp.ny;
+    const int nx = temp->nx;
+    const int ny = temp->ny;
 
     const int nx2 = ctx->vision_model.hparams.image_size;
     const int ny2 = ctx->vision_model.hparams.image_size;
@@ -783,10 +804,10 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
                 const int j10 = 3 * (y1 * nx + x0) + c;
                 const int j11 = 3 * (y1 * nx + x1) + c;
 
-                const float v00 = temp.data[j00];
-                const float v01 = temp.data[j01];
-                const float v10 = temp.data[j10];
-                const float v11 = temp.data[j11];
+                const float v00 = temp->data[j00];
+                const float v01 = temp->data[j01];
+                const float v10 = temp->data[j10];
+                const float v11 = temp->data[j11];
 
                 const float v0 = v00 * (1.0f - dx) + v01 * dx;
                 const float v1 = v10 * (1.0f - dx) + v11 * dx;
@@ -801,6 +822,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
             }
         }
     }
+    clip_image_u8_free(temp);
 
     return true;
 }
@@ -940,7 +962,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         }
 
         // quantize only 2D tensors
-        quantize &= (cur->n_dims == 2);
+        quantize &= (ggml_n_dims(cur) == 2);
 
         if (quantize) {
             new_type = type;
@@ -1013,7 +1035,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             fout.put(0);
         }
 
-        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), cur->n_dims, quantize,
+        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
                orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
     }
 
@@ -1047,16 +1069,16 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     return true;
 }
 
-int clip_n_mmproj_embd(struct clip_ctx * ctx) {
+int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     return ctx->vision_model.mm_2_b->ne[0];
 }
 
-int clip_n_patches(struct clip_ctx * ctx) {
+int clip_n_patches(const struct clip_ctx * ctx) {
     auto & params = ctx->vision_model.hparams;
 
     return (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
 }
 
-size_t clip_embd_nbytes(struct clip_ctx * ctx) {
+size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
     return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }

examples/llava/clip.h

@@ -1,7 +1,22 @@
 #ifndef CLIP_H
 #define CLIP_H
 
-#include "ggml.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define CLIP_API __declspec(dllexport)
+#        else
+#            define CLIP_API __declspec(dllimport)
+#        endif
+#    else
+#        define CLIP_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define CLIP_API
+#endif
 
 struct clip_ctx;
 
@@ -20,19 +35,20 @@ struct clip_vision_hparams {
     float eps;
 };
 
-struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
+/** load mmproj model */
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
+/** free mmproj model */
+CLIP_API void clip_free(struct clip_ctx * ctx);
 
-void clip_free(struct clip_ctx * ctx);
-
-size_t clip_embd_nbytes(struct clip_ctx * ctx);
-int clip_n_patches(struct clip_ctx * ctx);
-int clip_n_mmproj_embd(struct clip_ctx * ctx);
+size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+int clip_n_patches(const struct clip_ctx * ctx);
+int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 // RGB uint8 image
 struct clip_image_u8 {
     int nx;
     int ny;
-    uint8_t * data;
+    uint8_t * data = NULL;
     size_t size;
 };
 
@@ -41,7 +57,7 @@ struct clip_image_u8 {
 struct clip_image_f32 {
     int nx;
     int ny;
-    float * data;
+    float * data = NULL;
     size_t size;
 };
 
@@ -57,7 +73,12 @@ struct clip_image_f32_batch {
 
 struct clip_image_u8 * make_clip_image_u8();
 struct clip_image_f32 * make_clip_image_f32();
-bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+CLIP_API void clip_image_u8_free(clip_image_u8 * img);
+CLIP_API void clip_image_f32_free(clip_image_f32 * img);
+CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
+CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
+
 bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square);
 bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec);
 

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

@@ -5,7 +5,7 @@ import json
 import torch
 import numpy as np
 from gguf import *
-from transformers import CLIPModel, CLIPProcessor
+from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel
 
 TEXT = "clip.text"
 VISION = "clip.vision"
@@ -51,7 +51,7 @@ def bytes_to_unicode():
     The reversible bpe codes work on unicode strings.
     This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
     When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
     To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
     And avoids mapping to whitespace/control characters the bpe code barfs on.
     """
@@ -78,11 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
                 help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
 ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
 
+# with proper
 args = ap.parse_args()
 
 
@@ -96,15 +104,22 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
-
-with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
-    vocab = json.load(f)
-    tokens = [key for key in vocab]
+if args.clip_model_is_vision:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
 
 with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
     config = json.load(f)
-    v_hparams = config["vision_config"]
-    t_hparams = config["text_config"]
+    if args.clip_model_is_vision:
+        v_hparams = config
+        t_hparams = None
+    else:
+        v_hparams = config["vision_config"]
+        t_hparams = config["text_config"]
 
 # possible data types
 #   ftype == 0 -> float32
@@ -117,9 +132,12 @@ ftype = 1
 if args.use_f32:
     ftype = 0
 
-
-model = CLIPModel.from_pretrained(dir_model)
-processor = CLIPProcessor.from_pretrained(dir_model)
+if args.clip_model_is_vision:
+    model = CLIPVisionModel.from_pretrained(dir_model)
+    processor = None
+else:
+    model = CLIPModel.from_pretrained(dir_model)
+    processor = CLIPProcessor.from_pretrained(dir_model)
 
 fname_middle = None
 has_text_encoder = True
@@ -128,13 +146,13 @@ has_llava_projector = False
 if args.text_only:
     fname_middle = "text-"
     has_vision_encoder = False
-elif args.vision_only:
-    fname_middle = "vision-"
-    has_text_encoder = False
 elif args.llava_projector is not None:
     fname_middle = "mmproj-"
     has_text_encoder = False
     has_llava_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
 else:
     fname_middle = ""
 
@@ -182,8 +200,12 @@ if has_vision_encoder:
     block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
 
-    image_mean = processor.image_processor.image_mean if args.image_mean is None else args.image_mean
-    image_std = processor.image_processor.image_std if args.image_std is None else args.image_std
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
     fout.add_array("clip.vision.image_mean", image_mean)
     fout.add_array("clip.vision.image_std", image_std)
 

examples/llava/llava-cli.cpp

@@ -0,0 +1,314 @@
+#include "ggml.h"
+#include "common.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+
+#include "base64.hpp"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+            fprintf(stderr, "%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
+    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+    return true;
+}
+
+// TODO: use common/sampling.h
+static llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
+    auto & sparams = params.sparams;
+
+    // out of user input, sample next token
+    const float   temp      = sparams.temp;
+    const int32_t top_k     = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
+    const float   top_p     = sparams.top_p;
+    const float   tfs_z     = sparams.tfs_z;
+    const float   typical_p = sparams.typical_p;
+    // const int32_t repeat_last_n   = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
+    // const float   repeat_penalty  = sparams.repeat_penalty;
+    // const float   alpha_presence  = sparams.presence_penalty;
+    // const float   alpha_frequency = sparams.frequency_penalty;
+    const int     mirostat     = sparams.mirostat;
+    const float   mirostat_tau = sparams.mirostat_tau;
+    const float   mirostat_eta = sparams.mirostat_eta;
+    // const bool    penalize_nl     = sparams.penalize_nl;
+
+    llama_token id = 0;
+    {
+        auto logits  = llama_get_logits(ctx_llama);
+        auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
+
+        // Apply params.logit_bias map
+        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
+            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 };
+
+        if (temp <= 0) {
+              // Greedy sampling
+            id = llama_sample_token_greedy(ctx_llama, &candidates_p);
+        } else {
+            if (mirostat == 1) {
+                static float mirostat_mu = 2.0f * mirostat_tau;
+                const  int mirostat_m    = 100;
+                llama_sample_temp(ctx_llama, &candidates_p, temp);
+                id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+            } else if (mirostat == 2) {
+                static float mirostat_mu = 2.0f * mirostat_tau;
+                llama_sample_temp(ctx_llama, &candidates_p, temp);
+                id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+            } else {
+                  // Temperature sampling
+                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
+                llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1);
+                llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1);
+                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
+                llama_sample_temp(ctx_llama, &candidates_p, temp);
+                id = llama_sample_token(ctx_llama, &candidates_p);
+            }
+        }
+    }
+
+    return id;
+}
+
+static const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
+    int id = sample_id(ctx_llama, params);
+    static std::string ret;
+    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
+static const char* IMG_BASE64_TAG_END = "\">";
+
+static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
+    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
+    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
+}
+
+static bool prompt_contains_image(const std::string& prompt) {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    return (begin != std::string::npos);
+}
+
+// replaces the base64 image tag in the prompt with `replacement`
+static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
+    size_t img_base64_str_start, img_base64_str_end;
+    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
+    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
+        fprintf(stderr, "%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        return NULL;
+    }
+
+    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
+    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
+    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
+
+    auto required_bytes = base64::required_encode_size(base64_str.size());
+    auto img_bytes = std::vector<unsigned char>(required_bytes);
+    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
+    if (!embed) {
+        fprintf(stderr, "%s: could not load image from base64 string.\n", __func__);
+        return NULL;
+    }
+
+    return embed;
+}
+
+static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    if (begin == std::string::npos || end == std::string::npos) {
+        return prompt;
+    }
+    auto pre = prompt.substr(0, begin);
+    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
+    return pre + replacement + post;
+}
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    printf("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    printf("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
+
+    // load and preprocess the image
+    llava_image_embed * embed = NULL;
+    auto prompt = params->prompt;
+    if (prompt_contains_image(prompt)) {
+        if (!params->image.empty()) {
+            printf("using base64 encoded image instead of command line image path\n");
+        }
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
+        if (!embed) {
+            fprintf(stderr, "%s: can't load image from prompt\n", __func__);
+            return NULL;
+        }
+        params->prompt = remove_image_from_prompt(prompt);
+    } else {
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str());
+        if (!embed) {
+            fprintf(stderr, "%s: is %s really an image file?\n", __func__, params->image.c_str());
+            return NULL;
+        }
+    }
+
+    return embed;
+}
+
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) {
+    int n_past = 0;
+
+    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama));
+
+    // llava chat format is "<system_prompt>\nUSER:<image_embeddings>\n<textual_prompt>\nASSISTANT:"
+    eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos);
+    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
+    eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false);
+
+    // generate the response
+
+    printf("\n");
+
+    for (int i = 0; i < max_tgt_len; i++) {
+        const char * tmp = sample(ctx_llava->ctx_llama, *params, &n_past);
+        if (strcmp(tmp, "</s>") == 0) break;
+
+        printf("%s", tmp);
+        fflush(stdout);
+    }
+
+    printf("\n");
+}
+
+
+static struct llava_context * llava_init(gpt_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+
+    llama_backend_init(params->numa);
+
+    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return NULL;
+    }
+
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->ctx_clip = ctx_clip;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+        gpt_print_usage(argc, argv, params);
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    auto ctx_llava = llava_init(&params);
+    if (ctx_llava == NULL) {
+        fprintf(stderr, "%s: error: failed to init llava\n", __func__);
+        return 1;
+    }
+
+    auto image_embed = load_image(ctx_llava, &params);
+
+    // process the prompt
+    process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+    llama_print_timings(ctx_llava->ctx_llama);
+
+    llava_image_embed_free(image_embed);
+    llava_free(ctx_llava);
+    return 0;
+}

examples/llava/llava-surgery.py

@@ -16,13 +16,29 @@ checkpoint = torch.load(path)
 mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
 
 # store these tensors in a new dictionary and torch.save them
-projector = {name: checkpoint[name] for name in mm_tensors}
+projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
 # remove these tensors from the checkpoint and save it again
 for name in mm_tensors:
     del checkpoint[name]
 
+# BakLLaVA models contain CLIP tensors in it
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/llava.clip")
+
+    # remove these tensors
+    for name in clip_tensors:
+        del checkpoint[name]
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+
 torch.save(checkpoint, path)
 
 print("Done!")

examples/llava/llava-utils.h

@@ -1,145 +0,0 @@
-#pragma once
-
-// this one and clip lib will be eventually merged to a single lib, let's keep it this way for now
-
-#include "common.h"
-#include "llama.h"
-
-#include <cstdio>
-#include <cstdlib>
-#include <vector>
-
-inline bool eval_image_embd(llama_context * ctx_llama, float * embd, int N, int n_batch, int * n_past) {
-    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
-
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = N - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        llama_batch batch = {int32_t(n_eval), nullptr, (embd+i*n_embd), nullptr, nullptr, nullptr, *n_past, 1, 0, };
-        if (llama_decode(ctx_llama, batch)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-inline bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
-    int N = (int) tokens.size();
-    for (int i = 0; i < N; i += n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-inline bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(ctx_llama, tokens, 1, n_past);
-}
-
-inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
-    std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
-    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
-    return true;
-}
-
-// TODO: use common/sampling.h
-inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
-      // out of user input, sample next token
-    const float   temp      = params.sampling_params.temp;
-    const int32_t top_k     = params.sampling_params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : params.sampling_params.top_k;
-    const float   top_p     = params.sampling_params.top_p;
-    const float   tfs_z     = params.sampling_params.tfs_z;
-    const float   typical_p = params.sampling_params.typical_p;
-      // const int32_t repeat_last_n   = params.sampling_params.repeat_last_n < 0 ? n_ctx : params.sampling_params.repeat_last_n;
-      // const float   repeat_penalty  = params.sampling_params.repeat_penalty;
-      // const float   alpha_presence  = params.sampling_params.presence_penalty;
-      // const float   alpha_frequency = params.sampling_params.frequency_penalty;
-    const int     mirostat     = params.sampling_params.mirostat;
-    const float   mirostat_tau = params.sampling_params.mirostat_tau;
-    const float   mirostat_eta = params.sampling_params.mirostat_eta;
-      // const bool    penalize_nl     = params.sampling_params.penalize_nl;
-
-    llama_token id = 0;
-    {
-        auto logits  = llama_get_logits(ctx_llama);
-        auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
-
-          // Apply params.logit_bias map
-        for (auto it = params.sampling_params.logit_bias.begin(); it != params.sampling_params.logit_bias.end(); it++) {
-            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 };
-
-          // TODO: Apply penalties
-          // float nl_logit = logits[llama_token_nl(ctx)];
-          // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), 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(ctx)] = nl_logit;
-          // }
-
-        if (temp <= 0) {
-              // Greedy sampling
-            id = llama_sample_token_greedy(ctx_llama, &candidates_p);
-        } else {
-            if (mirostat == 1) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                const  int mirostat_m    = 100;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-            } else if (mirostat == 2) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-            } else {
-                  // Temperature sampling
-                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
-                llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1);
-                llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1);
-                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token(ctx_llama, &candidates_p);
-            }
-        }
-    }
-
-    return id;
-}
-
-inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
-    int id = sample_id(ctx_llama, params);
-    static std::string ret;
-    if (id == llama_token_eos(ctx_llama)) {
-        ret = "</s>";
-    } else {
-        ret = llama_token_to_piece(ctx_llama, id);
-    }
-    eval_id(ctx_llama, id, n_past);
-    return ret.c_str();
-}

examples/llava/llava.cpp

@@ -1,164 +1,163 @@
 #include "clip.h"
-#include "llava-utils.h"
 #include "common.h"
 #include "llama.h"
+#include "llava.h"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
 
-static void show_additional_info(int /*argc*/, char ** argv) {
-    printf("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    printf("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
-}
+#include "base64.hpp"
 
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        show_additional_info(argc, argv);
-        return 1;
+static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
+    clip_image_f32 * img_res = make_clip_image_f32();
+    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
+        fprintf(stderr, "%s: unable to preprocess image\n", __func__);
+        clip_image_f32_free(img_res);
+        return false;
     }
 
-    if (params.mmproj.empty() || params.image.empty()) {
-        gpt_print_usage(argc, argv, params);
-        show_additional_info(argc, argv);
-        return 1;
-    }
-
-    const char * clip_path = params.mmproj.c_str();
-    const char * img_path = params.image.c_str();
-
-    if (params.prompt.empty()) {
-        params.prompt = "describe the image in detail.";
-    }
-
-    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
-
-    // load and preprocess the image
-    clip_image_u8 img;
-    clip_image_f32 img_res;
-
-    if (!clip_image_load_from_file(img_path, &img)) {
-        fprintf(stderr, "%s: is %s really an image file?\n", __func__, img_path);
-
-        clip_free(ctx_clip);
-        return 1;
-    }
-
-    if (!clip_image_preprocess(ctx_clip, &img, &img_res, /*pad2square =*/ true)) {
-        fprintf(stderr, "%s: unable to preprocess %s\n", __func__, img_path);
-
-        clip_free(ctx_clip);
-        return 1;
-    }
-
-    int n_img_pos  = clip_n_patches(ctx_clip);
-    int n_img_embd = clip_n_mmproj_embd(ctx_clip);
-
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
-
-    if (!image_embd) {
-        fprintf(stderr, "Unable to allocate memory for image embeddings\n");
-
-        return 1;
-    }
+    *n_img_pos = clip_n_patches(ctx_clip);
 
     const int64_t t_img_enc_start_us = ggml_time_us();
-    if (!clip_image_encode(ctx_clip, params.n_threads, &img_res, image_embd)) {
+    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+    clip_image_f32_free(img_res);
+    if (!encoded) {
         fprintf(stderr, "Unable to encode image\n");
 
-        return 1;
+        return false;
     }
+
     const int64_t t_img_enc_end_us = ggml_time_us();
+    float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
-    // we get the embeddings, free up the memory required for CLIP
-    clip_free(ctx_clip);
+    printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
 
-    llama_backend_init(params.numa);
-
-    llama_model_params model_params              = llama_model_default_params();
-                       model_params.n_gpu_layers = params.n_gpu_layers;
-                       model_params.main_gpu     = params.main_gpu;
-                       model_params.tensor_split = params.tensor_split;
-                       model_params.use_mmap     = params.use_mmap;
-                       model_params.use_mlock    = params.use_mlock;
-
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
-    if (model == NULL) {
-        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
-        return 1;
-    }
-
-    llama_context_params ctx_params = llama_context_default_params();
-
-    ctx_params.n_ctx           = params.n_ctx < 2048 ? 2048 : params.n_ctx; // we need a longer context size to process image embeddings
-    ctx_params.n_threads       = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
-    ctx_params.seed            = params.seed;
-
-    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
-
-    if (ctx_llama == NULL) {
-        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
-        return 1;
-    }
-
-    // make sure that the correct mmproj was used, i.e., compare apples to apples
-    const int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
-
-    if (n_img_embd != n_llama_embd) {
-        printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_img_embd, n_llama_embd);
-
-        llama_free(ctx_llama);
-        llama_free_model(model);
-        llama_backend_free();
-        free(image_embd);
-
-        return 1;
-    }
-
-    // process the prompt
-    // llava chat format is "<system_prompt>USER: <image_embeddings>\n<textual_prompt>\nASSISTANT:"
-
-    int n_past = 0;
-
-    const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
-
-    eval_string(ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params.n_batch, &n_past, true);
-    eval_image_embd(ctx_llama, image_embd, n_img_pos, params.n_batch, &n_past);
-    eval_string(ctx_llama, (params.prompt + "\nASSISTANT:").c_str(), params.n_batch, &n_past, false);
-
-    // generate the response
-
-    printf("\n");
-    printf("prompt: '%s'\n", params.prompt.c_str());
-    printf("\n");
-
-    for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(ctx_llama, params, &n_past);
-        if (strcmp(tmp, "</s>") == 0) break;
-
-        printf("%s", tmp);
-        fflush(stdout);
-    }
-
-    printf("\n");
-
-    {
-        const float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
-
-        printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / n_img_pos);
-    }
-
-    llama_print_timings(ctx_llama);
-
-    llama_free(ctx_llama);
-    llama_free_model(model);
-    llama_backend_free();
-    free(image_embd);
-
-    return 0;
+    return true;
+}
+
+bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
+        // make sure that the correct mmproj was used, i.e., compare apples to apples
+    int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
+    auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
+    if (n_image_embd != n_llama_embd) {
+        printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        return false;
+    }
+    return true;
+}
+
+static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
+    if (!image_embd) {
+        fprintf(stderr, "Unable to allocate memory for image embeddings\n");
+        free(image_embd);
+        return false;
+    }
+
+    int n_img_pos;
+    if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
+        fprintf(stderr, "%s: cannot encode image, aborting\n", __func__);
+        free(image_embd);
+        return false;
+    }
+    *image_embd_out = image_embd;
+    *n_img_pos_out = n_img_pos;
+
+    return true;
+}
+
+bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
+
+    for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
+        int n_eval = image_embed->n_image_pos - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        if (llama_decode(ctx_llama, batch)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+    clip_image_u8 * img = make_clip_image_u8();
+    if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
+        clip_image_u8_free(img);
+        fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);
+        return NULL;
+    }
+
+    float* image_embed = NULL;
+    int n_image_pos = 0;
+    bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
+    if (!image_embed_result) {
+        clip_image_u8_free(img);
+        fprintf(stderr, "%s: coulnd't embed the image\n", __func__);
+        return NULL;
+    }
+
+    clip_image_u8_free(img);
+    auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    result->embed = image_embed;
+    result->n_image_pos = n_image_pos;
+    return result;
+}
+
+static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
+    auto file = fopen(path, "rb");
+    if (file == NULL) {
+        fprintf(stderr, "%s: can't read file %s\n", __func__, path);
+        return false;
+    }
+
+    fseek(file, 0, SEEK_END);
+    auto fileSize = ftell(file);
+    fseek(file, 0, SEEK_SET);
+
+    auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
+    if (buffer == NULL) {
+        fprintf(stderr, "%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        perror("Memory allocation error");
+        fclose(file);
+        return false;
+    }
+    errno = 0;
+    size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
+    if (ferror(file)) {
+        die_fmt("read error: %s", strerror(errno));
+    }
+    if (ret != (size_t) fileSize) {
+        die("unexpectedly reached end of file");
+    }
+    fclose(file); // Close the file
+
+    *bytesOut = buffer;
+    *sizeOut = fileSize;
+    return true;
+}
+
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+    unsigned char* image_bytes;
+    long image_bytes_length;
+    auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
+    if (!loaded) {
+        fprintf(stderr, "%s: failed to load %s\n", __func__, image_path);
+        return NULL;
+    }
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    free(image_bytes);
+
+    return embed;
+}
+
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
+    free(embed->embed);
+    free(embed);
 }

examples/llava/llava.h

@@ -0,0 +1,50 @@
+#ifndef LLAVA_H
+#define LLAVA_H
+
+#include "ggml.h"
+
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAVA_API __declspec(dllexport)
+#        else
+#            define LLAVA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAVA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAVA_API
+#endif
+
+struct clip_ctx;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct llava_image_embed {
+    float * embed;
+    int n_image_pos;
+};
+
+/** sanity check for clip <-> llava embed size match */
+LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
+
+/** build an image embed from image file bytes */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
+/** build an image embed from a path to an image filename */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
+/** free an embedding made with llava_image_embed_make_* */
+
+/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
+LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif