fix some spaces added by IDE in math op

ggml-ci

.devops/nix/docker.nix

@@ -1,37 +0,0 @@
-{
-  lib,
-  dockerTools,
-  buildEnv,
-  llama-cpp,
-  interactive ? true,
-  coreutils,
-}:
-
-# A tar that can be fed into `docker load`:
-#
-# $ nix build .#llamaPackages.docker
-# $ docker load < result
-
-# For details and variations cf.
-# - https://nixos.org/manual/nixpkgs/unstable/#ssec-pkgs-dockerTools-buildLayeredImage
-# - https://discourse.nixos.org/t/a-faster-dockertools-buildimage-prototype/16922
-# - https://nixery.dev/
-
-# Approximate (compressed) sizes, at the time of writing, are:
-#
-# .#llamaPackages.docker: 125M;
-# .#llamaPackagesCuda.docker: 537M;
-# .#legacyPackages.aarch64-linux.llamaPackagesXavier.docker: 415M.
-
-dockerTools.buildLayeredImage {
-  name = llama-cpp.pname;
-  tag = "latest";
-
-  contents =
-    [ llama-cpp ]
-    ++ lib.optionals interactive [
-      coreutils
-      dockerTools.binSh
-      dockerTools.caCertificates
-    ];
-}

.devops/nix/package.nix

@@ -1,6 +1,5 @@
 {
   lib,
-  glibc,
   config,
   stdenv,
   mkShell,
@@ -31,11 +30,6 @@
   useRocm ? config.rocmSupport,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
-
-  # It's necessary to consistently use backendStdenv when building with CUDA support,
-  # otherwise we get libstdc++ errors downstream.
-  effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
-  enableStatic ? effectiveStdenv.hostPlatform.isStatic
 }@inputs:
 
 let
@@ -47,7 +41,10 @@ let
     versionOlder
     ;
 
+  # It's necessary to consistently use backendStdenv when building with CUDA support,
+  # otherwise we get libstdc++ errors downstream.
   stdenv = throw "Use effectiveStdenv instead";
+  effectiveStdenv = if useCuda then cudaPackages.backendStdenv else inputs.stdenv;
 
   suffices =
     lib.optionals useBlas [ "BLAS" ]
@@ -170,9 +167,6 @@ effectiveStdenv.mkDerivation (
         # TODO: Replace with autoAddDriverRunpath
         # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
         cudaPackages.autoAddOpenGLRunpathHook
-      ]
-      ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
-        glibc.static
       ];
 
     buildInputs =
@@ -187,7 +181,7 @@ effectiveStdenv.mkDerivation (
       [
         (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
+        (cmakeBool "BUILD_SHARED_LIBS" true)
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
@@ -196,7 +190,6 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
-        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -262,11 +255,11 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL || useVulkan) lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin) || (useVulkan && effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.devops/nix/scope.nix

@@ -12,8 +12,5 @@ lib.makeScope newScope (
   self: {
     inherit llamaVersion;
     llama-cpp = self.callPackage ./package.nix { };
-    docker = self.callPackage ./docker.nix { };
-    docker-min = self.callPackage ./docker.nix { interactive = false; };
-    sif = self.callPackage ./sif.nix { };
   }
 )

.devops/nix/sif.nix

@@ -1,27 +0,0 @@
-{
-  lib,
-  singularity-tools,
-  llama-cpp,
-  bashInteractive,
-  interactive ? false,
-}:
-
-let
-  optionalInt = cond: x: if cond then x else 0;
-in
-singularity-tools.buildImage rec {
-  inherit (llama-cpp) name;
-  contents = [ llama-cpp ] ++ lib.optionals interactive [ bashInteractive ];
-
-  # These are excessive (but safe) for most variants. Building singularity
-  # images requires superuser privileges, so we build them inside a VM in a
-  # writable image of pre-determined size.
-  #
-  # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846
-  #
-  # Expected image sizes:
-  # - cpu/blas: 150M,
-  # - cuda, all gencodes: 560M,
-  diskSize = 4096 + optionalInt llama-cpp.useRocm 16384;
-  memSize = diskSize;
-}

.github/ISSUE_TEMPLATE/bug.md

@@ -7,5 +7,3 @@ assignees: ''
 ---
 
 Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
-
-If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/workflows/build.yml

@@ -145,28 +145,6 @@ jobs:
           cd build
           ctest -L main --verbose
 
-  ubuntu-22-cmake-vulkan:
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libvulkan-dev
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake -DLLAMA_VULKAN=ON ..
-          cmake --build . --config Release -j $(nproc)
-
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -333,7 +311,6 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
-            -DLLAMA_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -362,7 +339,6 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
-            -DLLAMA_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -427,8 +403,6 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'arm64'
-            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -524,7 +498,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -695,7 +669,8 @@ jobs:
         run: |
           cd examples/llama.android
 
-          ./gradlew build --no-daemon
+          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
+          ./gradlew build --no-daemon -Pskip-armeabi-v7a
 
 #  freeBSD-latest:
 #    runs-on: macos-12

.github/workflows/nix-ci-aarch64.yml

@@ -19,6 +19,7 @@ on:
 
 jobs:
   nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -36,8 +37,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -45,7 +46,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Show all output paths
       run: >
           nix run github:nix-community/nix-eval-jobs

.github/workflows/nix-ci.yml

@@ -23,8 +23,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,6 +37,7 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
+    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -50,8 +51,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -59,7 +60,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

.github/workflows/python-check-requirements.yml

@@ -3,14 +3,12 @@ name: Python check requirements.txt
 on:
   push:
     paths:
-      - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
       - 'requirements.txt'
       - 'requirements/*.txt'
   pull_request:
     paths:
-      - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
       - 'requirements.txt'
@@ -28,4 +26,4 @@ jobs:
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script
-        run:  bash scripts/check-requirements.sh
+        run:  bash scripts/check-requirements.sh nocleanup

.github/workflows/server.yml

@@ -1,135 +0,0 @@
-# Server build and tests
-name: Server
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  schedule:
-    -  cron: '0 0 * * *'
-
-jobs:
-  server:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug]
-        include:
-          - build_type: Release
-            sanitizer: ""
-          - build_type: Debug
-            sanitizer: THREAD
-            disabled_on_pr: true
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
-
-    container:
-      image: ubuntu:latest
-      ports:
-        - 8888
-      options: --cpus 4
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
-      - name: Dependencies
-        id: depends
-        run: |
-          apt-get update
-          apt-get -y install \
-            build-essential \
-            git \
-            cmake \
-            python3-pip \
-            wget \
-            language-pack-en
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. \
-              -DLLAMA_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r examples/server/tests/requirements.txt
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
-
-
-  server-windows:
-    runs-on: windows-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake ..  -DLLAMA_BUILD_SERVER=ON -DCMAKE_BUILD_TYPE=Release ;
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
-
-      - name: Python setup
-        id: setup_python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r examples/server/tests/requirements.txt
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
-        run: |
-          cd examples/server/tests
-          behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd examples/server/tests
-          behave.exe --stop --no-skipped --no-capture --tags slow

.github/workflows/tidy-post.yml

@@ -0,0 +1,20 @@
+name: clang-tidy review post comments
+
+on:
+  workflow_dispatch:
+    workflows: ["clang-tidy-review"]
+    types:
+      - completed
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: ZedThree/clang-tidy-review/post@v0.13.0
+        # lgtm_comment_body, max_comments, and annotations need to be set on the posting workflow in a split setup
+        with:
+          # adjust options as necessary
+          lgtm_comment_body: ''
+          annotations: false
+          max_comments: 25

.github/workflows/tidy-review.yml

@@ -0,0 +1,23 @@
+name: clang-tidy-review
+
+on:
+  pull_request:
+    branches:
+      - master
+
+jobs:
+  clang-tidy-review:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v3
+
+    - uses: ZedThree/clang-tidy-review@v0.13.0
+      id: review
+      with:
+        lgtm_comment_body: ''
+        build_dir: build
+        cmake_command: cmake . -B build -DCMAKE_EXPORT_COMPILE_COMMANDS=on
+        split_workflow: true
+
+    - uses: ZedThree/clang-tidy-review/upload@v0.13.0

.gitignore

@@ -25,8 +25,6 @@
 .vscode/
 .idea/
 
-ggml-metal-embed.metal
-
 lcov-report/
 gcovr-report/
 
@@ -47,7 +45,6 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
-/gritlm
 /imatrix
 /infill
 /libllama.so

CMakeLists.txt

@@ -110,15 +110,12 @@ option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
-option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
-set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
 option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
-set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeline parallelism")
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
@@ -148,7 +145,13 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
-add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
+if (LLAMA_FATAL_WARNINGS)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+        add_compile_options(-Werror)
+    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/WX)
+    endif()
+endif()
 
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -200,70 +203,40 @@ if (LLAMA_METAL)
         add_compile_definitions(GGML_METAL_NDEBUG)
     endif()
 
-    # copy ggml-common.h and ggml-metal.metal to bin directory
-    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
-    if (LLAMA_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
-        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-
-        # merge ggml-common.h and ggml-metal.metal into a single file
-        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-
-        add_custom_command(
-            OUTPUT ${METALLIB_EMBED_ASM}
-            COMMAND echo "Embedding Metal library"
-            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
-            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
-            DEPENDS ggml-metal.metal ggml-common.h
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
-    else()
-        if (LLAMA_METAL_SHADER_DEBUG)
-            # custom command to do the following:
-            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-            #
-            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-            #       disabling fast math is needed in order to pass tests/test-backend-ops
-            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-            set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        else()
-            set(XC_FLAGS -O3)
+    if (LLAMA_METAL_SHADER_DEBUG)
+        # custom command to do the following:
+        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+        #
+        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+        set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        if (LLAMA_QKK_64)
+            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
         endif()
 
         add_custom_command(
             OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
             COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
             COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-            DEPENDS ggml-metal.metal ggml-common.h
+            DEPENDS ggml-metal.metal
             COMMENT "Compiling Metal kernels"
-            )
+        )
 
         add_custom_target(
             ggml-metal ALL
             DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            )
-    endif() # LLAMA_METAL_EMBED_LIBRARY
+        )
+    endif()
 
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
@@ -544,10 +517,6 @@ if (LLAMA_HIPBLAS)
 endif()
 
 if (LLAMA_SYCL)
-    if (NOT LLAMA_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$")
-        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA")
-    endif()
-
     if ( NOT DEFINED ENV{ONEAPI_ROOT})
         message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
     endif()
@@ -569,9 +538,6 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
-    if (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-    endif()
 
     set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
@@ -579,11 +545,7 @@ if (LLAMA_SYCL)
     if (WIN32)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
     else()
-        if (LLAMA_SYCL_TARGET STREQUAL "INTEL")
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-        elseif (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl pthread m dl onemkl)
-        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
     endif()
 endif()
 
@@ -779,30 +741,28 @@ function(get_flags CCID CCVER)
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
+    elseif (CCID MATCHES "Intel")
+        if (NOT LLAMA_SYCL)
+            # enable max optimization level when using Intel compiler
+            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            add_link_options(-fuse-ld=lld -static-intel)
+        endif()
     endif()
 
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
 
-if (LLAMA_FATAL_WARNINGS)
-    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        list(APPEND C_FLAGS   -Werror)
-        list(APPEND CXX_FLAGS -Werror)
-    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
-        add_compile_options(/WX)
-    endif()
-endif()
-
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                                  -Werror=implicit-int -Werror=implicit-function-declaration)
-        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                          -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
 
-        list(APPEND C_FLAGS   ${WARNING_FLAGS})
-        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -818,10 +778,9 @@ endif()
 set(CUDA_CXX_FLAGS "")
 
 if (LLAMA_CUBLAS)
-    set(CUDA_FLAGS -use_fast_math)
-
-    if (LLAMA_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
+    if (NOT MSVC)
+        list(APPEND CUDA_FLAGS -Wno-pedantic)
     endif()
 
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
@@ -855,11 +814,7 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
-    endif()
-
-    if (NOT MSVC)
-        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
     endif()
 endif()
 
@@ -958,16 +913,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
-            endif()
+            # Raspberry Pi 2
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
@@ -1150,8 +1099,6 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
-            unicode.h
-            unicode.cpp
             )
 
 target_include_directories(llama PUBLIC .)

Makefile

@@ -2,7 +2,7 @@
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	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 lookup passkey gritlm tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -97,10 +97,9 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -I. -Icommon
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -167,17 +166,13 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
-ifdef LLAMA_SCHED_MAX_COPIES
-	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
-endif
-
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -205,10 +200,6 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
-ifdef LLAMA_SERVER_SSL
-	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
-	MK_LDFLAGS += -lssl -lcrypto
-endif
 
 ifdef LLAMA_CODE_COVERAGE
 	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
@@ -225,10 +216,34 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
 ifeq ($(LLAMA_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
+	MK_CFLAGS += -Werror
 	MK_CXXFLAGS += -Werror
 endif
 
+ifeq ($(CC_IS_CLANG), 1)
+	# clang options
+	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
+	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+
+	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+else
+	# gcc options
+	MK_CFLAGS        += -Wdouble-promotion
+	MK_HOST_CXXFLAGS += -Wno-array-bounds
+
+	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
+		MK_HOST_CXXFLAGS += -Wno-format-truncation
+	endif
+	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
+		MK_HOST_CXXFLAGS += -Wextra-semi
+	endif
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -389,18 +404,10 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	MK_NVCCFLAGS += -use_fast_math
-ifdef LLAMA_FATAL_WARNINGS
-	MK_NVCCFLAGS += -Werror all-warnings
-endif # LLAMA_FATAL_WARNINGS
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
@@ -457,11 +464,11 @@ endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.h
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -546,30 +553,11 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
-ifdef LLAMA_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJS        += ggml-metal-embed.o
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal ggml-common.h
-	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"   >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"        >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"              >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"          >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -581,10 +569,9 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
 override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -592,7 +579,7 @@ override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
+CUDA_CXXFLAGS := $(GF_CXXFLAGS)
 endif
 
 #
@@ -611,7 +598,7 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
-CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
@@ -635,15 +622,12 @@ 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 $@
 
-ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-unicode.o: unicode.cpp unicode.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o
-
-llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
+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 common/log.h
@@ -732,17 +716,14 @@ embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-gritlm: examples/gritlm/gritlm.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -910,7 +891,3 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

Package.swift

@@ -31,7 +31,6 @@ let package = Package(
             sources: [
                 "ggml.c",
                 "llama.cpp",
-                "unicode.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",

README-sycl.md

@@ -1,7 +1,6 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
-- [News](#news)
 - [OS](#os)
 - [Intel GPU](#intel-gpu)
 - [Docker](#docker)
@@ -26,21 +25,6 @@ The llama.cpp for SYCL is used to support Intel GPUs.
 
 For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 
-## News
-
-- 2024.3
-  - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
-  - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
-  - Support detecting all GPUs with level-zero and same top **Max compute units**.
-  - Support OPs
-    - hardsigmoid
-    - hardswish
-    - pool2d
-
-- 2024.1
-  - Create SYCL backend for Intel GPU.
-  - Support Windows build
-
 ## OS
 
 |OS|Status|Verified|
@@ -73,29 +57,6 @@ For iGPU, please make sure the shared memory from host memory is enough. For lla
 
 For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
 
-## Nvidia GPU
-
-### Verified
-
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Ampere Series| Support| A100|
-
-### oneMKL
-
-The current oneMKL release does not contain the oneMKL cuBlas backend.
-As a result for Nvidia GPU's oneMKL must be built from source.
-
-```
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-mkdir build
-cd build
-cmake -G Ninja .. -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON
-ninja
-// Add paths as necessary
-```
-
 ## Docker
 
 Note:
@@ -209,9 +170,6 @@ source /opt/intel/oneapi/setvars.sh
 # Or, for FP32:
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
-# For Nvidia GPUs
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
 # Build example/main only
 #cmake --build . --config Release --target main
 
@@ -314,7 +272,7 @@ Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact
 
 a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
 
-Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
+Recommend to install to default folder: **/opt/intel/oneapi**.
 
 Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
 
@@ -491,7 +449,6 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 |-|-|-|
 |GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
 |GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
-|ZES_ENABLE_SYSMAN| 0 (default) or 1|Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer|
 
 ## Known Issue
 
@@ -501,10 +458,6 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
   Solution: add **--no-mmap** or **--mmap 0**.
 
-- Split-mode: [row] is not supported
-
-  It's on developing.
-
 ## Q&A
 
 - Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.

README.md

@@ -8,19 +8,16 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-### Recent API changes
-
-- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
-- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
-- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
-- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
-
 ### Hot topics
 
-- Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
-- Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
-- Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
-- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
+- Remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD: https://github.com/ggerganov/llama.cpp/pull/5240
+- Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
+  - [SYCL backend](README-sycl.md) is ready (1/28/2024), support Linux/Windows in Intel GPUs (iGPU, Arc/Flex/Max series)
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
+- Collecting Apple Silicon performance stats:
+  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
+  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
+- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
 
@@ -64,7 +61,7 @@ variety of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
-- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
+- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
 - Vulkan, SYCL, and (partial) OpenCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@@ -110,21 +107,16 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
 - [x] [InternLM2](https://huggingface.co/models?search=internlm2)
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
-- [x] [Gemma](https://ai.google.dev/gemma)
-- [x] [Mamba](https://github.com/state-spaces/mamba)
 
 **Multimodal models:**
 
-- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
+- [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)
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 
-**HTTP server**
-
-[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
 **Bindings:**
 
@@ -153,7 +145,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
-- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
 - [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
@@ -165,9 +156,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
-- [Msty](https://msty.app) (proprietary)
-- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 
 ---
 
@@ -198,7 +186,7 @@ llm_load_print_meta: vocab type     = SPM
 llm_load_print_meta: n_vocab        = 32000
 llm_load_print_meta: n_merges       = 0
 llm_load_print_meta: n_ctx_train    = 4096
-llm_load_print_meta: n_ctx          = 512
+llm_load_print_meta: kv_size        = 512
 llm_load_print_meta: n_embd         = 5120
 llm_load_print_meta: n_head         = 40
 llm_load_print_meta: n_head_kv      = 40
@@ -226,7 +214,7 @@ llama_new_context_with_model: compute buffer total size =   75.41 MB
 
 system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
 sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
-generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0
+generate: kv_size = 512, n_batch = 512, n_predict = 400, n_keep = 0
 
 
  Building a website can be done in 10 simple steps:
@@ -793,7 +781,7 @@ And after 4.45 hours, you will have the final perplexity.
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
-In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
+In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
 Here is an example of a few-shot interaction, invoked with the command
 
@@ -857,7 +845,7 @@ Sample run:
 ```
 == Running in interactive mode. ==
  - Press Ctrl+C to interject at any time.
- - Press Return to return control to LLaMA.
+ - Press Return to return control to LLaMa.
  - If you want to submit another line, end your input in '\'.
 
  Below is an instruction that describes a task. Write a response that appropriately completes the request.
@@ -904,9 +892,6 @@ First, install the essential packages for termux:
 pkg install clang wget git cmake
 ```
 Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
-
-You can execute the following commands on your computer to avoid downloading the NDK to your mobile. Of course, you can also do this in Termux.
-
 ```
 $ mkdir build-android
 $ cd build-android
@@ -915,28 +900,7 @@ $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROI
 $ make
 ```
 Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
-```
-
-Here is a demo of an interactive session running on Pixel 5 phone:
+Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 

awq-py/README.md

@@ -0,0 +1,116 @@
+# AWQ: Activation-aware Weight Quantization for LLM - version apply to llamacpp
+[[Paper](https://arxiv.org/abs/2306.00978)][[Original Repo](https://github.com/mit-han-lab/llm-awq)][[Easy-to-use Repo](https://github.com/casper-hansen/AutoAWQ)]
+
+**Supported models:**
+
+- [X] LLaMA
+- [x] LLaMA 2
+- [X] MPT
+- [X] Mistral AI v0.1
+- [ ] Bloom
+- [ ] Mixtral MoE
+
+**TODO:**
+- [x] Update version work with both MPT and MPT-AWQ model
+- [ ] Add OPT model
+- [ ] Add Bloom model
+- [ ] Add Mixtral MoE
+- [ ] Support w3, w2
+
+
+## Contents
+
+- [Install](##Install)
+- [Convert](##Convert)
+- [Quantize](##Quantize)
+- [Test](##Test)
+- [Benchmark](##Benchmark)
+- [Results](##Results)
+
+## Install
+Install requirements
+```bash
+pip install -r requirements.txt
+```
+Get the pre-computed AWQ search results for multiple model families, including LLaMA, LLaMA2, MPT, OPT
+```bash
+git clone https://huggingface.co/datasets/mit-han-lab/awq-model-zoo awq_cache
+```
+
+## Convert
+Example for llama model
+```bash
+# For llama7b and llama2 models
+python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
+# For mistral and mpt models
+python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/mpt-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
+```
+
+## Quantize
+```bash
+# We only benchmark and confirm the results on q4_0, q4_1, and q2_k types.
+./quantize models/llama_7b_fp16.gguf models/llama_7b_q4_0.gguf q4_0
+```
+
+## Test
+```bash
+# For all models.
+./build/bin/main -m models/llama_7b_q4_0.gguf -n 128 --prompt "Once upon a time"
+```
+
+## Benchmark
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+```bash
+# For llama and llama2, and mistral models.
+./perplexity -m models/llama_7b_q4_0.gguf -f datasets/wikitext-2-raw/wiki.test.raw
+```
+
+## Results
+Results are run on OpenBLAS (CPU) and CuBLAS (GPU) for fair comparison
+We use three types of llamacpp quantization methods to work with our version, including q4_0, q4_1, and q2_k
+
+### Llama 7B (Build with OpenBLAS)
+
+| Model      | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-----------:|--------------|-------:|-------:|-------:|-------:|
+|Llama 7B    | perplexity   | 5.9066 | 6.1214 | 6.0643 | 6.5808 |
+|Llama 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama 7B| perplexity   | 5.9175 | 6.0252 | 5.9987 | 6.3692 |
+|AWQ-LLama 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Llama2 7B (Build with CuBLAS)
+
+| Model       | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|------------:|--------------|-------:|-------:|-------:|-------:|
+|Llama2 7B    | perplexity   | 5.8664 | 6.0260 | 6.0656 | 6.4496 |
+|Llama2 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama2 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama2 7B| perplexity   | 5.8801 | 6.0054 | 5.9849 | 6.3650 |
+|AWQ-LLama2 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama2 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Mistral 7B v0.1 (Build with CuBLAS)
+
+| Model        | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-------------:|--------------|-------:|-------:|-------:|-------:|
+|Mistral 7B    | perplexity   | 5.6931 | 5.8202 | 5.8268 | 6.1645 |
+|Mistral 7B    | file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|Mistral 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-Mistral 7B| perplexity   | 5.6934 | 5.8020 | 5.7691 | 6.0426 |
+|AWQ-Mistral 7B| file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|AWQ-Mistral 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+### MPT 7B (Build with OpenBLAS)
+
+| Model    | Measure      | F16    | Q4_0   | Q4_1   | Q2_K    |
+|---------:|--------------|-------:|-------:|-------:|--------:|
+|MPT 7B    | perplexity   | 8.4369 | 8.7956 | 8.6265 | 11.4913 |
+|MPT 7B    | file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|MPT 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |
+|AWQ-MPT 7B| perplexity   | 8.4944 | 8.7053 |  8.6750 | 10.2873|
+|AWQ-MPT 7B| file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|AWQ-MPT 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |

awq-py/awq/apply_awq.py

@@ -0,0 +1,254 @@
+"""
+Implements the AWQ for llama.cpp use cases.
+Original paper: https://arxiv.org/abs/2306.00978
+
+This code is based on versions of the AWQ implementation found in the following repositories:
+* https://github.com/mit-han-lab/llm-awq
+* https://github.com/casper-hansen/AutoAWQ
+"""
+
+import os
+import torch
+import torch.nn as nn
+
+from transformers import AutoModelForCausalLM, AutoConfig
+from transformers.models.bloom.modeling_bloom import BloomGelu
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
+from transformers.activations import GELUActivation
+
+
+class ScaledActivation(nn.Module):
+    """
+    ScaledActivation module wraps an existing activation function and applies a
+    scale factor to its output.
+
+    Args:
+        module (nn.Module): The activation function to be scaled.
+        scales (torch.Tensor): A tensor of size (num_features,) containing the initial
+            scale factors for each feature.
+
+    Returns:
+        torch.Tensor: The scaled output of the activation function.
+    """
+
+    def __init__(self, module, scales):
+        super().__init__()
+        self.act = module
+        self.scales = nn.Parameter(scales.data)
+
+    def forward(self, x):
+        return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
+
+
+def set_op_by_name(layer, name, new_module):
+    """
+    Set the new module for given module's name.
+
+    Args:
+        layer (nn.Module): The layer in which to replace the submodule.
+        name (str): The path to the submodule to be replaced, using dot notation
+            to access nested modules.
+        new_module (nn.Module): The new module to replace the existing one.
+    """
+    levels = name.split(".")
+    if len(levels) > 1:
+        mod_ = layer
+        for l_idx in range(len(levels) - 1):
+            if levels[l_idx].isdigit():
+                mod_ = mod_[int(levels[l_idx])]
+            else:
+                mod_ = getattr(mod_, levels[l_idx])
+        setattr(mod_, levels[-1], new_module)
+    else:
+        setattr(layer, name, new_module)
+
+
+def get_op_by_name(module, op_name):
+    """
+    Retrieves a submodule within a given layer based on its name.
+
+    Args:
+        module (nn.Module): The layer containing the submodule to find.
+        op_name (str): The name of the submodule.
+
+    Returns:
+        nn.Module: The requested submodule found within the given layer.
+
+    Raises:
+        ValueError: If the specified submodule cannot be found within the layer.
+    """
+    for name, m in module.named_modules():
+        if name == op_name:
+            return m
+    raise ValueError(f"Cannot find op {op_name} in module {module}")
+
+
+@torch.no_grad()
+def scale_ln_fcs(ln, fcs, scales):
+    """
+    Scales the weights of a LayerNorm and a list of fully-connected layers proportionally.
+
+    Args:
+        ln (nn.LayerNorm): The LayerNorm module to be scaled.
+        fcs (List[nn.Linear]): A list of fully-connected layers to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+
+    if not isinstance(fcs, list):
+        fcs = [fcs]
+
+    scales = scales.to(ln.weight.device)
+
+    ln.weight.div_(scales)
+    if hasattr(ln, "bias") and ln.bias is not None:
+        ln.bias.div_(scales)
+
+    for fc in fcs:
+        fc.weight.mul_(scales.view(1, -1))
+
+    for p in ln.parameters():
+        assert torch.isnan(p).sum() == 0
+    for fc in fcs:
+        for p in fc.parameters():
+            assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_fc_fc(fc1, fc2, scales):
+    """
+    Scales the weights of two fully-connected layers in a specific pattern.
+
+    Args:
+        fc1 (nn.Linear): The first fully-connected layer to be scaled.
+        fc2 (nn.Linear): The second fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+    assert isinstance(fc1, nn.Linear)
+    assert isinstance(fc2, nn.Linear)
+
+    scales = scales.to(fc1.weight.device)
+
+    fc1.weight[-scales.size(0):].div_(scales.view(-1, 1))
+    if fc1.bias is not None:
+        fc1.bias.div_(scales.view(-1))
+
+    fc2.weight.mul_(scales.view(1, -1))
+
+    for p in fc1.parameters():
+        assert torch.isnan(p).sum() == 0
+    for p in fc2.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_gelu_fc(gelu, fc, scales):
+    """
+    Scales the weight of a GELU activation and a fully-connected layer proportionally.
+
+    Args:
+        gelu (Union[nn.GELU, BloomGelu, GELUActivation]): The GELU activation module to be scaled.
+        fc (nn.Linear): The fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+
+    Raises:
+        TypeError: If the `gelu` module is not of type `nn.GELU`, `BloomGelu`, or `GELUActivation`.
+        TypeError: If the `fc` module is not of type `nn.Linear`.
+    """
+    assert isinstance(gelu, (nn.GELU, BloomGelu, GELUActivation))
+    assert isinstance(fc, nn.Linear)
+
+    fc.weight.mul_(scales.view(1, -1).to(fc.weight.device))
+
+    for p in fc.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+def apply_scale(module, scales_list, input_feat_dict=None):
+    """
+    Applies different scaling strategies to layers based on their type and hierarchy within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layers to be scaled.
+        scales_list (List[Tuple[str, List[str], torch.Tensor]]): A list of tuples containing:
+            * prev_op_name (str): The name of the preceding operation or module,
+                relative to which the layers to be scaled are located.
+            * layer_names (List[str]): A list of names of the layers to be scaled, relative to the preceding operation.
+            * scales (torch.Tensor): A 1D tensor of size (num_features,) containing the scaling factors for each feature.
+        input_feat_dict (Optional[Dict[str, torch.Tensor]]): A dictionary mapping layer names to their corresponding
+            input features (optional).
+    """
+    for prev_op_name, layer_names, scales in scales_list:
+        prev_op = get_op_by_name(module, prev_op_name)
+        layers = [get_op_by_name(module, name) for name in layer_names]
+
+        prev_op.cuda()
+        for layer in layers:
+            layer.cuda()
+        scales.cuda()
+
+        if isinstance(prev_op, nn.Linear):
+            assert len(layers) == 1
+            scale_fc_fc(prev_op, layers[0], scales)
+        elif isinstance(prev_op, (nn.LayerNorm, LlamaRMSNorm)) or "rmsnorm" in str(prev_op.__class__).lower():
+            scale_ln_fcs(prev_op, layers, scales)
+        elif isinstance(prev_op, (nn.GELU, BloomGelu, GELUActivation)):
+            new_module = ScaledActivation(prev_op, scales)
+            set_op_by_name(module, prev_op_name, new_module)
+            scale_gelu_fc(prev_op, layers[0], scales)
+        else:
+            raise NotImplementedError(f"prev_op {type(prev_op)} not supported yet!")
+
+        # apply the scaling to input feat if given; prepare it for clipping
+        if input_feat_dict is not None:
+            for layer_name in layer_names:
+                inp = input_feat_dict[layer_name]
+                inp.div_(scales.view(1, -1).to(inp.device))
+
+        prev_op.cpu()
+        for layer in layers:
+            layer.cpu()
+        scales.cpu()
+
+
+@torch.no_grad()
+def apply_clip(module, clip_list):
+    """
+    Applies element-wise clipping to the weight of a specific layer within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layer to be clipped.
+        clip_list (List[Tuple[str, torch.Tensor]]): A list of tuples containing:
+            * name (str): The name of the layer to be clipped, relative to the root of the module.
+            * max_val (torch.Tensor): A 1D or 2D tensor defining the upper bound for each element of the layer's weight.
+    """
+    for name, max_val in clip_list:
+        layer = get_op_by_name(module, name)
+        layer.cuda()
+        max_val = max_val.to(layer.weight.device)
+        org_shape = layer.weight.shape
+        layer.weight.data = layer.weight.data.reshape(*max_val.shape[:2], -1)
+        layer.weight.data = torch.clamp(layer.weight.data, -max_val, max_val)
+        layer.weight.data = layer.weight.data.reshape(org_shape)
+        layer.cpu()
+
+
+def add_scale_weights(model_path, scale_path, tmp_path):
+    """
+    Adds pre-computed Activation Weight Quantization (AWQ) results to a model,
+    including scaling factors and clipping bounds.
+
+    Args:
+        model_path (str): Path to the pre-trained model to be equipped with AWQ.
+        scale_path (str): Path to the AWQ scale factors (.pt file).
+        tmp_path (str): Path to the temporary directory where the equipped model will be saved.
+    """
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, config=config, trust_remote_code=True
+    )
+    model.eval()
+    awq_results = torch.load(str(scale_path), map_location="cpu")
+    apply_scale(model, awq_results["scale"])
+    apply_clip(model, awq_results["clip"])
+    model.save_pretrained(str(tmp_path))
+    os.system(f"cp {str(model_path)}/tokenizer* {str(tmp_path)}")

awq-py/requirements.txt

@@ -0,0 +1,2 @@
+torch>=2.1.1
+transformers>=4.32.0

build.zig

@@ -115,7 +115,6 @@ pub fn build(b: *std.build.Builder) !void {
     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 unicode = make.obj("unicode", "unicode.cpp");
     const llama = make.obj("llama", "llama.cpp");
     const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
@@ -124,16 +123,15 @@ pub fn build(b: *std.build.Builder) !void {
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, 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, unicode, common, buildinfo, 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, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -45,8 +45,7 @@ fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
     if [ -z ${ONEAPI_ROOT} ]; then
-        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
-        echo "source /opt/intel/oneapi/setvars.sh"
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
 
@@ -273,19 +272,19 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
@@ -344,17 +343,17 @@ function gg_run_open_llama_3b_v2 {
     python3 ../convert-lora-to-ggml.py ${path_lora}
 
     # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
     compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
     compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
     compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     set +e

common/CMakeLists.txt

@@ -19,12 +19,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         endif()
     endif()
 
-    if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
-    else()
-        message(WARNING "Git index not found in git repository.")
-        set(GIT_INDEX "")
-    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 "")

common/common.cpp

@@ -258,11 +258,19 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             }
             sparams.top_k = std::stoi(argv[i]);
         } else if (arg == "-c" || arg == "--ctx-size") {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.kv_size = std::stoi(argv[i]);
+            fprintf(stderr, "warning: -c,--ctx-size option is deprecated, use --kv-size instead");
+        } else if (arg == "-kv" || arg == "--kv-size" || arg == "--kv_size") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.n_ctx = std::stoi(argv[i]);
+            params.kv_size = std::stoi(argv[i]);
         } else if (arg == "--grp-attn-n" || arg == "-gan") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -295,9 +303,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::string value(argv[i]);
-            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
-            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
-            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
+            /**/ 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) {
@@ -335,22 +343,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--pooling") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::string value(argv[i]);
-            /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; }
-            else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; }
-            else if (value == "cls")  { params.pooling_type = LLAMA_POOLING_TYPE_CLS; }
-            else { invalid_param = true; break; }
-        } else if (arg == "--defrag-thold" || arg == "-dt") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.defrag_thold = std::stof(argv[i]);
         } else if (arg == "--samplers") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -483,12 +475,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_batch = std::stoi(argv[i]);
-        } else if (arg == "-ub" || arg == "--ubatch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_ubatch = std::stoi(argv[i]);
         } else if (arg == "--keep") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -519,6 +505,12 @@ bool gpt_params_parse_ex(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;
@@ -568,34 +560,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.lora_base = argv[i];
-        } else if (arg == "--control-vector") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.control_vectors.push_back({ 1.0f, argv[i], });
-        } else if (arg == "--control-vector-scaled") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            const char * fname = argv[i];
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.control_vectors.push_back({ std::stof(argv[i]), fname, });
-        } else if (arg == "--control-vector-layer-range") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.control_vector_layer_start = std::stoi(argv[i]);
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.control_vector_layer_end = std::stoi(argv[i]);
         } else if (arg == "--mmproj") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -674,15 +638,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             }
             std::string arg_next = argv[i];
             if (arg_next == "none") {
-                params.split_mode = LLAMA_SPLIT_MODE_NONE;
+                params.split_mode = LLAMA_SPLIT_NONE;
             } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
+                params.split_mode = LLAMA_SPLIT_LAYER;
             } else if (arg_next == "row") {
-#ifdef GGML_USE_SYCL
-                fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n");
-                exit(1);
-#endif // GGML_USE_SYCL
-                params.split_mode = LLAMA_SPLIT_MODE_ROW;
+                params.split_mode = LLAMA_SPLIT_ROW;
             } else {
                 invalid_param = true;
                 break;
@@ -885,15 +845,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             sep++;
             if (strncmp(sep, "int:", 4) == 0) {
                 sep += 4;
-                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+                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_TYPE_FLOAT;
+                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_TYPE_BOOL;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
                 if (std::strcmp(sep, "true") == 0) {
                     kvo.bool_value = true;
                 } else if (std::strcmp(sep, "false") == 0) {
@@ -1010,10 +970,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -bf FNAME, --binary-file FNAME\n");
     printf("                        binary file containing multiple choice tasks.\n");
     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  logical maximum batch size (default: %d)\n", params.n_batch);
-    printf("  -ub N, --ubatch-size N\n");
-    printf("                        physical maximum batch size (default: %d)\n", params.n_ubatch);
+    printf("  -kv N, --kv-size N    Specify the total size of the KV cache (default: %d)\n", params.kv_size);
+    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 \';\'\n");
     printf("                        (default: %s)\n", sampler_type_names.c_str());
     printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
@@ -1022,7 +980,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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.penalty_last_n);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = kv_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);
@@ -1054,14 +1012,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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("  --pooling {none,mean,cls}\n");
-    printf("                        pooling type for embeddings, use model default if unspecified\n");
-    printf("  -dt N, --defrag-thold N\n");
-    printf("                        KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
     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("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
-    printf("  --all-logits          return logits for all tokens in the batch (default: disabled)\n");
+    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");
     printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
     printf("  --winogrande          compute Winogrande score over random tasks from datafile supplied with -f\n");
@@ -1074,6 +1028,7 @@ 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");
@@ -1123,12 +1078,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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");
-    printf("  --control-vector FNAME\n");
-    printf("                        add a control vector\n");
-    printf("  --control-vector-scaled FNAME S\n");
-    printf("                        add a control vector with user defined scaling S\n");
-    printf("  --control-vector-layer-range START END\n");
-    printf("                        layer range to apply the control vector(s) to, start and end inclusive\n");
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
     printf("  -md FNAME, --model-draft FNAME\n");
@@ -1328,15 +1277,14 @@ static ggml_type kv_cache_type_from_str(const std::string & 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_seq_max         = params.n_parallel;
+    cparams.kv_size           = params.kv_size;
     cparams.n_batch           = params.n_batch;
-    cparams.n_ubatch          = params.n_ubatch;
     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.embeddings        = params.embedding;
+    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;
@@ -1345,8 +1293,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     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.pooling_type      = params.pooling_type;
-    cparams.defrag_thold      = params.defrag_thold;
     cparams.offload_kqv       = !params.no_kv_offload;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
@@ -1394,30 +1340,6 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         return std::make_tuple(nullptr, nullptr);
     }
 
-    if (!params.control_vectors.empty()) {
-        if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
-        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
-
-        const auto cvec = llama_control_vector_load(params.control_vectors);
-        if (cvec.n_embd == -1) {
-            llama_free(lctx);
-            llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
-        }
-
-        int err = llama_control_vector_apply(lctx,
-                                             cvec.data.data(),
-                                             cvec.data.size(),
-                                             cvec.n_embd,
-                                             params.control_vector_layer_start,
-                                             params.control_vector_layer_end);
-        if (err) {
-            llama_free(lctx);
-            llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
-        }
-    }
-
     for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
         const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]);
         float lora_scale = std::get<1>(params.lora_adapter[i]);
@@ -1446,7 +1368,6 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         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_clear(lctx);
-        llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }
 
@@ -1745,7 +1666,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cfg_scale: %f # default: 1.0\n", sparams.cfg_scale);
     fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
-    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "kv_size: %d # default: 512\n", params.kv_size);
     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.penalty_freq);
@@ -1803,6 +1724,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
     fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
+    fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
     fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
@@ -1854,17 +1776,17 @@ 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_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+        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_seq_max) {
+    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_seq_max; j++) {
+        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))]);
@@ -1877,14 +1799,14 @@ 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_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+        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_seq_max) {
-        for (int j = 0; j < view.n_seq_max; j++) {
+    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; }
@@ -1903,11 +1825,11 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
 
     c_curr = view.cells;
     cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+    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_seq_max; j++) {
+        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]) : '+');
@@ -1920,188 +1842,3 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
 
     printf("\n=== Done dumping\n");
 }
-
-void llama_embd_normalize(const float * inp, float * out, int n) {
-    double sum = 0.0;
-    for (int i = 0; i < n; i++) {
-        sum += inp[i] * inp[i];
-    }
-    sum = sqrt(sum);
-
-    const float norm = sum > 0.0 ? 1.0f / sum : 0.0f;
-
-    for (int i = 0; i < n; i++) {
-        out[i] = inp[i] * norm;
-    }
-}
-
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){
-    double sum  = 0.0;
-    double sum1 = 0.0;
-    double sum2 = 0.0;
-
-    for (int i = 0; i < n; i++) {
-        sum  += embd1[i] * embd2[i];
-        sum1 += embd1[i] * embd1[i];
-        sum2 += embd2[i] * embd2[i];
-    }
-
-    return sum / (sqrt(sum1) * sqrt(sum2));
-}
-
-//
-// Control vector utils
-//
-
-static llama_control_vector_data llama_control_vector_load_one(const llama_control_vector_load_info & load_info) {
-    int32_t n_tensors;
-
-    size_t n_bytes = 0;
-
-    uint32_t max_direction_layer = 0;
-
-    llama_control_vector_data result = { -1, {} };
-
-    // calculate size of ctx needed for tensors, ensure tensors are f32, and find max layer
-    {
-        struct ggml_init_params meta_params = {
-            /* .mem_size   = */ ggml_tensor_overhead() * 128 + ggml_graph_overhead(),
-            /* .mem_buffer = */ nullptr,
-            /* .no_alloc   = */ true,
-        };
-        ggml_context * meta_ctx = ggml_init(meta_params);
-        struct gguf_init_params meta_gguf_params = {
-            /* .no_alloc = */ true,
-            /* .ctx      = */ &meta_ctx,
-        };
-        struct gguf_context * meta_ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params);
-        if (!meta_ctx_gguf) {
-            fprintf(stderr, "%s: failed to load control vector from %s\n", __func__, load_info.fname.c_str());
-            ggml_free(meta_ctx);
-            return result;
-        }
-
-        n_tensors = gguf_get_n_tensors(meta_ctx_gguf);
-        for (int i = 0; i < n_tensors; i++) {
-            std::string name = gguf_get_tensor_name(meta_ctx_gguf, i);
-
-            // split on '.'
-            size_t dotpos = name.find('.');
-            if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") {
-                try {
-                    uint32_t layer = std::stoi(name.substr(dotpos + 1));
-                    if (layer == 0) {
-                        fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str());
-                        ggml_free(meta_ctx);
-                        gguf_free(meta_ctx_gguf);
-                        return result;
-                    }
-                    if (layer > max_direction_layer) {
-                        max_direction_layer = layer;
-                    }
-                } catch (...) {
-                    fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str());
-                    ggml_free(meta_ctx);
-                    gguf_free(meta_ctx_gguf);
-                    return result;
-                }
-            }
-
-            struct ggml_tensor * tensor_meta = ggml_get_tensor(meta_ctx, name.c_str());
-            if (tensor_meta->type != GGML_TYPE_F32 || ggml_n_dims(tensor_meta) != 1) {
-                fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str());
-                ggml_free(meta_ctx);
-                gguf_free(meta_ctx_gguf);
-                return result;
-            }
-            if (result.n_embd == -1) {
-                result.n_embd = ggml_nelements(tensor_meta);
-            } else if (ggml_nelements(tensor_meta) != result.n_embd) {
-                fprintf(stderr, "%s: direction tensor sizes mismatched in %s\n", __func__, load_info.fname.c_str());
-                ggml_free(meta_ctx);
-                gguf_free(meta_ctx_gguf);
-                return result;
-            }
-            n_bytes += ggml_nbytes(tensor_meta);
-        }
-        ggml_free(meta_ctx);
-        gguf_free(meta_ctx_gguf);
-    }
-
-    if (n_tensors == 0) {
-        fprintf(stderr, "%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str());
-        return result;
-    }
-
-    // load and scale tensors into final control vector context
-    struct ggml_init_params ggml_params = {
-        /* .mem_size   = */ ggml_tensor_overhead() * n_tensors + n_bytes,
-        /* .mem_buffer = */ nullptr,
-        /* .no_alloc   = */ false,
-    };
-    struct ggml_context * ctx = ggml_init(ggml_params);
-
-    struct gguf_init_params params = {
-        /*.no_alloc = */ false,
-        /*.ctx      = */ &ctx,
-    };
-    struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), params);
-    if (!ctx_gguf) {
-        fprintf(stderr, "%s: failed to load control vector from %s\n", __func__, load_info.fname.c_str());
-        ggml_free(ctx);
-        return result;
-    }
-
-    // do not store data for layer 0 (it's not used)
-    result.data.resize(result.n_embd * max_direction_layer);
-
-    for (uint32_t il = 1; il <= max_direction_layer; il++) {
-        const std::string name = "direction." + std::to_string(il);
-        const ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str());
-
-        float * dst = result.data.data() + result.n_embd * (il - 1);
-
-        if (tensor) {
-            const float * src = (const float *) tensor->data;
-            for (int j = 0; j < result.n_embd; j++) {
-                dst[j] = src[j] * load_info.strength;
-            }
-        } else {
-            for (int j = 0; j < result.n_embd; j++) {
-                dst[j] = 0.0f;
-            }
-        }
-    }
-
-    return result;
-}
-
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos) {
-    llama_control_vector_data result = { -1, {} };
-
-    for (const auto & info : load_infos) {
-        auto cur = llama_control_vector_load_one(info);
-
-        if (cur.n_embd == -1) {
-            return result;
-        }
-        if (result.n_embd != -1 && (result.n_embd != cur.n_embd || result.data.size() != cur.data.size())) {
-            fprintf(stderr, "%s: control vector in %s does not match previous vector dimensions\n", __func__, info.fname.c_str());
-            return result;
-        }
-
-        if (result.n_embd == -1) {
-            result = std::move(cur);
-        } else {
-            for (size_t i = 0; i < cur.data.size(); i++) {
-                result.data[i] += cur.data[i];
-            }
-        }
-    }
-
-    if (result.n_embd == -1) {
-        fprintf(stderr, "%s: no vectors passed\n", __func__);
-    }
-
-    return result;
-}

common/common.h

@@ -37,34 +37,31 @@ extern char const *LLAMA_COMMIT;
 extern char const *LLAMA_COMPILER;
 extern char const *LLAMA_BUILD_TARGET;
 
-struct llama_control_vector_load_info;
-
-int32_t get_num_physical_cores();
-
 //
 // CLI argument parsing
 //
+int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+    uint32_t seed                 = -1;    // RNG seed
 
     int32_t n_threads             = get_num_physical_cores();
     int32_t n_threads_draft       = -1;
     int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_threads_batch_draft = -1;
     int32_t n_predict             = -1;    // new tokens to predict
-    int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t kv_size               = 512;   // KV Cache 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               = 5;     // number of tokens to draft during speculative decoding
+    int32_t n_draft               = 8;     // 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)
-    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    llama_split_mode split_mode   = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
@@ -78,12 +75,8 @@ struct gpt_params {
     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
-    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
-
-    ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
-
-    llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-    llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED;
+    ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -106,11 +99,6 @@ struct gpt_params {
     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
 
-    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
-
-    int32_t control_vector_layer_start = -1; // layer range for control vector
-    int32_t control_vector_layer_end   = -1; // layer range for control vector
-
     int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     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)
@@ -126,6 +114,7 @@ struct gpt_params {
 
     bool   kl_divergence   = false; // compute KL-divergence
 
+    bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     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
@@ -269,32 +258,3 @@ 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);
-
-//
-// Embedding utils
-//
-
-void llama_embd_normalize(const float * inp, float * out, int n);
-
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
-
-//
-// Control vector utils
-//
-
-struct llama_control_vector_data {
-    int n_embd;
-
-    // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
-    std::vector<float> data;
-};
-
-struct llama_control_vector_load_info {
-    float strength;
-
-    std::string fname;
-};
-
-// Load control vectors, scale each by strength, and add them together.
-// On error, returns {-1, empty}
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);

common/grammar-parser.cpp

@@ -278,22 +278,6 @@ namespace grammar_parser {
             while (*pos) {
                 pos = parse_rule(state, pos);
             }
-            // Validate the state to ensure that all rules are defined
-            for (const auto & rule : state.rules) {
-                for (const auto & elem : rule) {
-                    if (elem.type == LLAMA_GRETYPE_RULE_REF) {
-                        // Ensure that the rule at that location exists
-                        if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) {
-                            // Get the name of the rule that is missing
-                            for (const auto & kv : state.symbol_ids) {
-                                if (kv.second == elem.value) {
-                                    throw std::runtime_error("Undefined rule identifier '" + kv.first + "'");
-                                }
-                            }
-                        }
-                    }
-                }
-            }
             return state;
         } catch (const std::exception & err) {
             fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());

common/log.h

@@ -297,7 +297,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #ifndef _MSC_VER
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
-    #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
 #endif
 
 // Main TEE macro.
@@ -311,7 +311,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #ifndef _MSC_VER
     #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
-    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
 #endif
 
 // LOG macro variants with auto endline.
@@ -319,8 +319,8 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else
-    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
-    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n")
 #endif
 
 // INTERNAL, DO NOT USE

common/sampling.cpp

@@ -17,13 +17,6 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
             return nullptr;
         }
 
-        // Ensure that there is a "root" node.
-        if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
-            fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
-            delete result;
-            return nullptr;
-        }
-
         std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
 
         result->grammar = llama_grammar_init(
@@ -273,7 +266,7 @@ static llama_token llama_sampling_sample_impl(
             //    }
             //}
 
-            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
         }
     }
 
@@ -302,77 +295,6 @@ static llama_token llama_sampling_sample_impl(
     return id;
 }
 
-static llama_token_data_array llama_sample_probability_distribution_impl(
-                  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 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 bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
-
-    // Get a pointer to the logits
-    float * logits = llama_get_logits_ith(ctx_main, idx);
-
-    // Declare original_logits at the beginning of the function scope
-    std::vector<float> original_logits;
-
-    // apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    if (ctx_cfg) {
-        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
-        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
-    }
-
-    cur.clear();
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
-
-    // apply penalties
-    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
-    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
-    if (penalty_tokens_used_size) {
-        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-
-        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
-                penalty_tokens_used_size, 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(llama_get_model(ctx_main))) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    // apply grammar checks
-    if (ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    }
-
-    llama_sample_softmax(ctx_main, &cur_p);
-    return cur_p;
-}
-
 llama_token llama_sampling_sample(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
@@ -382,14 +304,6 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
-llama_token_data_array llama_sampling_probability_distribution(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx) {
-    return llama_sample_probability_distribution_impl(ctx_sampling,ctx_main, ctx_cfg, idx);
-}
-
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -131,13 +131,6 @@ llama_token llama_sampling_sample(
         struct llama_context * ctx_cfg,
         int idx = 0);
 
-// returns the probability that token of given id will be sampled
-llama_token_data_array llama_sampling_probability_distribution(
-        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,

common/train.cpp

@@ -31,7 +31,7 @@ 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_TYPE_ADAM);
+    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;
 
@@ -556,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
     std::string opt_type;
     GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
     if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
-        opt->params.type = GGML_OPT_TYPE_ADAM;
+        opt->params.type = GGML_OPT_ADAM;
 
         GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
         GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@@ -568,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
         copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
         copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
     } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
-        opt->params.type = GGML_OPT_TYPE_LBFGS;
+        opt->params.type = GGML_OPT_LBFGS;
 
         GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
         GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@@ -603,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
     gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
 
     switch (opt->params.type) {
-        case GGML_OPT_TYPE_ADAM:
+        case GGML_OPT_ADAM:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
                 gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
@@ -622,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
                     gguf_add_tensor(fctx, opt->adam.pf);
                 }
             } break;
-        case GGML_OPT_TYPE_LBFGS:
+        case GGML_OPT_LBFGS:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
                 gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);

convert-hf-to-gguf.py

@@ -8,10 +8,9 @@ import json
 import os
 import re
 import sys
-from abc import ABC, abstractmethod
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterator, Sequence, TypeVar, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
 
 import numpy as np
 import torch
@@ -37,12 +36,7 @@ class SentencePieceTokenTypes(IntEnum):
     BYTE = 6
 
 
-AnyModel = TypeVar("AnyModel", bound="type[Model]")
-
-
-class Model(ABC):
-    _model_classes: dict[str, type[Model]] = {}
-
+class Model:
     def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
         self.dir_model = dir_model
         self.ftype = ftype
@@ -53,14 +47,10 @@ class Model(ABC):
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
+        self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
 
-    @property
-    @abstractmethod
-    def model_arch(self) -> gguf.MODEL_ARCH:
-        pass
-
     def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
         key = next((k for k in keys if k in self.hparams), None)
         if key is not None:
@@ -106,11 +96,9 @@ class Model(ABC):
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
 
-        if (rope_theta := self.hparams.get("rope_theta")) is not None:
-            self.gguf_writer.add_rope_freq_base(rope_theta)
         if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
             self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
-        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_eps(f_norm_eps)
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
@@ -186,22 +174,51 @@ class Model(ABC):
         with open(dir_model / "config.json", "r", encoding="utf-8") as f:
             return json.load(f)
 
-    @classmethod
-    def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
-        assert names
-
-        def func(modelcls: type[Model]):
-            for name in names:
-                cls._model_classes[name] = modelcls
-            return modelcls
-        return func
-
-    @classmethod
-    def from_model_architecture(cls, arch):
-        try:
-            return cls._model_classes[arch]
-        except KeyError:
-            raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
+    @staticmethod
+    def from_model_architecture(model_architecture):
+        if model_architecture == "GPTNeoXForCausalLM":
+            return GPTNeoXModel
+        if model_architecture == "BloomForCausalLM":
+            return BloomModel
+        if model_architecture == "MPTForCausalLM":
+            return MPTModel
+        if model_architecture in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return BaichuanModel
+        if model_architecture in ("FalconForCausalLM", "RWForCausalLM"):
+            return FalconModel
+        if model_architecture == "GPTBigCodeForCausalLM":
+            return StarCoderModel
+        if model_architecture == "GPTRefactForCausalLM":
+            return RefactModel
+        if model_architecture == "PersimmonForCausalLM":
+            return PersimmonModel
+        if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+            return StableLMModel
+        if model_architecture == "QWenLMHeadModel":
+            return QwenModel
+        if model_architecture == "Qwen2ForCausalLM":
+            return Model
+        if model_architecture == "MixtralForCausalLM":
+            return MixtralModel
+        if model_architecture == "GPT2LMHeadModel":
+            return GPT2Model
+        if model_architecture == "PhiForCausalLM":
+            return Phi2Model
+        if model_architecture == "PlamoForCausalLM":
+            return PlamoModel
+        if model_architecture == "CodeShellForCausalLM":
+            return CodeShellModel
+        if model_architecture == "OrionForCausalLM":
+            return OrionModel
+        if model_architecture == "InternLM2ForCausalLM":
+            return InternLM2Model
+        if model_architecture == "MiniCPMForCausalLM":
+            return MiniCPMModel
+        if model_architecture == "BertModel":
+            return BertModel
+        if model_architecture == "NomicBertModel":
+            return NomicBertModel
+        return Model
 
     def _is_model_safetensors(self) -> bool:
         return Model.count_model_parts(self.dir_model, ".safetensors") > 0
@@ -216,6 +233,53 @@ class Model(ABC):
             return ("pytorch_model.bin",)
         return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
 
+    def _get_model_architecture(self) -> gguf.MODEL_ARCH:
+        arch = self.hparams["architectures"][0]
+        if arch == "GPTNeoXForCausalLM":
+            return gguf.MODEL_ARCH.GPTNEOX
+        if arch == "BloomForCausalLM":
+            return gguf.MODEL_ARCH.BLOOM
+        if arch == "MPTForCausalLM":
+            return gguf.MODEL_ARCH.MPT
+        if arch in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return gguf.MODEL_ARCH.BAICHUAN
+        if arch in ("FalconForCausalLM", "RWForCausalLM"):
+            return gguf.MODEL_ARCH.FALCON
+        if arch == "GPTBigCodeForCausalLM":
+            return gguf.MODEL_ARCH.STARCODER
+        if arch == "GPTRefactForCausalLM":
+            return gguf.MODEL_ARCH.REFACT
+        if arch == "PersimmonForCausalLM":
+            return gguf.MODEL_ARCH.PERSIMMON
+        if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+            return gguf.MODEL_ARCH.STABLELM
+        if arch == "QWenLMHeadModel":
+            return gguf.MODEL_ARCH.QWEN
+        if arch == "Qwen2ForCausalLM":
+            return gguf.MODEL_ARCH.QWEN2
+        if arch == "MixtralForCausalLM":
+            return gguf.MODEL_ARCH.LLAMA
+        if arch == "GPT2LMHeadModel":
+            return gguf.MODEL_ARCH.GPT2
+        if arch == "PhiForCausalLM":
+            return gguf.MODEL_ARCH.PHI2
+        if arch == "PlamoForCausalLM":
+            return gguf.MODEL_ARCH.PLAMO
+        if arch == "CodeShellForCausalLM":
+            return gguf.MODEL_ARCH.CODESHELL
+        if arch == "OrionForCausalLM":
+            return gguf.MODEL_ARCH.ORION
+        if arch == "InternLM2ForCausalLM":
+            return gguf.MODEL_ARCH.INTERNLM2
+        if arch == "MiniCPMForCausalLM":
+            return gguf.MODEL_ARCH.MINICPM
+        if arch == "BertModel":
+            return gguf.MODEL_ARCH.BERT
+        if arch == "NomicBertModel":
+            return gguf.MODEL_ARCH.NOMIC_BERT
+
+        raise NotImplementedError(f'Architecture "{arch}" not supported!')
+
     def _set_vocab_gpt2(self):
         dir_model = self.dir_model
         hparams = self.hparams
@@ -383,10 +447,7 @@ class Model(ABC):
         special_vocab.add_to_gguf(self.gguf_writer)
 
 
-@Model.register("GPTNeoXForCausalLM")
 class GPTNeoXModel(Model):
-    model_arch = gguf.MODEL_ARCH.GPTNEOX
-
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
 
@@ -403,10 +464,7 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
 
-@Model.register("BloomForCausalLM")
 class BloomModel(Model):
-    model_arch = gguf.MODEL_ARCH.BLOOM
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name("Bloom")
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
@@ -498,10 +556,7 @@ class BloomModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
-@Model.register("MPTForCausalLM")
 class MPTModel(Model):
-    model_arch = gguf.MODEL_ARCH.MPT
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
         self.gguf_writer.add_name(self.dir_model.name)
@@ -563,11 +618,13 @@ class MPTModel(Model):
 
             self.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":
+                self.gguf_writer.add_tensor("output.weight", data)
+
 
-@Model.register("OrionForCausalLM")
 class OrionModel(Model):
-    model_arch = gguf.MODEL_ARCH.ORION
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -598,8 +655,6 @@ class OrionModel(Model):
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
-        # note: config provides rms norm but it is actually layer norm
-        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
     def write_tensors(self):
@@ -646,10 +701,7 @@ class OrionModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
 class BaichuanModel(Model):
-    model_arch = gguf.MODEL_ARCH.BAICHUAN
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -764,10 +816,7 @@ class BaichuanModel(Model):
         return weights[r * n_part:r * n_part + r, ...]
 
 
-@Model.register("FalconForCausalLM", "RWForCausalLM")
 class FalconModel(Model):
-    model_arch = gguf.MODEL_ARCH.FALCON
-
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_hidden_layers")
         if block_count is None:
@@ -860,10 +909,7 @@ class FalconModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("GPTBigCodeForCausalLM")
 class StarCoderModel(Model):
-    model_arch = gguf.MODEL_ARCH.STARCODER
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -878,10 +924,7 @@ class StarCoderModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
 
-@Model.register("GPTRefactForCausalLM")
 class RefactModel(Model):
-    model_arch = gguf.MODEL_ARCH.REFACT
-
     def set_gguf_parameters(self):
         hidden_dim = self.hparams["n_embd"]
         inner_dim = 4 * hidden_dim
@@ -965,10 +1008,7 @@ class RefactModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("PersimmonForCausalLM")
 class PersimmonModel(Model):
-    model_arch = gguf.MODEL_ARCH.PERSIMMON
-
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
         head_count = self.hparams["num_attention_heads"]
@@ -991,6 +1031,7 @@ class PersimmonModel(Model):
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -1016,10 +1057,7 @@ class PersimmonModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
 class StableLMModel(Model):
-    model_arch = gguf.MODEL_ARCH.STABLELM
-
     def set_vocab(self):
         if (self.dir_model / "tokenizer.json").is_file():
             self._set_vocab_gpt2()
@@ -1036,25 +1074,18 @@ class StableLMModel(Model):
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
-        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
-        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
-@Model.register("MixtralForCausalLM")
 class MixtralModel(Model):
-    model_arch = gguf.MODEL_ARCH.LLAMA
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
 
-@Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
-    model_arch = gguf.MODEL_ARCH.MINICPM
-
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
         self.gguf_writer.add_name("MiniCPM")
@@ -1131,10 +1162,7 @@ class MiniCPMModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("QWenLMHeadModel")
 class QwenModel(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN
-
     @staticmethod
     def token_bytes_to_string(b):
         from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
@@ -1214,15 +1242,7 @@ class QwenModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("Qwen2ForCausalLM")
-class Qwen2Model(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN2
-
-
-@Model.register("GPT2LMHeadModel")
 class GPT2Model(Model):
-    model_arch = gguf.MODEL_ARCH.GPT2
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_block_count(self.hparams["n_layer"])
@@ -1284,10 +1304,7 @@ class GPT2Model(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
-@Model.register("PhiForCausalLM")
 class Phi2Model(Model):
-    model_arch = gguf.MODEL_ARCH.PHI2
-
     def set_gguf_parameters(self):
         block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
@@ -1309,10 +1326,7 @@ class Phi2Model(Model):
         self.gguf_writer.add_add_bos_token(False)
 
 
-@Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
-    model_arch = gguf.MODEL_ARCH.PLAMO
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -1391,10 +1405,7 @@ class PlamoModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("CodeShellForCausalLM")
 class CodeShellModel(Model):
-    model_arch = gguf.MODEL_ARCH.CODESHELL
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -1459,10 +1470,7 @@ class CodeShellModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
-@Model.register("InternLM2ForCausalLM")
 class InternLM2Model(Model):
-    model_arch = gguf.MODEL_ARCH.INTERNLM2
-
     def set_vocab(self):
         # (TODO): Is there a better way?
         # Copy from _set_vocab_sentencepiece, The only difference is that we will treat the character
@@ -1634,10 +1642,7 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
-@Model.register("BertModel")
 class BertModel(Model):
-    model_arch = gguf.MODEL_ARCH.BERT
-
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.vocab_size = None
@@ -1647,17 +1652,16 @@ class BertModel(Model):
         self.gguf_writer.add_causal_attention(False)
 
         # get pooling path
+        with open(self.dir_model / "modules.json", encoding="utf-8") as f:
+            modules = json.load(f)
         pooling_path = None
-        module_path = self.dir_model / "modules.json"
-        if module_path.is_file():
-            with open(module_path, encoding="utf-8") as f:
-                modules = json.load(f)
-            for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
-                    pooling_path = mod["path"]
-                    break
+        for mod in modules:
+            if mod["type"] == "sentence_transformers.models.Pooling":
+                pooling_path = mod["path"]
+                break
 
         # get pooling type
+        pooling_type = gguf.PoolingType.NONE
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
@@ -1667,7 +1671,8 @@ class BertModel(Model):
                 pooling_type = gguf.PoolingType.CLS
             else:
                 raise NotImplementedError("Only MEAN and CLS pooling types supported")
-            self.gguf_writer.add_pooling_type(pooling_type)
+
+        self.gguf_writer.add_pooling_type(pooling_type.value)
 
     def set_vocab(self):
         path = self.dir_model
@@ -1743,10 +1748,7 @@ class BertModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("NomicBertModel")
 class NomicBertModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
-
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
 
@@ -1783,188 +1785,6 @@ class NomicBertModel(BertModel):
             yield name, data
 
 
-@Model.register("GemmaForCausalLM")
-class GemmaModel(Model):
-    model_arch = gguf.MODEL_ARCH.GEMMA
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-    def set_gguf_parameters(self):
-        hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
-        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
-        self.gguf_writer.add_key_length(hparams["head_dim"])
-        self.gguf_writer.add_value_length(hparams["head_dim"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
-            if name.endswith("norm.weight"):
-                data_torch = data_torch + 1
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("Starcoder2ForCausalLM")
-class StarCoder2Model(Model):
-    model_arch = gguf.MODEL_ARCH.STARCODER2
-
-
-@Model.register("MambaForCausalLM", "MambaLMHeadModel")
-class MambaModel(Model):
-    model_arch = gguf.MODEL_ARCH.MAMBA
-
-    def set_vocab(self):
-        vocab_size = self.hparams["vocab_size"]
-        # Round vocab size to next multiple of 8
-        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
-        # pad using ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
-        self.hparams["vocab_size"] = vocab_size
-
-        if (self.dir_model / "tokenizer.json").is_file():
-            self._set_vocab_gpt2()
-        else:
-            # Use the GPT-NeoX tokenizer when no tokenizer files are present
-            tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
-            print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
-            neox_reader = gguf.GGUFReader(tokenizer_path, "r")
-
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
-            self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
-            self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
-            self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
-            self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
-            self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
-            self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
-            self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
-
-    def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size", "d_model"])
-        d_conv  = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
-        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
-        # ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
-        dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
-        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
-        # Fail early for models which don't have a block expansion factor of 2
-        assert d_inner == 2 * d_model
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
-        self.gguf_writer.add_embedding_length(d_model)
-        self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
-        self.gguf_writer.add_ssm_conv_kernel(d_conv)
-        self.gguf_writer.add_ssm_inner_size(d_inner)
-        self.gguf_writer.add_ssm_state_size(d_state)
-        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
-        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams["n_layer"]
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        tok_embd = None
-        tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight"
-        output_name   = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT]     + ".weight"
-
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            if name.endswith(".A_log"):
-                print("A_log --> A ==> " + new_name)
-                data_torch = -torch.exp(data_torch)
-
-            # assuming token_embd.weight is seen before output.weight
-            if tok_embd is not None and new_name == output_name:
-                if torch.equal(tok_embd, data_torch):
-                    print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
-                    continue
-            if new_name == tok_embd_name:
-                tok_embd = data_torch
-
-            data = data_torch.squeeze().numpy()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.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 self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert big float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
 ###### CONVERSION LOGIC ######
 
 

convert-llama-ggml-to-gguf.py

@@ -373,7 +373,7 @@ def handle_metadata(cfg, hp):
         raise ValueError('Unable to load metadata')
     vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
     vocab_factory = convert.VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), cfg.model_metadata_dir)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
     return params, vocab, special_vocab
 
@@ -398,8 +398,8 @@ def handle_args():
                         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")
-    parser.add_argument("--vocabtype", default="spm,hfft",
-                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
+    parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
     return parser.parse_args()
 
 

convert.py

@@ -1282,32 +1282,35 @@ def load_some_model(path: Path) -> ModelPlus:
 
 
 class VocabFactory:
-    _FILES = {"spm": "tokenizer.model", "bpe": "vocab.json", "hfft": "tokenizer.json"}
-
     def __init__(self, path: Path):
         self.path = path
-        self.file_paths = self._detect_files()
-        print(f"Found vocab files: {self.file_paths}")
+        self.files: dict[str, Path | None] = {
+            "tokenizer.model": None,
+            "vocab.json": None,
+            "tokenizer.json": None,
+        }
+        self._detect_files()
 
-    def _detect_files(self) -> dict[str, Path | None]:
-        def locate(file: str) -> Path | None:
-            if (path := self.path / file).exists():
-                return path
-            if (path := self.path.parent / file).exists():
-                return path
-            return None
+    def _detect_files(self):
+        for file in self.files.keys():
+            file_path = self.path / file
+            parent_file_path = self.path.parent / file
+            if file_path.exists():
+                self.files[file] = file_path
+            elif parent_file_path.exists():
+                self.files[file] = parent_file_path
+        print(f"Found vocab files: {self.files}")
 
-        return {vt: locate(f) for vt, f in self._FILES.items()}
-
-    def _select_file(self, vocab_types: list[str]) -> tuple[str, Path]:
-        for vtype in vocab_types:
-            try:
-                path = self.file_paths[vtype]
-            except KeyError:
-                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
-            if path is not None:
-                return vtype, path
-        raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")
+    def _select_file(self, vocabtype: str | None) -> Path:
+        if vocabtype in ["spm", "bpe"]:
+            for file_key in self.files.keys():
+                if (file := self.files[file_key]) is not None:
+                    return file
+            raise FileNotFoundError(f"{vocabtype} vocab not found.")
+        if vocabtype == "hfft":
+            # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file
+            return self.path
+        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
     def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
         load_merges = vocabtype == "bpe"
@@ -1319,30 +1322,30 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
-        vocab_type, path = self._select_file(vocab_types)
-        print(f"Loading vocab file {path!r}, type {vocab_type!r}")
+    def load_vocab(self, vocabtype: str, model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+        path = self._select_file(vocabtype)
+        print(f"Loading vocab file '{path}', type '{vocabtype}'")
 
         added_tokens_path = path.parent / "added_tokens.json"
         vocab: Vocab
-        if vocab_type == "bpe":
+        if vocabtype == "bpe":
             vocab = BpeVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "spm":
+        elif vocabtype == "spm":
             vocab = SentencePieceVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "hfft":
+        elif vocabtype == "hfft":
             vocab = HfVocab(
-                path.parent, added_tokens_path if added_tokens_path.exists() else None
+                path, added_tokens_path if added_tokens_path.exists() else None
             )
         else:
-            raise ValueError(vocab_type)
+            raise ValueError(f"Unsupported vocabulary type {vocabtype}")
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
-            vocab_type,
+            vocabtype,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1376,13 +1379,15 @@ def main(args_in: list[str] | None = None) -> None:
     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")
+    vocab_types = ["spm", "bpe", "hfft"]
+    parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
+    parser.add_argument("--awq-path",     type=Path,              help="Path to scale awq cache file", default=None)
     parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
     parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
+    parser.add_argument("--vocab-type",   choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
     parser.add_argument("--outfile",      type=Path,              help="path to write to; default: based on input")
     parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
@@ -1392,6 +1397,18 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
+    if args.awq_path:
+        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
+        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
+        tmp_model_path = args.model / "weighted_model"
+        if tmp_model_path.is_dir():
+            print(f"{tmp_model_path} exists as a weighted model.")
+        else:
+            tmp_model_path.mkdir(parents=True, exist_ok=True)
+            print("Saving new weighted model ...")
+            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
+            print(f"Saved weighted model at {tmp_model_path}.")
+        args.model = tmp_model_path
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1431,7 +1448,7 @@ def main(args_in: list[str] | None = None) -> None:
     model_parent_path = model_plus.paths[0].parent
     vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
     vocab_factory = VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type.split(","), model_parent_path)
+    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type, model_parent_path)
 
     if args.vocab_only:
         if not args.outfile:

examples/CMakeLists.txt

@@ -20,7 +20,6 @@ else()
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(finetune)
-    add_subdirectory(gritlm)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)

examples/Miku.sh

@@ -7,11 +7,11 @@ USER_NAME="${USER_NAME:-Anon}"
 
 # Uncomment and adjust to the number of CPU cores you want to use.
 #N_THREAD="${N_THREAD:-4}"
-CTX_SIZE="${CTX_SIZE:-4096}"
+KV_SIZE="${KV_SIZE:-4096}"
 N_PREDICTS="${N_PREDICTS:-4096}"
 
 GEN_OPTIONS=(--batch_size 1024
---ctx_size "$CTX_SIZE"
+--kv_size "$KV_SIZE"
 --keep -1
 --repeat_last_n 256
 --repeat_penalty 1.17647

examples/alpaca.sh

@@ -10,7 +10,7 @@ cd ..
 ./main -m ./models/alpaca.13b.ggmlv3.q8_0.bin \
        --color \
        -f ./prompts/alpaca.txt \
-       --ctx_size 2048 \
+       --kv_size 2048 \
        -n -1 \
        -ins -b 256 \
        --top_k 10000 \

examples/baby-llama/baby-llama.cpp

@@ -532,16 +532,16 @@ static struct ggml_tensor * forward(
                 // Vcur shape [n_embd, N, 1, 1]
                 struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wv, cur), n_embd, N)));
 
-                // kv_self.k shape [n_embd * n_ctx * n_layer, 1]
-                // kv_self.v shape [n_embd * n_ctx * n_layer, 1]
+                // kv_self.k shape [n_embd * kv_size * n_layer, 1]
+                // kv_self.v shape [n_embd * kv_size * n_layer, 1]
                 // k         shape [n_embd * N, 1]   == kv_self.k[:,n_past:n_past+N,il,0]
                 // v         shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0]
 
                 /* {
-                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*kv_size + n_past));
                     struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
-                            (   n_ctx)*ggml_element_size(kv_self.v),
-                            (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                            (   kv_size)*ggml_element_size(kv_self.v),
+                            (il*kv_size)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
 
                     // important: storing RoPE-ed version of K in the KV cache!
                     ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
@@ -560,7 +560,7 @@ static struct ggml_tensor * forward(
                         Qcur,
                         0, 2, 1, 3);
 
-            // kv_self.k shape [n_embd * n_ctx * n_layer, 1]
+            // kv_self.k shape [n_embd * kv_size * n_layer, 1]
             // K shape [n_embd/n_head, n_past + N, n_head, 1]
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
@@ -780,16 +780,16 @@ static struct ggml_tensor * forward_batch(
 
                 assert_shape_3d(Vcur, N, n_embd, n_batch);
 
-                // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
-                // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer]
+                // kv_self.k shape [n_embd * kv_size * n_batch * n_layer]
+                // kv_self.v shape [kv_size * n_embd * n_batch * n_layer]
                 // k         shape [n_embd * N, n_batch]   == kv_self.k[:,n_past:n_past+N,:,il]
                 // v         shape [N, n_embd, n_batch, 1] == kv_self.v[:,n_past:n_past+N,:,il]
 
                 /* {
-                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*kv_size + n_past));
                     struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
-                            (   n_ctx)*ggml_element_size(kv_self.v),
-                            (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                            (   kv_size)*ggml_element_size(kv_self.v),
+                            (il*kv_size)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
 
                     // important: storing RoPE-ed version of K in the KV cache!
                     ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
@@ -817,7 +817,7 @@ static struct ggml_tensor * forward_batch(
                         0, 2, 1, 3);
             assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
 
-            // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
+            // kv_self.k shape [n_embd * kv_size * n_batch * n_layer]
             // K shape [n_embd/n_head, n_past + N, n_head, n_batch]
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
@@ -855,7 +855,7 @@ static struct ggml_tensor * forward_batch(
             assert_shape_4d(KQ_soft_max, n_past + N, N, n_head, n_batch);
 
             // split cached V into n_head heads
-            // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer]
+            // kv_self.v shape [kv_size * n_embd * n_batch * n_layer]
             // V shape [n_past + N, n_embd/n_head, n_head, n_batch] == kv_self.v[:(n_past+N),:,:,il]
             struct ggml_tensor * V =
                 ggml_view_4d(ctx0, vc,
@@ -1082,16 +1082,16 @@ static struct ggml_tensor * forward_lora(
                                                                 cur)),
                                                         n_embd, N)));
 
-                // kv_self.k shape [n_embd * n_ctx * n_layer, 1]
-                // kv_self.v shape [n_embd * n_ctx * n_layer, 1]
+                // kv_self.k shape [n_embd * kv_size * n_layer, 1]
+                // kv_self.v shape [n_embd * kv_size * n_layer, 1]
                 // k         shape [n_embd * N, 1]   == kv_self.k[:,n_past:n_past+N,il,0]
                 // v         shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0]
 
                 /* {
-                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*kv_size + n_past));
                     struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
-                            (   n_ctx)*ggml_element_size(kv_self.v),
-                            (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                            (   kv_size)*ggml_element_size(kv_self.v),
+                            (il*kv_size)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
 
                     // important: storing RoPE-ed version of K in the KV cache!
                     ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
@@ -1110,7 +1110,7 @@ static struct ggml_tensor * forward_lora(
                         Qcur,
                         0, 2, 1, 3);
 
-            // kv_self.k shape [n_embd * n_ctx * n_layer, 1]
+            // kv_self.k shape [n_embd * kv_size * n_layer, 1]
             // K shape [n_embd/n_head, n_past + N, n_head, 1]
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
@@ -1470,7 +1470,7 @@ int main(int argc, char ** argv) {
 /*
     struct llama_model_lora model_lora;
     // model.hparams.n_vocab = 6;
-    // model.hparams.n_ctx   = 64;
+    // model.hparams.kv_size = 64;
     // model.hparams.n_embd  = 128;
     // model.hparams.n_mult  = 2;
     // model.hparams.n_head  = 8;
@@ -1478,7 +1478,7 @@ int main(int argc, char ** argv) {
     // model.hparams.n_rot   = model.hparams.n_embd / model.hparams.n_head;
 
     model_lora.hparams.n_vocab = 16;
-    model_lora.hparams.n_ctx   = 32;
+    model_lora.hparams.kv_size = 32;
     model_lora.hparams.n_embd  = 256;
     model_lora.hparams.n_mult  = 2;
     model_lora.hparams.n_head  = 16;
@@ -1533,28 +1533,27 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        struct ggml_cgraph * gf = NULL;
-        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
+        ggml_cgraph gf = {};
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
-        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
+        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
         // struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
-        ggml_build_forward_expand(gf, e);
-        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(&gf, e);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
-        ggml_build_forward_expand(gf, e);
-        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(&gf, e);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1601,14 +1600,13 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            struct ggml_cgraph * gf = NULL;
-            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
+            ggml_cgraph gf = {};
 
             int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
+            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
 
-            ggml_build_forward_expand(gf, logits);
-            ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+            ggml_build_forward_expand(&gf, logits);
+            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

examples/batched-bench/batched-bench.cpp

@@ -32,15 +32,16 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>\n" , argv[0]);
         printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
-        printf("  example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        printf("  example: %s ggml-model-f16.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
         return 1 ;
     }
 
     int n_kv_max     = 2048;
     int is_pp_shared = 0;
     int n_gpu_layers = 0;
+    int mmq          = 0;
 
     std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
     std::vector<int> n_tg = { 128, 256, };
@@ -64,15 +65,19 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 6) {
-        n_pp = parse_list(argv[5]);
+        mmq = std::atoi(argv[5]);
     }
 
     if (argc >= 7) {
-        n_tg = parse_list(argv[6]);
+        n_pp = parse_list(argv[6]);
     }
 
     if (argc >= 8) {
-        n_pl = parse_list(argv[7]);
+        n_tg = parse_list(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_pl = parse_list(argv[8]);
     }
 
     // init LLM
@@ -99,15 +104,13 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed      = 1234;
-    ctx_params.n_ctx     = n_kv_max;
+    ctx_params.kv_size   = n_kv_max;
     ctx_params.n_batch   = 512;
+    ctx_params.mul_mat_q = mmq;
 
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
-    // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
-
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
     if (ctx == NULL) {
@@ -138,8 +141,6 @@ int main(int argc, char ** argv) {
                 LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
             }
-
-            llama_synchronize(ctx);
         }
 
         return true;
@@ -158,7 +159,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
@@ -179,10 +180,10 @@ int main(int argc, char ** argv) {
 
                 llama_batch_clear(batch);
 
-                for (int i = 0; i < pp; ++i) {
-                    for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                        llama_batch_add(batch, 0, i, { j }, 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;
 
@@ -197,7 +198,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_cache_seq_cp(ctx, 0, i, 0, pp);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

@@ -38,7 +38,7 @@ let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_par
 
 var context_params = llama_context_default_params()
 context_params.seed = 1234
-context_params.n_ctx = n_kv_req
+context_params.kv_size = n_kv_req
 context_params.n_batch = UInt32(max(n_len, n_parallel))
 context_params.n_threads = 8
 context_params.n_threads_batch = 8
@@ -53,12 +53,12 @@ defer {
     llama_free(context)
 }
 
-let n_ctx = llama_n_ctx(context)
+let kv_size = llama_kv_size(context)
 
-print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
+print("\nn_len = \(n_len), kv_size = \(kv_size), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
 
-if n_kv_req > n_ctx {
-    print("error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", n_kv_req)
+if n_kv_req > kv_size {
+    print("error: n_kv_req (%d) > kv_size, the required KV cache size is not big enough\n", n_kv_req)
     exit(1)
 }
 

examples/batched/README.md

@@ -7,7 +7,7 @@ The example demonstrates batched generation from a given prompt
 
 ...
 
-main: n_len = 32, n_ctx = 2048, n_parallel = 4, n_kv_req = 113
+main: n_len = 32, kv_size = 2048, n_parallel = 4, n_kv_req = 113
 
  Hello my name is
 

examples/batched/batched.cpp

@@ -78,9 +78,8 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed  = 1234;
-    ctx_params.n_ctx = n_kv_req;
+    ctx_params.kv_size = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_seq_max       = n_parallel;
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
@@ -91,14 +90,14 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int n_ctx    = llama_n_ctx(ctx);
+    const int kv_size    = llama_kv_size(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, kv_size = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, kv_size, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
-    if (n_kv_req > n_ctx) {
-        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
-        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+    if (n_kv_req > kv_size) {
+        LOG_TEE("%s: error: n_kv_req (%d) > kv_size, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_TEE("%s:        either reduce n_parallel or increase kv_size\n", __func__);
         return 1;
     }
 
@@ -133,7 +132,7 @@ int main(int argc, char ** argv) {
     // assign the system KV cache to all parallel sequences
     // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
     for (int32_t i = 1; i < n_parallel; ++i) {
-        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+        llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
     }
 
     if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -139,8 +139,8 @@ int main(int argc, char ** argv)
 
     std::vector<llama_token> tokens_list = llama_tokenize(ctx, params.prompt, true);
 
-    const size_t max_context_size     = llama_n_ctx( ctx );
-    const size_t max_tokens_list_size = max_context_size - 4 ;
+    const size_t max_kv_size          = llama_kv_size(ctx);
+    const size_t max_tokens_list_size = max_kv_size - 4 ;
 
     if (tokens_list.size() > max_tokens_list_size)
     {

examples/benchmark/benchmark-matmult.cpp

@@ -128,20 +128,20 @@ int main(int argc, char ** argv)  {
     // TODO: perform the bench for all types or for a user specified type
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
-    size_t ctx_size = 0;
-    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;
+    size_t kv_size = 0;
+    kv_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
+    kv_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
+    kv_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
+    kv_size += ggml_row_size(qtype, sizex*sizey);
+    kv_size += ggml_row_size(qtype, sizex*sizey);
+    kv_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
+    kv_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
+    kv_size += 1024*1024*16;
 
-    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
+    printf("Allocating Memory of size %zi bytes, %zi MB\n", kv_size, (kv_size/1024/1024));
 
     struct ggml_init_params params = {
-        /*.mem_size   =*/ ctx_size,
+        /*.mem_size   =*/ kv_size,
         /*.mem_buffer =*/ NULL,
         /* no_alloc   =*/ 0
     };
@@ -189,10 +189,12 @@ int main(int argc, char ** argv)  {
 
     int32_t nelements = sizex*sizey;
 
+    std::vector<int64_t> hist_cur(1 << 4, 0);
+
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
     struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -205,7 +207,7 @@ int main(int argc, char ** argv)  {
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
     struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

examples/chat-13B.bat

@@ -15,7 +15,7 @@ rem Adjust to the number of CPU cores you want to use.
 rem if not defined N_THREAD set "N_THREAD=8"
 rem Number of tokens to predict (made it larger than default because we want a long interaction)
 if not defined N_PREDICTS set "N_PREDICTS=2048"
-if not defined GEN_OPTIONS set "GEN_OPTIONS=--ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
+if not defined GEN_OPTIONS set "GEN_OPTIONS=--kv_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
 
 rem Default main script paths
 set "DEFAULT_MAIN_SCRIPT_PATHS=main.exe build\bin\main.exe"

examples/chat-13B.sh

@@ -15,8 +15,8 @@ N_THREAD="${N_THREAD:-8}"
 N_PREDICTS="${N_PREDICTS:-2048}"
 
 # Note: you can also override the generation options by specifying them on the command line:
-# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
-GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
+# For example, override the context size by doing: ./chatLLaMa --kv_size 1024
+GEN_OPTIONS="${GEN_OPTIONS:---kv_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
 
 DATE_TIME=$(date +%H:%M)
 DATE_YEAR=$(date +%Y)

examples/chat-persistent.sh

@@ -27,9 +27,9 @@ SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+
 SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
 SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
 
-CTX_SIZE=2048
-CTX_ROTATE_POINT=$((CTX_SIZE * 3 / 5)) # REVIEW
-OPTS=(--model "$MODEL" --ctx_size "$CTX_SIZE" --repeat_last_n 256 "$@")
+KV_SIZE=2048
+KV_ROTATE_POINT=$((KV_SIZE * 3 / 5)) # REVIEW
+OPTS=(--model "$MODEL" --kv_size "$KV_SIZE" --repeat_last_n 256 "$@")
 
 # An unbuffered `tail -c+N`
 skip_bytes() {
@@ -84,7 +84,7 @@ n_tokens=0
 
 while read -e line; do
     # Limit generation to remaining context, with a buffer and estimating 2 chars/token for input
-    n_predict=$((CTX_SIZE - n_tokens - ${#line} / 2 - 32))
+    n_predict=$((KV_SIZE - n_tokens - ${#line} / 2 - 32))
 
     # Swap prompts when we're about to run out of context
     if ((n_predict <= 0)); then
@@ -97,11 +97,11 @@ while read -e line; do
         cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
 
         n_tokens=0
-        n_predict=$((CTX_SIZE / 2))
+        n_predict=$((KV_SIZE / 2))
     fi
 
     echo " ${line}" >>"$CUR_PROMPT_FILE"
-    if ((n_tokens > CTX_ROTATE_POINT)); then
+    if ((n_tokens > KV_ROTATE_POINT)); then
         echo " ${line}" >>"$NEXT_PROMPT_FILE"
     fi
 
@@ -139,7 +139,7 @@ while read -e line; do
 
     n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
 
-    if ((n_tokens > CTX_ROTATE_POINT)); then
+    if ((n_tokens > KV_ROTATE_POINT)); then
         tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"
     fi
 

examples/chat-vicuna.sh

@@ -15,8 +15,8 @@ N_THREAD="${N_THREAD:-8}"
 N_PREDICTS="${N_PREDICTS:-2048}"
 
 # Note: you can also override the generation options by specifying them on the command line:
-# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
-GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
+# For example, override the context size by doing: ./chatLLaMa --kv_size 1024
+GEN_OPTIONS="${GEN_OPTIONS:---kv_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
 
 DATE_TIME=$(date +%H:%M)
 DATE_YEAR=$(date +%Y)

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

@@ -226,7 +226,7 @@ struct llama_vocab {
 
 struct my_llama_hparams {
     uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
+    uint32_t kv_size   = 512;   // this is provided as user input?
     uint32_t n_embd  = 4096;
     uint32_t n_ff    = 11008;
     uint32_t n_mult  = 4;
@@ -326,7 +326,7 @@ struct train_params {
 
 static void print_params(struct my_llama_hparams * params) {
     printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
+    printf("%s: kv_size: %u\n", __func__, params->kv_size);
     printf("%s: n_embd:  %u\n", __func__, params->n_embd);
     printf("%s: n_mult:  %u\n", __func__, params->n_mult);
     printf("%s: n_head:  %u\n", __func__, params->n_head);
@@ -732,7 +732,7 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
     gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
 
-    gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
+    gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.kv_size);
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
     gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
     gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
@@ -937,7 +937,7 @@ int main(int argc, char ** argv) {
 
     struct my_llama_model model;
     model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
+    model.hparams.kv_size = params.n_ctx;
     model.hparams.n_embd  = config.dim; //params.n_embd;
     model.hparams.n_ff    = config.hidden_dim;
     model.hparams.n_mult  = 32;//params.n_mult;

examples/embedding/embedding.cpp

@@ -19,7 +19,18 @@ static std::vector<std::string> split_lines(const std::string & s) {
 
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
     for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
+        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+    }
+}
+
+static void normalize(float * vec, float * out, int n) {
+    float norm = 0;
+    for (int i = 0; i < n; i++) {
+        norm += vec[i] * vec[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n; i++) {
+        out[i] = vec[i] / norm;
     }
 }
 
@@ -33,23 +44,11 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
-    for (int i = 0; i < batch.n_tokens; i++) {
-        if (!batch.logits[i]) {
-            continue;
-        }
-
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        if (embd == NULL) {
-            embd = llama_get_embeddings_ith(ctx, i);
-            if (embd == NULL) {
-                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
-                continue;
-            }
-        }
-
-        float * out = output + batch.seq_id[i][0] * n_embd;
-        llama_embd_normalize(embd, out, n_embd);
+    // normalize on copy
+    for (int k = 0; k < n_seq; k++) {
+        float * emb = llama_get_embeddings_ith(ctx, k);
+        float * out = output + k * n_embd;
+        normalize(emb, out, n_embd);
     }
 }
 
@@ -89,11 +88,11 @@ int main(int argc, char ** argv) {
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
-    const int n_ctx = llama_n_ctx(ctx);
+    const int kv_size = llama_kv_size(ctx);
 
-    if (n_ctx > n_ctx_train) {
+    if (kv_size > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, n_ctx);
+                __func__, n_ctx_train, kv_size);
     }
 
     // print system information
@@ -107,7 +106,7 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch >= params.n_ctx);
+    GGML_ASSERT(params.n_batch == params.kv_size);
 
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
@@ -133,7 +132,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
 
     // allocate output
     const int n_embd = llama_n_embd(model);
@@ -146,7 +145,6 @@ int main(int argc, char ** argv) {
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
         auto & inp = inputs[k];
-
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity
@@ -167,26 +165,15 @@ int main(int argc, char ** argv) {
     float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd);
 
-    // print the first part of the embeddings
-    fprintf(stdout, "\n");
-    for (int j = 0; j < n_prompts; j++) {
-        fprintf(stdout, "embedding %d: ", j);
-        for (int i = 0; i < std::min(16, n_embd); i++) {
-            fprintf(stdout, "%f ", emb[j * n_embd + i]);
+    // print first 3 embeddings
+    for (int j = 0; j < std::min(3, n_prompts); j++) {
+        fprintf(stderr, "embedding %d: ", j);
+        for (int i = 0; i < n_embd; i++) {
+            fprintf(stderr, "%f ", emb[j * n_embd + i]);
         }
-        fprintf(stdout, "\n");
-    }
-
-    // print cosine similarity matrix
-    fprintf(stdout, "\n");
-    printf("cosine similarity matrix:\n\n");
-    for (int i = 0; i < n_prompts; i++) {
-        for (int j = 0; j < n_prompts; j++) {
-            float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-            fprintf(stdout, "%6.2f ", sim);
-        }
-        fprintf(stdout, "\n");
+        fprintf(stderr, "\n\n");
     }
+    fprintf(stderr, "\n");
 
     // clean up
     llama_print_timings(ctx);

examples/finetune/finetune.cpp

@@ -16,7 +16,7 @@
 
 struct my_llama_hparams {
     uint32_t n_vocab    = 32000;
-    uint32_t n_ctx      = 512;
+    uint32_t kv_size    = 512;
     uint32_t n_embd     = 4096;
     uint32_t n_ff       = 11008;
     uint32_t n_head     = 32;
@@ -190,7 +190,7 @@ static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
 
 static void print_params(struct my_llama_hparams * params) {
     printf("%s: n_vocab               : %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx                 : %u\n", __func__, params->n_ctx);
+    printf("%s: kv_size               : %u\n", __func__, params->kv_size);
     printf("%s: n_embd                : %u\n", __func__, params->n_embd);
     printf("%s: n_ff                  : %u\n", __func__, params->n_ff);
     printf("%s: n_head                : %u\n", __func__, params->n_head);
@@ -250,7 +250,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
     };
 
     GGUF_GET_KEY(ctx, hparams->n_embd,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_EMBEDDING_LENGTH));
-    GGUF_GET_KEY(ctx, hparams->n_ctx,          gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->kv_size,        gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH));
     GGUF_GET_KEY(ctx, hparams->n_ff,           gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_FEED_FORWARD_LENGTH));
     GGUF_GET_KEY(ctx, hparams->n_head,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
     GGUF_GET_KEY(ctx, hparams->n_layer,        gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_BLOCK_COUNT));
@@ -268,7 +268,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
     }
 }
 
-static void init_model(struct llama_model * input, struct my_llama_model * model, const char * fn_model, uint32_t n_ctx) {
+static void init_model(struct llama_model * input, struct my_llama_model * model, const char * fn_model, uint32_t kv_size) {
     auto & hparams = model->hparams;
 
     std::vector<char> tn_buf;
@@ -298,7 +298,7 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         gguf_free(mctx);
     }
     hparams.n_vocab = llama_n_vocab(input);
-    hparams.n_ctx = n_ctx;
+    hparams.kv_size = kv_size;
 
     // get tensors from llama_model (possibly mmapped)
     model->tok_embeddings = llama_get_model_tensor(input, tn(LLM_TENSOR_TOKEN_EMBD));
@@ -529,7 +529,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     const int n_past = 0;
     const int N = n_tokens;
     const auto & hparams  = model->hparams;
-    const int n_ctx       = hparams.n_ctx;
+    const int kv_size     = hparams.kv_size;
     const int n_vocab     = hparams.n_vocab;
     const int n_embd      = hparams.n_embd;
     const int n_layer     = hparams.n_layer;
@@ -558,13 +558,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     ggml_set_input(KQ_pos);
 
     // rope has so much parameters that we make a custom function for it
-    auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
+    auto rope = [ctx, KQ_pos, n_rot, kv_size, rope_freq_base, rope_freq_scale]
                 (struct ggml_tensor * t) -> struct ggml_tensor * {
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
         return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+            t, KQ_pos, n_rot, rope_mode, kv_size, 0,
             rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
         );
     };
@@ -848,7 +848,7 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch);
     gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype);
 
-    gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH),              model->hparams.n_ctx);
+    gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH),              model->hparams.kv_size);
     gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH),            model->hparams.n_embd);
     gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH),         model->hparams.n_ff);
     gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT),        model->hparams.n_head);
@@ -1531,7 +1531,7 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_output         = n_rank_output;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    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;
@@ -1554,9 +1554,9 @@ int main(int argc, char ** argv) {
     bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
 
     if (existed) {
-        // overwrite last n_ctx with user provided n_ctx
+        // overwrite last kv_size with user provided kv_size
         if (params.common.custom_n_ctx) {
-            model.hparams.n_ctx = params.common.n_ctx;
+            model.hparams.kv_size = params.common.n_ctx;
         }
 
         const bool opt_param_count_changed = (
@@ -1625,7 +1625,7 @@ int main(int argc, char ** argv) {
     printf("%s: opt_size  = %zu bytes (%.1f MB)\n", __func__, ggml_get_mem_size(opt->ctx), (float) ggml_get_mem_size(opt->ctx) / (1024.0f*1024.0f));
     printf("%s: opt iter %d\n", __func__, opt->iter);
 
-    int n_tokens = model.hparams.n_ctx;
+    int n_tokens = model.hparams.kv_size;
     int n_vocab  = model.hparams.n_vocab;
     int n_batch  = params.common.n_batch;
 

examples/gpt4all.sh

@@ -10,6 +10,6 @@ cd ..
 ./main --color --instruct --threads 4 \
        --model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \
        --file ./prompts/alpaca.txt \
-       --batch_size 8 --ctx_size 2048 -n -1 \
+       --batch_size 8 --kv_size 2048 -n -1 \
        --repeat_last_n 64 --repeat_penalty 1.3 \
        --n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95

examples/gritlm/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET gritlm)
-add_executable(${TARGET} gritlm.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gritlm/gritlm.cpp

@@ -1,215 +0,0 @@
-#include "common.h"
-#include "llama.h"
-
-#include <string>
-#include <vector>
-
-// #define GRIT_DEBUG
-
-static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
-    std::vector<std::vector<float>> result;
-
-    const llama_model * mdl = llama_get_model(ctx);
-
-    llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    for (uint64_t i = 0; i < sentences.size(); i++) {
-        llama_batch_clear(batch);
-
-        const std::string input_string = instruction + sentences[i];
-
-        std::vector<llama_token> inputs = llama_tokenize(mdl, input_string, true, false);
-
-        const int32_t n_toks = inputs.size();
-
-        // GritLM seems to have EOS = ""
-        // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_token_eos(mdl));
-
-        // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
-
-#ifdef GRIT_DEBUG
-        // debug tokens - should be matching as referenced in the GritLM sample
-        std::for_each(inputs.begin(), inputs.end(), [&ctx](llama_token t) {
-            std::printf("[%u:%s]", t, llama_token_to_piece(ctx, t).c_str());
-        });
-        std::printf("\n");
-#endif
-
-        // add input to batch (this increments n_tokens)
-        for (int32_t j = 0; j < n_toks; j++) {
-            llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
-        }
-
-        // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
-        llama_set_causal_attn(ctx, false);
-
-        // run model
-        llama_decode(ctx, batch);
-
-        // get embedding dimensions
-        uint64_t n_embd = llama_n_embd(mdl);
-
-        // allocate embedding output
-        std::vector<float> emb_unorm(n_embd, 0.0f);
-
-        // sum up all token embeddings
-        for (int32_t k = n_inst; k < n_toks; k++) {
-            float * emb = llama_get_embeddings_ith(ctx, k);
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] += emb[j];
-            }
-        }
-
-        // divide by number of tokens (mean pooling)
-        {
-            const uint64_t n_sent = n_toks - n_inst;
-
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] /= n_sent;
-            }
-        }
-
-        std::vector<float> emb_norm(emb_unorm.size());
-        llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
-        result.push_back(emb_norm);
-
-#ifdef GRIT_DEBUG
-        // print out emb_norm
-        std::printf("embedding %ld: ", i);
-        for (uint64_t j = 0; j < n_embd; j++) {
-            std::printf("%.5f ", emb_norm[j]);
-        }
-        std::printf("\n\n");
-#endif
-    }
-
-    llama_batch_free(batch);
-
-    return result;
-}
-
-static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
-    std::string result;
-
-    const llama_model * mdl = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(mdl);
-
-    llama_kv_cache_clear(ctx);
-    llama_set_causal_attn(ctx, true);
-    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
-    int32_t i_current_token = 0;
-
-    while (true) {
-        llama_batch_clear(bat);
-        auto n_inputs = (int32_t)inputs.size();
-        for (int32_t i = 0; i < n_inputs; i++) {
-            llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
-        }
-        inputs.clear();
-
-        llama_decode(ctx, bat);
-        auto logits = llama_get_logits_ith(ctx, bat.n_tokens - 1);
-
-        auto candidates = std::vector<llama_token_data>(llama_n_vocab(mdl));
-        auto n_candidates = (int32_t)candidates.size();
-        for (int32_t token = 0; token < n_candidates; token++) {
-            candidates[token] = llama_token_data{ token, logits[token], 0.0f };
-        }
-        auto candidates_p = llama_token_data_array{ candidates.data(), candidates.size(), false };
-
-        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
-        if (token == eos_token) {
-            break;
-        }
-
-        std::string piece = llama_token_to_piece(ctx, token);
-        if (stream) {
-            std::printf("%s", piece.c_str());
-            std::fflush(stdout);
-        }
-
-        inputs.push_back(token);
-
-        result += piece;
-    }
-
-    if (stream) {
-        std::printf("\n");
-    }
-
-    llama_batch_free(bat);
-
-    return result;
-}
-
-static std::string gritlm_instruction(const std::string & instruction) {
-    return !instruction.empty() ? "<|user|>\n" + instruction + "\n<|embed|>\n" : "<|embed|>\n";
-}
-
-int main(int argc, char * argv[]) {
-    gpt_params params;
-    if (!gpt_params_parse(argc, argv, params)) {
-        return 1;
-    }
-
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
-    llama_backend_init();
-
-    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
-
-    // create new context - set to embedding mode
-    cparams.embeddings = true;
-    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
-
-    // ### Embedding/Representation ###
-    // samples taken from: https://github.com/ContextualAI/gritlm#basic
-    {
-        const std::string instruction = "Given a scientific paper title, retrieve the paper's abstract";
-
-        const std::vector<std::string> queries = {
-            "Bitcoin: A Peer-to-Peer Electronic Cash System",
-            "Generative Representational Instruction Tuning",
-        };
-
-        const std::vector<std::string> documents = {
-            "A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
-            "All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
-        };
-
-        // No need to add instruction for retrieval documents
-        const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
-        const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
-
-        const int n_embd = llama_n_embd(mdl);
-
-        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
-        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
-
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[0].c_str(), cosine_sim_q1_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
-    }
-
-    // ### Generation ###
-    // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
-    {
-        const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
-        std::string response = generate(ctx, prompt, true);
-    }
-
-    llama_free(ctx);
-    llama_free_model(mdl);
-    llama_backend_free();
-
-    return 0;
-}

examples/imatrix/imatrix.cpp

@@ -325,7 +325,7 @@ static void process_logits(
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl, int from_chunk) {
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    const int n_ctx = llama_n_ctx(ctx);
+    const int kv_size  = llama_kv_size(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
@@ -336,17 +336,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
     if (from_chunk > 0) {
-        if (size_t((from_chunk + 2)*n_ctx) >= tokens.size()) {
+        if (size_t((from_chunk + 2)*kv_size) >= tokens.size()) {
             fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, from_chunk);
             return false;
         }
-        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, from_chunk, from_chunk*n_ctx);
-        tokens.erase(tokens.begin(), tokens.begin() + from_chunk*n_ctx);
+        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, from_chunk, from_chunk * kv_size);
+        tokens.erase(tokens.begin(), tokens.begin() + from_chunk * kv_size);
     }
 
-    if (int(tokens.size()) < 2*n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
-                n_ctx);
+    if (int(tokens.size()) < 2*kv_size) {
+        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2 * kv_size,
+                kv_size);
         fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
         return false;
     }
@@ -359,7 +359,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
         prob_history.resize(tokens.size());
     }
 
-    const int n_chunk_max = tokens.size() / n_ctx;
+    const int n_chunk_max = tokens.size() / kv_size;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
     const int n_vocab = llama_n_vocab(llama_get_model(ctx));
@@ -373,16 +373,16 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
-    const int num_batches = (n_ctx + n_batch - 1) / n_batch;
+    const int num_batches = (kv_size + n_batch - 1) / n_batch;
 
     std::vector<float> logits;
     if (compute_ppl && num_batches > 1) {
-        logits.reserve((size_t)n_ctx * n_vocab);
+        logits.reserve((size_t)kv_size * n_vocab);
     }
 
     for (int i = 0; i < n_chunk; ++i) {
-        const int start =     i * n_ctx;
-        const int end   = start + n_ctx;
+        const int start = i * kv_size;
+        const int end   = start + kv_size;
 
         std::vector<float> logits;
 
@@ -431,11 +431,11 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
         }
 
         if (compute_ppl) {
-            const int first = n_ctx/2;
+            const int first = kv_size / 2;
             const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
-            process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+            process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, kv_size - 1 - first,
                     workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
-            count += n_ctx - first - 1;
+            count += kv_size - first - 1;
 
             printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
             fflush(stdout);
@@ -553,7 +553,7 @@ int main(int argc, char ** argv) {
     }
 
     params.logits_all = true;
-    params.n_batch = std::min(params.n_batch, params.n_ctx);
+    params.n_batch = std::min(params.n_batch, params.kv_size);
 
     print_build_info();
 
@@ -593,9 +593,9 @@ int main(int argc, char ** argv) {
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
-    if (params.n_ctx > n_ctx_train) {
+    if (params.kv_size > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, params.n_ctx);
+                __func__, n_ctx_train, params.kv_size);
     }
 
     // print system information

examples/infill/README.md

@@ -14,7 +14,8 @@ In this section, we cover the most commonly used options for running the `infill
 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
--   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-c N`, `--ctx-size N`: Deprecated, use `--kv-size` instead.
+-   `-kv N`, `--kv-size N`: Specify the total size of the KV cache for the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 
 ## Input Prompts
 

examples/infill/infill.cpp

@@ -135,9 +135,9 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    if (params.n_ctx != 0 && params.n_ctx < 8) {
+    if (params.kv_size != 0 && params.kv_size < 8) {
         LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
-        params.n_ctx = 8;
+        params.kv_size = 8;
     }
     if (params.instruct) {
         printf("\n************\n");
@@ -225,12 +225,12 @@ int main(int argc, char ** argv) {
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
-    const int n_ctx = llama_n_ctx(ctx);
-    LOG("n_ctx: %d\n", n_ctx);
+    const int kv_size     = llama_kv_size(ctx);
+    LOG("kv_size: %d\n", kv_size);
 
-    if (n_ctx > n_ctx_train) {
+    if (kv_size > n_ctx_train) {
         LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, n_ctx);
+                __func__, n_ctx_train, kv_size);
     }
 
     // print system information
@@ -291,8 +291,8 @@ int main(int argc, char ** argv) {
         LOG("guidance_offset:     %s", log_tostr(guidance_offset));
     }
 
-    if ((int) embd_inp.size() > n_ctx - 4) {
-        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+    if ((int) embd_inp.size() > kv_size - 4) {
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), kv_size - 4);
         return 1;
     }
 
@@ -366,7 +366,7 @@ int main(int argc, char ** argv) {
         }
     }
     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("generate: kv_size = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", kv_size, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
     LOG_TEE("\n#####  Infill mode  #####\n\n");
@@ -378,10 +378,10 @@ int main(int argc, char ** argv) {
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
-            control_message = " - To return control to LLaMA, end your input with '\\'.\n"
+            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
-            control_message = " - Press Return to return control to LLaMA.\n"
+            control_message = " - Press Return to return control to LLaMa.\n"
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
@@ -416,9 +416,9 @@ int main(int argc, char ** argv) {
     while (n_remain != 0 || params.interactive) {
         // predict
         if (!embd.empty()) {
-            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // Note: kv_size - 4 here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
-            int max_embd_size = n_ctx - 4;
+            int max_embd_size = kv_size - 4;
 
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
             if ((int) embd.size() > max_embd_size) {
@@ -434,8 +434,8 @@ int main(int argc, char ** argv) {
             // infinite text generation via context swapping
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)
-            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+            // - take half of the last (kv_size - n_keep) tokens and recompute the logits in batches
+            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > kv_size) {
                 if (params.n_predict == -2) {
                     LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
@@ -444,11 +444,11 @@ int main(int argc, char ** argv) {
                 const int n_left    = n_past - params.n_keep - 1;
                 const int n_discard = n_left/2;
 
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+                LOG("context full, swapping: n_past = %d, n_left = %d, kv_size = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, kv_size, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json-schema-to-grammar.py

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

examples/llama-bench/README.md

@@ -35,6 +35,7 @@ options:
   -mg, --main-gpu <i>                 (default: 0)
   -nkvo, --no-kv-offload <0|1>        (default: 0)
   -mmp, --mmap <0|1>                  (default: 1)
+  -mmq, --mul-mat-q <0|1>             (default: 1)
   -ts, --tensor_split <ts0/ts1/..>    (default: 0)
   -r, --repetitions <n>               (default: 5)
   -o, --output <csv|json|md|sql>      (default: md)

examples/llama-bench/llama-bench.cpp

@@ -123,15 +123,20 @@ static std::string get_gpu_info() {
     }
 #endif
 #ifdef GGML_USE_SYCL
-    int count = ggml_backend_sycl_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_sycl_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
+    int device_list[GGML_SYCL_MAX_DEVICES];
+    ggml_sycl_get_gpu_list(device_list, GGML_SYCL_MAX_DEVICES);
+
+    for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
+        if (device_list[i] >0 ){
+            char buf[128];
+            ggml_sycl_get_device_description(i, buf, sizeof(buf));
+            id += buf;
             id += "/";
         }
     }
+    if (id.length() >2 ) {
+        id.pop_back();
+    }
 #endif
     // TODO: other backends
     return id;
@@ -152,9 +157,9 @@ static const char * output_format_str(output_formats format) {
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_MODE_NONE:  return "none";
-        case LLAMA_SPLIT_MODE_LAYER: return "layer";
-        case LLAMA_SPLIT_MODE_ROW:   return "row";
+        case LLAMA_SPLIT_NONE:  return "none";
+        case LLAMA_SPLIT_LAYER: return "layer";
+        case LLAMA_SPLIT_ROW:   return "row";
         default: GGML_ASSERT(!"invalid split mode");
     }
 }
@@ -164,7 +169,6 @@ struct cmd_params {
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
-    std::vector<int> n_ubatch;
     std::vector<ggml_type> type_k;
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
@@ -172,9 +176,9 @@ struct cmd_params {
     std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
     std::vector<bool> no_kv_offload;
+    std::vector<bool> mul_mat_q;
     std::vector<std::vector<float>> tensor_split;
     std::vector<bool> use_mmap;
-    std::vector<bool> embeddings;
     int reps;
     bool verbose;
     output_formats output_format;
@@ -184,18 +188,17 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_batch       */ {2048},
-    /* n_ubatch      */ {512},
+    /* n_batch       */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
-    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
+    /* mul_mat_q     */ {true},
     /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
     /* use_mmap      */ {true},
-    /* embeddings    */ {false},
     /* reps          */ 5,
     /* verbose       */ false,
     /* output_format */ MARKDOWN
@@ -210,7 +213,6 @@ 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("  -ub N, --ubatch-size <n>            (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").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());
@@ -219,8 +221,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
-    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
-    printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
     printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
     printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
@@ -300,13 +302,6 @@ 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 == "-ub" || arg == "--ubatch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<int>(argv[i], split_delim);
-            params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
         } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -363,11 +358,11 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             for (const auto & m : p) {
                 llama_split_mode mode;
                 if (m == "none") {
-                    mode = LLAMA_SPLIT_MODE_NONE;
+                    mode = LLAMA_SPLIT_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_MODE_LAYER;
+                    mode = LLAMA_SPLIT_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_MODE_ROW;
+                    mode = LLAMA_SPLIT_ROW;
                 } else {
                     invalid_param = true;
                     break;
@@ -388,6 +383,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
+        } else if (arg == "-mmq" || arg == "--mul-mat-q") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.mul_mat_q.insert(params.mul_mat_q.end(), p.begin(), p.end());
         } else if (arg == "-mmp" || arg == "--mmap") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -395,13 +397,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
-        } else if (arg == "-embd" || arg == "--embeddings") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -465,16 +460,15 @@ 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.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
     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.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
+    if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
-    if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
 
     return params;
@@ -485,7 +479,6 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
-    int n_ubatch;
     ggml_type type_k;
     ggml_type type_v;
     int n_threads;
@@ -493,9 +486,9 @@ struct cmd_params_instance {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
-    bool embeddings;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -521,13 +514,12 @@ struct cmd_params_instance {
     llama_context_params to_llama_cparams() const {
         llama_context_params cparams = llama_context_default_params();
 
-        cparams.n_ctx = n_prompt + n_gen;
+        cparams.kv_size = n_prompt + n_gen;
         cparams.n_batch = n_batch;
-        cparams.n_ubatch = n_ubatch;
         cparams.type_k = type_k;
         cparams.type_v = type_v;
+        cparams.mul_mat_q = mul_mat_q;
         cparams.offload_kqv = !no_kv_offload;
-        cparams.embeddings = embeddings;
 
         return cparams;
     }
@@ -543,11 +535,10 @@ 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 & mmp : params.use_mmap)
-    for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
-    for (const auto & nub : params.n_ubatch)
     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 & nkvo : params.no_kv_offload)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -559,7 +550,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
-                /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
@@ -567,9 +557,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
-                /* .embeddings   = */ embd,
             };
             instances.push_back(instance);
         }
@@ -583,7 +573,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
-                /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
@@ -591,9 +580,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
-                /* .embeddings   = */ embd,
             };
             instances.push_back(instance);
         }
@@ -620,7 +609,6 @@ struct test {
     uint64_t model_size;
     uint64_t model_n_params;
     int n_batch;
-    int n_ubatch;
     int n_threads;
     ggml_type type_k;
     ggml_type type_v;
@@ -628,9 +616,9 @@ struct test {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
-    bool embeddings;
     int n_prompt;
     int n_gen;
     std::string test_time;
@@ -644,7 +632,6 @@ struct test {
         model_size = llama_model_size(lmodel);
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
-        n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
         type_k = inst.type_k;
         type_v = inst.type_v;
@@ -652,9 +639,9 @@ struct test {
         split_mode = inst.split_mode;
         main_gpu = inst.main_gpu;
         no_kv_offload = inst.no_kv_offload;
+        mul_mat_q = inst.mul_mat_q;
         tensor_split = inst.tensor_split;
         use_mmap = inst.use_mmap;
-        embeddings = inst.embeddings;
         n_prompt = inst.n_prompt;
         n_gen = inst.n_gen;
         // RFC 3339 date-time format
@@ -723,11 +710,10 @@ struct test {
             "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_ubatch",
-            "n_threads", "type_k", "type_v",
+            "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload",
-            "tensor_split", "use_mmap", "embeddings",
+            "mul_mat_q", "tensor_split", "use_mmap",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -738,8 +724,7 @@ struct test {
     enum field_type {STRING, BOOL, INT, FLOAT};
 
     static field_type get_field_type(const std::string & field) {
-        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
-            field == "n_threads" ||
+        if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -748,7 +733,7 @@ struct test {
         }
         if (field == "cuda" || field == "opencl"  || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
-            field == "use_mmap" || field == "embeddings") {
+            field == "mul_mat_q" || field == "use_mmap") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -779,11 +764,10 @@ struct test {
             std::to_string(metal), std::to_string(sycl), 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_ubatch),
-            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            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), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload),
-            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
+            std::to_string(mul_mat_q), tensor_split_str, std::to_string(use_mmap),
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -947,15 +931,15 @@ struct markdown_printer : public printer {
         if (field == "n_threads") {
             return "threads";
         }
+        if (field == "mul_mat_q") {
+            return "mmq";
+        }
         if (field == "no_kv_offload") {
             return "nkvo";
         }
         if (field == "use_mmap") {
             return "mmap";
         }
-        if (field == "embeddings") {
-            return "embd";
-        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -978,9 +962,6 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
-        if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) {
-            fields.emplace_back("n_ubatch");
-        }
         if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
             fields.emplace_back("type_k");
         }
@@ -993,6 +974,9 @@ struct markdown_printer : public printer {
         if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
             fields.emplace_back("split_mode");
         }
+        if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
+            fields.emplace_back("mul_mat_q");
+        }
         if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
             fields.emplace_back("no_kv_offload");
         }
@@ -1002,9 +986,6 @@ struct markdown_printer : public printer {
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }
-        if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
-            fields.emplace_back("embeddings");
-        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");
 
@@ -1120,32 +1101,25 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    llama_set_n_threads(ctx, n_threads, n_threads);
-
-    //std::vector<llama_token> tokens(n_prompt, llama_token_bos(llama_get_model(ctx)));
-    //llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt, n_past, 0));
-    //GGML_UNUSED(n_batch);
-
     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);
+
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
         llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
         n_processed += n_tokens;
     }
-
-    llama_synchronize(ctx);
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_set_n_threads(ctx, n_threads, n_threads);
-
     llama_token token = llama_token_bos(llama_get_model(ctx));
 
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     for (int i = 0; i < n_gen; i++) {
         llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
-        llama_synchronize(ctx);
     }
 }
 
@@ -1234,8 +1208,7 @@ int main(int argc, char ** argv) {
 
         // warmup run
         if (t.n_prompt > 0) {
-            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             test_gen(ctx, 1, 0, t.n_threads);
@@ -1251,7 +1224,6 @@ int main(int argc, char ** argv) {
             if (t.n_gen > 0) {
                 test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
             }
-
             uint64_t t_ns = get_time_ns() - t_start;
             t.samples_ns.push_back(t_ns);
         }

examples/llama.android/app/build.gradle.kts

@@ -21,8 +21,12 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
+            // Workaround for https://github.com/llvm/llvm-project/issues/65820
+            // affecting armeabi-v7a. Skip armeabi-v7a when invoked with
+            // -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
+            if (project.hasProperty("skip-armeabi-v7a")) {
+                abiFilters += listOf("arm64-v8a", "x86_64", "x86")
+            }
         }
         externalNativeBuild {
             cmake {

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

@@ -33,45 +33,6 @@ jclass la_int_var;
 jmethodID la_int_var_value;
 jmethodID la_int_var_inc;
 
-std::string cached_token_chars;
-
-bool is_valid_utf8(const char * string) {
-    if (!string) {
-        return true;
-    }
-
-    const unsigned char * bytes = (const unsigned char *)string;
-    int num;
-
-    while (*bytes != 0x00) {
-        if ((*bytes & 0x80) == 0x00) {
-            // U+0000 to U+007F
-            num = 1;
-        } else if ((*bytes & 0xE0) == 0xC0) {
-            // U+0080 to U+07FF
-            num = 2;
-        } else if ((*bytes & 0xF0) == 0xE0) {
-            // U+0800 to U+FFFF
-            num = 3;
-        } else if ((*bytes & 0xF8) == 0xF0) {
-            // U+10000 to U+10FFFF
-            num = 4;
-        } else {
-            return false;
-        }
-
-        bytes += 1;
-        for (int i = 1; i < num; ++i) {
-            if ((*bytes & 0xC0) != 0x80) {
-                return false;
-            }
-            bytes += 1;
-        }
-    }
-
-    return true;
-}
-
 static void log_callback(ggml_log_level level, const char * fmt, void * data) {
     if (level == GGML_LOG_LEVEL_ERROR)     __android_log_print(ANDROID_LOG_ERROR, TAG, fmt, data);
     else if (level == GGML_LOG_LEVEL_INFO) __android_log_print(ANDROID_LOG_INFO, TAG, fmt, data);
@@ -334,8 +295,6 @@ Java_com_example_llama_Llm_completion_1init(
         jint n_len
     ) {
 
-    cached_token_chars.clear();
-
     const auto text = env->GetStringUTFChars(jtext, 0);
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
@@ -413,16 +372,8 @@ Java_com_example_llama_Llm_completion_1loop(
     }
 
     auto new_token_chars = llama_token_to_piece(context, new_token_id);
-    cached_token_chars += new_token_chars;
-
-    jstring new_token = nullptr;
-    if (is_valid_utf8(cached_token_chars.c_str())) {
-        new_token = env->NewStringUTF(cached_token_chars.c_str());
-        LOGi("cached: %s, new_token_chars: `%s`, id: %d", cached_token_chars.c_str(), new_token_chars.c_str(), new_token_id);
-        cached_token_chars.clear();
-    } else {
-        new_token = env->NewStringUTF("");
-    }
+    LOGi("new_token_chars: `%s`", new_token_chars.c_str());
+    auto new_token = env->NewStringUTF(new_token_chars.c_str());
 
     llama_batch_clear(*batch);
     llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);

examples/llama.android/app/src/main/java/com/example/llama/Llm.kt

@@ -71,7 +71,7 @@ class Llm {
         batch: Long,
         nLen: Int,
         ncur: IntVar
-    ): String?
+    ): String
 
     private external fun kv_cache_clear(context: Long)
 
@@ -115,7 +115,7 @@ class Llm {
                 val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
                 while (ncur.value <= nlen) {
                     val str = completion_loop(state.context, state.batch, nlen, ncur)
-                    if (str == null) {
+                    if (str.isEmpty()) {
                         break
                     }
                     emit(str)

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

@@ -68,8 +68,8 @@ actor LlamaContext {
         print("Using \(n_threads) threads")
 
         var ctx_params = llama_context_default_params()
-        ctx_params.seed  = 1234
-        ctx_params.n_ctx = 2048
+        ctx_params.seed    = 1234
+        ctx_params.kv_size = 2048
         ctx_params.n_threads       = UInt32(n_threads)
         ctx_params.n_threads_batch = UInt32(n_threads)
 
@@ -112,13 +112,13 @@ actor LlamaContext {
         tokens_list = tokenize(text: text, add_bos: true)
         temporary_invalid_cchars = []
 
-        let n_ctx = llama_n_ctx(context)
+        let kv_size = llama_kv_size(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)")
+        print("\n n_len = \(n_len), kv_size = \(kv_size), 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")
+        if n_kv_req > kv_size {
+            print("error: n_kv_req > kv_size, the required KV cache size is not big enough")
         }
 
         for id in tokens_list {
@@ -221,7 +221,6 @@ actor LlamaContext {
             if llama_decode(context, batch) != 0 {
                 print("llama_decode() failed during prompt")
             }
-            llama_synchronize(context)
 
             let t_pp_end = ggml_time_us()
 
@@ -241,7 +240,6 @@ actor LlamaContext {
                 if llama_decode(context, batch) != 0 {
                     print("llama_decode() failed during text generation")
                 }
-                llama_synchronize(context)
             }
 
             let t_tg_end = ggml_time_us()

examples/llama2-13b.sh

@@ -9,7 +9,7 @@ cd ..
 
 ./main -m models/available/Llama2/13B/llama-2-13b.ggmlv3.q4_0.bin \
        --color \
-       --ctx_size 2048 \
+       --kv_size 2048 \
        -n -1 \
        -ins -b 256 \
        --top_k 10000 \

examples/llama2.sh

@@ -9,7 +9,7 @@ cd ..
 
 ./main -m models/available/Llama2/7B/llama-2-7b.ggmlv3.q4_0.bin \
        --color \
-       --ctx_size 2048 \
+       --kv_size 2048 \
        -n -1 \
        -ins -b 256 \
        --top_k 10000 \

examples/llava/README.md

@@ -59,47 +59,14 @@ python ./convert.py ../llava-v1.5-7b --skip-unknown
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
 ## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
 
-2) Install the required Python packages:
-
-```sh
-pip install -r examples/llava/requirements.txt
-```
-
-3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
+1) Backup your pth/safetensor model files as llava-surgery modifies them
+2) Use `python llava-surgery-v2.py -C -m /path/to/hf-model` which also supports llava-1.5 variants pytorch as well as safetensor models:
 - you will find a llava.projector and a llava.clip file in your model directory
-
-4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-5) Create the visual gguf model:
-```console
-python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
+3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory (https://huggingface.co/cmp-nct/llava-1.6-gguf/blob/main/config_vit.json) and rename it to config.json.
+4) Create the visual gguf model: `python ./examples/llava/convert-image-encoder-to-gguf.py -m ../path/to/vit --llava-projector ../path/to/llava.projector --output-dir ../path/to/output --clip-model-is-vision`
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-6) Then convert the model to gguf format:
-```console
-python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-7) And finally we can run the llava-cli using the 1.6 model version:
-```console
-./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
-```
-
+5) Everything else as usual: convert.py the hf model, quantize as needed
 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
 **note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
 

examples/llava/clip.cpp

@@ -616,9 +616,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             KQ = ggml_soft_max_inplace(ctx0, KQ);
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
             KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size);
-            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            KQV = ggml_cont(ctx0, ggml_permute(ctx0, KQV, 0, 2, 1, 3));
 
-            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
+            cur = ggml_cpy(ctx0, KQV, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size));
         }
 
         // attention output
@@ -1862,6 +1862,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
     std::vector<uint8_t> work(512);
     std::vector<float> conv_buf(512);
+    std::vector<int64_t> hist_all(1 << 4, 0);
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 
@@ -1916,7 +1917,48 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             }
             new_data = work.data();
 
-            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr);
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
+            switch (new_type) {
+                case GGML_TYPE_Q4_0: {
+                    new_size = ggml_quantize_q4_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_1: {
+                    new_size = ggml_quantize_q4_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_0: {
+                    new_size = ggml_quantize_q5_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_1: {
+                    new_size = ggml_quantize_q5_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q8_0: {
+                    new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q2_K: {
+                    new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q3_K: {
+                    new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_K: {
+                    new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_K: {
+                    new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q6_K: {
+                    new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                default: {
+                    fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
+                    return false;
+                }
+            }
+
+            for (size_t j = 0; j < hist_cur.size(); ++j) {
+                hist_all[j] += hist_cur[j];
+            }
         } else {
             new_type = cur->type;
             new_data = cur->data;
@@ -1951,6 +1993,17 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     {
         printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
         printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+
+        int64_t sum_all = 0;
+        for (size_t i = 0; i < hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (size_t i = 0; i < hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
     }
 
     return true;

examples/llava/llava-cli.cpp

@@ -230,7 +230,7 @@ static struct llava_context * llava_init(gpt_params * params) {
     }
 
     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
+    ctx_params.kv_size              = params->kv_size < 2048 ? 2048 : params->kv_size; // we need a longer context size to process image embeddings
 
     llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 

examples/llava/llava-surgery-v2.py

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

examples/llava/llava-surgery.py

@@ -25,6 +25,9 @@ if len(clip_tensors) > 0:
     clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
     torch.save(clip, f"{args.model}/llava.clip")
 
+    # remove these tensors
+    for name in clip_tensors:
+        del checkpoint[name]
 
     # added tokens should be removed to be able to convert Mistral models
     if os.path.exists(f"{args.model}/added_tokens.json"):
@@ -32,6 +35,7 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
+    torch.save(checkpoint, path)
 
 print("Done!")
 print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")

examples/llava/llava.cpp

@@ -103,15 +103,15 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     const size_t num_images = num_patches_width * num_patches_height + 1;
 
     // TODO: size calculation is not calculated - it's only tens of MB
-    size_t ctx_size = 0;
+    size_t kv_size = 0;
 
     {
-        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
-        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+        kv_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        kv_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
     }
 
     struct ggml_init_params params {
-        /*.mem_size   =*/ ctx_size,
+        /*.mem_size   =*/ kv_size,
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
     };
@@ -152,7 +152,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
     model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
-    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
+    if (newline_tmp->backend != GGML_BACKEND_CPU) {
         if (newline_tmp->buffer == NULL) {
             printf("newline_tmp tensor buffer is NULL\n");
         }
@@ -311,7 +311,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
     if (!image_embd) {
         fprintf(stderr, "Unable to allocate memory for image embeddings\n");

examples/llava/llava.h

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

examples/lookahead/lookahead.cpp

@@ -73,8 +73,8 @@ int main(int argc, char ** argv) {
     inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
     all = inp;
 
-    const int max_context_size     = llama_n_ctx(ctx);
-    const int max_tokens_list_size = max_context_size - 4;
+    const int max_kv_size          = llama_kv_size(ctx);
+    const int max_tokens_list_size = max_kv_size - 4;
 
     if ((int) inp.size() > max_tokens_list_size) {
         fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
@@ -117,7 +117,7 @@ int main(int argc, char ** argv) {
     // seq_id == 0           : the current input token
     // seq_id [1, W]         : tokens from the past N - 1 Jacobi iterations
     // seq_id [W + 1, W + G] : verification n-grams
-    llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
+    llama_batch batch = llama_batch_init(params.kv_size, 0, W + G + 1);
 
     // target model sampling context
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);

examples/lookup/lookup.cpp

@@ -47,8 +47,8 @@ int main(int argc, char ** argv){
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
 
-    const int max_context_size     = llama_n_ctx(ctx);
-    const int max_tokens_list_size = max_context_size - 4;
+    const int max_kv_size          = llama_kv_size(ctx);
+    const int max_tokens_list_size = max_kv_size - 4;
 
     if ((int) inp.size() > max_tokens_list_size) {
         fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
@@ -86,7 +86,7 @@ int main(int argc, char ** argv){
 
     std::vector<llama_token> draft;
 
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+    llama_batch batch_tgt = llama_batch_init(params.kv_size, 0, 1);
 
     // debug
     struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);

examples/main/README.md

@@ -70,7 +70,8 @@ In this section, we cover the most commonly used options for running the `main`
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
 -   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
--   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-c N`, `--ctx-size N`: Deprecated, use `--kv-size` instead.
+-   `-kv N`, `--kv-size N`: Set the size of the KV cache for the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 
 ## Input Prompts
 
@@ -134,15 +135,15 @@ By understanding and utilizing these interaction options, you can create engagin
 
 During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.
 
-### Context Size
+### KV Context Size
 
-The `--ctx-size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
+The `--kv-size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
 
--   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
+-   `-c N, --kv-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
 ### Extended Context Size
 
-Some fine-tuned models have extended the context length by scaling RoPE. For example, if the original pre-trained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+Some fine-tuned models have extended the context length by scaling RoPE. For example, if the original pre-trained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--kv-size` to 32768 (32k) and `--rope-scale` to 8.
 
 -   `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
 
@@ -152,7 +153,7 @@ The `--keep` option allows users to retain the original prompt when the model ru
 
 -   `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
 
-By utilizing context management options like `--ctx-size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
+By utilizing context management options like `--kv-size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
 
 ## Generation Flags
 
@@ -181,12 +182,12 @@ Example usage: `--temp 0.5`
 ### Repeat Penalty
 
 -   `--repeat-penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
--   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = kv-size).
 -   `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
 
 The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
 
-The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
+The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`kv-size`).
 
 Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases.
 

examples/main/main.cpp

@@ -157,9 +157,9 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    if (params.n_ctx != 0 && params.n_ctx < 8) {
-        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
-        params.n_ctx = 8;
+    if (params.kv_size != 0 && params.kv_size < 8) {
+        LOG_TEE("%s: warning: minimum KV size is 8, using minimum size.\n", __func__);
+        params.kv_size = 8;
     }
 
     if (params.rope_freq_base != 0.0) {
@@ -208,12 +208,12 @@ int main(int argc, char ** argv) {
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
-    const int n_ctx = llama_n_ctx(ctx);
-    LOG("n_ctx: %d\n", n_ctx);
+    const int kv_size     = llama_kv_size(ctx);
+    LOG("kv_size: %d\n", kv_size);
 
-    if (n_ctx > n_ctx_train) {
+    if (kv_size > n_ctx_train) {
         LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, n_ctx);
+                __func__, n_ctx_train, kv_size);
     }
 
     // print system information
@@ -233,7 +233,7 @@ int main(int argc, char ** argv) {
             LOG_TEE("%s: The session file is empty. A new session will be initialized.\n", __func__);
         } else {
             // The file exists and is not empty
-            session_tokens.resize(n_ctx);
+            session_tokens.resize(kv_size);
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                 LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
@@ -289,8 +289,8 @@ int main(int argc, char ** argv) {
         LOG("guidance_offset:     %s", log_tostr(guidance_offset));
     }
 
-    if ((int) embd_inp.size() > n_ctx - 4) {
-        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+    if ((int) embd_inp.size() > kv_size - 4) {
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), kv_size - 4);
         return 1;
     }
 
@@ -334,8 +334,6 @@ int main(int argc, char ** argv) {
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
         params.n_keep = (int)embd_inp.size();
-    } else {
-        params.n_keep += add_bos; // always keep the BOS token
     }
 
     // prefix & suffix for instruct mode
@@ -385,8 +383,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > add_bos) {
-            LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
                 LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
@@ -452,7 +450,7 @@ int main(int argc, char ** argv) {
     }
     LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
     LOG_TEE("sampling order: \n%s\n", llama_sampling_order_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("generate: kv_size = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", kv_size, params.n_batch, params.n_predict, params.n_keep);
 
     // group-attention state
     // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
@@ -465,7 +463,7 @@ int main(int argc, char ** argv) {
         GGML_ASSERT(ga_n > 0                    && "grp_attn_n must be positive");                     // NOLINT
         GGML_ASSERT(ga_w % ga_n == 0            && "grp_attn_w must be a multiple of grp_attn_n");     // NOLINT
       //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of grp_attn_w");    // NOLINT
-      //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
+      //GGML_ASSERT(kv_size >= n_ctx_train * ga_n && "kv_size must be at least n_ctx_train * grp_attn_n"); // NOLINT
         LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
     }
     LOG_TEE("\n\n");
@@ -511,22 +509,14 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    // tokenized antiprompts
-    std::vector<std::vector<llama_token>> antiprompt_ids;
-
-    antiprompt_ids.reserve(params.antiprompt.size());
-    for (const std::string & antiprompt : params.antiprompt) {
-        antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
-    }
-
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (!embd.empty()) {
-            // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
+            // Note: (kv_size - 4) here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
-            int max_embd_size = n_ctx - 4;
+            int max_embd_size = kv_size - 4;
 
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
             if ((int) embd.size() > max_embd_size) {
@@ -543,21 +533,21 @@ int main(int argc, char ** argv) {
                 // infinite text generation via context shifting
                 // if we run out of context:
                 // - take the n_keep first tokens from the original prompt (via n_past)
-                // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                // - take half of the last (kv_size - n_keep) tokens and recompute the logits in batches
+                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > kv_size) {
                     if (params.n_predict == -2) {
                         LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep;
+                    const int n_left    = n_past - params.n_keep - 1;
                     const int n_discard = n_left/2;
 
-                    LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                            n_past, n_left, n_ctx, params.n_keep, n_discard);
+                    LOG("context full, swapping: n_past = %d, n_left = %d, kv_size = %d, n_keep = %d, n_discard = %d\n",
+                        n_past, n_left, kv_size, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -584,9 +574,9 @@ int main(int argc, char ** argv) {
                     LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 
@@ -676,7 +666,7 @@ int main(int argc, char ** argv) {
                 LOG("n_past = %d\n", n_past);
                 // Display total tokens alongside total time
                 if (params.n_print > 0 && n_past % params.n_print == 0) {
-                    LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
+                    LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, kv_size);
                 }
             }
 
@@ -777,18 +767,6 @@ int main(int argc, char ** argv) {
                     }
                 }
 
-                // check for reverse prompt using special tokens
-                llama_token last_token = llama_sampling_last(ctx_sampling);
-                for (std::vector<llama_token> ids : antiprompt_ids) {
-                    if (ids.size() == 1 && last_token == ids[0]) {
-                        if (params.interactive) {
-                            is_interacting = true;
-                        }
-                        is_antiprompt = true;
-                        break;
-                    }
-                }
-
                 if (is_antiprompt) {
                     LOG("found antiprompt: %s\n", last_output.c_str());
                 }
@@ -878,7 +856,6 @@ int main(int argc, char ** argv) {
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
                     const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
                     const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
-
                     LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
 
                     embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());

examples/parallel/parallel.cpp

@@ -107,9 +107,6 @@ int main(int argc, char ** argv) {
     // number of simultaneous "clients" to simulate
     const int32_t n_clients = params.n_parallel;
 
-    // dedicate one sequence to the system prompt
-    params.n_parallel += 1;
-
     // requests to simulate
     const int32_t n_seq = params.n_sequences;
 
@@ -155,7 +152,7 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "\n\n");
     fflush(stderr);
 
-    const int n_ctx = llama_n_ctx(ctx);
+    const int kv_size = llama_kv_size(ctx);
 
     std::vector<client> clients(n_clients);
     for (size_t i = 0; i < clients.size(); ++i) {
@@ -172,7 +169,7 @@ int main(int argc, char ** argv) {
 
     // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
     // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
-    llama_batch batch = llama_batch_init(n_ctx, 0, 1);
+    llama_batch batch = llama_batch_init(kv_size, 0, 1);
 
     int32_t n_total_prompt = 0;
     int32_t n_total_gen    = 0;
@@ -199,8 +196,8 @@ int main(int argc, char ** argv) {
         }
 
         // assign the system KV cache to all parallel sequences
-        for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+        for (int32_t i = 1; i < n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
         }
 
         LOG_TEE("\n");
@@ -224,17 +221,15 @@ int main(int argc, char ** argv) {
 
             client.i_batch = batch.n_tokens;
 
-            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
+            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true);
 
             client.n_decoded += 1;
         }
 
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
-            for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, -1, -1);
-                // but keep the system prompt
-                llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+            for (int i = 0; i < n_clients; ++i) {
+                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
             }
 
             LOG_TEE("%s: clearing the KV cache\n", __func__);
@@ -260,7 +255,7 @@ int main(int argc, char ** argv) {
                     tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
+                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false);
                     }
 
                     // extract the logits only for the last token
@@ -371,8 +366,7 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
-                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -92,7 +92,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed    = seed;
-    ctx_params.n_ctx   = llama_n_ctx_train(model)*n_grp + n_keep;
+    ctx_params.kv_size   = llama_n_ctx_train(model)*n_grp + n_keep;
     ctx_params.n_batch = 512;
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
@@ -121,12 +121,12 @@ int main(int argc, char ** argv) {
     // total length of the sequences including the prompt
     const int n_len = n_tokens_all + n_predict;
 
-    const int n_ctx       = llama_n_ctx(ctx) - n_keep;
-    const int n_kv_req    = llama_n_ctx(ctx);
+    const int kv_size     = llama_kv_size(ctx) - n_keep;
+    const int n_kv_req    = llama_kv_size(ctx);
     const int n_batch     = ctx_params.n_batch;
     const int n_batch_grp = ctx_params.n_batch/n_grp;
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
+    LOG_TEE("\n%s: n_len = %d, kv_size = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, kv_size, n_kv_req, n_grp, n_batch);
 
     // print the prompt token-by-token
 
@@ -140,17 +140,16 @@ int main(int argc, char ** argv) {
     int n_past = 0;
 
     // fill the KV cache
-    for (int i = 0; i < n_ctx; i += n_batch) {
+    for (int i = 0; i < kv_size; i += n_batch) {
         if (i > 0 && n_grp > 1) {
             // if SelfExtend is enabled, we compress the position from the last batch by a factor of n_grp
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past -= bd;
         }
 
         llama_batch_clear(batch);
@@ -175,17 +174,15 @@ int main(int argc, char ** argv) {
         }
     }
 
-    for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
+    for (int i = kv_size; i < n_tokens_all; i += n_batch) {
         const int n_discard = n_batch;
 
         LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, kv_size, -n_discard);
 
-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past -= n_discard;
 
         llama_batch_clear(batch);
 
@@ -206,17 +203,15 @@ int main(int argc, char ** argv) {
     }
 
     {
-        const int n_discard = n_past - n_ctx + n_predict;
+        const int n_discard = n_past - kv_size + n_predict;
 
         if (n_discard > 0) {
             LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, kv_size, -n_discard);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past -= n_discard;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -320,11 +320,11 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
-    const int n_ctx = llama_n_ctx(ctx);
+    const int kv_size = llama_kv_size(ctx);
 
-    if (int(tokens.size()) < 2*n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
-                n_ctx);
+    if (int(tokens.size()) < 2*kv_size) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n", __func__, 2 * kv_size,
+                kv_size);
         fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
         return {std::move(tokens), 0., {}, {}};
     }
@@ -340,13 +340,13 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         return {tokens, -1, logit_history, prob_history};
     }
 
-    const int calc_chunk = n_ctx;
+    const int calc_chunk = kv_size;
 
     fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
 
     if (int(tokens.size()) <= calc_chunk) {
-        fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
-                tokens.size(), n_ctx, params.ppl_stride);
+        fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n", __func__,
+                tokens.size(), kv_size, params.ppl_stride);
         return {tokens, -1, logit_history, prob_history};
     }
 
@@ -414,8 +414,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
-        //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
-        for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {
+        //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.kv_size - params.ppl_stride + start, params.kv_size + start);
+        for (int j = kv_size - params.ppl_stride - 1; j < kv_size - 1; ++j) {
 
             // Calculate probability of next token, given the previous ones.
             const std::vector<float> tok_logits(
@@ -442,7 +442,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     return {tokens, std::exp(nll / count), logit_history, prob_history};
 }
 
-static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) {
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     if (params.ppl_stride > 0) {
         return perplexity_v2(ctx, params);
     }
@@ -453,6 +453,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // BOS tokens will be added for each chunk before eval
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const int kv_size  = llama_kv_size(ctx);
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -463,7 +464,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         }
         fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str());
         logits_stream.write("_logits_", 8);
-        logits_stream.write(reinterpret_cast<const char *>(&n_ctx), sizeof(n_ctx));
+        logits_stream.write(reinterpret_cast<const char *>(&kv_size), sizeof(kv_size));
     }
 
     auto tim1 = std::chrono::high_resolution_clock::now();
@@ -474,9 +475,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     auto tim2 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
-    if (int(tokens.size()) < 2*n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
-                n_ctx);
+    if (int(tokens.size()) < 2*kv_size) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n", __func__, 2 * kv_size,
+                kv_size);
         fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
         return {std::move(tokens), 0., {}, {}};
     }
@@ -487,7 +488,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     std::vector<float> prob_history;
     prob_history.resize(tokens.size());
 
-    const int n_chunk_max = tokens.size() / n_ctx;
+    const int n_chunk_max = tokens.size() / kv_size;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
     const int n_vocab = llama_n_vocab(llama_get_model(ctx));
@@ -497,20 +498,14 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     double nll = 0.0;
     double nll2 = 0.0;
 
-    const int num_batches = (n_ctx + n_batch - 1) / n_batch;
-    const int n_seq = std::max(1, n_batch / n_ctx);
-
-    GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0);
-    GGML_ASSERT(params.n_ctx == n_seq * n_ctx);
-
-    llama_batch batch = llama_batch_init(std::min(n_batch, n_ctx*n_seq), 0, 1);
+    const int num_batches = (kv_size + n_batch - 1) / n_batch;
 
     std::vector<float> logits;
     if (num_batches > 1) {
-        logits.reserve((size_t)n_ctx * n_vocab);
+        logits.reserve((size_t)kv_size * n_vocab);
     }
 
-    fprintf(stderr, "%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
@@ -518,30 +513,14 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     if (!params.logits_file.empty()) {
         logits_stream.write((const char *)&n_vocab, sizeof(n_vocab));
         logits_stream.write((const char *)&n_chunk, sizeof(n_chunk));
-        logits_stream.write((const char *)tokens.data(), n_chunk*n_ctx*sizeof(tokens[0]));
+        logits_stream.write((const char *)tokens.data(), n_chunk * kv_size * sizeof(tokens[0]));
         const int nv = 2*((n_vocab + 1)/2) + 4;
-        log_probs.resize(n_ctx * nv);
+        log_probs.resize(kv_size * nv);
     }
 
-    // We get the logits for all the tokens in the context window (params.n_ctx)
-    // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
-    // calculate the perplexity over the last half of the window (so the model always has
-    // some context to predict the token).
-    //
-    // We rely on the fact that attention in the forward pass only looks at previous
-    // tokens here, so the logits returned for each token are an accurate representation
-    // of what the model would have predicted at that point.
-    //
-    // Example, we have a context window of 512, we will compute perplexity for each of the
-    // last 256 tokens.  Then, we split the input up into context window size chunks to
-    // process the entire prompt.
-    const int first = n_ctx/2;
-
-    for (int i = 0; i < n_chunk; i += n_seq) {
-        const int start =     i * n_ctx;
-        const int end   = start + n_ctx;
-
-        const int n_seq_batch = std::min(n_seq, n_chunk - i);
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start = i * kv_size;
+        const int end   = start + kv_size;
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -552,50 +531,34 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
-            batch.n_tokens = 0;
-            for (int seq = 0; seq < n_seq_batch; seq++) {
-                int seq_start = batch_start + seq*n_ctx;
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
 
-                // save original token and restore it after eval
-                const auto token_org = tokens[seq_start];
-
-                // add BOS token for the first batch of each chunk
-                if (add_bos && j == 0) {
-                    tokens[seq_start] = llama_token_bos(llama_get_model(ctx));
-                }
-
-                for (int k = 0; k < batch_size; ++k) {
-                    const int idx = seq*n_ctx + k;
-                    batch.token[idx] = tokens[seq_start + k];
-                    batch.pos[idx] = j*n_batch + k;
-                    batch.n_seq_id[idx] = 1;
-                    batch.seq_id[idx][0] = seq;
-                    batch.logits[idx] = batch.pos[idx] >= first ? 1 : 0;
-                }
-                batch.n_tokens += batch_size;
-
-                // restore the original token in case it was set to BOS
-                tokens[seq_start] = token_org;
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            if (llama_decode(ctx, batch)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
 
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
             if (num_batches > 1) {
                 const auto * batch_logits = llama_get_logits(ctx);
                 logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
             }
         }
 
+        const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
-            llama_synchronize(ctx);
-            const auto t_end = std::chrono::high_resolution_clock::now();
             const float t_total = std::chrono::duration<float>(t_end - t_start).count();
             fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
-            int total_seconds = (int)(t_total*n_chunk/n_seq);
+            int total_seconds = (int)(t_total * n_chunk);
             if (total_seconds >= 60*60) {
                 fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
@@ -603,31 +566,37 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
-        for (int seq = 0; seq < n_seq_batch; seq++) {
-            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx);
-            llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;
-            if (!params.logits_file.empty()) {
-                process_logits(logits_stream, n_vocab, all_logits + first*n_vocab,
-                        tokens_data, n_ctx - 1 - first,
-                        workers, log_probs, nll, nll2);
-            } else {
-                process_logits(n_vocab, all_logits + first*n_vocab,
-                        tokens_data, n_ctx - 1 - first,
-                        workers, nll, nll2,
-                        logit_history.data() + start + seq*n_ctx + first,
-                        prob_history.data()  + start + seq*n_ctx + first);
-            }
-            count += n_ctx - first - 1;
+        // We get the logits for all the tokens in the context window (params.kv_size)
+        // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
+        // calculate the perplexity over the last half of the window (so the model always has
+        // some context to predict the token).
+        //
+        // We rely on the fact that attention in the forward pass only looks at previous
+        // tokens here, so the logits returned for each token are an accurate representation
+        // of what the model would have predicted at that point.
+        //
+        // Example, we have a context window of 512, we will compute perplexity for each of the
+        // last 256 tokens.  Then, we split the input up into context window size chunks to
+        // process the entire prompt.
+        const int first = kv_size/2;
+        const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+        if (!params.logits_file.empty()) {
+            process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, kv_size - 1 - first,
+                    workers, log_probs, nll, nll2);
+        } else {
+            process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, kv_size - 1 - first,
+                    workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+        }
+        count += kv_size - first - 1;
 
-            // perplexity is e^(average negative log-likelihood)
-            if (params.ppl_output_type == 0) {
-                printf("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
-            } else {
-                double av = nll/count;
-                double av2 = nll2/count - av*av;
-                if (av2 > 0) av2 = sqrt(av2/(count-1));
-                printf("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
-            }
+        // perplexity is e^(average negative log-likelihood)
+        if (params.ppl_output_type == 0) {
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        } else {
+            double av = nll/count;
+            double av2 = nll2/count - av*av;
+            if (av2 > 0) av2 = sqrt(av2/(count-1));
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*kv_size, std::exp(nll / count), av, av2);
         }
         fflush(stdout);
 
@@ -646,8 +615,6 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         printf("Unexpected negative standard deviation of log(prob)\n");
     }
 
-    llama_batch_free(batch);
-
     return {tokens, ppl, logit_history, prob_history};
 }
 
@@ -838,16 +805,16 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     double acc = 0.0f;
 
     const int n_vocab = llama_n_vocab(llama_get_model(ctx));
-    const int n_ctx   = llama_n_ctx(ctx);
+    const int kv_size = llama_kv_size(ctx);
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 32;
-    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
+    const int max_seq = 4*max_tasks_per_batch;
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(kv_size, 0, max_seq);
 
     std::vector<float> tok_logits(n_vocab);
-    std::vector<float> batch_logits(n_vocab*n_ctx);
+    std::vector<float> batch_logits(n_vocab*kv_size);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
     std::vector<float> eval_results;
@@ -865,7 +832,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         // each task has 4 unique seuqnce ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
-        while (n_cur + (int) hs_data[i1].required_tokens <= n_ctx) {
+        while (n_cur + (int) hs_data[i1].required_tokens <= kv_size) {
             auto & hs_cur = hs_data[i1];
 
             const int s0 = 4*(i1 - i0);
@@ -1115,16 +1082,16 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     fprintf(stderr, "%s : calculating winogrande score over selected tasks.\n", __func__);
 
     const int n_vocab = llama_n_vocab(llama_get_model(ctx));
-    const int n_ctx   = llama_n_ctx(ctx);
+    const int kv_size = llama_kv_size(ctx);
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 128;
-    const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
+    const int max_seq = 2*max_tasks_per_batch;
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(kv_size, 0, max_seq);
 
     std::vector<float> tok_logits(n_vocab);
-    std::vector<float> batch_logits(n_vocab*n_ctx);
+    std::vector<float> batch_logits(n_vocab*kv_size);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
     std::vector<float> eval_results;
@@ -1141,7 +1108,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
 
         llama_batch_clear(batch);
 
-        while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
+        while (n_cur + (int) data[i1].required_tokens <= kv_size) {
             const int s0 = 2*(i1 - i0);
             if (s0 + 2 > max_seq) {
                 break;
@@ -1467,16 +1434,16 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     printf("\ntask\tacc_norm\n");
 
     const int n_vocab = llama_n_vocab(llama_get_model(ctx));
-    const int n_ctx   = llama_n_ctx(ctx);
+    const int kv_size = llama_kv_size(ctx);
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 32;
-    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
+    const int max_seq = 4*max_tasks_per_batch;
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(kv_size, 0, max_seq);
 
     std::vector<float> tok_logits(n_vocab);
-    std::vector<float> batch_logits(n_vocab*n_ctx);
+    std::vector<float> batch_logits(n_vocab*kv_size);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
     std::vector<float> eval_results;
@@ -1500,7 +1467,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         int s0 = 0;
-        while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) {
+        while (n_cur + (int) tasks[i1].required_tokens <= kv_size) {
             auto& cur_task = tasks[i1];
 
             int num_answers = cur_task.seq_tokens.size();
@@ -1653,11 +1620,11 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
         }
     }
 
-    uint32_t n_ctx;
-    in.read((char *)&n_ctx, sizeof(n_ctx));
-    if (n_ctx > llama_n_ctx(ctx)) {
-        fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
-                __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
+    uint32_t kv_size;
+    in.read((char *)&kv_size, sizeof(kv_size));
+    if (kv_size > llama_kv_size(ctx)) {
+        fprintf(stderr, "%s: %s has been computed with %u, while the current KV Cache size is %d. Increase it with -kv and retry\n",
+                __func__, params.logits_file.c_str(), kv_size, params.kv_size);
     }
 
     int n_vocab, n_chunk;
@@ -1671,22 +1638,22 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
         fprintf(stderr, "%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
     }
 
-    std::vector<llama_token> tokens(n_ctx * n_chunk);
+    std::vector<llama_token> tokens(kv_size * n_chunk);
     if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) {
         fprintf(stderr, "%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
         return;
     }
 
     const int n_batch = params.n_batch;
-    const int num_batches = (n_ctx + n_batch - 1)/n_batch;
+    const int num_batches = (kv_size + n_batch - 1)/n_batch;
     const int nv = 2*((n_vocab + 1)/2) + 4;
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
 
-    std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
-    std::vector<float> kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
+    std::vector<uint16_t> log_probs_uint16(size_t(kv_size - 1 - kv_size/2) * nv);
+    std::vector<float> kld_values(size_t(kv_size - 1 - kv_size /2)*n_chunk);
     std::vector<float> logits;
     if (num_batches > 1) {
-        logits.reserve(n_ctx * n_vocab);
+        logits.reserve(kv_size * n_vocab);
     }
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
@@ -1705,8 +1672,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     auto kld_ptr = kld_values.data();
 
     for (int i = 0; i < n_chunk; ++i) {
-        const int start =     i * n_ctx;
-        const int end   = start + n_ctx;
+        const int start = i * kv_size;
+        const int end   = start + kv_size;
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -1759,11 +1726,11 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
             printf("\nchunk        PPL          ln(PPL(Q)/PPL(base))          KL-Divergence           Same top\n");
         }
 
-        const int first = n_ctx/2;
+        const int first = kv_size/2;
         const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
-        process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+        process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, kv_size - 1 - first,
                 workers, log_probs_uint16, kld, kld_ptr);
-        kld_ptr += n_ctx - 1 - first;
+        kld_ptr += kv_size - 1 - first;
 
         auto ppl           = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
         auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count);
@@ -1815,29 +1782,18 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
+    params.n_batch = 512;
     if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
     params.logits_all = true;
-
-    const int32_t n_ctx = params.n_ctx;
-
-    const bool ppl = !params.hellaswag && !params.winogrande && !params.multiple_choice && !params.kl_divergence;
-    if (ppl) {
-        int n_seq = std::max(1, params.n_batch / n_ctx);
-        int32_t n_kv = n_seq * n_ctx;
-        params.n_parallel = n_seq;
-        params.n_ctx = n_kv;
-        params.n_batch = std::min(params.n_batch, n_kv);
-    } else {
-        params.n_batch = std::min(params.n_batch, params.n_ctx);
-    }
+    params.n_batch = std::min(params.n_batch, params.kv_size);
 
     if (params.ppl_stride > 0) {
-        fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
-                params.n_ctx, params.n_ctx + params.ppl_stride/2);
-        params.n_ctx += params.ppl_stride/2;
+        fprintf(stderr, "Will perform strided perplexity calculation -> adjusting KV size from %d to %d\n",
+                params.kv_size, params.kv_size + params.ppl_stride / 2);
+        params.kv_size += params.ppl_stride/2;
     }
 
     print_build_info();
@@ -1859,9 +1815,6 @@ int main(int argc, char ** argv) {
     llama_model * model;
     llama_context * ctx;
 
-    // ensure there's at least enough seq_ids for HellaSwag
-    params.n_parallel = std::max(4, params.n_parallel);
-
     // load the model and apply lora adapter, if any
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == NULL) {
@@ -1870,9 +1823,9 @@ int main(int argc, char ** argv) {
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
-    if (params.n_ctx > n_ctx_train) {
+    if (params.kv_size > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, params.n_ctx);
+                __func__, n_ctx_train, params.kv_size);
     }
 
     // print system information
@@ -1891,7 +1844,7 @@ int main(int argc, char ** argv) {
     } else if (params.kl_divergence) {
         kl_divergence(ctx, params);
     } else {
-        results = perplexity(ctx, params, n_ctx);
+        results = perplexity(ctx, params);
     }
 
     llama_print_timings(ctx);

examples/quantize-stats/quantize-stats.cpp

@@ -319,7 +319,7 @@ int main(int argc, char ** argv) {
         }
 
         auto cparams = llama_context_default_params();
-        cparams.n_ctx      = 256;
+        cparams.kv_size    = 256;
         cparams.seed       = 1;
 
         ctx = llama_new_context_with_model(model, cparams);

examples/quantize/quantize.cpp

@@ -23,21 +23,15 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
     { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
     { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
-    { "IQ2_S",  LLAMA_FTYPE_MOSTLY_IQ2_S,  " 2.5  bpw quantization",            },
-    { "IQ2_M",  LLAMA_FTYPE_MOSTLY_IQ2_M,  " 2.7  bpw quantization",            },
     { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
-    { "IQ3_S",  LLAMA_FTYPE_MOSTLY_IQ3_S,  " 3.44 bpw quantization",            },
-    { "IQ3_M",  LLAMA_FTYPE_MOSTLY_IQ3_M,  " 3.66 bpw quantization mix",        },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
-    { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization"   ,          },
+    { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
     { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
-    { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", },
-    { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", },
     { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
     { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
     { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
@@ -295,7 +289,6 @@ int main(int argc, char ** argv) {
     }
 
     if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
          params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
         fprintf(stderr, "\n===============================================================================================\n");
         fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");

examples/server-embd.py

@@ -1,34 +0,0 @@
-import asyncio
-import requests
-import numpy as np
-
-n = 8
-
-result = []
-
-async def requests_post_async(*args, **kwargs):
-    return await asyncio.to_thread(requests.post, *args, **kwargs)
-
-async def main():
-    model_url = "http://127.0.0.1:6900"
-    responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
-        url= f"{model_url}/embedding",
-        json= {"content": str(0)*1024}
-    ) for i in range(n)])
-
-    for response in responses:
-        embedding = response.json()["embedding"]
-        print(embedding[-8:])
-        result.append(embedding)
-
-asyncio.run(main())
-
-# compute cosine similarity
-
-for i in range(n-1):
-    for j in range(i+1, n):
-        embedding1 = np.array(result[i])
-        embedding2 = np.array(result[j])
-        similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
-        print(f"Similarity between {i} and {j}: {similarity:.2f}")
-

examples/server-llama2-13B.sh

@@ -12,7 +12,7 @@ PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat-system.txt}
 N_THREAD="${N_THREAD:-12}"
 
 # Note: you can also override the generation options by specifying them on the command line:
-GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 4096 --batch-size 1024}"
+GEN_OPTIONS="${GEN_OPTIONS:---kv_size 4096 --batch-size 1024}"
 
 
 # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS

examples/server/CMakeLists.txt

@@ -1,18 +1,12 @@
 set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
-if (LLAMA_SERVER_SSL)
-    find_package(OpenSSL REQUIRED)
-    target_link_libraries(${TARGET} PRIVATE OpenSSL::SSL OpenSSL::Crypto)
-    target_compile_definitions(${TARGET} PRIVATE CPPHTTPLIB_OPENSSL_SUPPORT)
-endif()
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -1,27 +1,15 @@
-# LLaMA.cpp HTTP Server
+# llama.cpp/example/server
 
-Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/yhirose/cpp-httplib), [nlohmann::json](https://github.com/nlohmann/json) and **llama.cpp**.
+This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
 
-Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
-
-**Features:**
- * LLM inference of F16 and quantum models on GPU and CPU
- * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
- * Parallel decoding with multi-user support
- * Continuous batching
- * Multimodal (wip)
- * Monitoring endpoints
-
-The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
-
-**Command line options:**
+Command line options:
 
 - `--threads N`, `-t N`: Set the number of threads to use during generation.
 - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
-- `--threads-http N`: number of threads in the http server pool to process requests (default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`)
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
 - `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
-- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
+- `-c N`, `--ctx-size N`: Deprecated, use `--kv-size` instead.
+- `-kv N`, `--kv-size N`: Specify the total size of the KV cache. This corresponds to the total amount of tokens that can be stored across all independent sequences / slots. `llama.cpp` implements a "unified" cache strategy, the KV cache size is actually shared across all sequences. It's allowed to have sequences with more than `T` tokens as long as the sum of all tokens does not exceed `P*T`. The default is 512.
 - `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 - `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 - `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
@@ -42,26 +30,18 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
 - `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 - `--port`: Set the port to listen. Default: `8080`.
-- `--path`: path from which to serve static files (default: disabled)
+- `--path`: path from which to serve static files (default examples/server/public)
 - `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
 - `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
 - `--embedding`: Enable embedding extraction, Default: disabled.
-- `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
+- `-np N`, `--parallel N`: Set the number of slots / sequences for process requests (default: 1). Each sequence can have a maximum of `T` tokens, use together with `--kv-size`.
 - `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
 - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 - `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
-- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
+- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
-- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
-- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- `--log-disable`: Output logs to stdout only, default: enabled.
-- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
-
-**If compiled with `LLAMA_SERVER_SSL=ON`**
-- `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key
-- `--ssl-cert-file FNAME`: path to file a PEM-encoded SSL certificate
 
 ## Build
 
@@ -79,28 +59,6 @@ server is build alongside everything else from the root of the project
   cmake --build . --config Release
   ```
 
-## Build with SSL
-
-server can also be built with SSL support using OpenSSL 3
-
-- Using `make`:
-
-  ```bash
-  # NOTE: For non-system openssl, use the following:
-  #   CXXFLAGS="-I /path/to/openssl/include"
-  #   LDFLAGS="-L /path/to/openssl/lib"
-  make LLAMA_SERVER_SSL=true server
-  ```
-
-- Using `CMake`:
-
-  ```bash
-  mkdir build
-  cd build
-  cmake .. -DLLAMA_SERVER_SSL=ON
-  make server
-  ```
-
 ## Quick Start
 
 To get started right away, run the following command, making sure to use the correct path for the model you have:
@@ -123,10 +81,10 @@ You can consume the endpoints with Postman or NodeJS with axios library. You can
 ### Docker
 
 ```bash
-docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
+docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
 
 # or, with CUDA:
-docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
+docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
 ```
 
 ## Testing with CURL
@@ -140,12 +98,6 @@ curl --request POST \
     --data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
 ```
 
-## Advanced testing
-
-We implemented a [server test framework](./tests/README.md) using human-readable scenario.
-
-*Before submitting an issue, please try to reproduce it with this format.*
-
 ## Node JS Test
 
 You need to have [Node.js](https://nodejs.org/en) installed.
@@ -183,27 +135,20 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - 503 -> `{"status": "loading model"}` if the model is still being loaded.
-  - 500 -> `{"status": "error"}` if the model failed to load.
-  - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
-
-  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+  - `{"status": "loading model"}` if the model is still being loaded.
+  - `{"status": "error"}` if the model failed to load.
+  - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
+  - `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available
 
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
-    `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, if `cache_prompt` is `true`, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. A `BOS` token is inserted at the start, if all of the following conditions are true:
-
-      - The prompt is a string or an array with the first element given as a string
-      - The model's `tokenizer.ggml.add_bos_token` metadata is `true`
-      - The system prompt is empty
+    `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. If the prompt is a string or an array with the first element given as a string, a `bos` token is inserted in the front like `main` does.
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled).
+    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
 
     `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -229,7 +174,7 @@ node index.js
 
     `repeat_penalty`: Control the repetition of token sequences in the generated text (default: 1.1).
 
-    `repeat_last_n`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+    `repeat_last_n`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = kv-size).
 
     `penalize_nl`: Penalize newline tokens when applying the repeat penalty (default: true).
 
@@ -261,7 +206,7 @@ node index.js
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
+    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
@@ -294,7 +239,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.).
+- `generation_settings`: The provided options above excluding `prompt` but including `kv_size`, `model`
 - `model`: The path to the model loaded with `-m`
 - `prompt`: The provided `prompt`
 - `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
@@ -304,7 +249,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
 - `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
-- `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
+- `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the KV size (`kv_size`)
 
 - **POST** `/tokenize`: Tokenize a given text.
 
@@ -312,7 +257,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `content`: Set the text to tokenize.
 
-    Note that a special `BOS` token is never inserted.
+    Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
 - **POST** `/detokenize`: Convert tokens to text.
 
@@ -356,7 +301,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `default_generation_settings` - the default generation settings for the `/completion` endpoint, has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint.
+- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
 
     *Options:*
 
@@ -459,14 +404,14 @@ Notice that each `probs` is an array of length `n_probs`.
         "mirostat_eta": 0.10000000149011612,
         "mirostat_tau": 5.0,
         "model": "llama-2-7b-32k-instruct.Q2_K.gguf",
-        "n_ctx": 2048,
+        "kv_size": 2048,
         "n_keep": 0,
         "n_predict": 100000,
         "n_probs": 0,
         "next_token": {
             "has_next_token": true,
             "n_remain": -1,
-            "n_decoded": 0,
+            "num_tokens_predicted": 0,
             "stopped_eos": false,
             "stopped_limit": false,
             "stopped_word": false,
@@ -503,18 +448,6 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
-
-Available metrics:
-- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
-- `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
-- `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
-- `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
-- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
-- `llamacpp:kv_cache_tokens`: KV-cache tokens.
-- `llamacpp:requests_processing`: Number of request processing.
-- `llamacpp:requests_deferred`: Number of request deferred.
-
 ## More examples
 
 ### Change system prompt on runtime
@@ -556,55 +489,26 @@ Run with bash:
 bash chat.sh
 ```
 
-### OAI-like API
+### API like OAI
 
-The HTTP server supports OAI-like API: https://github.com/openai/openai-openapi
+API example using Python Flask: [api_like_OAI.py](api_like_OAI.py)
+This example must be used with server.cpp
 
-### API errors
-
-Server returns error in the same format as OAI: https://github.com/openai/openai-openapi
-
-Example of an error:
-
-```json
-{
-    "error": {
-        "code": 401,
-        "message": "Invalid API Key",
-        "type": "authentication_error"
-    }
-}
+```sh
+python api_like_OAI.py
 ```
 
-Apart from error types supported by OAI, we also have custom types that are specific to functionalities of llama.cpp:
+After running the API server, you can use it in Python by setting the API base URL.
 
-**When /metrics or /slots endpoint is disabled**
-
-```json
-{
-    "error": {
-        "code": 501,
-        "message": "This server does not support metrics endpoint.",
-        "type": "not_supported_error"
-    }
-}
+```python
+openai.api_base = "http://<Your api-server IP>:port"
 ```
 
-**When the server receives invalid grammar via */completions endpoint**
-
-```json
-{
-    "error": {
-        "code": 400,
-        "message": "Failed to parse grammar",
-        "type": "invalid_request_error"
-    }
-}
-```
+Then you can utilize llama.cpp as an OpenAI's **chat.completion** or **text_completion** API
 
 ### Extending or building alternative Web Front End
 
-You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.
+The default location for the static files is `examples/server/public`. You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.
 
 Read the documentation in `/completion.js` to see convenient ways to access llama.
 

examples/server/api_like_OAI.py

@@ -0,0 +1,228 @@
+#!/usr/bin/env python3
+import argparse
+from flask import Flask, jsonify, request, Response
+import urllib.parse
+import requests
+import time
+import json
+
+
+app = Flask(__name__)
+slot_id = -1
+
+parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
+parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.')
+parser.add_argument("--user-name", type=str, help="USER name in chat completions(default: 'USER: ')", default="USER: ")
+parser.add_argument("--ai-name", type=str, help="ASSISTANT name in chat completions(default: 'ASSISTANT: ')", default="ASSISTANT: ")
+parser.add_argument("--system-name", type=str, help="SYSTEM name in chat completions(default: 'ASSISTANT's RULE: ')", default="ASSISTANT's RULE: ")
+parser.add_argument("--stop", type=str, help="the end of response in chat completions(default: '</s>')", default="</s>")
+parser.add_argument("--llama-api", type=str, help="Set the address of server.cpp in llama.cpp(default: http://127.0.0.1:8080)", default='http://127.0.0.1:8080')
+parser.add_argument("--api-key", type=str, help="Set the api key to allow only few user(default: NULL)", default="")
+parser.add_argument("--host", type=str, help="Set the ip address to listen.(default: 127.0.0.1)", default='127.0.0.1')
+parser.add_argument("--port", type=int, help="Set the port to listen.(default: 8081)", default=8081)
+
+args = parser.parse_args()
+
+def is_present(json, key):
+    try:
+        buf = json[key]
+    except KeyError:
+        return False
+    if json[key] == None:
+        return False
+    return True
+
+#convert chat to prompt
+def convert_chat(messages):
+
+    system_n = args.system_name
+    user_n = args.user_name
+    ai_n = args.ai_name
+    stop = args.stop
+
+    prompt = "" + args.chat_prompt + stop
+
+    for line in messages:
+        if (line["role"] == "system"):
+            prompt += f"{system_n}{line['content']}{stop}"
+        if (line["role"] == "user"):
+            prompt += f"{user_n}{line['content']}{stop}"
+        if (line["role"] == "assistant"):
+            prompt += f"{ai_n}{line['content']}{stop}"
+    prompt += ai_n.rstrip()
+
+    return prompt
+
+def make_postData(body, chat=False, stream=False):
+    postData = {}
+    if (chat):
+        postData["prompt"] = convert_chat(body["messages"])
+    else:
+        postData["prompt"] = body["prompt"]
+    if(is_present(body, "temperature")): postData["temperature"] = body["temperature"]
+    if(is_present(body, "top_k")): postData["top_k"] = body["top_k"]
+    if(is_present(body, "top_p")): postData["top_p"] = body["top_p"]
+    if(is_present(body, "max_tokens")): postData["n_predict"] = body["max_tokens"]
+    if(is_present(body, "presence_penalty")): postData["presence_penalty"] = body["presence_penalty"]
+    if(is_present(body, "frequency_penalty")): postData["frequency_penalty"] = body["frequency_penalty"]
+    if(is_present(body, "repeat_penalty")): postData["repeat_penalty"] = body["repeat_penalty"]
+    if(is_present(body, "mirostat")): postData["mirostat"] = body["mirostat"]
+    if(is_present(body, "mirostat_tau")): postData["mirostat_tau"] = body["mirostat_tau"]
+    if(is_present(body, "mirostat_eta")): postData["mirostat_eta"] = body["mirostat_eta"]
+    if(is_present(body, "seed")): postData["seed"] = body["seed"]
+    if(is_present(body, "grammar")): postData["grammar"] = body["grammar"]
+    if(is_present(body, "logit_bias")): postData["logit_bias"] = [[int(token), body["logit_bias"][token]] for token in body["logit_bias"].keys()]
+    if (args.stop != ""):
+        postData["stop"] = [args.stop]
+    else:
+        postData["stop"] = []
+    if(is_present(body, "stop")): postData["stop"] += body["stop"]
+    postData["n_keep"] = -1
+    postData["stream"] = stream
+    postData["cache_prompt"] = True
+    postData["slot_id"] = slot_id
+    return postData
+
+def make_resData(data, chat=False, promptToken=[]):
+    resData = {
+        "id": "chatcmpl" if (chat) else "cmpl",
+        "object": "chat.completion" if (chat) else "text_completion",
+        "created": int(time.time()),
+        "truncated": data["truncated"],
+        "model": "LLaMA_CPP",
+        "usage": {
+            "prompt_tokens": data["tokens_evaluated"],
+            "completion_tokens": data["tokens_predicted"],
+            "total_tokens": data["tokens_evaluated"] + data["tokens_predicted"]
+        }
+    }
+    if (len(promptToken) != 0):
+        resData["promptToken"] = promptToken
+    if (chat):
+        #only one choice is supported
+        resData["choices"] = [{
+            "index": 0,
+            "message": {
+                "role": "assistant",
+                "content": data["content"],
+            },
+            "finish_reason": "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+        }]
+    else:
+        #only one choice is supported
+        resData["choices"] = [{
+            "text": data["content"],
+            "index": 0,
+            "logprobs": None,
+            "finish_reason": "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+        }]
+    return resData
+
+def make_resData_stream(data, chat=False, time_now = 0, start=False):
+    resData = {
+        "id": "chatcmpl" if (chat) else "cmpl",
+        "object": "chat.completion.chunk" if (chat) else "text_completion.chunk",
+        "created": time_now,
+        "model": "LLaMA_CPP",
+        "choices": [
+            {
+                "finish_reason": None,
+                "index": 0
+            }
+        ]
+    }
+    slot_id = data.get("slot_id")
+    if (chat):
+        if (start):
+            resData["choices"][0]["delta"] =  {
+                "role": "assistant"
+            }
+        else:
+            resData["choices"][0]["delta"] =  {
+                "content": data["content"]
+            }
+            if (data["stop"]):
+                resData["choices"][0]["finish_reason"] = "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+    else:
+        resData["choices"][0]["text"] = data["content"]
+        if (data["stop"]):
+            resData["choices"][0]["finish_reason"] = "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+
+    return resData
+
+
+@app.route('/chat/completions', methods=['POST', 'OPTIONS'])
+@app.route('/v1/chat/completions', methods=['POST', 'OPTIONS'])
+def chat_completions():
+    if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
+        return Response(status=403)
+    if request.method == 'OPTIONS':
+        return Response(headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+    body = request.get_json()
+    stream = False
+    tokenize = False
+    if(is_present(body, "stream")): stream = body["stream"]
+    if(is_present(body, "tokenize")): tokenize = body["tokenize"]
+    postData = make_postData(body, chat=True, stream=stream)
+
+    promptToken = []
+    if (tokenize):
+        tokenData = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/tokenize"), data=json.dumps({"content": postData["prompt"]})).json()
+        promptToken = tokenData["tokens"]
+
+    if (not stream):
+        data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData))
+        print(data.json())
+        resData = make_resData(data.json(), chat=True, promptToken=promptToken)
+        return jsonify(resData)
+    else:
+        def generate():
+            data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
+            time_now = int(time.time())
+            resData = make_resData_stream({}, chat=True, time_now=time_now, start=True)
+            yield 'data: {}\n\n'.format(json.dumps(resData))
+            for line in data.iter_lines():
+                if line:
+                    decoded_line = line.decode('utf-8')
+                    resData = make_resData_stream(json.loads(decoded_line[6:]), chat=True, time_now=time_now)
+                    yield 'data: {}\n\n'.format(json.dumps(resData))
+        return Response(generate(), mimetype='text/event-stream', headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+
+
+@app.route('/completions', methods=['POST', 'OPTIONS'])
+@app.route('/v1/completions', methods=['POST', 'OPTIONS'])
+def completion():
+    if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
+        return Response(status=403)
+    if request.method == 'OPTIONS':
+        return Response(headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+    body = request.get_json()
+    stream = False
+    tokenize = False
+    if(is_present(body, "stream")): stream = body["stream"]
+    if(is_present(body, "tokenize")): tokenize = body["tokenize"]
+    postData = make_postData(body, chat=False, stream=stream)
+
+    promptToken = []
+    if (tokenize):
+        tokenData = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/tokenize"), data=json.dumps({"content": postData["prompt"]})).json()
+        promptToken = tokenData["tokens"]
+
+    if (not stream):
+        data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData))
+        print(data.json())
+        resData = make_resData(data.json(), chat=False, promptToken=promptToken)
+        return jsonify(resData)
+    else:
+        def generate():
+            data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
+            time_now = int(time.time())
+            for line in data.iter_lines():
+                if line:
+                    decoded_line = line.decode('utf-8')
+                    resData = make_resData_stream(json.loads(decoded_line[6:]), chat=False, time_now=time_now)
+                    yield 'data: {}\n\n'.format(json.dumps(resData))
+        return Response(generate(), mimetype='text/event-stream', headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+
+if __name__ == '__main__':
+    app.run(args.host, port=args.port)

examples/server/bench/README.md

@@ -1,88 +0,0 @@
-### Server benchmark tools
-
-Benchmark is using [k6](https://k6.io/).
-
-##### Install k6
-
-Follow instruction from: https://k6.io/docs/get-started/installation/
-
-Example for ubuntu:
-```shell
-snap install k6
-```
-
-#### Download a dataset
-
-This dataset was originally proposed in [vLLM benchmarks](https://github.com/vllm-project/vllm/blob/main/benchmarks/README.md).
-
-```shell
-wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-```
-
-#### Download a model
-Example for PHI-2
-
-```shell
-../../../scripts/hf.sh --repo ggml-org/models --file phi-2/ggml-model-q4_0.gguf
-```
-
-#### Start the server
-The server must answer OAI Chat completion requests on `http://localhost:8080/v1` or according to the environment variable `SERVER_BENCH_URL`.
-
-Example:
-```shell
-server --host localhost --port 8080 \
-  --model ggml-model-q4_0.gguf \
-  --cont-batching \
-  --metrics \
-  --parallel 8 \
-  --batch-size 512 \
-  --ctx-size 4096 \
-  --log-format text \
-  -ngl 33
-```
-
-#### Run the benchmark
-
-For 500 chat completions request with 8 concurrent users during maximum 10 minutes, run:
-```shell
-k6 run script.js --duration 10m --iterations 500 --vus 8
-```
-
-The benchmark values can be overridden with:
-- `SERVER_BENCH_URL` server url prefix for chat completions, default `http://localhost:8080/v1`
-- `SERVER_BENCH_N_PROMPTS` total prompts to randomly select in the benchmark, default `480`
-- `SERVER_BENCH_MODEL_ALIAS` model alias to pass in the completion request, default `my-model`
-- `SERVER_BENCH_MAX_TOKENS` max tokens to predict, default: `512`
-- `SERVER_BENCH_DATASET` path to the benchmark dataset file
-- `SERVER_BENCH_MAX_PROMPT_TOKENS` maximum prompt tokens to filter out in the dataset: default `1024`
-- `SERVER_BENCH_MAX_CONTEXT` maximum context size of the completions request to filter out in the dataset: prompt + predicted tokens, default `2048`
-
-Note: the local tokenizer is just a string space split, real number of tokens will differ.
-
-Or with [k6 options](https://k6.io/docs/using-k6/k6-options/reference/):
-
-```shell
-SERVER_BENCH_N_PROMPTS=500 k6 run script.js --duration 10m --iterations 500 --vus 8
-```
-
-To [debug http request](https://k6.io/docs/using-k6/http-debugging/) use `--http-debug="full"`.
-
-#### Metrics
-
-Following metrics are available computed from the OAI chat completions response `usage`:
-- `llamacpp_tokens_second` Trend of `usage.total_tokens / request duration`
-- `llamacpp_prompt_tokens` Trend of `usage.prompt_tokens`
-- `llamacpp_prompt_tokens_total_counter` Counter of `usage.prompt_tokens`
-- `llamacpp_completion_tokens` Trend of `usage.completion_tokens`
-- `llamacpp_completion_tokens_total_counter` Counter of `usage.completion_tokens`
-- `llamacpp_completions_truncated_rate` Rate of completions truncated, i.e. if `finish_reason === 'length'`
-- `llamacpp_completions_stop_rate` Rate of completions stopped by the model, i.e. if `finish_reason === 'stop'`
-
-The script will fail if too many completions are truncated, see `llamacpp_completions_truncated_rate`.
-
-K6 metrics might be compared against [server metrics](../README.md), with:
-
-```shell
-curl http://localhost:8080/metrics
-```

examples/server/bench/script.js

@@ -1,120 +0,0 @@
-import http from 'k6/http'
-import {check, sleep} from 'k6'
-import {SharedArray} from 'k6/data'
-import {Counter, Rate, Trend} from 'k6/metrics'
-import exec from 'k6/execution';
-
-// Server chat completions prefix
-const server_url = __ENV.SERVER_BENCH_URL ? __ENV.SERVER_BENCH_URL : 'http://localhost:8080/v1'
-
-// Number of total prompts in the dataset - default 10m / 10 seconds/request * number of users
-const n_prompt = __ENV.SERVER_BENCH_N_PROMPTS ? parseInt(__ENV.SERVER_BENCH_N_PROMPTS) : 600 / 10 * 8
-
-// Model name to request
-const model = __ENV.SERVER_BENCH_MODEL_ALIAS ? __ENV.SERVER_BENCH_MODEL_ALIAS : 'my-model'
-
-// Dataset path
-const dataset_path = __ENV.SERVER_BENCH_DATASET ? __ENV.SERVER_BENCH_DATASET : './ShareGPT_V3_unfiltered_cleaned_split.json'
-
-// Max tokens to predict
-const max_tokens = __ENV.SERVER_BENCH_MAX_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_TOKENS) : 512
-
-// Max prompt tokens
-const n_prompt_tokens = __ENV.SERVER_BENCH_MAX_PROMPT_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_PROMPT_TOKENS) : 1024
-
-// Max slot context
-const n_ctx_slot = __ENV.SERVER_BENCH_MAX_CONTEXT ? parseInt(__ENV.SERVER_BENCH_MAX_CONTEXT) : 2048
-
-export function setup() {
-    console.info(`Benchmark config: server_url=${server_url} n_prompt=${n_prompt} model=${model} dataset_path=${dataset_path} max_tokens=${max_tokens}`)
-}
-
-const data = new SharedArray('conversations', function () {
-    const tokenizer = (message) => message.split(/[\s,'".?]/)
-
-    return JSON.parse(open(dataset_path))
-        // Filter out the conversations with less than 2 turns.
-        .filter(data => data["conversations"].length >= 2)
-        .filter(data => data["conversations"][0]["from"] === "human")
-        .map(data => {
-            return {
-                prompt: data["conversations"][0]["value"],
-                n_prompt_tokens: tokenizer(data["conversations"][0]["value"]).length,
-                n_completion_tokens: tokenizer(data["conversations"][1]["value"]).length,
-            }
-        })
-        // Filter out too short sequences
-        .filter(conv => conv.n_prompt_tokens >= 4 && conv.n_completion_tokens >= 4)
-        // Filter out too long sequences.
-        .filter(conv => conv.n_prompt_tokens <= n_prompt_tokens && conv.n_prompt_tokens + conv.n_completion_tokens <= n_ctx_slot)
-        // Keep only first n prompts
-        .slice(0, n_prompt)
-})
-
-const llamacpp_prompt_tokens = new Trend('llamacpp_prompt_tokens')
-const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens')
-const llamacpp_tokens_second = new Trend('llamacpp_tokens_second')
-
-const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter')
-const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter')
-
-const llamacpp_completions_truncated_rate = new Rate('llamacpp_completions_truncated_rate')
-const llamacpp_completions_stop_rate = new Rate('llamacpp_completions_stop_rate')
-
-export const options = {
-    thresholds: {
-        llamacpp_completions_truncated_rate: [
-            // more than 80% of truncated input will abort the test
-            {threshold: 'rate < 0.8', abortOnFail: true, delayAbortEval: '1m'},
-        ],
-    },
-    duration: '10m',
-    vus: 8,
-}
-
-export default function () {
-    const conversation = data[exec.scenario.iterationInInstance % data.length]
-    const payload = {
-        "messages": [
-            {
-                "role": "system",
-                "content": "You are ChatGPT, an AI assistant.",
-            },
-            {
-                "role": "user",
-                "content": conversation.prompt,
-            }
-        ],
-        "model": model,
-        "stream": false,
-        "max_tokens": max_tokens
-    }
-
-    const body = JSON.stringify(payload)
-
-    let res = http.post(`${server_url}/chat/completions`, body, {
-        headers: {'Content-Type': 'application/json'},
-        timeout: '300s'
-    })
-
-    check(res, {'success completion': (r) => r.status === 200})
-
-    if (res.status === 200) {
-        const completions = res.json()
-
-        llamacpp_prompt_tokens.add(completions.usage.prompt_tokens)
-        llamacpp_prompt_tokens_total_counter.add(completions.usage.prompt_tokens)
-
-        llamacpp_completion_tokens.add(completions.usage.completion_tokens)
-        llamacpp_completion_tokens_total_counter.add(completions.usage.completion_tokens)
-
-        llamacpp_completions_truncated_rate.add(completions.choices[0].finish_reason === 'length')
-        llamacpp_completions_stop_rate.add(completions.choices[0].finish_reason === 'stop')
-
-        llamacpp_tokens_second.add(completions.usage.total_tokens / res.timings.duration * 1.e3)
-    } else {
-        console.error(`response: ${res.body} request=${payload}`)
-    }
-
-    sleep(0.3)
-}

examples/server/completion.js.hpp

@@ -231,256 +231,255 @@ unsigned char completion_js[] = {
   0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
   0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72,
-  0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f,
-  0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e,
-  0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67,
-  0x65, 0x2e, 0x69, 0x6e, 0x63, 0x6c, 0x75, 0x64, 0x65, 0x73, 0x28, 0x27,
-  0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61, 0x69, 0x6c,
-  0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x3d,
+  0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x2e, 0x69, 0x6e, 0x63, 0x6c, 0x75, 0x64, 0x65, 0x73,
+  0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61,
+  0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x2f, 0x2f, 0x20, 0x54, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x61,
   0x6e, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x74, 0x6f, 0x20, 0x62,
   0x65, 0x20, 0x63, 0x61, 0x75, 0x67, 0x68, 0x74, 0x20, 0x62, 0x79, 0x20,
   0x75, 0x70, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x20, 0x63, 0x61, 0x6c,
   0x6c, 0x65, 0x72, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72,
-  0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72,
-  0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61,
-  0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72,
-  0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
-  0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x5b, 0x24, 0x7b, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e,
-  0x63, 0x6f, 0x64, 0x65, 0x7d, 0x20, 0x2d, 0x20, 0x24, 0x7b, 0x72, 0x65,
-  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x74,
-  0x79, 0x70, 0x65, 0x7d, 0x5d, 0x3a, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x6d, 0x65,
-  0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x65, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x2e, 0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20,
-  0x24, 0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72,
-  0x6f, 0x72, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77,
+  0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x27,
+  0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61, 0x69, 0x6c,
+  0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c,
+  0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x3a, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e,
+  0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72,
+  0x6f, 0x72, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61,
+  0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x3a, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e,
+  0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20,
-  0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61,
-  0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72,
-  0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20,
-  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65,
-  0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
-  0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20,
-  0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20,
-  0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x73,
-  0x63, 0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a,
-  0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70,
-  0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d,
-  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70,
-  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65,
-  0x72, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c,
-  0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f,
-  0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65,
-  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69,
-  0x6c, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f,
-  0x2f, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65,
-  0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
-  0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d,
-  0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
-  0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20,
-  0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
-  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77,
-  0x20, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
-  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63,
-  0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66,
-  0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
-  0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63,
-  0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
-  0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65,
-  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73,
-  0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e,
-  0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65,
-  0x6e, 0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22,
-  0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d,
-  0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
-  0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75,
-  0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67,
-  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64,
-  0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
-  0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
-  0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
-  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
-  0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61,
+  0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65,
+  0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x45,
+  0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20,
+  0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x65, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
+  0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a,
+  0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
+  0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x61,
+  0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x79, 0x6f,
+  0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x73, 0x63, 0x72,
+  0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f,
+  0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72,
+  0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65,
+  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f, 0x0a,
+  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28,
+  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28,
+  0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75,
+  0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x2e,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70,
+  0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
+  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b,
+  0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20,
+  0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
+  0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28,
+  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d,
+  0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72,
+  0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c,
+  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e,
+  0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
+  0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
+  0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74,
+  0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
   0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61,
   0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77,
   0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74,
-  0x28, 0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20,
+  0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20,
   0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d,
-  0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
-  0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75,
-  0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64,
-  0x6f, 0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61,
-  0x69, 0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
-  0x74, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d,
-  0x29, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c,
-  0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74,
-  0x75, 0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73,
-  0x65, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c,
-  0x76, 0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63,
-  0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78,
-  0x74, 0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73,
-  0x20, 0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74,
-  0x20, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65,
-  0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65,
-  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65,
-  0x6e, 0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20,
-  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77,
-  0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72,
-  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72,
-  0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72,
-  0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
-  0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72,
-  0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f,
-  0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20,
-  0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67,
-  0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77,
-  0x20, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79,
-  0x6e, 0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c,
-  0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69,
-  0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75,
-  0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28,
-  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e,
-  0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f,
-  0x6c, 0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63,
-  0x68, 0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74,
-  0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a,
-  0x0a, 0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70,
-  0x72, 0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f,
-  0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c,
-  0x65, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
-  0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c,
-  0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72,
+  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
+  0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
+  0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74,
+  0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65, 0x6e,
+  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74,
+  0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64,
+  0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69,
+  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20,
+  0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69,
+  0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
+  0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63,
+  0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43,
+  0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22,
+  0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20,
+  0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e,
+  0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
+  0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74,
+  0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f, 0x6e,
+  0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
+  0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
+  0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
+  0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b,
+  0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20,
+  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
+  0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20,
+  0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65,
+  0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d,
+  0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e,
+  0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20, 0x6e,
+  0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x73,
+  0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f, 0x0a,
+  0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a,
+  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28,
+  0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
+  0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a,
+  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d,
+  0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
+  0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74,
+  0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d,
+  0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
+  0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20,
+  0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d,
+  0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x50,
+  0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63,
+  0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20, 0x72,
+  0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
   0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61,
-  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c,
-  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63,
-  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29,
-  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c,
-  0x62, 0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b,
-  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47,
-  0x65, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c,
-  0x20, 0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74,
-  0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54,
-  0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75,
-  0x6c, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e,
-  0x67, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78,
-  0x74, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64,
-  0x20, 0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f,
-  0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20,
-  0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d,
-  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67,
-  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70,
-  0x73, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65,
-  0x74, 0x63, 0x68, 0x28, 0x22, 0x2f, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x22,
-  0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20,
-  0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
-  0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d,
-  0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x64, 0x65, 0x66, 0x61, 0x75,
-  0x6c, 0x74, 0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
-  0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
+  0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72,
+  0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
+  0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20,
+  0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x2f,
+  0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72, 0x65,
+  0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a, 0x65,
+  0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74,
+  0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
+  0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f,
+  0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61,
+  0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65, 0x74,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x69,
+  0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68, 0x69,
+  0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c, 0x20,
+  0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x20,
+  0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x20,
+  0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73,
+  0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20,
+  0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65, 0x6e,
+  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x20,
+  0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63,
+  0x68, 0x28, 0x22, 0x2f, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x22, 0x29, 0x2e,
+  0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20, 0x72, 0x2e,
+  0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
   0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
-  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x70,
+  0x72, 0x6f, 0x70, 0x73, 0x2e, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74,
+  0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67,
+  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
+  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
 };
-unsigned int completion_js_len = 5796;
+unsigned int completion_js_len = 5782;

examples/server/oai.hpp

@@ -0,0 +1,227 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+#include "utils.hpp"
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+
+using json = nlohmann::json;
+
+inline static json oaicompat_completion_params_parse(
+    const json &body, /* openai api json semantics */
+    const std::string &chat_template)
+{
+    json llama_params;
+    std::string formatted_prompt = chat_template == "chatml"
+        ? format_chatml(body["messages"])  // OpenAI 'messages' to chatml (with <|im_start|>,...)
+        : format_llama2(body["messages"]); // OpenAI 'messages' to llama2 (with [INST],...)
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
+    llama_params["prompt"]            = formatted_prompt;
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
+    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
+    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
+    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
+    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
+    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
+    llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
+    llama_params["seed"]              = json_value(body, "seed", LLAMA_DEFAULT_SEED);
+    llama_params["stream"]            = json_value(body, "stream", false);
+    llama_params["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
+    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
+
+    if (body.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body.contains("stop") && body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
+{
+    json result = response.result_json;
+
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res =
+        json{{"choices", choices},
+            {"created", t},
+            {"model",
+                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+            {"usage",
+                json{{"completion_tokens", num_tokens_predicted},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
+            {"id", gen_chatcmplid()}};
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
+    json result = response.result_json;
+
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({response.result_json});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word", false);
+    bool stopped_eos    = json_value(result, "stopped_eos", false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content", std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json{{"choices", choices},
+                    {"created", t},
+                    {"id", gen_chatcmplid()},
+                    {"model", modelname},
+                    {"object", "chat.completion.chunk"}};
+
+    return std::vector<json>({ret});
+}
+
+inline static json format_embeddings_response_oaicompat(const json &request, const json &embeddings)
+{
+    json res =
+        json{
+            {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", "list"},
+            {"usage",
+                json{{"prompt_tokens", 0},
+                     {"total_tokens", 0}}},
+            {"data", embeddings}
+        };
+    return res;
+}
+

examples/server/public/completion.js

@@ -96,18 +96,18 @@ export async function* llama(prompt, params = {}, config = {}) {
             }
           }
           if (result.error) {
-            try {
-              result.error = JSON.parse(result.error);
-              if (result.error.message.includes('slot unavailable')) {
-                // Throw an error to be caught by upstream callers
-                throw new Error('slot unavailable');
-              } else {
-                console.error(`llama.cpp error [${result.error.code} - ${result.error.type}]: ${result.error.message}`);
-              }
-            } catch(e) {
-              console.error(`llama.cpp error ${result.error}`)
+            result.error = JSON.parse(result.error);
+            if (result.error.content.includes('slot unavailable')) {
+              // Throw an error to be caught by upstream callers
+              throw new Error('slot unavailable');
+            } else {
+              console.error(`llama.cpp error: ${result.error.content}`);
             }
           }
+          if (result.error) {
+            result.error = JSON.parse(result.error);
+            console.error(`llama.cpp error: ${result.error.content}`);
+          }
         }
       }
     }

examples/server/tests/README.md

@@ -1,67 +0,0 @@
-# Server tests
-
-Python based server tests scenario using [BDD](https://en.wikipedia.org/wiki/Behavior-driven_development)
-and [behave](https://behave.readthedocs.io/en/latest/):
-
-* [issues.feature](./features/issues.feature) Pending issues scenario
-* [parallel.feature](./features/parallel.feature) Scenario involving multi slots and concurrent requests
-* [security.feature](./features/security.feature) Security, CORS and API Key
-* [server.feature](./features/server.feature) Server base scenario: completion, embedding, tokenization, etc...
-
-Tests target GitHub workflows job runners with 4 vCPU.
-
-Requests are
-using [aiohttp](https://docs.aiohttp.org/en/stable/client_reference.html), [asyncio](https://docs.python.org/fr/3/library/asyncio.html)
-based http client.
-
-Note: If the host architecture inference speed is faster than GitHub runners one, parallel scenario may randomly fail.
-To mitigate it, you can increase values in `n_predict`, `kv_size`.
-
-### Install dependencies
-
-`pip install -r requirements.txt`
-
-### Run tests
-
-1. Build the server
-
-```shell
-cd ../../..
-mkdir build
-cd build
-cmake ../
-cmake --build . --target server
-```
-
-2. Start the test: `./tests.sh`
-
-It's possible to override some scenario steps values with environment variables:
-
-| variable                 | description                                                                                    |
-|--------------------------|------------------------------------------------------------------------------------------------|
-| `PORT`                   | `context.server_port` to set the listening port of the server during scenario, default: `8080` |
-| `LLAMA_SERVER_BIN_PATH`  | to change the server binary path, default: `../../../build/bin/server`                         |
-| `DEBUG`                  | "ON" to enable steps and server verbose mode `--verbose`                                       |
-| `SERVER_LOG_FORMAT_JSON` | if set switch server logs to json format                                                       |
-| `N_GPU_LAYERS`           | number of model layers to offload to VRAM `-ngl --n-gpu-layers`                                |
-
-### Run @bug, @wip or @wrong_usage annotated scenario
-
-Feature or Scenario must be annotated with `@llama.cpp` to be included in the default scope.
-
-- `@bug` annotation aims to link a scenario with a GitHub issue.
-- `@wrong_usage` are meant to show user issue that are actually an expected behavior
-- `@wip` to focus on a scenario working in progress
-- `@slow` heavy test, disabled by default
-
-To run a scenario annotated with `@bug`, start:
-
-```shell
-DEBUG=ON ./tests.sh --no-skipped --tags bug
-```
-
-After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.
-
-```shell
-./tests.sh --no-skipped --tags bug,wrong_usage || echo "should failed but compile"
-```

examples/server/tests/features/embeddings.feature

@@ -1,95 +0,0 @@
-@llama.cpp
-@embeddings
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models
-    And   a model alias bert-bge-small
-    And   42 as server seed
-    And   2 slots
-    And   1024 as batch size
-    And   1024 as ubatch size
-    And   2048 KV cache size
-    And   embeddings extraction
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Embedding
-    When embeddings are computed for:
-    """
-    What is the capital of Bulgaria ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility
-    Given a model bert-bge-small
-    When an OAI compatible embeddings computation request for:
-    """
-    What is the capital of Spain ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility with multiple inputs
-    Given a model bert-bge-small
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    When an OAI compatible embeddings computation request for multiple inputs
-    Then embeddings are generated
-
-  Scenario: Multi users embeddings
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    Given concurrent embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: Multi users OAI compatibility embeddings
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    And a prompt:
-      """
-      What is the biggest US city ?
-      """
-    And a prompt:
-      """
-      What is the capital of Bulgaria ?
-      """
-    And   a model bert-bge-small
-    Given concurrent OAI embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: All embeddings should be the same
-    Given 10 fixed prompts
-    And   a model bert-bge-small
-    Given concurrent OAI embedding requests
-    Then all embeddings are the same

examples/server/tests/features/environment.py

@@ -1,100 +0,0 @@
-import errno
-import os
-import socket
-import subprocess
-import time
-from contextlib import closing
-import signal
-
-
-def before_scenario(context, scenario):
-    context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
-    if context.debug:
-        print("DEBUG=ON\n")
-    print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m\n")
-    port = 8080
-    if 'PORT' in os.environ:
-        port = int(os.environ['PORT'])
-    if is_server_listening("localhost", port):
-        assert False, "Server already started"
-
-
-def after_scenario(context, scenario):
-    if context.server_process is None:
-        return
-    if scenario.status == "failed":
-        if 'GITHUB_ACTIONS' in os.environ:
-            print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
-            if os.path.isfile('llama.log'):
-                with closing(open('llama.log', 'r')) as f:
-                    for line in f:
-                        print(line)
-        if not is_server_listening(context.server_fqdn, context.server_port):
-            print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")
-
-    if not pid_exists(context.server_process.pid):
-        assert False, f"Server not running pid={context.server_process.pid} ..."
-
-    server_graceful_shutdown(context)
-
-    # Wait few for socket to free up
-    time.sleep(0.05)
-
-    attempts = 0
-    while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
-        server_kill(context)
-        time.sleep(0.1)
-        attempts += 1
-        if attempts > 5:
-            server_kill_hard(context)
-
-
-def server_graceful_shutdown(context):
-    print(f"shutting down server pid={context.server_process.pid} ...\n")
-    if os.name == 'nt':
-        os.kill(context.server_process.pid, signal.CTRL_C_EVENT)
-    else:
-        os.kill(context.server_process.pid, signal.SIGINT)
-
-
-def server_kill(context):
-    print(f"killing server pid={context.server_process.pid} ...\n")
-    context.server_process.kill()
-
-
-def server_kill_hard(context):
-    pid = context.server_process.pid
-    path = context.server_path
-
-    print(f"Server dangling exits, hard killing force {pid}={path}...\n")
-    if os.name == 'nt':
-        process = subprocess.check_output(['taskkill', '/F', '/pid', str(pid)]).decode()
-        print(process)
-    else:
-        os.kill(-pid, signal.SIGKILL)
-
-
-def is_server_listening(server_fqdn, server_port):
-    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-        result = sock.connect_ex((server_fqdn, server_port))
-        _is_server_listening = result == 0
-        if _is_server_listening:
-            print(f"server is listening on {server_fqdn}:{server_port}...\n")
-        return _is_server_listening
-
-
-def pid_exists(pid):
-    """Check whether pid exists in the current process table."""
-    if pid < 0:
-        return False
-    if os.name == 'nt':
-        output = subprocess.check_output(['TASKLIST', '/FI', f'pid eq {pid}']).decode()
-        print(output)
-        return "No tasks are running" not in output
-    else:
-        try:
-            os.kill(pid, 0)
-        except OSError as e:
-            return e.errno == errno.EPERM
-        else:
-            return True

examples/server/tests/features/issues.feature

@@ -1,5 +0,0 @@
-# List of ongoing issues
-# run with: DEBUG=ON ./tests.sh --no-skipped --tags bug
-@bug
-Feature: Issues
-  # No confirmed issue at the moment

examples/server/tests/features/parallel.feature

@@ -1,101 +0,0 @@
-@llama.cpp
-@parallel
-Feature: Parallel
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   42 as server seed
-    And   128 as batch size
-    And   256 KV cache size
-    And   2 slots
-    And   continuous batching
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Multi users completion
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And <n_predict> max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | n_predict |
-      | 128       |
-
-  Scenario Outline: Multi users OAI completions compatibility
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-  Scenario Outline: Multi users OAI completions compatibility no v1
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests no v1
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-
-  Scenario:  Multi users with total number of tokens to predict exceeds the KV Cache size #3969
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    And 128 max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted

examples/server/tests/features/passkey.feature

@@ -1,55 +0,0 @@
-# run with: ./tests.sh --no-skipped --tags passkey
-@passkey
-@slow
-Feature: Passkey / Self-extend with context shift
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-
-  # Generates a long text of junk and inserts a secret passkey number inside it.
-  # Then we query the LLM for the secret passkey.
-  # see #3856 and #4810
-  Scenario Outline: Passkey
-    Given a model file <hf_file> from HF repo <hf_repo>
-    And   <n_batch> as batch size
-    And   <n_junk> as number of junk
-    And   <n_predicted> server max tokens to predict
-    And   42 as seed
-    And   <n_ctx> KV cache size
-    And   1 slots
-    And   <n_ga> group attention factor to extend context size through self-extend
-    And   <n_ga_w> group attention width to extend context size through self-extend
-    # Can be override with N_GPU_LAYERS
-    And   <ngl> GPU offloaded layers
-    Then  the server is starting
-    Then  the server is healthy
-    Given available models
-    Then  model 0 is trained on <n_ctx_train> tokens context
-    Given a prefix prompt:
-    """
-    here is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.
-    """
-    And a passkey prompt template:
-    """
-    The pass key is <passkey> Remember it. <passkey> is the pass key.
-    """
-    And a junk suffix prompt:
-    """
-    The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.
-    """
-    And a suffix prompt:
-    """
-    What is the pass key? The pass key is
-    """
-    Given a "<passkey>" passkey challenge prompt with the passkey inserted every <i_pos> junk
-    And  a completion request with no api error
-    Then <n_predicted> tokens are predicted matching <re_content>
-
-    Examples:
-      | hf_repo                         | hf_file                     | n_ctx_train | ngl | n_ctx | n_batch | n_ga | n_ga_w | n_junk | i_pos | passkey | n_predicted | re_content     |
-      | TheBloke/phi-2-GGUF             | phi-2.Q4_K_M.gguf           | 2048        | 5   | 8192  | 512     | 4    | 512    | 250    | 50    | 42      | 1           | 42             |
-      | TheBloke/phi-2-GGUF             | phi-2.Q4_K_M.gguf           | 2048        | 5   | 8192  | 512     | 2    | 512    | 250    | 50    | 42      | 1           | \b((?!42)\w)+\b  |
-      #| TheBloke/Llama-2-7B-GGUF        | llama-2-7b.Q2_K.gguf        | 4096        | 3   | 16384 | 512     | 4    | 512    | 500    | 300   | 1234    | 5           | 1234           |
-      #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768       | 2   | 16384 | 512     | 4    | 512    | 500    | 100   | 0987    | 5           | 0
-      # 987           |
-

examples/server/tests/features/security.feature

@@ -1,52 +0,0 @@
-@llama.cpp
-@security
-Feature: Security
-
-  Background: Server startup with an api key defined
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a server api key llama.cpp
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Completion with some user api key
-    Given a prompt test
-    And   a user api key <api_key>
-    And   4 max tokens to predict
-    And   a completion request with <api_error> api error
-
-    Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackeme   | raised    |
-      |           | raised    |
-
-  Scenario Outline: OAI Compatibility
-    Given a system prompt test
-    And   a user prompt test
-    And   a model test
-    And   2 max tokens to predict
-    And   streaming is disabled
-    And   a user api key <api_key>
-    Given an OAI compatible chat completions request with <api_error> api error
-
-    Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackme    | raised    |
-
-
-  Scenario Outline: CORS Options
-    Given a user api key llama.cpp
-    When  an OPTIONS request is sent from <origin>
-    Then  CORS header <cors_header> is set to <cors_header_value>
-
-    Examples: Headers
-      | origin          | cors_header                      | cors_header_value |
-      | localhost       | Access-Control-Allow-Origin      | localhost         |
-      | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
-      | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | POST              |
-      | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |