increase igpu cluster limit (#6159)

* gguf-split: split and merge gguf files per tensor

* gguf-split: build with make toolchain

* gguf-split: rename `--split-tensors-size` to `--split-max-tensors`. Set general.split_count KV to all split

* split : minor style + fix compile warnings

* gguf-split: remove --upload not implemented

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.devops/main-intel.Dockerfile

@@ -1,8 +1,8 @@
 ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
-ARG UBUNTU_VERSION=22.04
 
-FROM intel/hpckit:$ONEAPI_VERSION as build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
+ARG LLAMA_SYCL_F16=OFF
 RUN apt-get update && \
     apt-get install -y git
 
@@ -10,16 +10,18 @@ WORKDIR /app
 
 COPY . .
 
-# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
 RUN mkdir build && \
     cd build && \
-    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
-    cmake --build . --config Release --target main server
+    if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build . --config Release --target main
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
 COPY --from=build /app/build/bin/main /main
-COPY --from=build /app/build/bin/server /server
 
 ENV LC_ALL=C.utf8
 

.devops/main-vulkan.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake --build . --config Release --target main
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/main /main && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

.devops/nix/docker.nix

@@ -0,0 +1,37 @@
+{
+  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

@@ -1,5 +1,6 @@
 {
   lib,
+  glibc,
   config,
   stdenv,
   mkShell,
@@ -13,19 +14,28 @@
   cudaPackages,
   darwin,
   rocmPackages,
+  vulkan-headers,
+  vulkan-loader,
   clblast,
   useBlas ? builtins.all (x: !x) [
     useCuda
     useMetalKit
     useOpenCL
     useRocm
+    useVulkan
   ],
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
   useMpi ? false, # Increases the runtime closure size by ~700M
   useOpenCL ? false,
   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
@@ -37,10 +47,7 @@ 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" ]
@@ -48,7 +55,8 @@ let
     ++ lib.optionals useMetalKit [ "MetalKit" ]
     ++ lib.optionals useMpi [ "MPI" ]
     ++ lib.optionals useOpenCL [ "OpenCL" ]
-    ++ lib.optionals useRocm [ "ROCm" ];
+    ++ lib.optionals useRocm [ "ROCm" ]
+    ++ lib.optionals useVulkan [ "Vulkan" ];
 
   pnameSuffix =
     strings.optionalString (suffices != [ ])
@@ -108,6 +116,11 @@ let
     hipblas
     rocblas
   ];
+
+  vulkanBuildInputs = [
+    vulkan-headers
+    vulkan-loader
+  ];
 in
 
 effectiveStdenv.mkDerivation (
@@ -157,6 +170,9 @@ 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 =
@@ -164,13 +180,14 @@ effectiveStdenv.mkDerivation (
       ++ optionals useCuda cudaBuildInputs
       ++ optionals useMpi [ mpi ]
       ++ optionals useOpenCL [ clblast ]
-      ++ optionals useRocm rocmBuildInputs;
+      ++ optionals useRocm rocmBuildInputs
+      ++ optionals useVulkan vulkanBuildInputs;
 
     cmakeFlags =
       [
         (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" true)
+        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
@@ -178,6 +195,8 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
+        (cmakeBool "LLAMA_VULKAN" useVulkan)
+        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -218,6 +237,7 @@ effectiveStdenv.mkDerivation (
         useMpi
         useOpenCL
         useRocm
+        useVulkan
         ;
 
       shell = mkShell {

.devops/nix/scope.nix

@@ -12,5 +12,8 @@ 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

@@ -0,0 +1,27 @@
+{
+  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;
+}

.devops/server-intel.Dockerfile

@@ -1,8 +1,8 @@
 ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
-ARG UBUNTU_VERSION=22.04
 
-FROM intel/hpckit:$ONEAPI_VERSION as build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
+ARG LLAMA_SYCL_F16=OFF
 RUN apt-get update && \
     apt-get install -y git
 
@@ -10,13 +10,16 @@ WORKDIR /app
 
 COPY . .
 
-# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
 RUN mkdir build && \
     cd build && \
-    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
-    cmake --build . --config Release --target main server
+    if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build . --config Release --target server
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
 COPY --from=build /app/build/bin/server /server
 

.devops/server-vulkan.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake --build . --config Release --target server
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/server /server && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.flake8

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

.github/ISSUE_TEMPLATE/bug.md

@@ -7,3 +7,5 @@ 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

@@ -37,6 +37,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           CC=gcc-8 make -j $(nproc)
 
@@ -46,6 +48,28 @@ jobs:
           CC=gcc-8 make tests -j $(nproc)
           make test -j $(nproc)
 
+  ubuntu-focal-make-curl:
+    runs-on: ubuntu-20.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
+
+      - name: Build
+        id: make_build
+        env:
+          LLAMA_FATAL_WARNINGS: 1
+          LLAMA_CURL: 1
+        run: |
+          CC=gcc-8 make -j $(nproc)
+
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
 
@@ -65,7 +89,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -74,40 +98,40 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-  ubuntu-latest-cmake-sanitizer:
-    runs-on: ubuntu-latest
-
-    continue-on-error: true
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
+#  ubuntu-latest-cmake-sanitizer:
+#    runs-on: ubuntu-latest
+#
+#    continue-on-error: true
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#        build_type: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v3
+#
+#      - name: Dependencies
+#        id: depends
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          mkdir build
+#          cd build
+#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+#
+#      - name: Test
+#        id: cmake_test
+#        run: |
+#          cd build
+#          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -143,6 +167,28 @@ 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
 
@@ -184,6 +230,47 @@ jobs:
           cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
           cmake --build . --config Release -j $(nproc)
 
+  ubuntu-22-cmake-sycl-fp16:
+    runs-on: ubuntu-22.04
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON ..
+          cmake --build . --config Release -j $(nproc)
+
   # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
   #       how to debug it.
   #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
@@ -203,6 +290,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
@@ -236,7 +325,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_METAL=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -266,6 +355,7 @@ 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 \
@@ -294,6 +384,7 @@ 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 \
@@ -356,6 +447,10 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute'
             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
@@ -406,7 +501,7 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute' }}
+        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
@@ -451,7 +546,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -565,6 +660,31 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
+  windows-latest-cmake-sycl:
+    runs-on: windows-latest
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
+
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
+      - name: Install
+        run:  scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Build
+        id: cmake_build
+        run:  examples/sycl/win-build-sycl.bat
+
   ios-xcode-build:
     runs-on: macos-latest
 
@@ -597,8 +717,7 @@ jobs:
         run: |
           cd examples/llama.android
 
-          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
-          ./gradlew build --no-daemon -Pskip-armeabi-v7a
+          ./gradlew build --no-daemon
 
 #  freeBSD-latest:
 #    runs-on: macos-12

.github/workflows/close-issue.yml

@@ -0,0 +1,23 @@
+name: Close inactive issues
+on:
+  schedule:
+    - cron: "42 0 * * *"
+
+jobs:
+  close-issues:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v5
+        with:
+          exempt-issue-labels: "refactor,help wanted,good first issue,research"
+          days-before-issue-stale: 30
+          days-before-issue-close: 14
+          stale-issue-label: "stale"
+          close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale."
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
+          operations-per-run: 1000
+          repo-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/editorconfig.yml

@@ -1,6 +1,12 @@
 name: EditorConfig Checker
 
 on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
   push:
     branches:
       - master

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

@@ -19,7 +19,6 @@ on:
 
 jobs:
   nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -37,8 +36,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -46,7 +45,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - 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: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,7 +37,6 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
-    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -51,8 +50,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -60,7 +59,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

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

@@ -3,12 +3,14 @@ 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'
@@ -26,4 +28,4 @@ jobs:
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script
-        run:  bash scripts/check-requirements.sh nocleanup
+        run:  bash scripts/check-requirements.sh

.github/workflows/python-lint.yml

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

.github/workflows/server.yml

@@ -0,0 +1,151 @@
+# 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:
+        # TODO: temporary disabled due to linux kernel issues
+        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [UNDEFINED]
+        build_type: [Debug]
+        include:
+          - build_type: Release
+            sanitizer: ""
+            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 \
+            libcurl4-openssl-dev
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. \
+              -DLLAMA_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=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: libCURL
+        id: get_libcurl
+        env:
+          CURL_VERSION: 8.6.0_6
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
+          mkdir $env:RUNNER_TEMP/libcurl
+          tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
+          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: Copy Libcurl
+        id: prepare_libcurl
+        run: |
+          cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
+
+      - 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

@@ -1,20 +0,0 @@
-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

@@ -1,23 +0,0 @@
-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

@@ -23,11 +23,15 @@
 .clang-tidy
 .vs/
 .vscode/
+.idea/
+
+ggml-metal-embed.metal
 
 lcov-report/
 gcovr-report/
 
 build*
+cmake-build-*
 out/
 tmp/
 
@@ -43,6 +47,7 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/gritlm
 /imatrix
 /infill
 /libllama.so
@@ -89,3 +94,4 @@ examples/jeopardy/results.txt
 
 poetry.lock
 poetry.toml
+nppBackup

CMakeLists.txt

@@ -55,6 +55,9 @@ option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"
 option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
 option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
+# build
+option(LLAMA_FATAL_WARNINGS             "llama: enable -Werror flag"                            OFF)
+
 # sanitizers
 option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
 option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
@@ -79,7 +82,7 @@ if (NOT MSVC)
 endif()
 
 if (WIN32)
-    option(LLAMA_WIN_VER                     "llama: Windows Version"                           0x602)
+    set(LLAMA_WIN_VER "0x602" CACHE STRING "llama: Windows Version")
 endif()
 
 # 3rd party libs
@@ -96,29 +99,34 @@ option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
+option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
+option(LLAMA_VULKAN_CHECK_RESULTS            "llama: run Vulkan op checks"                      OFF)
+option(LLAMA_VULKAN_DEBUG                    "llama: enable Vulkan debug output"                OFF)
+option(LLAMA_VULKAN_VALIDATE                 "llama: enable Vulkan validation"                  OFF)
+option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"                          OFF)
 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})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
-
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
-if (LLAMA_PERF)
-    add_definitions(-DGGML_PERF)
-endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -126,6 +134,7 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
+
 if (LLAMA_SYCL)
     set(CMAKE_CXX_STANDARD 17)
 else()
@@ -136,9 +145,12 @@ set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
+
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -147,17 +159,17 @@ endif()
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
-        link_libraries(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
     endif()
 
     if (LLAMA_SANITIZE_ADDRESS)
         add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries(-fsanitize=address)
+        link_libraries     (-fsanitize=address)
     endif()
 
     if (LLAMA_SANITIZE_UNDEFINED)
         add_compile_options(-fsanitize=undefined)
-        link_libraries(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
     endif()
 endif()
 
@@ -189,40 +201,70 @@ if (LLAMA_METAL)
         add_compile_definitions(GGML_METAL_NDEBUG)
     endif()
 
-    # get full path to the file
-    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
-    # copy ggml-metal.metal to bin directory
+    # copy ggml-common.h and ggml-metal.metal to bin directory
+    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
-    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 ${XC_FLAGS} -DQK_K=64)
+    if (LLAMA_METAL_EMBED_LIBRARY)
+        enable_language(ASM)
+        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
+
+        set(METALLIB_COMMON "${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)
         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
-            DEPENDS ggml-metal.metal
+            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
             COMMENT "Compiling Metal kernels"
-        )
+            )
 
         add_custom_target(
             ggml-metal ALL
             DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
-    endif()
+            )
+    endif() # LLAMA_METAL_EMBED_LIBRARY
 
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
@@ -294,14 +336,17 @@ if (LLAMA_BLAS)
         endif()
 
         message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
         add_compile_options(${BLAS_LINKER_FLAGS})
+
         add_compile_definitions(GGML_USE_OPENBLAS)
+
         if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
             add_compile_definitions(GGML_BLAS_USE_MKL)
         endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
+        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -326,9 +371,6 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
         add_compile_definitions(GGML_USE_CUBLAS)
-#        if (LLAMA_CUDA_CUBLAS)
-#            add_compile_definitions(GGML_CUDA_CUBLAS)
-#        endif()
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -383,15 +425,20 @@ if (LLAMA_MPI)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
+
         set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
+        set(GGML_SOURCES_MPI ggml-mpi.c)
+
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
+
         if (NOT MSVC)
             add_compile_options(-Wno-cast-qual)
         endif()
+
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
+
         # Even if you're only using the C header, C++ programs may bring in MPI
         # C++ functions, so more linkage is needed
         if (MPI_CXX_FOUND)
@@ -426,15 +473,25 @@ if (LLAMA_VULKAN)
         set(GGML_HEADERS_VULKAN ggml-vulkan.h)
         set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
 
-        add_library(ggml-vulkan STATIC ggml-vulkan.cpp ggml-vulkan.h)
-        if (BUILD_SHARED_LIBS)
-            set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
-        endif()
-        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
-
         add_compile_definitions(GGML_USE_VULKAN)
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+        if (LLAMA_VULKAN_CHECK_RESULTS)
+            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+        endif()
+
+        if (LLAMA_VULKAN_DEBUG)
+            add_compile_definitions(GGML_VULKAN_DEBUG)
+        endif()
+
+        if (LLAMA_VULKAN_VALIDATE)
+            add_compile_definitions(GGML_VULKAN_VALIDATE)
+        endif()
+
+        if (LLAMA_VULKAN_RUN_TESTS)
+            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+        endif()
+
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -446,56 +503,66 @@ if (LLAMA_HIPBLAS)
     if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
     endif()
+
     if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
     endif()
 
-    find_package(hip)
-    find_package(hipblas)
-    find_package(rocblas)
+    find_package(hip     REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
 
-    if (${hipblas_FOUND} AND ${hip_FOUND})
-        message(STATUS "HIP and hipBLAS found")
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
-        if (LLAMA_HIP_UMA)
-            add_compile_definitions(GGML_HIP_UMA)
-        endif()
-        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
-        if (BUILD_SHARED_LIBS)
-            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
-        endif()
-        if (LLAMA_CUDA_FORCE_DMMV)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
-        endif()
-        if (LLAMA_CUDA_FORCE_MMQ)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
-        endif()
-        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
-        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+    message(STATUS "HIP and hipBLAS found")
 
-        if (LLAMA_STATIC)
-            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-        endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
-    else()
-        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    set(GGML_HEADERS_ROCM ggml-cuda.h)
+    set(GGML_SOURCES_ROCM ggml-cuda.cu)
+
+    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+
+    if (LLAMA_HIP_UMA)
+        add_compile_definitions(GGML_HIP_UMA)
     endif()
+
+    if (LLAMA_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (LLAMA_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+
+    set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+
+    if (LLAMA_STATIC)
+        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+    endif()
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
+    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()
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
+
+    message(STATUS "SYCL found")
+
+    add_compile_definitions(GGML_USE_SYCL)
+
     if (LLAMA_SYCL_F16)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
-    add_compile_definitions(GGML_USE_SYCL)
 
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
@@ -503,11 +570,22 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${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.h ggml-sycl.h)
+    set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+    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()
+    endif()
 endif()
 
 if (LLAMA_KOMPUTE)
@@ -519,61 +597,61 @@ if (LLAMA_KOMPUTE)
     endif()
 
     function(compile_shader)
-      set(options)
-      set(oneValueArgs)
-      set(multiValueArgs SOURCES)
-      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-      foreach(source ${compile_shader_SOURCES})
-        get_filename_component(filename ${source} NAME)
-        set(spv_file ${filename}.spv)
-        add_custom_command(
-            OUTPUT ${spv_file}
-            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-              COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            COMMENT "Compiling ${source} to ${spv_file}"
-        )
+        set(options)
+        set(oneValueArgs)
+        set(multiValueArgs SOURCES)
+        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+        foreach(source ${compile_shader_SOURCES})
+            get_filename_component(filename ${source} NAME)
+            set(spv_file ${filename}.spv)
+            add_custom_command(
+                OUTPUT ${spv_file}
+                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                COMMENT "Compiling ${source} to ${spv_file}"
+                )
 
-        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-        set(FILE_NAME "shader${RAW_FILE_NAME}")
-        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        if(CMAKE_GENERATOR MATCHES "Visual Studio")
-            add_custom_command(
-              OUTPUT ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              DEPENDS ${spv_file} xxd
-              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-            )
-        else()
-            add_custom_command(
-              OUTPUT ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              DEPENDS ${spv_file} xxd
-              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-            )
-        endif()
-      endforeach()
+            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+            set(FILE_NAME "shader${RAW_FILE_NAME}")
+            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+            if(CMAKE_GENERATOR MATCHES "Visual Studio")
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                    )
+            else()
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                    )
+            endif()
+        endforeach()
     endfunction()
 
     if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
@@ -583,66 +661,66 @@ if (LLAMA_KOMPUTE)
 
         # Compile our shaders
         compile_shader(SOURCES
-          kompute-shaders/op_scale.comp
-          kompute-shaders/op_scale_8.comp
-          kompute-shaders/op_add.comp
-          kompute-shaders/op_addrow.comp
-          kompute-shaders/op_mul.comp
-          kompute-shaders/op_silu.comp
-          kompute-shaders/op_relu.comp
-          kompute-shaders/op_gelu.comp
-          kompute-shaders/op_softmax.comp
-          kompute-shaders/op_norm.comp
-          kompute-shaders/op_rmsnorm.comp
-          kompute-shaders/op_diagmask.comp
-          kompute-shaders/op_mul_mat_mat_f32.comp
-          kompute-shaders/op_mul_mat_f16.comp
-          kompute-shaders/op_mul_mat_q8_0.comp
-          kompute-shaders/op_mul_mat_q4_0.comp
-          kompute-shaders/op_mul_mat_q4_1.comp
-          kompute-shaders/op_mul_mat_q6_k.comp
-          kompute-shaders/op_getrows_f16.comp
-          kompute-shaders/op_getrows_q4_0.comp
-          kompute-shaders/op_getrows_q4_1.comp
-          kompute-shaders/op_getrows_q6_k.comp
-          kompute-shaders/op_rope_f16.comp
-          kompute-shaders/op_rope_f32.comp
-          kompute-shaders/op_cpy_f16_f16.comp
-          kompute-shaders/op_cpy_f16_f32.comp
-          kompute-shaders/op_cpy_f32_f16.comp
-          kompute-shaders/op_cpy_f32_f32.comp
+            kompute-shaders/op_scale.comp
+            kompute-shaders/op_scale_8.comp
+            kompute-shaders/op_add.comp
+            kompute-shaders/op_addrow.comp
+            kompute-shaders/op_mul.comp
+            kompute-shaders/op_silu.comp
+            kompute-shaders/op_relu.comp
+            kompute-shaders/op_gelu.comp
+            kompute-shaders/op_softmax.comp
+            kompute-shaders/op_norm.comp
+            kompute-shaders/op_rmsnorm.comp
+            kompute-shaders/op_diagmask.comp
+            kompute-shaders/op_mul_mat_mat_f32.comp
+            kompute-shaders/op_mul_mat_f16.comp
+            kompute-shaders/op_mul_mat_q8_0.comp
+            kompute-shaders/op_mul_mat_q4_0.comp
+            kompute-shaders/op_mul_mat_q4_1.comp
+            kompute-shaders/op_mul_mat_q6_k.comp
+            kompute-shaders/op_getrows_f16.comp
+            kompute-shaders/op_getrows_q4_0.comp
+            kompute-shaders/op_getrows_q4_1.comp
+            kompute-shaders/op_getrows_q6_k.comp
+            kompute-shaders/op_rope_f16.comp
+            kompute-shaders/op_rope_f32.comp
+            kompute-shaders/op_cpy_f16_f16.comp
+            kompute-shaders/op_cpy_f16_f32.comp
+            kompute-shaders/op_cpy_f32_f16.comp
+            kompute-shaders/op_cpy_f32_f32.comp
         )
 
         # Create a custom target for our generated shaders
         add_custom_target(generated_shaders DEPENDS
-          shaderop_scale.h
-          shaderop_scale_8.h
-          shaderop_add.h
-          shaderop_addrow.h
-          shaderop_mul.h
-          shaderop_silu.h
-          shaderop_relu.h
-          shaderop_gelu.h
-          shaderop_softmax.h
-          shaderop_norm.h
-          shaderop_rmsnorm.h
-          shaderop_diagmask.h
-          shaderop_mul_mat_mat_f32.h
-          shaderop_mul_mat_f16.h
-          shaderop_mul_mat_q8_0.h
-          shaderop_mul_mat_q4_0.h
-          shaderop_mul_mat_q4_1.h
-          shaderop_mul_mat_q6_k.h
-          shaderop_getrows_f16.h
-          shaderop_getrows_q4_0.h
-          shaderop_getrows_q4_1.h
-          shaderop_getrows_q6_k.h
-          shaderop_rope_f16.h
-          shaderop_rope_f32.h
-          shaderop_cpy_f16_f16.h
-          shaderop_cpy_f16_f32.h
-          shaderop_cpy_f32_f16.h
-          shaderop_cpy_f32_f32.h
+            shaderop_scale.h
+            shaderop_scale_8.h
+            shaderop_add.h
+            shaderop_addrow.h
+            shaderop_mul.h
+            shaderop_silu.h
+            shaderop_relu.h
+            shaderop_gelu.h
+            shaderop_softmax.h
+            shaderop_norm.h
+            shaderop_rmsnorm.h
+            shaderop_diagmask.h
+            shaderop_mul_mat_mat_f32.h
+            shaderop_mul_mat_f16.h
+            shaderop_mul_mat_q8_0.h
+            shaderop_mul_mat_q4_0.h
+            shaderop_mul_mat_q4_1.h
+            shaderop_mul_mat_q6_k.h
+            shaderop_getrows_f16.h
+            shaderop_getrows_q4_0.h
+            shaderop_getrows_q4_1.h
+            shaderop_getrows_q6_k.h
+            shaderop_rope_f16.h
+            shaderop_rope_f32.h
+            shaderop_cpy_f16_f16.h
+            shaderop_cpy_f16_f32.h
+            shaderop_cpy_f32_f16.h
+            shaderop_cpy_f32_f32.h
         )
 
         # Create a custom command that depends on the generated_shaders
@@ -655,8 +733,10 @@ if (LLAMA_KOMPUTE)
 
         # Add the stamp to the main sources to ensure dependency tracking
         set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        set(GGML_HEADERS_KOMPUTE ggml-kompute.h   ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+
         add_compile_definitions(GGML_USE_KOMPUTE)
+
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
     else()
@@ -664,6 +744,18 @@ if (LLAMA_KOMPUTE)
     endif()
 endif()
 
+if (LLAMA_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
+
+if (LLAMA_PERF)
+    add_compile_definitions(GGML_PERF)
+endif()
+
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -688,28 +780,30 @@ 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)
-        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 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(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
-        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+        list(APPEND C_FLAGS   ${WARNING_FLAGS})
+        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -725,9 +819,10 @@ endif()
 set(CUDA_CXX_FLAGS "")
 
 if (LLAMA_CUBLAS)
-    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
-    if (NOT MSVC)
-        list(APPEND CUDA_FLAGS -Wno-pedantic)
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (LLAMA_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
     endif()
 
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
@@ -761,7 +856,11 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+        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)
     endif()
 endif()
 
@@ -788,9 +887,9 @@ if (LLAMA_CCACHE)
     if (LLAMA_CCACHE_FOUND)
         set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
         set(ENV{CCACHE_SLOPPINESS} time_macros)
-        message(STATUS "Using ccache")
+        message(STATUS "ccache found, compilation results will be cached. Disable with LLAMA_CCACHE=OFF.")
     else()
-        message(STATUS "Warning: ccache not found - consider installing it or use LLAMA_CCACHE=OFF")
+        message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with LLAMA_CCACHE=OFF")
     endif ()
 endif()
 
@@ -800,6 +899,7 @@ execute_process(
     ERROR_VARIABLE output
     OUTPUT_QUIET
 )
+
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
 endif()
@@ -809,10 +909,10 @@ endif()
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 if (MSVC)
-  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
-  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
 else ()
-  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+    set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
 
 if (NOT MSVC)
@@ -829,14 +929,26 @@ endif()
 
 set(ARCH_FLAGS "")
 
-if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
+if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+    (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+     CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
     message(STATUS "ARM detected")
     if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
-        add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
@@ -847,15 +959,23 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
             list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            # Raspberry Pi 2
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            if ("${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()
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Android arm64-v8a
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
     endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
+elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
     message(STATUS "x86 detected")
     if (MSVC)
         # instruction set detection for MSVC only
@@ -992,11 +1112,6 @@ endif()
 
 # ggml
 
-if (GGML_USE_CPU_HBM)
-    add_definitions(-DGGML_USE_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
-endif()
-
 add_library(ggml OBJECT
             ggml.c
             ggml.h
@@ -1008,22 +1123,22 @@ add_library(ggml OBJECT
             ggml-quants.h
             ${GGML_SOURCES_CUDA}    ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL}  ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
             ${GGML_SOURCES_METAL}   ${GGML_HEADERS_METAL}
             ${GGML_SOURCES_MPI}     ${GGML_HEADERS_MPI}
             ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
             ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
-target_compile_features(ggml PUBLIC c_std_11) # don't bump
+target_compile_features   (ggml PUBLIC c_std_11) # don't bump
+
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
-if (GGML_USE_CPU_HBM)
-    target_link_libraries(ggml PUBLIC memkind)
-endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
     add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
@@ -1036,10 +1151,13 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
+            unicode.h
+            unicode.cpp
             )
 
 target_include_directories(llama PUBLIC .)
-target_compile_features(llama PUBLIC cxx_std_11) # don't bump
+target_compile_features   (llama PUBLIC cxx_std_11) # don't bump
+
 target_link_libraries(llama PRIVATE
     ggml
     ${LLAMA_EXTRA_LIBS}
@@ -1090,7 +1208,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}" "${GGML_HEADERS_VULKAN}"
+        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
         "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

Makefile

@@ -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 tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -97,9 +97,10 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS = -I. -Icommon
-MK_CFLAGS   = -std=c11   -fPIC
-MK_CXXFLAGS = -std=c++11 -fPIC
+MK_CPPFLAGS  = -I. -Icommon
+MK_CFLAGS    = -std=c11   -fPIC
+MK_CXXFLAGS  = -std=c++11 -fPIC
+MK_NVCCFLAGS = -std=c++11
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -109,8 +110,21 @@ MK_NVCCFLAGS  += -O3
 else
 MK_CFLAGS     += -O3
 MK_CXXFLAGS   += -O3
+MK_NVCCFLAGS  += -O3
 endif
 
+ifndef LLAMA_NO_CCACHE
+CCACHE := $(shell which ccache)
+ifdef CCACHE
+export CCACHE_SLOPPINESS = time_macros
+$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.)
+CC    := $(CCACHE) $(CC)
+CXX   := $(CCACHE) $(CXX)
+else
+$(info I ccache not found. Consider installing it for faster compilation.)
+endif # CCACHE
+endif # LLAMA_NO_CCACHE
+
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
@@ -153,13 +167,17 @@ 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 ($(UNAME_S),Linux)
-		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
+		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -187,6 +205,10 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(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 ''
@@ -202,6 +224,11 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
+ifeq ($(LLAMA_FATAL_WARNINGS),1)
+	MK_CFLAGS   += -Werror
+	MK_CXXFLAGS += -Werror
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -362,10 +389,18 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
+	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
 	OBJS         += ggml-cuda.o
-	MK_NVCCFLAGS  = -use_fast_math
+	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
@@ -373,9 +408,9 @@ ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
 endif # LLAMA_DEBUG
 ifdef LLAMA_CUDA_NVCC
-	NVCC = $(LLAMA_CUDA_NVCC)
+	NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
 else
-	NVCC = nvcc
+	NVCC = $(CCACHE) nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
@@ -422,11 +457,11 @@ 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-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.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 $(NVCCFLAGS) -Xcompiler "$(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) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -457,6 +492,18 @@ ifdef LLAMA_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
 endif
 
+ifdef LLAMA_VULKAN_DEBUG
+	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
+endif
+
+ifdef LLAMA_VULKAN_VALIDATE
+	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
+endif
+
+ifdef LLAMA_VULKAN_RUN_TESTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
+endif
+
 ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_VULKAN
@@ -470,7 +517,7 @@ ifdef LLAMA_HIPBLAS
 		ROCM_PATH	?= /opt/rocm
 		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
-	HIPCC                   ?= $(ROCM_PATH)/bin/hipcc
+	HIPCC                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
@@ -499,11 +546,30 @@ 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
+ggml-metal.o: ggml-metal.m ggml-metal.h ggml.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
@@ -515,9 +581,10 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-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 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 NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -525,7 +592,12 @@ override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(GF_CXXFLAGS)
+CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
+endif
+
+ifdef LLAMA_CURL
+override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
+override LDFLAGS  := $(LDFLAGS) -lcurl
 endif
 
 #
@@ -540,8 +612,19 @@ $(info I CFLAGS:    $(CFLAGS))
 $(info I CXXFLAGS:  $(CXXFLAGS))
 $(info I NVCCFLAGS: $(NVCCFLAGS))
 $(info I LDFLAGS:   $(LDFLAGS))
-$(info I CC:        $(shell $(CC) --version | head -n 1))
-$(info I CXX:       $(shell $(CXX) --version | head -n 1))
+$(info I CC:        $(shell $(CC)   --version | head -n 1))
+$(info I CXX:       $(shell $(CXX)  --version | head -n 1))
+ifdef LLAMA_CUBLAS
+$(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # LLAMA_CUBLAS
 $(info )
 
 #
@@ -557,12 +640,15 @@ 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-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
+unicode.o: unicode.cpp unicode.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
+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
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
@@ -586,99 +672,147 @@ train.o: common/train.cpp common/train.h
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
+libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
+	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
+
 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	find examples pocs -type f -name "*.o" -delete
 
 #
 # Examples
 #
 
+# $< is the first prerequisite, i.e. the source file.
+# Explicitly compile this to an object file so that it can be cached with ccache.
+# The source file is then filtered out from $^ (the list of all prerequisites) and the object file is added instead.
+
+# Helper function that replaces .c, .cpp, and .cu file endings with .o:
+GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
+
 main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
 infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tokenize: examples/tokenize/tokenize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(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) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
+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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+gguf-split: examples/gguf-split/gguf-split.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)
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
 llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
 
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
@@ -686,7 +820,7 @@ swift: examples/batched.swift
 endif
 
 common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
-	@sh scripts/build-info.sh $(CC) > $@.tmp
+	@sh scripts/build-info.sh "$(CC)" > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
 	else \
@@ -703,7 +837,8 @@ build-info.o: common/build-info.cpp
 tests: $(TEST_TARGETS)
 
 benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 run-benchmark-matmult: benchmark-matmult
 	./$@
@@ -711,58 +846,80 @@ run-benchmark-matmult: benchmark-matmult
 .PHONY: run-benchmark-matmult swift
 
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-opt: tests/test-opt.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-c.o: tests/test-c.c llama.h
 	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
 
 tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(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

@@ -13,17 +13,32 @@ let package = Package(
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
-    dependencies: [
-        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
-    ],
     targets: [
         .target(
             name: "llama",
-            dependencies: ["ggml"],
             path: ".",
-            exclude: ["ggml-metal.metal"],
+            exclude: [
+               "cmake",
+               "examples",
+               "scripts",
+               "models",
+               "tests",
+               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
+               "Makefile"
+            ],
             sources: [
+                "ggml.c",
                 "llama.cpp",
+                "unicode.cpp",
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "ggml-quants.c",
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
             ],
             publicHeadersPath: "spm-headers",
             cSettings: [

README-sycl.md

@@ -0,0 +1,541 @@
+# llama.cpp for SYCL
+
+- [Background](#background)
+- [News](#news)
+- [OS](#os)
+- [Intel GPU](#intel-gpu)
+- [Docker](#docker)
+- [Linux](#linux)
+- [Windows](#windows)
+- [Environment Variable](#environment-variable)
+- [Known Issue](#known-issue)
+- [Q&A](#q&a)
+- [Todo](#todo)
+
+## Background
+
+SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
+
+oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
+
+Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
+
+To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
+
+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|
+|-|-|-|
+|Linux|Support|Ubuntu 22.04, Fedora Silverblue 39|
+|Windows|Support|Windows 11|
+
+
+## Intel GPU
+
+### Verified
+
+|Intel GPU| Status | Verified Model|
+|-|-|-|
+|Intel Data Center Max Series| Support| Max 1550|
+|Intel Data Center Flex Series| Support| Flex 170|
+|Intel Arc Series| Support| Arc 770, 730M|
+|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
+|Intel iGPU| Support| iGPU in i5-1250P, i7-1260P, i7-1165G7|
+
+Note: If the EUs (Execution Unit) in iGPU is less than 80, the inference speed will be too slow to use.
+
+### Memory
+
+The memory is a limitation to run LLM on GPUs.
+
+When run llama.cpp, there is print log to show the applied memory on GPU. You could know how much memory to be used in your case. Like `llm_load_tensors:            buffer size =  3577.56 MiB`.
+
+For iGPU, please make sure the shared memory from host memory is enough. For llama-2-7b.Q4_0, recommend the host memory is 8GB+.
+
+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:
+- Only docker on Linux is tested. Docker on WSL may not work.
+- You may need to install Intel GPU driver on the host machine (See the [Linux](#linux) section to know how to do that)
+
+### Build the image
+
+You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
+
+
+```sh
+# For F16:
+#docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
+
+# Or, for F32:
+docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile .
+
+# Note: you can also use the ".devops/main-server.Dockerfile", which compiles the "server" example
+```
+
+### Run
+
+```sh
+# Firstly, find all the DRI cards:
+ls -la /dev/dri
+# Then, pick the card that you want to use.
+
+# For example with "/dev/dri/card1"
+docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+
+## Linux
+
+### Setup Environment
+
+1. Install Intel GPU driver.
+
+a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
+
+Note: for iGPU, please install the client GPU driver.
+
+b. Add user to group: video, render.
+
+```sh
+sudo usermod -aG render username
+sudo usermod -aG video username
+```
+
+Note: re-login to enable it.
+
+c. Check
+
+```sh
+sudo apt install clinfo
+sudo clinfo -l
+```
+
+Output (example):
+
+```
+Platform #0: Intel(R) OpenCL Graphics
+ `-- Device #0: Intel(R) Arc(TM) A770 Graphics
+
+
+Platform #0: Intel(R) OpenCL HD Graphics
+ `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
+```
+
+2. Install Intel® oneAPI Base toolkit.
+
+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: **/opt/intel/oneapi**.
+
+Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
+
+b. Check
+
+```sh
+source /opt/intel/oneapi/setvars.sh
+
+sycl-ls
+```
+
+There should be one or more level-zero devices. Please confirm that at least one GPU is present, like **[ext_oneapi_level_zero:gpu:0]**.
+
+Output (example):
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
+
+```
+
+2. Build locally:
+
+Note:
+- You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
+- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+
+```sh
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+# For FP16:
+#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
+# 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
+
+# Or, build all binary
+cmake --build . --config Release -v
+
+cd ..
+```
+
+or
+
+```sh
+./examples/sycl/build.sh
+```
+
+### Run
+
+1. Put model file to folder **models**
+
+You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+
+2. Enable oneAPI running environment
+
+```
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. List device ID
+
+Run without parameter:
+
+```sh
+./build/bin/ls-sycl-device
+
+# or running the "main" executable and look at the output log:
+
+./build/bin/main
+```
+
+Check the ID in startup log, like:
+
+```
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+
+```
+
+|Attribute|Note|
+|-|-|
+|compute capability 1.3|Level-zero running time, recommended |
+|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+
+4. Set device ID and execute llama.cpp
+
+Set device ID = 0 by **GGML_SYCL_DEVICE=0**
+
+```sh
+GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+or run by script:
+
+```sh
+./examples/sycl/run_llama2.sh
+```
+
+Note:
+
+- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+
+
+5. Check the device ID in output
+
+Like:
+```
+Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+```
+
+## Windows
+
+### Setup Environment
+
+1. Install Intel GPU driver.
+
+Please install Intel GPU driver by official guide: [Install GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+
+Note: **The driver is mandatory for compute function**.
+
+2. Install Visual Studio.
+
+Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact oneAPI environment enabling in Windows.
+
+3. Install Intel® oneAPI Base toolkit.
+
+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**.
+
+Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
+
+b. Enable oneAPI running environment:
+
+- In Search, input 'oneAPI'.
+
+Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
+
+- In Run:
+
+In CMD:
+```
+"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
+```
+
+c. Check GPU
+
+In oneAPI command line:
+
+```
+sycl-ls
+```
+
+There should be one or more level-zero devices. Please confirm that at least one GPU is present, like **[ext_oneapi_level_zero:gpu:0]**.
+
+Output (example):
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu:1] Intel(R) OpenCL, 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Iris(R) Xe Graphics OpenCL 3.0 NEO  [31.0.101.5186]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
+```
+
+4. Install cmake & make
+
+a. Download & install cmake for Windows: https://cmake.org/download/
+
+b. Download & install mingw-w64 make for Windows provided by w64devkit
+
+- Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+
+- Extract `w64devkit` on your pc.
+
+- Add the **bin** folder path in the Windows system PATH environment, like `C:\xxx\w64devkit\bin\`.
+
+### Build locally:
+
+In oneAPI command line window:
+
+```
+mkdir -p build
+cd build
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+::  for FP16
+::  faster for long-prompt inference
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+
+::  for FP32
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+
+
+::  build example/main only
+::  make main
+
+::  build all binary
+make -j
+cd ..
+```
+
+or
+
+```
+.\examples\sycl\win-build-sycl.bat
+```
+
+Note:
+
+- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+
+### Run
+
+1. Put model file to folder **models**
+
+You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+
+2. Enable oneAPI running environment
+
+- In Search, input 'oneAPI'.
+
+Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
+
+- In Run:
+
+In CMD:
+```
+"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
+```
+
+3. List device ID
+
+Run without parameter:
+
+```
+build\bin\ls-sycl-device.exe
+
+or
+
+build\bin\main.exe
+```
+
+Check the ID in startup log, like:
+
+```
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+
+```
+
+|Attribute|Note|
+|-|-|
+|compute capability 1.3|Level-zero running time, recommended |
+|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+
+4. Set device ID and execute llama.cpp
+
+Set device ID = 0 by **set GGML_SYCL_DEVICE=0**
+
+```
+set GGML_SYCL_DEVICE=0
+build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0
+```
+or run by script:
+
+```
+.\examples\sycl\win-run-llama2.bat
+```
+
+Note:
+
+- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+
+
+5. Check the device ID in output
+
+Like:
+```
+Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+```
+
+## Environment Variable
+
+#### Build
+
+|Name|Value|Function|
+|-|-|-|
+|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
+|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
+|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
+|CMAKE_CXX_COMPILER|icpx (Linux), icx (Windows)|use icpx/icx for SYCL code path|
+
+#### Running
+
+
+|Name|Value|Function|
+|-|-|-|
+|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
+
+- Hang during startup
+
+  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
+
+  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`.
+
+  Miss to enable oneAPI running environment.
+
+  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
+
+- In Windows, no result, not error.
+
+  Miss to enable oneAPI running environment.
+
+- Meet compile error.
+
+  Remove folder **build** and try again.
+
+- I can **not** see **[ext_oneapi_level_zero:gpu:0]** afer install GPU driver in Linux.
+
+  Please run **sudo sycl-ls**.
+
+  If you see it in result, please add video/render group to your ID:
+
+  ```
+  sudo usermod -aG render username
+  sudo usermod -aG video username
+  ```
+
+  Then **relogin**.
+
+  If you do not see it, please check the installation GPU steps again.
+
+## Todo
+
+- Support multiple cards.

README.md

@@ -6,15 +6,21 @@
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
-Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
+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
 
-- 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
+- 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
 
 ----
 
@@ -30,17 +36,14 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
         <li><a href="#get-the-code">Get the Code</a></li>
         <li><a href="#build">Build</a></li>
         <li><a href="#blas-build">BLAS Build</a></li>
-        <li><a href="#prepare-data--run">Prepare Data & Run</a></li>
+        <li><a href="#prepare-and-quantize">Prepare and Quantize</a></li>
+        <li><a href="#run-the-quantized-model">Run the quantized model</a></li>
         <li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
         <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
-        <li><a href="#instruction-mode-with-alpaca">Instruction mode with Alpaca</a></li>
-        <li><a href="#using-openllama">Using OpenLLaMA</a></li>
-        <li><a href="#using-gpt4all">Using GPT4All</a></li>
-        <li><a href="#using-pygmalion-7b--metharme-7b">Using Pygmalion 7B & Metharme 7B</a></li>
-        <li><a href="#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data">Obtaining the Facebook LLaMA original model and Stanford Alpaca model data</a></li>
-        <li><a href="#verifying-the-model-files">Verifying the model files</a></li>
+        <li><a href="#instruct-mode">Instruct mode</a></li>
+        <li><a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a></li>
         <li><a href="#seminal-papers-and-background-on-the-models">Seminal papers and background on the models</a></li>
         <li><a href="#perplexity-measuring-model-quality">Perplexity (measuring model quality)</a></li>
         <li><a href="#android">Android</a></li>
@@ -55,18 +58,20 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ## Description
 
-The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quantization on a MacBook
+The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
+variety of hardware - locally and in the cloud.
 
-- Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
+- 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
-- Mixed F16 / F32 precision
-- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
-- CUDA, Metal, OpenCL, SYCL GPU backend support
+- 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
+- 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
 
-The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
-Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves
-as the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
+Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022), the project has
+improved significantly thanks to many contributions. It is the main playground for developing new features for the
+[ggml](https://github.com/ggerganov/ggml) library.
 
 **Supported platforms:**
 
@@ -74,45 +79,53 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] Linux
 - [X] Windows (via CMake)
 - [X] Docker
+- [X] FreeBSD
 
 **Supported models:**
 
+Typically finetunes of the base models below are supported as well.
+
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
+- [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
+- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [X] Falcon
-- [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
-- [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
-- [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
-- [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
-- [X] [Pygmalion/Metharme](#using-pygmalion-7b--metharme-7b)
-- [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
-- [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
 - [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
 - [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
-- [X] [StableLM-3b-4e1t](https://github.com/ggerganov/llama.cpp/pull/3586)
+- [X] [StableLM models](https://huggingface.co/stabilityai)
 - [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
 - [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
-- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
+- [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
 - [x] [GPT-2](https://huggingface.co/gpt2)
+- [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)
+- [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
 
 **Multimodal models:**
 
-- [x] [Llava 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
-- [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
+- [x] [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] [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:**
 
@@ -120,6 +133,8 @@ as the main playground for developing new features for the [ggml](https://github
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
+- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
+- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
@@ -133,19 +148,34 @@ as the main playground for developing new features for the [ggml](https://github
 
 **UI:**
 
-- [nat/openplayground](https://github.com/nat/openplayground)
-- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
-- [withcatai/catai](https://github.com/withcatai/catai)
-- [semperai/amica](https://github.com/semperai/amica)
-- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
-- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+Unless otherwise noted these projects are open-source with permissive licensing:
+
 - [iohub/collama](https://github.com/iohub/coLLaMA)
+- [janhq/jan](https://github.com/janhq/jan) (AGPL)
+- [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)
+- [ollama/ollama](https://github.com/ollama/ollama)
+- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
+- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
+- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [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)
 
 ---
 
 Here is a typical run using LLaMA v2 13B on M2 Ultra:
 
-```java
+```
 $ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
 I llama.cpp build info:
 I UNAME_S:  Darwin
@@ -229,7 +259,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the end-to-end binary build and model conversion steps for the LLaMA-7B model.
+Here are the end-to-end binary build and model conversion steps for most supported models.
 
 ### Get the Code
 
@@ -390,28 +420,28 @@ Building the program with BLAS support may lead to some performance improvements
 
   Check [BLIS.md](docs/BLIS.md) for more information.
 
+- #### SYCL
+  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+
+  llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+
+  For detailed info, please refer to [llama.cpp for SYCL](README-sycl.md).
+
 - #### Intel oneMKL
+  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./README-sycl.md).
+
   - Using manual oneAPI installation:
     By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
       ```bash
       mkdir build
       cd build
-      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-runtime docker image, only required for manual installation
+      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
       cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
       cmake --build . --config Release
       ```
 
   - Using oneAPI docker image:
-    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-runtime](https://hub.docker.com/r/intel/oneapi-runtime)
-
-      ```bash
-      mkdir build
-      cd build
-      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
-      cmake --build . --config Release
-      ```
-
-  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni.
+    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
 
   Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
 
@@ -598,43 +628,87 @@ Building the program with BLAS support may lead to some performance improvements
 
   You can get a list of platforms and devices from the `clinfo -l` command, etc.
 
-- #### SYCL
+- #### Vulkan
 
-  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+  **With docker**:
 
-  llama.cpp based on SYCL is used to support Intel GPU (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+  You don't need to install Vulkan SDK. It will be installed inside the container.
 
-  For detailed info, please refer to [llama.cpp for SYCL](README_sycl.md).
+  ```sh
+  # Build the image
+  docker build -t llama-cpp-vulkan -f .devops/main-vulkan.Dockerfile .
 
+  # Then, use it:
+  docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+  ```
 
-### Prepare Data & Run
+  **Without docker**:
+
+  Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
+
+  For example, on Ubuntu 22.04 (jammy), use the command below:
+
+  ```bash
+  wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
+  wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+  apt update -y
+  apt-get install -y vulkan-sdk
+  # To verify the installation, use the command below:
+  vulkaninfo
+  ```
+
+  Alternatively your package manager might be able to provide the appropiate libraries. For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
+
+  Then, build llama.cpp using the cmake command below:
+
+  ```bash
+  mkdir -p build
+  cd build
+  cmake .. -DLLAMA_VULKAN=1
+  cmake --build . --config Release
+  # Test the output binary (with "-ngl 33" to offload all layers to GPU)
+  ./bin/main -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+
+  # You should see in the output, ggml_vulkan detected your GPU. For example:
+  # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
+  ```
+
+### Prepare and Quantize
+
+To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
 
 ```bash
-# obtain the original LLaMA model weights and place them in ./models
+# obtain the official LLaMA model weights and place them in ./models
 ls ./models
-65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
+llama-2-7b tokenizer_checklist.chk tokenizer.model
 # [Optional] for models using BPE tokenizers
 ls ./models
-65B 30B 13B 7B vocab.json
+<folder containing weights and tokenizer json> vocab.json
+# [Optional] for PyTorch .bin models like Mistral-7B
+ls ./models
+<folder containing weights and tokenizer json>
 
 # install Python dependencies
 python3 -m pip install -r requirements.txt
 
-# convert the 7B model to ggml FP16 format
-python3 convert.py models/7B/
+# convert the model to ggml FP16 format
+python3 convert.py models/mymodel/
 
 # [Optional] for models using BPE tokenizers
-python convert.py models/7B/ --vocabtype bpe
+python convert.py models/mymodel/ --vocab-type bpe
 
-# quantize the model to 4-bits (using q4_0 method)
-./quantize ./models/7B/ggml-model-f16.gguf ./models/7B/ggml-model-q4_0.gguf q4_0
+# quantize the model to 4-bits (using Q4_K_M method)
+./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
 
-# update the gguf filetype to current if older version is unsupported by another application
-./quantize ./models/7B/ggml-model-q4_0.gguf ./models/7B/ggml-model-q4_0-v2.gguf COPY
+# update the gguf filetype to current version if older version is now unsupported
+./quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+```
 
+### Run the quantized model
 
-# run the inference
-./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
+```bash
+# start inference on a gguf model
+./main -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
 ```
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
@@ -655,7 +729,7 @@ From the unzipped folder, open a terminal/cmd window here and place a pre-conver
 
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
-| Model | Original size | Quantized size (4-bit) |
+| Model | Original size | Quantized size (Q4_0) |
 |------:|--------------:|-----------------------:|
 |    7B |         13 GB |                 3.9 GB |
 |   13B |         24 GB |                 7.8 GB |
@@ -682,9 +756,21 @@ Several quantization methods are supported. They differ in the resulting model d
 |   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
 
 - [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
-- recent k-quants improvements
+- recent k-quants improvements and new i-quants
   - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
   - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
+  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
+  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
+  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
+  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
+  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
+  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
+  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
+  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
+  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
+  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
 
 ### Perplexity (measuring model quality)
 
@@ -696,7 +782,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 
 #### How to run
 
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
 2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```
@@ -709,7 +795,7 @@ 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
 
@@ -759,9 +845,9 @@ The `grammars/` folder contains a handful of sample grammars. To write your own,
 
 For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
 
-### Instruction mode with Alpaca
+### Instruct mode
 
-1. First, download the `ggml` Alpaca model into the `./models` folder
+1. First, download and place the `ggml` model into the `./models` folder
 2. Run the `main` tool like this:
 
 ```
@@ -773,7 +859,7 @@ 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.
@@ -787,50 +873,6 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 >
 ```
 
-### Using [OpenLLaMA](https://github.com/openlm-research/open_llama)
-
-OpenLLaMA is an openly licensed reproduction of Meta's original LLaMA model. It uses the same architecture and is a drop-in replacement for the original LLaMA weights.
-
-- Download the [3B](https://huggingface.co/openlm-research/open_llama_3b), [7B](https://huggingface.co/openlm-research/open_llama_7b), or [13B](https://huggingface.co/openlm-research/open_llama_13b) model from Hugging Face.
-- Convert the model to ggml FP16 format using `python convert.py <path to OpenLLaMA directory>`
-
-### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
-
-*Note: these instructions are likely obsoleted by the GGUF update*
-
-- Obtain the `tokenizer.model` file from LLaMA model and put it to `models`
-- Obtain the `added_tokens.json` file from Alpaca model and put it to `models`
-- Obtain the `gpt4all-lora-quantized.bin` file from GPT4All model and put it to `models/gpt4all-7B`
-- It is distributed in the old `ggml` format which is now obsoleted
-- You have to convert it to the new format using `convert.py`:
-
-```bash
-python3 convert.py models/gpt4all-7B/gpt4all-lora-quantized.bin
-```
-
-- You can now use the newly generated `models/gpt4all-7B/ggml-model-q4_0.bin` model in exactly the same way as all other models
-
-- The newer GPT4All-J model is not yet supported!
-
-### Using Pygmalion 7B & Metharme 7B
-
-- Obtain the [LLaMA weights](#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data)
-- Obtain the [Pygmalion 7B](https://huggingface.co/PygmalionAI/pygmalion-7b/) or [Metharme 7B](https://huggingface.co/PygmalionAI/metharme-7b) XOR encoded weights
-- Convert the LLaMA model with [the latest HF convert script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py)
-- Merge the XOR files with the converted LLaMA weights by running the [xor_codec](https://huggingface.co/PygmalionAI/pygmalion-7b/blob/main/xor_codec.py) script
-- Convert to `ggml` format using the `convert.py` script in this repo:
-```bash
-python3 convert.py pygmalion-7b/ --outtype q4_1
-```
-> The Pygmalion 7B & Metharme 7B weights are saved in [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format) precision. If you wish to convert to `ggml` without quantizating, please specify the `--outtype` as `f32` instead of `f16`.
-
-
-### Obtaining the Facebook LLaMA original model and Stanford Alpaca model data
-
-- **Under no circumstances should IPFS, magnet links, or any other links to model downloads be shared anywhere in this repository, including in issues, discussions, or pull requests. They will be immediately deleted.**
-- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
-- Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
-
 ### Obtaining and using the Facebook LLaMA 2 model
 
 - Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
@@ -842,20 +884,6 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
   - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
   - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
 
-### Verifying the model files
-
-Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
-- The following python script will verify if you have all possible latest files in your self-installed `./models` subdirectory:
-
-```bash
-# run the verification script
-./scripts/verify-checksum-models.py
-```
-
-- On linux or macOS it is also possible to run the following commands to verify if you have all possible latest files in your self-installed `./models` subdirectory:
-    - On Linux: `sha256sum --ignore-missing -c SHA256SUMS`
-    - on macOS: `shasum -a 256 --ignore-missing -c SHA256SUMS`
-
 ### Seminal papers and background on the models
 
 If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
@@ -878,6 +906,9 @@ 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
@@ -886,7 +917,28 @@ $ 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 the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
+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:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
@@ -944,7 +996,7 @@ We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 

README_sycl.md

@@ -1,252 +0,0 @@
-# llama.cpp for SYCL
-
-[Background](#background)
-
-[OS](#os)
-
-[Intel GPU](#intel-gpu)
-
-[Linux](#linux)
-
-[Environment Variable](#environment-variable)
-
-[Known Issue](#known-issue)
-
-[Todo](#todo)
-
-## Background
-
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
-
-oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
-
-Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
-
-To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
-
-The llama.cpp for SYCL is used to support Intel GPUs.
-
-For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
-
-## OS
-
-|OS|Status|Verified|
-|-|-|-|
-|Linux|Support|Ubuntu 22.04|
-|Windows|Ongoing| |
-
-
-## Intel GPU
-
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Intel Data Center Max Series| Support| Max 1550|
-|Intel Data Center Flex Series| Support| Flex 170|
-|Intel Arc Series| Support| Arc 770|
-|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
-|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7|
-
-
-## Linux
-
-### Setup Environment
-
-1. Install Intel GPU driver.
-
-a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
-
-Note: for iGPU, please install the client GPU driver.
-
-b. Add user to group: video, render.
-
-```
-sudo usermod -aG render username
-sudo usermod -aG video username
-```
-
-Note: re-login to enable it.
-
-c. Check
-
-```
-sudo apt install clinfo
-sudo clinfo -l
-```
-
-Output (example):
-
-```
-Platform #0: Intel(R) OpenCL Graphics
- `-- Device #0: Intel(R) Arc(TM) A770 Graphics
-
-
-Platform #0: Intel(R) OpenCL HD Graphics
- `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
-```
-
-2. Install Intel® oneAPI Base toolkit.
-
-
-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: **/opt/intel/oneapi**.
-
-Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
-
-b. Check
-
-```
-source /opt/intel/oneapi/setvars.sh
-
-sycl-ls
-```
-
-There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**.
-
-Output (example):
-```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
-[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
-[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
-[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
-
-```
-
-2. Build locally:
-
-```
-mkdir -p build
-cd build
-source /opt/intel/oneapi/setvars.sh
-
-#for FP16
-#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference
-
-#for FP32
-cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-#build example/main only
-#cmake --build . --config Release --target main
-
-#build all binary
-cmake --build . --config Release -v
-
-```
-
-or
-
-```
-./examples/sycl/build.sh
-```
-
-Note:
-
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
-
-### Run
-
-1. Put model file to folder **models**
-
-2. Enable oneAPI running environment
-
-```
-source /opt/intel/oneapi/setvars.sh
-```
-
-3. List device ID
-
-Run without parameter:
-
-```
-./build/bin/ls-sycl-device
-
-or
-
-./build/bin/main
-```
-
-Check the ID in startup log, like:
-
-```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-
-```
-
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
-
-4. Set device ID and execute llama.cpp
-
-Set device ID = 0 by **GGML_SYCL_DEVICE=0**
-
-```
-GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
-```
-or run by script:
-
-```
-./examples/sycl/run_llama2.sh
-```
-
-Note:
-
-- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
-
-
-5. Check the device ID in output
-
-Like：
-```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
-```
-
-
-## Environment Variable
-
-#### Build
-
-|Name|Value|Function|
-|-|-|-|
-|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
-|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
-|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
-|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path|
-
-#### Running
-
-
-|Name|Value|Function|
-|-|-|-|
-|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|
-
-## Known Issue
-
-- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
-
-  Miss to enable oneAPI running environment.
-
-  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
-
-
-- Hang during startup
-
-  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
-
-  Solution: add **--no-mmap**.
-
-## Todo
-
-- Support to build in Windows.
-
-- Support multiple cards.

SHA256SUMS

@@ -1,40 +0,0 @@
-700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
-666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
-ec2f2d1f0dfb73b72a4cbac7fa121abbe04c37ab327125a38248f930c0f09ddf  models/7B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_1.bin
-7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265  models/7B/params.json
-745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
-d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
-2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
-fad169e6f0f575402cf75945961cb4a8ecd824ba4da6be2af831f320c4348fa5  models/13B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_1.bin
-4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f  models/13B/params.json
-e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/consolidated.00.pth
-4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff  models/30B/consolidated.01.pth
-24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
-1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
-7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
-d2a441403944819492ec8c2002cc36fa38468149bfb4b7b4c52afc7bd9a7166d  models/30B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_1.bin
-2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb  models/30B/params.json
-135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe  models/65B/consolidated.00.pth
-9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde  models/65B/consolidated.01.pth
-e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770  models/65B/consolidated.02.pth
-73176ffb426b40482f2aa67ae1217ef79fbbd1fff5482bae5060cdc5a24ab70e  models/65B/consolidated.03.pth
-882e6431d0b08a8bc66261a0d3607da21cbaeafa96a24e7e59777632dbdac225  models/65B/consolidated.04.pth
-a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/consolidated.05.pth
-72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
-d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
-60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
-cde053439fa4910ae454407e2717cc46cc2c2b4995c00c93297a2b52e790fa92  models/65B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_1.bin
-999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b  models/65B/params.json
-9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347  models/tokenizer.model

awq-py/README.md

@@ -1,116 +0,0 @@
-# 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

@@ -1,254 +0,0 @@
-"""
-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

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

build.zig

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

ci/run.sh

@@ -33,7 +33,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA=""
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
@@ -45,7 +45,8 @@ 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:\n source /opt/intel/oneapi/setvars.sh"
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
+        echo "source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
 
@@ -219,7 +220,7 @@ function gg_run_open_llama_3b_v2 {
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
     head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
@@ -272,19 +273,19 @@ 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 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/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/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.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/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
@@ -343,17 +344,17 @@ 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 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
+    (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
     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 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
+    (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
     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 2 ) 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 1 ) 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
@@ -401,7 +402,7 @@ function gg_run_open_llama_7b_v2 {
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
     path_models="../models-mnt/open-llama/7B-v2"
@@ -568,6 +569,54 @@ function gg_sum_open_llama_7b_v2 {
     #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
+# bge-small
+
+function gg_run_embd_bge_small {
+    cd ${SRC}
+
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+
+    gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/bge-small"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+
+    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+
+    set +e
+}
+
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -591,6 +640,8 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    test $ret -eq 0 && gg_run embd_bge_small
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

common/CMakeLists.txt

@@ -19,7 +19,12 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         endif()
     endif()
 
-    set(GIT_INDEX "${GIT_DIR}/index")
+    if(EXISTS "${GIT_DIR}/index")
+        set(GIT_INDEX "${GIT_DIR}/index")
+    else()
+        message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
+    endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
     set(GIT_INDEX "")
@@ -63,6 +68,17 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
+set(LLAMA_COMMON_EXTRA_LIBS build_info)
+
+# Use curl to download model url
+if (LLAMA_CURL)
+    find_package(CURL REQUIRED)
+    add_definitions(-DLLAMA_USE_CURL)
+    include_directories(${CURL_INCLUDE_DIRS})
+    find_library(CURL_LIBRARY curl REQUIRED)
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+endif ()
+
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)
+target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)

common/common.h

@@ -37,51 +37,59 @@ 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                           = -1;    // RNG seed
+    uint32_t seed                 = LLAMA_DEFAULT_SEED; // 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                         = 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                         = 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_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[LLAMA_MAX_DEVICES] = {0};   // how split tensors should be distributed across GPUs
-    int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
-    int32_t grp_attn_n                      = 1;     // group-attention factor
-    int32_t grp_attn_w                      = 512;   // group-attention width
-    int32_t n_print                         = -1;    // print token count every n tokens (-1 = disabled)
-    float   rope_freq_base                  = 0.0f;  // RoPE base frequency
-    float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
-    float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor                = 1.0f;  // YaRN magnitude scaling factor
-    float   yarn_beta_fast                  = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow                  = 1.0f;  // YaRN high correction dim
-    int32_t yarn_orig_ctx                   = 0;     // YaRN original context length
-    int8_t  rope_scaling_type               = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
-                                                                              //       pinging @cebtenzzre
+    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 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_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_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
+    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.
+    int32_t grp_attn_n            = 1;     // group-attention factor
+    int32_t grp_attn_w            = 512;   // group-attention width
+    int32_t n_print               = -1;    // print token count every n tokens (-1 = disabled)
+    float   rope_freq_base        = 0.0f;  // RoPE base frequency
+    float   rope_freq_scale       = 0.0f;  // RoPE frequency scaling factor
+    float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
+    float   yarn_attn_factor      = 1.0f;  // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
+    float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
+    int32_t yarn_orig_ctx         = 0;     // YaRN original context length
+    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
 
     // // sampling parameters
     struct llama_sampling_params sparams;
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_url         = ""; // model url to download
     std::string model_draft       = "";                              // draft model for speculative decoding
     std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
@@ -99,6 +107,11 @@ 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)
@@ -114,7 +127,6 @@ 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
@@ -135,7 +147,6 @@ struct gpt_params {
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
@@ -163,10 +174,13 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 void process_escapes(std::string& input);
 
 //
-// String parsing
+// String utils
 //
 
-std::string parse_samplers_input(std::string input);
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
+std::vector<std::string> string_split(std::string input, char separator);
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
 
 //
 // Model utils
@@ -178,6 +192,9 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model,
+                                                         struct llama_model_params     params);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -256,3 +273,32 @@ 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

@@ -278,6 +278,22 @@ 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

@@ -17,6 +17,13 @@ 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(
@@ -103,15 +110,10 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
 std::string llama_sampling_order_print(const llama_sampling_params & params) {
     std::string result = "CFG -> Penalties ";
     if (params.mirostat == 0) {
-        for (auto s : params.samplers_sequence) {
-            switch (s) {
-                case 'k': result += "-> top_k "; break;
-                case 'f': result += "-> tfs_z "; break;
-                case 'y': result += "-> typical_p "; break;
-                case 'p': result += "-> top_p "; break;
-                case 'm': result += "-> min_p "; break;
-                case 't': result += "-> temp "; break;
-                default : break;
+        for (auto sampler_type : params.samplers_sequence) {
+            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            if (!sampler_type_name.empty()) {
+                result += "-> " + sampler_type_name + " ";
             }
         }
     } else {
@@ -126,27 +128,25 @@ static void sampler_queue(
                    struct llama_context * ctx_main,
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
-                                 size_t & min_keep) {
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
+                                 size_t   min_keep) {
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
-    const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
+    const int32_t       top_k             = params.top_k;
     const float         top_p             = params.top_p;
     const float         min_p             = params.min_p;
     const float         tfs_z             = params.tfs_z;
     const float         typical_p         = params.typical_p;
-    const std::string & samplers_sequence = params.samplers_sequence;
+    const std::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
 
-    for (auto s : samplers_sequence) {
-        switch (s){
-            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
-            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
-            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
-            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
-            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't':
+    for (auto sampler_type : samplers_sequence) {
+        switch (sampler_type) {
+            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case llama_sampler_type::TEMPERATURE:
                 if (dynatemp_range > 0) {
                     float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
                     float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
@@ -256,7 +256,7 @@ static llama_token llama_sampling_sample_impl(
             id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
             // temperature sampling
-            size_t min_keep = std::max(1, params.n_probs);
+            size_t min_keep = std::max(1, params.min_keep);
 
             sampler_queue(ctx_main, params, cur_p, min_keep);
 
@@ -273,7 +273,7 @@ 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());
         }
     }
 
@@ -302,6 +302,77 @@ 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,
@@ -311,6 +382,14 @@ 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

@@ -8,10 +8,21 @@
 #include <vector>
 #include <unordered_map>
 
+// sampler types
+enum class llama_sampler_type : char {
+    TOP_K       = 'k',
+    TOP_P       = 'p',
+    MIN_P       = 'm',
+    TFS_Z       = 'f',
+    TYPICAL_P   = 'y',
+    TEMPERATURE = 't'
+};
+
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
     int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
     int32_t     top_k                 = 40;       // <= 0 to use vocab size
     float       top_p                 = 0.95f;    // 1.0 = disabled
     float       min_p                 = 0.05f;    // 0.0 = disabled
@@ -21,14 +32,22 @@ typedef struct llama_sampling_params {
     float       dynatemp_range        = 0.00f;    // 0.0 = disabled
     float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
     int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
+    float       penalty_repeat        = 1.00f;    // 1.0 = disabled
     float       penalty_freq          = 0.00f;    // 0.0 = disabled
     float       penalty_present       = 0.00f;    // 0.0 = disabled
     int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float       mirostat_tau          = 5.00f;    // target entropy
     float       mirostat_eta          = 0.10f;    // learning rate
-    bool        penalize_nl           = true;     // consider newlines as a repeatable token
-    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
+    bool        penalize_nl           = false;     // consider newlines as a repeatable token
+
+    std::vector<llama_sampler_type> samplers_sequence = {
+        llama_sampler_type::TOP_K,
+        llama_sampler_type::TFS_Z,
+        llama_sampler_type::TYPICAL_P,
+        llama_sampler_type::TOP_P,
+        llama_sampler_type::MIN_P,
+        llama_sampler_type::TEMPERATURE
+    };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 
@@ -112,6 +131,13 @@ 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_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_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_ADAM;
+        opt->params.type = GGML_OPT_TYPE_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_LBFGS;
+        opt->params.type = GGML_OPT_TYPE_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_ADAM:
+        case GGML_OPT_TYPE_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_LBFGS:
+        case GGML_OPT_TYPE_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);
@@ -1363,12 +1363,12 @@ bool consume_common_train_arg(
                 *invalid_param = true;
                 return true;
             }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-            params->n_gpu_layers = std::stoi(argv[i]);
-#else
-            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
-            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
-#endif
+            if (llama_supports_gpu_offload()) {
+                params->n_gpu_layers = std::stoi(argv[i]);
+            } else {
+                fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+            }
     } else if (arg == "-h" || arg == "--help") {
         params->print_usage = true;
         return true;

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, cfg.model_metadata_dir)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), 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", choices=["spm", "bpe"], default="spm",
-                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
+    parser.add_argument("--vocabtype", default="spm,hfft",
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
     return parser.parse_args()
 
 

convert-persimmon-to-gguf.py

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

convert.py

@@ -332,11 +332,14 @@ class Params:
 #
 
 class BpeVocab:
+    tokenizer_model = "gpt2"
+    name = "bpe"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        try:
+        if isinstance(self.bpe_tokenizer.get('model'), dict):
             self.vocab = self.bpe_tokenizer["model"]["vocab"]
-        except KeyError:
+        else:
             self.vocab = self.bpe_tokenizer
         added_tokens: dict[str, int]
         if fname_added_tokens is not None:
@@ -390,6 +393,9 @@ class BpeVocab:
 
 
 class SentencePieceVocab:
+    tokenizer_model = "llama"
+    name = "spm"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
         added_tokens: dict[str, int]
@@ -453,6 +459,9 @@ class SentencePieceVocab:
 
 
 class HfVocab:
+    tokenizer_model = "llama"
+    name = "hfft"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
         try:
             from transformers import AutoTokenizer
@@ -515,10 +524,14 @@ class HfVocab:
 
             # Yield token text, score, and type
             yield token_text, self.get_token_score(token_id), self.get_token_type(
-                token_id, self.special_ids  # Reuse already stored special IDs
+                token_id, token_text, self.special_ids  # Reuse already stored special IDs
             )
 
-    def get_token_type(self, token_id: int, special_ids: set[int]) -> gguf.TokenType:
+    def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType:
+        # Special case for byte tokens
+        if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
+            return gguf.TokenType.BYTE
+
         # Determine token type based on whether it's a special token
         return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
 
@@ -530,7 +543,7 @@ class HfVocab:
     def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         for text in self.added_tokens_list:
             if text in self.specials:
-                toktype = self.get_token_type(self.specials[text], self.special_ids)
+                toktype = self.get_token_type(self.specials[text], b'', self.special_ids)
                 score = self.get_token_score(self.specials[text])
             else:
                 toktype = gguf.TokenType.USER_DEFINED
@@ -549,7 +562,15 @@ class HfVocab:
         return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab"
+class NoVocab:
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab"
 
 
 #
@@ -931,8 +952,10 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
     # Handle special case where the model's vocab size is not set
     if params.n_vocab == -1:
         raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?"
+            f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}"
         )
+    if isinstance(vocab, NoVocab):
+        return  # model has no vocab
 
     # Check for a vocab size mismatch
     if params.n_vocab == vocab.vocab_size:
@@ -973,6 +996,7 @@ class OutputFile:
             name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
+        self.gguf.add_vocab_size          (params.n_vocab)
         self.gguf.add_context_length      (params.n_ctx)
         self.gguf.add_embedding_length    (params.n_embd)
         self.gguf.add_block_count         (params.n_layer)
@@ -1009,21 +1033,9 @@ class OutputFile:
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
-    def handle_tokenizer_model(self, vocab: Vocab) -> str:
-        # Map the vocab types to the supported tokenizer models
-        tokenizer_model = {
-            SentencePieceVocab: "llama",
-            HfVocab: "llama",
-            BpeVocab: "gpt2",
-        }.get(type(vocab))
-
-        # Block if vocab type is not predefined
-        if tokenizer_model is None:
-            raise ValueError("Unknown vocab type: Not supported")
-
-        return tokenizer_model
-
     def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
+        assert not isinstance(vocab, NoVocab)
+
         tokens = []
         scores = []
         toktypes = []
@@ -1039,11 +1051,8 @@ class OutputFile:
         return tokens, scores, toktypes
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
-        # Handle the tokenizer model
-        tokenizer_model = self.handle_tokenizer_model(vocab)
-
         # Ensure that tokenizer_model is added to the GGUF model
-        self.gguf.add_tokenizer_model(tokenizer_model)
+        self.gguf.add_tokenizer_model(vocab.tokenizer_model)
 
         # Extract model vocabulary for model conversion
         tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
@@ -1070,6 +1079,26 @@ class OutputFile:
     def write_tensor_info(self) -> None:
         self.gguf.write_ti_data_to_file()
 
+    def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
+                use_processpool_executor=True,
+            )
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(
+                f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
+            )
+            self.gguf.write_tensor_data(ndarray)
+
     def close(self) -> None:
         self.gguf.close()
 
@@ -1078,7 +1107,7 @@ class OutputFile:
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
         endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
     ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1116,8 +1145,11 @@ class OutputFile:
 
         # meta data
         of.add_meta_arch(params)
-        of.add_meta_vocab(vocab)
-        of.add_meta_special_vocab(svocab)
+        if isinstance(vocab, NoVocab):
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
+        else:
+            of.add_meta_vocab(vocab)
+            of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -1127,24 +1159,7 @@ class OutputFile:
         of.write_tensor_info()
 
         # tensor data
-        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
-        if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                use_processpool_executor=True,
-            )
-        else:
-            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
-
-        start = time.time()
-        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            elapsed = time.time() - start
-            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
-            padi = len(str(len(model)))
-            print(
-                f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
-            )
-            of.gguf.write_tensor_data(ndarray)
+        of.write_tensor_data(ftype, model, concurrency)
 
         of.close()
 
@@ -1152,9 +1167,9 @@ class OutputFile:
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
     wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
-    if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
+    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
         return GGMLFileType.AllF32
-    if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
+    if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
         return GGMLFileType.MostlyF16
     if output_type_str == "q8_0":
         return GGMLFileType.MostlyQ8_0
@@ -1169,7 +1184,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 
-def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
+def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
     should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
@@ -1195,7 +1210,11 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     for name, lazy_tensor in model.items():
         tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
         if name_new is None:
-            raise Exception(f"Unexpected tensor name: {name}")
+            if skip_unknown:
+                print(f"Unexpected tensor name: {name} - skipping")
+                continue
+            else:
+                raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1274,38 +1293,35 @@ 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.files: dict[str, Path | None] = {
-            "tokenizer.model": None,
-            "vocab.json": None,
-            "tokenizer.json": None,
-        }
-        self._detect_files()
+        self.file_paths = self._detect_files()
+        print(f"Found vocab files: {self.file_paths}")
 
-    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}")
+    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 _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}")
+        return {vt: locate(f) for vt, f in self._FILES.items()}
 
-    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
-        load_merges = vocabtype == "bpe"
+    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 _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab:
+        load_merges = vocab.name == "bpe"
         n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
         return gguf.SpecialVocab(
             model_parent_path,
@@ -1314,30 +1330,34 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    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}'")
+    def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
+        vocab_type, path = self._select_file(vocab_types)
+        print(f"Loading vocab file {path!r}, type {vocab_type!r}")
 
         added_tokens_path = path.parent / "added_tokens.json"
+        if vocab_type == "bpe":
+            return BpeVocab(
+                path, added_tokens_path if added_tokens_path.exists() else None
+            )
+        if vocab_type == "spm":
+            return SentencePieceVocab(
+                path, added_tokens_path if added_tokens_path.exists() else None
+            )
+        if vocab_type == "hfft":
+            return HfVocab(
+                path.parent, added_tokens_path if added_tokens_path.exists() else None
+            )
+        raise ValueError(vocab_type)
+
+    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
         vocab: Vocab
-        if vocabtype == "bpe":
-            vocab = BpeVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        elif vocabtype == "spm":
-            vocab = SentencePieceVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        elif vocabtype == "hfft":
-            vocab = HfVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
+        if len(vocab_types) == 1 and "no_vocab" in vocab_types:
+            vocab = NoVocab()
         else:
-            raise ValueError(f"Unsupported vocabulary type {vocabtype}")
+            vocab = self._create_vocab_by_path(vocab_types)
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
-            vocabtype,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1371,35 +1391,27 @@ 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")
-    vocab_types = ["spm", "bpe", "hfft"]
-    parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
-    parser.add_argument("--awq-path",    type=Path,              help="Path to scale awq cache file", default=None)
-    parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
-    parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
-    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
-    parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--vocab-type",  choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
-    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
-    parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
-    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
-    parser.add_argument("--big-endian",  action="store_true",    help="model is executed on big endian machine")
-    parser.add_argument("--pad-vocab",   action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
+    parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
+    parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
+    parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--no-vocab",     action="store_true",    help="store model without 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("--outfile",      type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency",  type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
+    parser.add_argument("--big-endian",   action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
-    if args.awq_path:
-        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.no_vocab:
+        if args.vocab_only:
+            raise ValueError("no need to specify --vocab-only if using --no-vocab")
+        args.vocab_type = "no_vocab"
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1439,7 +1451,7 @@ 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, model_parent_path)
+    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type.split(","), model_parent_path)
 
     if args.vocab_only:
         if not args.outfile:
@@ -1450,14 +1462,14 @@ def main(args_in: list[str] | None = None) -> None:
         print(f"Wrote {outfile}")
         return
 
-    if model_plus.vocab is not None and args.vocab_dir is None:
+    if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
         vocab = model_plus.vocab
 
     print(f"Vocab info: {vocab}")
     print(f"Special vocab info: {special_vocab}")
 
     model   = model_plus.model
-    model   = convert_model_names(model, params)
+    model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
     outfile = args.outfile or default_outfile(model_plus.paths, ftype)

examples/CMakeLists.txt

@@ -20,6 +20,8 @@ else()
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(finetune)
+    add_subdirectory(gritlm)
+    add_subdirectory(gguf-split)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
@@ -38,6 +40,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
+    add_subdirectory(gguf)
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)

examples/baby-llama/baby-llama.cpp

@@ -1533,27 +1533,28 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        ggml_cgraph gf = {};
+        struct ggml_cgraph * gf = NULL;
+        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
         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_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_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);
 
@@ -1600,13 +1601,14 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            ggml_cgraph gf = {};
+            struct ggml_cgraph * gf = NULL;
+            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
             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

@@ -32,16 +32,15 @@ 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] [MMQ] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <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 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        printf("  example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 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, };
@@ -65,30 +64,27 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 6) {
-        mmq = std::atoi(argv[5]);
+        n_pp = parse_list(argv[5]);
     }
 
     if (argc >= 7) {
-        n_pp = parse_list(argv[6]);
+        n_tg = parse_list(argv[6]);
     }
 
     if (argc >= 8) {
-        n_tg = parse_list(argv[7]);
-    }
-
-    if (argc >= 9) {
-        n_pl = parse_list(argv[8]);
+        n_pl = parse_list(argv[7]);
     }
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
     llama_model_params model_params = llama_model_default_params();
 
-    const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
+    const std::vector<float> t_split(llama_max_devices(), 0.0f);
 
     model_params.n_gpu_layers = n_gpu_layers;
     model_params.tensor_split = t_split.data();
@@ -105,11 +101,13 @@ int main(int argc, char ** argv) {
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = 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) {
@@ -140,6 +138,8 @@ 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 +158,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, 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("\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 +179,10 @@ int main(int argc, char ** argv) {
 
                 llama_batch_clear(batch);
 
-                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);
+                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);
+                    }
                 }
                 batch.logits[batch.n_tokens - 1] = true;
 
@@ -197,7 +197,7 @@ 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, 0, pp);
+                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

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

examples/batched/batched.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -79,6 +80,7 @@ int main(int argc, char ** argv) {
     ctx_params.seed  = 1234;
     ctx_params.n_ctx = 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;
 
@@ -91,7 +93,7 @@ int main(int argc, char ** argv) {
 
     const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
@@ -131,7 +133,7 @@ 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, 0, batch.n_tokens);
+        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
     }
 
     if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -119,7 +119,8 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/benchmark/benchmark-matmult.cpp

@@ -189,12 +189,10 @@ 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], hist_cur.data(), nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -207,7 +205,7 @@ 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], hist_cur.data(), nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

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

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

examples/embedding/embedding.cpp

@@ -7,6 +7,52 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static std::vector<std::string> split_lines(const std::string & s) {
+    std::string line;
+    std::vector<std::string> lines;
+    std::stringstream ss(s);
+    while (std::getline(ss, line)) {
+        lines.push_back(line);
+    }
+    return lines;
+}
+
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
+    }
+}
+
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+    // clear previous kv_cache values (irrelevant for embeddings)
+    llama_kv_cache_clear(ctx);
+
+    // run model
+    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
+    }
+
+    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);
+    }
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -29,7 +75,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -55,49 +102,103 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    int n_past = 0;
+    // split the prompt into lines
+    std::vector<std::string> prompts = split_lines(params.prompt);
 
-    // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch >= params.n_ctx);
 
+    // tokenize the prompts and trim
+    std::vector<std::vector<int32_t>> inputs;
+    for (const auto & prompt : prompts) {
+        auto inp = ::llama_tokenize(ctx, prompt, true, false);
+        if (inp.size() > n_batch) {
+            inp.resize(n_batch);
+        }
+        inputs.push_back(inp);
+    }
+
+    // add eos if not present
+    for (auto & inp : inputs) {
+        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+            inp.push_back(llama_token_eos(model));
+        }
+    }
+
+    // tokenization stats
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-        for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        for (int i = 0; i < (int) inputs.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            for (int j = 0; j < (int) inputs[i].size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
         }
-        fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)n_ctx) {
-        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), n_ctx);
-        return 1;
-    }
-
-    while (!embd_inp.empty()) {
-        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return 1;
-        }
-        n_past += n_tokens;
-        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
-    }
+    // initialize batch
+    const int n_prompts = prompts.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
+    // allocate output
     const int n_embd = llama_n_embd(model);
-    const auto * embeddings = llama_get_embeddings(ctx);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    float * emb = embeddings.data();
 
-    for (int i = 0; i < n_embd; i++) {
-        printf("%f ", embeddings[i]);
+    // break into batches
+    int p = 0; // number of prompts processed already
+    int s = 0; // number of prompts in current batch
+    for (int k = 0; k < n_prompts; k++) {
+        // clamp to n_batch tokens
+        auto & inp = inputs[k];
+
+        const uint64_t n_toks = inp.size();
+
+        // encode if at capacity
+        if (batch.n_tokens + n_toks > n_batch) {
+            float * out = emb + p * n_embd;
+            batch_decode(ctx, batch, out, s, n_embd);
+            llama_batch_clear(batch);
+            p += s;
+            s = 0;
+        }
+
+        // add to batch
+        batch_add_seq(batch, inp, s);
+        s += 1;
     }
-    printf("\n");
 
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
+
+    // print 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, "%9.6f ", 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");
+    }
+
+    // clean up
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/export-lora/export-lora.cpp

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

examples/finetune/README.md

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

examples/finetune/finetune.cpp

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

examples/gguf-split/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET gguf-split)
+add_executable(${TARGET} gguf-split.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/gguf-split/README.md

@@ -0,0 +1,9 @@
+## GGUF split Example
+
+CLI to split / merge GGUF files.
+
+**Command line options:**
+
+- `--split`: split GGUF to multiple GGUF, default operation.
+- `--split-max-tensors`: maximum tensors in each split: default(128)
+- `--merge`: merge multiple GGUF to a single GGUF.

examples/gguf-split/gguf-split.cpp

@@ -0,0 +1,489 @@
+#include "llama.h"
+#include "ggml.h"
+#include "common.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <fstream>
+#include <ios>
+#include <string>
+#include <vector>
+
+#include <stdio.h>
+#include <fcntl.h>
+#include <string.h>
+
+enum split_operation : uint8_t {
+    SPLIT_OP_SPLIT,
+    SPLIT_OP_MERGE,
+};
+
+static const char * const LLM_KV_GENERAL_SPLIT_I_SPLIT = "general.split";
+static const char * const LLM_KV_GENERAL_SPLIT_N_SPLIT = "general.split_count";
+
+static const int SPLIT_FILENAME_MAX = 256;
+
+static const char * const SPLIT_FILENAME_FORMAT = "%s-%05d-of-%05d.gguf";
+
+struct split_params {
+    split_operation operation = SPLIT_OP_SPLIT;
+    int n_split_tensors = 128;
+    std::string input;
+    std::string output;
+};
+
+static void split_print_usage(const char * executable) {
+    const split_params default_params;
+    printf("\n");
+    printf("usage: %s [options] GGUF_IN GGUF_OUT\n", executable);
+    printf("\n");
+    printf("Apply a GGUF operation on IN to OUT.");
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  --version             show version and build info\n");
+    printf("  --split               split GGUF to multiple GGUF (default)\n");
+    printf("  --split-max-tensors   max tensors in each split: default(%d)\n", default_params.n_split_tensors);
+    printf("  --merge               merge multiple GGUF to a single GGUF\n");
+    printf("\n");
+}
+
+static bool split_params_parse_ex(int argc, const char ** argv, split_params & params) {
+    std::string arg;
+    const std::string arg_prefix = "--";
+    bool invalid_param = false;
+
+    int arg_idx = 1;
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        arg = argv[arg_idx];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        bool arg_found = false;
+        if (arg == "-h" || arg == "--help") {
+            split_print_usage(argv[0]);
+            exit(0);
+        }
+        if (arg == "--version") {
+            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            exit(0);
+        }
+
+        if (arg == "--merge") {
+            arg_found = true;
+            params.operation = SPLIT_OP_MERGE;
+        }
+        if (arg == "--split") {
+            arg_found = true;
+            params.operation = SPLIT_OP_SPLIT;
+        }
+        if (arg == "--split-max-tensors") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            params.n_split_tensors = atoi(argv[arg_idx]);
+        }
+
+        if (!arg_found) {
+            throw std::invalid_argument("error: unknown argument: " + arg);
+        }
+    }
+
+    if (invalid_param) {
+        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
+    }
+
+    if (argc - arg_idx < 2) {
+        printf("%s: bad arguments\n", argv[0]);
+        split_print_usage(argv[0]);
+        return false;
+    }
+
+    params.input = argv[arg_idx++];
+    params.output = argv[arg_idx++];
+
+    return true;
+}
+
+static bool split_params_parse(int argc, const char ** argv, split_params & params) {
+    bool result = true;
+    try {
+        if (!split_params_parse_ex(argc, argv, params)) {
+            split_print_usage(argv[0]);
+            exit(1);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        split_print_usage(argv[0]);
+        exit(1);
+    }
+    return result;
+}
+
+static void zeros(std::ofstream & file, size_t n) {
+    char zero = 0;
+    for (size_t i = 0; i < n; ++i) {
+        file.write(&zero, 1);
+    }
+}
+
+static std::string split_file_name(const std::string & path, int i_split, int n_split) {
+    char f_split[SPLIT_FILENAME_MAX] = {0};
+    snprintf(f_split, sizeof(f_split), SPLIT_FILENAME_FORMAT, path.c_str(), i_split + 1, n_split);
+    return std::string(f_split);
+}
+
+struct split_strategy {
+    const split_params params;
+    std::ifstream & f_input;
+    struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_meta = NULL;
+    const int n_tensors;
+
+    const int n_split;
+    int i_split = 0;
+
+    int i_tensor = 0;
+
+    std::vector<uint8_t> read_data;
+
+    struct gguf_context * ctx_out;
+    std::ofstream fout;
+
+    split_strategy(const split_params & params,
+            std::ifstream & f_input,
+            struct gguf_context * ctx_gguf,
+            struct ggml_context * ctx_meta) :
+        params(params),
+        f_input(f_input),
+        ctx_gguf(ctx_gguf),
+        ctx_meta(ctx_meta),
+        n_tensors(gguf_get_n_tensors(ctx_gguf)),
+        n_split(std::ceil(1. * n_tensors / params.n_split_tensors)) {
+        }
+
+    bool should_split() const {
+        return i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
+    }
+
+    void split_start() {
+        ctx_out = gguf_init_empty();
+
+        // Save all metadata in first split only
+        if (i_split == 0) {
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+        gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_I_SPLIT, i_split);
+        gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, n_split);
+
+        // populate the original tensors, so we get an initial metadata
+        for (int i = i_split * params.n_split_tensors; i < n_tensors && i < (i_split + 1) * params.n_split_tensors; ++i) {
+            struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+            gguf_add_tensor(ctx_out, meta);
+        }
+
+        auto split_name = split_file_name(params.output, i_split, n_split);
+
+        fprintf(stderr, "%s: %s ...", __func__, split_name.c_str());
+        fout = std::ofstream(split_name, std::ios::binary);
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+        auto meta_size = gguf_get_meta_size(ctx_out);
+
+        // placeholder for the meta data
+        ::zeros(fout, meta_size);
+
+        i_split++;
+    }
+
+    void next_tensor() {
+        const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+        struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+        auto n_bytes = ggml_nbytes(t);
+
+        if (read_data.size() < n_bytes) {
+            read_data.resize(n_bytes);
+        }
+
+        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
+        f_input.seekg(offset);
+        f_input.read((char *)read_data.data(), n_bytes);
+
+        t->data = read_data.data();
+
+        // write tensor data + padding
+        fout.write((const char *)t->data, n_bytes);
+        zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+
+        i_tensor++;
+    }
+
+    void split_end() {
+        // go back to beginning of file and write the updated metadata
+        fout.seekp(0);
+        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+        gguf_get_meta_data(ctx_out, data.data());
+        fout.write((const char *)data.data(), data.size());
+
+        fout.close();
+        gguf_free(ctx_out);
+
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+};
+
+static void gguf_split(const split_params & split_params) {
+    struct ggml_context * ctx_meta = NULL;
+
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ &ctx_meta,
+    };
+
+    std::ifstream f_input(split_params.input.c_str(), std::ios::binary);
+    if (!f_input.is_open()) {
+        fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_params.input.c_str());
+        exit(1);
+    }
+
+    auto * ctx_gguf = gguf_init_from_file(split_params.input.c_str(), params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+        exit(1);
+    }
+
+    split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
+            __func__, split_params.input.c_str(),
+            split_file_name(split_params.output, strategy.i_split, strategy.n_split).c_str(),
+            split_params.n_split_tensors);
+
+    strategy.split_start();
+
+    while (strategy.i_tensor < strategy.n_tensors) {
+        strategy.next_tensor();
+        if (strategy.should_split()) {
+            strategy.split_end();
+            strategy.split_start();
+        }
+    }
+    strategy.split_end();
+
+    gguf_free(ctx_gguf);
+    f_input.close();
+
+    fprintf(stderr, "%s: %d gguf split written with a total of %d tensors.\n",
+            __func__, strategy.n_split, strategy.n_tensors);
+}
+
+static void gguf_merge(const split_params & split_params) {
+    fprintf(stderr, "%s: %s -> %s\n",
+            __func__, split_params.input.c_str(),
+            split_params.output.c_str());
+    int n_split = 1;
+    int total_tensors = 0;
+
+    auto * ctx_out = gguf_init_empty();
+    std::ofstream fout(split_params.output.c_str(), std::ios::binary);
+    fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+    std::vector<uint8_t> read_data;
+    std::vector<ggml_context *> ctx_metas;
+    std::vector<gguf_context *> ctx_ggufs;
+
+    std::string split_prefix;
+
+    // First pass to find KV and tensors metadata
+    for (int i_split = 0; i_split < n_split; i_split++) {
+        struct ggml_context * ctx_meta = NULL;
+
+        struct gguf_init_params params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ &ctx_meta,
+        };
+
+        auto split_name = split_params.input;
+        if (i_split > 0) {
+            split_name = split_file_name(split_prefix, i_split, n_split);
+        }
+        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_name.c_str());
+
+        auto * ctx_gguf = gguf_init_from_file(split_name.c_str(), params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+            exit(1);
+        }
+        ctx_ggufs.push_back(ctx_gguf);
+        ctx_metas.push_back(ctx_meta);
+
+        if (i_split == 0) {
+            auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_GENERAL_SPLIT_N_SPLIT);
+            if (key_n_split < 0) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain %s metadata\n",
+                        __func__,
+                        LLM_KV_GENERAL_SPLIT_N_SPLIT);
+                gguf_free(ctx_gguf);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+
+            n_split = gguf_get_val_u8(ctx_gguf, key_n_split);
+            if (n_split < 1) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain a valid split count %d\n",
+                        __func__,
+                        n_split);
+                gguf_free(ctx_gguf);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+
+            // Do not trigger merge if we try to merge again the output
+            gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, 0);
+
+            // Set metadata from the first split
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+
+        // Verify the file naming
+        {
+            int i_split_file = 0;
+            int n_split_file = 0;
+            const char * i_split_format = "-00000-of-00000.gguf";
+
+            if (split_name.size() < strlen(i_split_format)) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s\n", __func__, split_params.input.c_str());
+                for (auto * _ctx_gguf : ctx_ggufs) {
+                    gguf_free(_ctx_gguf);
+                }
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+
+            split_prefix = split_name.substr(0, split_name.size() - strlen(i_split_format));
+
+            const char * split_name_c_str = split_name.c_str();
+            int n_part = sscanf(&split_name_c_str[0] + split_prefix.size(), "-%d-of-%d", &i_split_file, &n_split_file);
+
+            if (n_part != 2 || i_split_file - 1 != i_split || n_split_file != n_split) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s"
+                                " i_split=%d i_split_file=%d"
+                                " n_split=%d n_split_file=%d\n", __func__,
+                        split_params.input.c_str(),
+                        i_split, i_split_file,
+                        n_split, n_split_file);
+                for (auto * _ctx_gguf : ctx_ggufs) {
+                    gguf_free(_ctx_gguf);
+                }
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+        }
+
+        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
+        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
+            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+            gguf_add_tensor(ctx_out, t);
+        }
+        total_tensors += n_tensors;
+
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+
+    // placeholder for the meta data
+    {
+        auto meta_size = gguf_get_meta_size(ctx_out);
+        ::zeros(fout, meta_size);
+    }
+
+    // Write tensors data
+    for (int i_split = 0; i_split < n_split; i_split++) {
+        auto split_name = split_file_name(split_prefix, i_split, n_split);
+        std::ifstream f_input(split_name.c_str(), std::ios::binary);
+        if (!f_input.is_open()) {
+            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_name.c_str());
+            for (auto * _ctx_gguf : ctx_ggufs) {
+                gguf_free(_ctx_gguf);
+            }
+            gguf_free(ctx_out);
+            fout.close();
+            exit(1);
+        }
+        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_name.c_str());
+
+        auto * ctx_gguf = ctx_ggufs[i_split];
+        auto * ctx_meta = ctx_metas[i_split];
+
+        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
+        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
+            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+
+            auto n_bytes = ggml_nbytes(t);
+
+            if (read_data.size() < n_bytes) {
+                read_data.resize(n_bytes);
+            }
+
+            auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
+            f_input.seekg(offset);
+            f_input.read((char *)read_data.data(), n_bytes);
+
+            // write tensor data + padding
+            fout.write((const char *)read_data.data(), n_bytes);
+            zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+        }
+
+        gguf_free(ctx_gguf);
+        ggml_free(ctx_meta);
+        f_input.close();
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+
+    {
+        // go back to beginning of file and write the updated metadata
+        fout.seekp(0);
+        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+        gguf_get_meta_data(ctx_out, data.data());
+        fout.write((const char *)data.data(), data.size());
+
+        fout.close();
+        gguf_free(ctx_out);
+    }
+
+    fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
+            __func__, split_params.output.c_str(), n_split, total_tensors);
+}
+
+int main(int argc, const char ** argv) {
+    if (argc < 3) {
+        split_print_usage(argv[0]);
+    }
+
+    split_params params;
+    split_params_parse(argc, argv, params);
+
+    switch (params.operation) {
+        case SPLIT_OP_SPLIT: gguf_split(params);
+            break;
+        case SPLIT_OP_MERGE: gguf_merge(params);
+            break;
+        default:split_print_usage(argv[0]);
+            exit(1);
+    }
+
+    return 0;
+}

examples/gguf/gguf.cpp

@@ -211,6 +211,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
                         fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
+                        gguf_free(ctx);
                         return false;
                     }
                 }

examples/gritlm/CMakeLists.txt

@@ -0,0 +1,5 @@
+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/README.md

@@ -0,0 +1,62 @@
+## Generative Representational Instruction Tuning (GRIT) Example
+[gritlm] a model which can generate embeddings as well as "normal" text
+generation depending on the instructions in the prompt.
+
+* Paper: https://arxiv.org/pdf/2402.09906.pdf
+
+### Retrieval-Augmented Generation (RAG) use case
+One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
+that we take documents that we want to use as context, to ground the large
+language model (LLM), and we create token embeddings for them. We then store
+these token embeddings in a vector database.
+
+When we perform a query, prompt the LLM, we will first create token embeddings
+for the query and then search the vector database to retrieve the most
+similar vectors, and return those documents so they can be passed to the LLM as
+context. Then the query and the context will be passed to the LLM which will
+have to _again_ create token embeddings for the query. But because gritlm is used
+the first query can be cached and the second query tokenization generation does
+not have to be performed at all.
+
+### Running the example
+Download a Grit model:
+```console
+$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf
+```
+
+Run the example using the downloaded model:
+```console
+$ ./gritlm -m gritlm-7b_q4_1.gguf
+
+Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
+Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
+Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
+Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
+
+Oh, brave adventurer, who dared to climb
+The lofty peak of Mt. Fuji in the night,
+When shadows lurk and ghosts do roam,
+And darkness reigns, a fearsome sight.
+
+Thou didst set out, with heart aglow,
+To conquer this mountain, so high,
+And reach the summit, where the stars do glow,
+And the moon shines bright, up in the sky.
+
+Through the mist and fog, thou didst press on,
+With steadfast courage, and a steadfast will,
+Through the darkness, thou didst not be gone,
+But didst climb on, with a steadfast skill.
+
+At last, thou didst reach the summit's crest,
+And gazed upon the world below,
+And saw the beauty of the night's best,
+And felt the peace, that only nature knows.
+
+Oh, brave adventurer, who dared to climb
+The lofty peak of Mt. Fuji in the night,
+Thou art a hero, in the eyes of all,
+For thou didst conquer this mountain, so bright.
+```
+
+[gritlm]: https://github.com/ContextualAI/gritlm

examples/gritlm/gritlm.cpp

@@ -0,0 +1,215 @@
+#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

@@ -36,6 +36,8 @@ public:
     void set_parameters(StatParams&& params) { m_params = std::move(params); }
     bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
     void save_imatrix() const;
+    bool load_imatrix(const char * file_name, bool add);
+    static bool load_imatrix(const char * file_name, std::unordered_map<std::string, Stats>& imatrix);
 private:
     std::unordered_map<std::string, Stats> m_stats;
     StatParams                             m_params;
@@ -54,13 +56,31 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
 
+    std::string wname;
+    {
+        // remove any prefix and suffixes from the name
+        // CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+        const char * p = strchr(src0->name, '#');
+        if (p != NULL) {
+            p = p + 1;
+            const char * q = strchr(p, '#');
+            if (q != NULL) {
+                wname = std::string(p, q - p);
+            } else {
+                wname = p;
+            }
+        } else {
+            wname = src0->name;
+        }
+    }
+
     // when ask is true, the scheduler wants to know if we are interested in data from this tensor
     // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
     if (ask) {
         if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
         if (t->op != GGML_OP_MUL_MAT) return false;
         if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
-        if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
+        if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
         return true;
     }
 
@@ -92,12 +112,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         // this is necessary to guarantee equal number of "ncall" for each tensor
         for (int ex = 0; ex < n_as; ++ex) {
             src0 = t->src[2 + ex];
-            auto& e = m_stats[src0->name];
+            auto& e = m_stats[wname];
             if (e.values.empty()) {
                 e.values.resize(src1->ne[0], 0);
             }
             else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
                 exit(1); //GGML_ASSERT(false);
             }
             // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
@@ -105,7 +125,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
             ++e.ncall;
             if (m_params.verbosity > 1) {
-                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
             }
             for (int row = 0; row < (int)src1->ne[1]; ++row) {
                 const int excur = m_ids[row*n_as + idx];
@@ -127,17 +147,17 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             }
         }
     } else {
-        auto& e = m_stats[src0->name];
+        auto& e = m_stats[wname];
         if (e.values.empty()) {
             e.values.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
             exit(1); //GGML_ASSERT(false);
         }
         ++e.ncall;
         if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
         }
         for (int row = 0; row < (int)src1->ne[1]; ++row) {
             const float * x = data + row * src1->ne[0];
@@ -189,6 +209,57 @@ void IMatrixCollector::save_imatrix(const char * fname) const {
     }
 }
 
+bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_map<std::string, Stats>& imatrix_data) {
+    std::ifstream in(imatrix_file, std::ios::binary);
+    if (!in) {
+        printf("%s: failed to open %s\n",__func__,imatrix_file);
+        return false;
+    }
+    int n_entries;
+    in.read((char*)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        printf("%s: no data in file %s\n", __func__, imatrix_file);
+        return false;
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len; in.read((char *)&len, sizeof(len));
+        std::vector<char> name_as_vec(len+1);
+        in.read((char *)name_as_vec.data(), len);
+        if (in.fail()) {
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file);
+            return false;
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto& e = imatrix_data[std::move(name)];
+        int ncall;
+        in.read((char*)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char *)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return false;
+        }
+        e.values.resize(nval);
+        in.read((char*)e.values.data(), nval*sizeof(float));
+        if (in.fail()) {
+            printf("%s: failed reading data for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return false;
+        }
+        e.ncall = ncall;
+    }
+    return true;
+}
+
+bool IMatrixCollector::load_imatrix(const char * file_name, bool add) {
+    if (!add) {
+        m_stats.clear();
+    }
+    return load_imatrix(file_name, m_stats);
+}
+
 static IMatrixCollector g_collector;
 
 static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
@@ -269,7 +340,7 @@ static void process_logits(
     }
 }
 
-static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl) {
+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);
@@ -282,6 +353,15 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
     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 (from_chunk > 0) {
+        if (size_t((from_chunk + 2)*n_ctx) >= 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);
+    }
+
     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);
@@ -402,7 +482,10 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
 int main(int argc, char ** argv) {
 
     StatParams sparams;
+    std::string prev_result_file;
+    std::string combine_files;
     bool compute_ppl = true;
+    int  from_chunk  = 0;
     std::vector<char*> args;
     args.push_back(argv[0]);
     int iarg = 1;
@@ -423,6 +506,13 @@ int main(int argc, char ** argv) {
             compute_ppl = false;
         } else if (arg == "--keep-imatrix") {
             sparams.keep_every = std::stoi(argv[++iarg]);
+        } else if (arg == "--continue-from") {
+            prev_result_file = argv[++iarg];
+        } else if (arg == "--combine") {
+            combine_files = argv[++iarg];
+        }
+        else if (arg == "--from-chunk") {
+            from_chunk = std::stoi(argv[++iarg]);
         } else {
             args.push_back(argv[iarg]);
         }
@@ -436,14 +526,50 @@ int main(int argc, char ** argv) {
         }
     }
 
+    g_collector.set_parameters(std::move(sparams));
+
+    if (!combine_files.empty()) {
+        std::vector<std::string> files;
+        size_t pos = 0;
+        while (true) {
+            auto new_pos = combine_files.find(',', pos);
+            if (new_pos != std::string::npos) {
+                files.emplace_back(combine_files.substr(pos, new_pos - pos));
+                pos = new_pos + 1;
+            } else {
+                files.emplace_back(combine_files.substr(pos));
+                break;
+            }
+        }
+        if (files.size() < 2) {
+            fprintf(stderr, "You must provide at least two comma separated files to use --combine\n");
+            return 1;
+        }
+        printf("Combining the following %d files\n", int(files.size()));
+        for (auto& file : files) {
+            printf("    %s\n", file.c_str());
+            if (!g_collector.load_imatrix(file.c_str(), true)) {
+                fprintf(stderr, "Failed to load %s\n", file.c_str());
+                return 1;
+            }
+        }
+        g_collector.save_imatrix();
+        return 0;
+    }
+
+    if (!prev_result_file.empty()) {
+        if (!g_collector.load_imatrix(prev_result_file.c_str(), false)) {
+            fprintf(stderr, "=============== Failed to load %s\n", prev_result_file.c_str());
+            return 1;
+        }
+    }
+
     gpt_params params;
     params.n_batch = 512;
     if (!gpt_params_parse(args.size(), args.data(), params)) {
         return 1;
     }
 
-    g_collector.set_parameters(std::move(sparams));
-
     params.logits_all = true;
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
@@ -460,7 +586,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model_params mparams = llama_model_params_from_gpt_params(params);
 
@@ -495,7 +622,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    bool OK = compute_imatrix(ctx, params, compute_ppl);
+    bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);
     if (!OK) {
         return 1;
     }

examples/infill/infill.cpp

@@ -202,7 +202,8 @@ int main(int argc, char ** argv) {
     std::mt19937 rng(params.seed);
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -377,10 +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";
         }
@@ -446,8 +447,8 @@ int main(int argc, char ** argv) {
                 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);
 
-                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);
+                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);
 
                 n_past -= n_discard;
 

examples/json-schema-to-grammar.py

@@ -87,7 +87,21 @@ 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')
-            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
+            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'
             return self._add_rule(rule_name, rule)
 
         else:

examples/llama-bench/README.md

@@ -23,19 +23,22 @@ usage: ./llama-bench [options]
 
 options:
   -h, --help
-  -m, --model <filename>            (default: models/7B/ggml-model-q4_0.gguf)
-  -p, --n-prompt <n>                (default: 512)
-  -n, --n-gen <n>                   (default: 128)
-  -b, --batch-size <n>              (default: 512)
-  --memory-f32 <0|1>                (default: 0)
-  -t, --threads <n>                 (default: 16)
-  -ngl N, --n-gpu-layers <n>        (default: 99)
-  -mg i, --main-gpu <i>             (default: 0)
-  -mmq, --mul-mat-q <0|1>           (default: 1)
-  -ts, --tensor_split <ts0/ts1/..>
-  -r, --repetitions <n>             (default: 5)
-  -o, --output <csv|json|md|sql>    (default: md)
-  -v, --verbose                     (default: 0)
+  -m, --model <filename>              (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                  (default: 512)
+  -n, --n-gen <n>                     (default: 128)
+  -b, --batch-size <n>                (default: 512)
+  -ctk <t>, --cache-type-k <t>        (default: f16)
+  -ctv <t>, --cache-type-v <t>        (default: f16)
+  -t, --threads <n>                   (default: 112)
+  -ngl, --n-gpu-layers <n>            (default: 99)
+  -sm, --split-mode <none|layer|row>  (default: layer)
+  -mg, --main-gpu <i>                 (default: 0)
+  -nkvo, --no-kv-offload <0|1>        (default: 0)
+  -mmp, --mmap <0|1>                  (default: 1)
+  -ts, --tensor_split <ts0/ts1/..>    (default: 0)
+  -r, --repetitions <n>               (default: 5)
+  -o, --output <csv|json|md|sql>      (default: md)
+  -v, --verbose                       (default: 0)
 
 Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
 ```
@@ -51,6 +54,10 @@ Each test is repeated the number of times given by `-r`, and the results are ave
 
 For a description of the other options, see the [main example](../main/README.md).
 
+Note:
+
+- When using SYCL backend, there would be hang issue in some cases. Please set `--mmp 0`.
+
 ## Examples
 
 ### Text generation with different models

examples/llama-bench/llama-bench.cpp

@@ -8,6 +8,7 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <cstdlib>
 #include <iterator>
 #include <map>
 #include <numeric>
@@ -20,6 +21,7 @@
 #include "llama.h"
 #include "common.h"
 #include "ggml-cuda.h"
+#include "ggml-sycl.h"
 
 // utils
 static uint64_t get_time_ns() {
@@ -102,6 +104,7 @@ static std::string get_cpu_info() {
                 }
             }
         }
+        fclose(f);
     }
 #endif
     // TODO: other platforms
@@ -111,10 +114,21 @@ static std::string get_cpu_info() {
 static std::string get_gpu_info() {
     std::string id;
 #ifdef GGML_USE_CUBLAS
-    int count = ggml_cuda_get_device_count();
+    int count = ggml_backend_cuda_get_device_count();
     for (int i = 0; i < count; i++) {
         char buf[128];
-        ggml_cuda_get_device_description(i, buf, sizeof(buf));
+        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#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) {
             id += "/";
@@ -140,9 +154,9 @@ static const char * output_format_str(output_formats format) {
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_NONE:  return "none";
-        case LLAMA_SPLIT_LAYER: return "layer";
-        case LLAMA_SPLIT_ROW:   return "row";
+        case LLAMA_SPLIT_MODE_NONE:  return "none";
+        case LLAMA_SPLIT_MODE_LAYER: return "layer";
+        case LLAMA_SPLIT_MODE_ROW:   return "row";
         default: GGML_ASSERT(!"invalid split mode");
     }
 }
@@ -152,6 +166,7 @@ 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;
@@ -159,8 +174,9 @@ 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::array<float, LLAMA_MAX_DEVICES>> tensor_split;
+    std::vector<std::vector<float>> tensor_split;
+    std::vector<bool> use_mmap;
+    std::vector<bool> embeddings;
     int reps;
     bool verbose;
     output_formats output_format;
@@ -170,16 +186,18 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_batch       */ {512},
+    /* n_batch       */ {2048},
+    /* n_ubatch      */ {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_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
-    /* mul_mat_q     */ {true},
-    /* tensor_split  */ {{}},
+    /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
+    /* use_mmap      */ {true},
+    /* embeddings    */ {false},
     /* reps          */ 5,
     /* verbose       */ false,
     /* output_format */ MARKDOWN
@@ -194,6 +212,7 @@ 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());
@@ -201,8 +220,9 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
     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("  -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("  -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("  -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");
@@ -282,6 +302,13 @@ 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;
@@ -338,11 +365,11 @@ 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_NONE;
+                    mode = LLAMA_SPLIT_MODE_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_LAYER;
+                    mode = LLAMA_SPLIT_MODE_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_ROW;
+                    mode = LLAMA_SPLIT_MODE_ROW;
                 } else {
                     invalid_param = true;
                     break;
@@ -363,13 +390,20 @@ 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") {
+        } else if (arg == "-mmp" || arg == "--mmap") {
             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());
+            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;
@@ -380,10 +414,10 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 const std::regex regex{R"([;/]+)"};
                 std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
                 std::vector<std::string> split_arg{it, {}};
-                GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+                GGML_ASSERT(split_arg.size() <= llama_max_devices());
 
-                std::array<float, LLAMA_MAX_DEVICES> tensor_split;
-                for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+                std::vector<float> tensor_split(llama_max_devices());
+                for (size_t i = 0; i < llama_max_devices(); ++i) {
                     if (i < split_arg.size()) {
                         tensor_split[i] = std::stof(split_arg[i]);
                     } else {
@@ -433,14 +467,16 @@ 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;
@@ -451,6 +487,7 @@ 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;
@@ -458,8 +495,9 @@ struct cmd_params_instance {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
-    bool mul_mat_q;
-    std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+    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();
@@ -468,6 +506,7 @@ struct cmd_params_instance {
         mparams.split_mode = split_mode;
         mparams.main_gpu = main_gpu;
         mparams.tensor_split = tensor_split.data();
+        mparams.use_mmap = use_mmap;
 
         return mparams;
     }
@@ -477,6 +516,7 @@ struct cmd_params_instance {
                n_gpu_layers == other.n_gpu_layers &&
                split_mode == other.split_mode &&
                main_gpu == other.main_gpu &&
+               use_mmap == other.use_mmap &&
                tensor_split == other.tensor_split;
     }
 
@@ -485,10 +525,11 @@ struct cmd_params_instance {
 
         cparams.n_ctx = 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;
     }
@@ -503,10 +544,12 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & sm : params.split_mode)
     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) {
@@ -518,6 +561,7 @@ 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,
@@ -525,8 +569,9 @@ 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);
         }
@@ -540,6 +585,7 @@ 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,
@@ -547,8 +593,9 @@ 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);
         }
@@ -563,7 +610,9 @@ struct test {
     static const bool cuda;
     static const bool opencl;
     static const bool vulkan;
+    static const bool kompute;
     static const bool metal;
+    static const bool sycl;
     static const bool gpu_blas;
     static const bool blas;
     static const std::string cpu_info;
@@ -573,6 +622,7 @@ 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;
@@ -580,8 +630,9 @@ struct test {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
-    bool mul_mat_q;
-    std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+    std::vector<float> tensor_split;
+    bool use_mmap;
+    bool embeddings;
     int n_prompt;
     int n_gen;
     std::string test_time;
@@ -595,6 +646,7 @@ 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;
@@ -602,8 +654,9 @@ 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
@@ -647,28 +700,36 @@ struct test {
         if (vulkan) {
             return "Vulkan";
         }
+        if (kompute) {
+            return "Kompute";
+        }
         if (metal) {
             return "Metal";
         }
+        if (sycl) {
+            return GGML_SYCL_NAME;
+        }
         if (gpu_blas) {
             return "GPU BLAS";
         }
         if (blas) {
             return "BLAS";
         }
+
         return "CPU";
     }
 
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "vulkan", "metal", "gpu_blas", "blas",
+            "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_threads", "type_k", "type_v",
+            "n_batch", "n_ubatch",
+            "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload",
-            "mul_mat_q", "tensor_split",
+            "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -679,15 +740,17 @@ 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_threads" ||
+        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
+            field == "n_threads" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
-        if (field == "cuda" || field == "opencl"  || field == "vulkan"|| field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
+        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") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -699,7 +762,7 @@ struct test {
     std::vector<std::string> get_values() const {
         std::string tensor_split_str;
         int max_nonzero = 0;
-        for (int i = 0; i < LLAMA_MAX_DEVICES; i++) {
+        for (size_t i = 0; i < llama_max_devices(); i++) {
             if (tensor_split[i] > 0) {
                 max_nonzero = i;
             }
@@ -714,13 +777,15 @@ struct test {
         }
         std::vector<std::string> values = {
             build_commit, std::to_string(build_number),
-            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
+            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(vulkan),
+            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_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            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_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload),
-            std::to_string(mul_mat_q), tensor_split_str,
+            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
             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())
@@ -743,9 +808,11 @@ const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
+const bool        test::kompute      = !!ggml_cpu_has_kompute();
 const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();
+const bool        test::sycl         = !!ggml_cpu_has_sycl();
 const std::string test::cpu_info     = get_cpu_info();
 const std::string test::gpu_info     = get_gpu_info();
 
@@ -882,12 +949,15 @@ 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";
         }
@@ -896,43 +966,49 @@ struct markdown_printer : public printer {
 
     void print_header(const cmd_params & params) override {
         // select fields to print
-        fields.push_back("model");
-        fields.push_back("size");
-        fields.push_back("params");
-        fields.push_back("backend");
+        fields.emplace_back("model");
+        fields.emplace_back("size");
+        fields.emplace_back("params");
+        fields.emplace_back("backend");
         bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
         if (!is_cpu_backend) {
-            fields.push_back("n_gpu_layers");
+            fields.emplace_back("n_gpu_layers");
         }
         if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
-            fields.push_back("n_threads");
+            fields.emplace_back("n_threads");
         }
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
-            fields.push_back("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.push_back("type_k");
+            fields.emplace_back("type_k");
         }
         if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
-            fields.push_back("type_v");
+            fields.emplace_back("type_v");
         }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
-            fields.push_back("main_gpu");
+            fields.emplace_back("main_gpu");
         }
         if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
-            fields.push_back("split_mode");
-        }
-        if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
-            fields.push_back("mul_mat_q");
+            fields.emplace_back("split_mode");
         }
         if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
-            fields.push_back("no_kv_offload");
+            fields.emplace_back("no_kv_offload");
         }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
-            fields.push_back("tensor_split");
+            fields.emplace_back("tensor_split");
         }
-        fields.push_back("test");
-        fields.push_back("t/s");
+        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");
 
         fprintf(fout, "|");
         for (const auto & field : fields) {
@@ -1046,25 +1122,40 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
-    int n_processed = 0;
-
     llama_set_n_threads(ctx, n_threads, n_threads);
 
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token> tokens(n_batch);
+
+    int n_processed = 0;
+
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
+        tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+        for (int i = 1; i < n_tokens; i++) {
+            tokens[i] = std::rand() % n_vocab;
+        }
         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_token token = llama_token_bos(llama_get_model(ctx));
-
     llama_set_n_threads(ctx, n_threads, n_threads);
 
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+
     for (int i = 0; i < n_gen; i++) {
         llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
+        llama_synchronize(ctx);
+        token = std::rand() % n_vocab;
     }
 }
 
@@ -1096,8 +1187,7 @@ int main(int argc, char ** argv) {
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
     }
-    bool numa = false;
-    llama_backend_init(numa);
+    llama_backend_init();
 
     // initialize printer
     std::unique_ptr<printer> p;
@@ -1154,7 +1244,8 @@ int main(int argc, char ** argv) {
 
         // warmup run
         if (t.n_prompt > 0) {
-            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
+            //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);
         }
         if (t.n_gen > 0) {
             test_gen(ctx, 1, 0, t.n_threads);
@@ -1170,6 +1261,7 @@ 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

@@ -21,12 +21,8 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // 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")
-            }
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
         }
         externalNativeBuild {
             cmake {

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

@@ -33,6 +33,45 @@ 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);
@@ -274,8 +313,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
-    llama_backend_init(numa);
+Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+    llama_backend_init();
 }
 
 extern "C"
@@ -295,6 +334,8 @@ 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);
@@ -372,8 +413,16 @@ Java_com_example_llama_Llm_completion_1loop(
     }
 
     auto new_token_chars = llama_token_to_piece(context, new_token_id);
-    LOGi("new_token_chars: `%s`", new_token_chars.c_str());
-    auto new_token = env->NewStringUTF(new_token_chars.c_str());
+    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("");
+    }
 
     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.isEmpty()) {
+                    if (str == null) {
                         break
                     }
                     emit(str)

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

@@ -51,7 +51,7 @@ actor LlamaContext {
     }
 
     static func create_context(path: String) throws -> LlamaContext {
-        llama_backend_init(false)
+        llama_backend_init()
         var model_params = llama_model_default_params()
 
 #if targetEnvironment(simulator)
@@ -221,6 +221,7 @@ actor LlamaContext {
             if llama_decode(context, batch) != 0 {
                 print("llama_decode() failed during prompt")
             }
+            llama_synchronize(context)
 
             let t_pp_end = ggml_time_us()
 
@@ -240,6 +241,7 @@ 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/llava/MobileVLM-README.md

@@ -111,17 +111,71 @@ llama_print_timings:        eval time =    1279.03 ms /    18 runs   (   71.06 m
 llama_print_timings:       total time =   34570.79 ms
 ```
 
+## Orin compile and run
+### compile
+```sh
+make LLAMA_CUBLAS=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
+```
+
+### run on Orin
+### case 1
+**input**
+```sh
+./llava-cli \
+    -m /data/local/tmp/ggml-model-q4_k.gguf \
+    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
+    --image /data/local/tmp/demo.jpeg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:" \
+    --n-gpu-layers 999
+```
+**output**
+```sh
+
+encode_image_with_clip: image encoded in   296.62 ms by CLIP (    2.06 ms per image patch)
+
+ Susan Wise Bauer
+
+llama_print_timings:        load time =    1067.64 ms
+llama_print_timings:      sample time =       1.53 ms /     6 runs   (    0.25 ms per token,  3934.43 tokens per second)
+llama_print_timings: prompt eval time =     306.84 ms /   246 tokens (    1.25 ms per token,   801.72 tokens per second)
+llama_print_timings:        eval time =      91.50 ms /     6 runs   (   15.25 ms per token,    65.58 tokens per second)
+llama_print_timings:       total time =    1352.63 ms /   252 tokens
+```
+
+### case 2
+**input**
+```sh
+./llava-cli \
+    -m /data/local/tmp/ggml-model-q4_k.gguf \
+    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:" \
+    --n-gpu-layers 999
+
+```
+**output**
+```sh
+encode_image_with_clip: image encoded in   302.15 ms by CLIP (    2.10 ms per image patch)
+
+ The image features a cat lying in the grass.
+
+llama_print_timings:        load time =    1057.07 ms
+llama_print_timings:      sample time =       3.27 ms /    11 runs   (    0.30 ms per token,  3360.83 tokens per second)
+llama_print_timings: prompt eval time =     213.60 ms /   232 tokens (    0.92 ms per token,  1086.14 tokens per second)
+llama_print_timings:        eval time =     166.65 ms /    11 runs   (   15.15 ms per token,    66.01 tokens per second)
+llama_print_timings:       total time =    1365.47 ms /   243 tokens
+```
+
 ## Minor shortcomings
 The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost.
 
 ## TODO
 
-- [ ] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid`
+- [x] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid`
 - [ ] Optimize LDP projector performance
 
       - Optimize the structure definition to avoid unnecessary memory rearrangements, to reduce the use of `ggml_permute_cpy`;
       - Optimize operator implementation (ARM CPU/NVIDIA GPU): such as depthwise conv, hardswish, hardsigmoid, etc.
-- [ ] run MobileVLM on `Jetson Orin`
+- [x] run MobileVLM on `Jetson Orin`
 - [ ] Support more model variants, such as `MobileVLM-3B`.
 
 

examples/llava/README.md

@@ -1,10 +1,12 @@
 # LLaVA
 
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
+as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
 
 The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
 and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
 models are available.
+For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
 
 After API is confirmed, more models will be supported / uploaded.
 
@@ -14,14 +16,15 @@ Build with cmake or run `make llava-cli` to build it.
 After building, run: `./llava-cli` to see the usage. For example:
 
 ```sh
-./llava-cli -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
+./llava-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
 
-## Model conversion
+## LLaVA 1.5
 
-- Clone `llava-v15-7b`` and `clip-vit-large-patch14-336`` locally:
+- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
@@ -29,28 +32,108 @@ git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
 git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 ```
 
-2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+2. Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
 python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
+4. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert-image-encoder-to-gguf -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```
 
-4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
+5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b
+python ./convert.py ../llava-v1.5-7b --skip-unknown
 ```
 
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
+## LLaVA 1.6 gguf conversion
+1) 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/
+```
+- 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
+```
+- 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
+```
+
+**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
+**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
+
+## llava-cli templating and llava-1.6 prompting
+
+llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
+For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
+
+**For Mistral and using llava-cli binary:**
+Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
+The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
+
+**For the 34B this should work:**
+Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
+
+
+## How to know if you are running in llava-1.5 or llava-1.6 mode
+
+When running llava-cli you will see a visual information right before the prompt is being processed:
+
+**Llava-1.5:**
+`encode_image_with_clip: image embedding created: 576 tokens`
+
+**Llava-1.6 (anything above 576):**
+`encode_image_with_clip: image embedding created: 2880 tokens`
+
+
+Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
+
+
+
+
 ## TODO
 
-- [ ] Support non-CPU backend for the image encoding part.
+- [x] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.h

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

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

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

examples/llava/llava-cli.cpp

@@ -34,7 +34,7 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
 
 static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
     std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
     eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
     return true;
 }
@@ -152,26 +152,32 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     size_t image_pos = prompt.find("<image>");
     if (image_pos != std::string::npos) {
         // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
-
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        // We replace \n with actual newlines in user_prompt, just in case -e was not used in templating string
-        size_t pos = 0;
-        while ((pos = user_prompt.find("\\n", pos)) != std::string::npos) {
-            user_prompt.replace(pos, 2, "\n");
-            pos += 1; // Advance past the replaced newline
-        }
-        while ((pos = system_prompt.find("\\n", pos)) != std::string::npos) {
-            system_prompt.replace(pos, 2, "\n");
-            pos += 1; // Advance past the replaced newline
-        }
-
         printf("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
         printf("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     } else {
         // llava-1.5 native mode
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     }
 
     eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
@@ -183,13 +189,17 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     fprintf(stderr, "\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-
+    std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
-
         printf("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
         fflush(stdout);
     }
 
@@ -208,7 +218,8 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
-    llama_backend_init(params->numa);
+    llama_backend_init();
+    llama_numa_init(params->numa);
 
     llama_model_params model_params = llama_model_params_from_gpt_params(*params);
 

examples/llava/llava-surgery-v2.py

@@ -0,0 +1,155 @@
+import argparse
+import glob
+import os
+import torch
+from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file
+
+# Function to determine if file is a SafeTensor file
+def is_safetensor_file(file_path):
+    return file_path.endswith('.safetensors')
+
+
+# Unified loading function
+def load_model(file_path):
+    if is_safetensor_file(file_path):
+        tensors = {}
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            for key in f.keys():
+                tensors[key] = f.get_tensor(key).clone()
+                # output shape
+                print(f"{key} : {tensors[key].shape}")
+        return tensors, 'safetensor'
+    else:
+        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
+
+
+# Unified saving function
+def save_model(model, file_path, file_type):
+    if file_type == 'safetensor':
+        # safe_save(model, file_path)
+        save_file(model, file_path)
+    else:
+        torch.save(model, file_path)
+
+
+# Adapted function to clean vision tower from checkpoint
+def clean_vision_tower_from_checkpoint(checkpoint_path):
+    checkpoint, file_type = load_model(checkpoint_path)
+    # file_type = 'pytorch'
+    model_path = os.path.dirname(checkpoint_path)
+    print(f"Searching for vision tower tensors in {checkpoint_path}")
+    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+
+    if len(clip_tensors) > 0:
+        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
+        # Adapted for file type
+        clip_path = os.path.join(model_path, "llava.clip")
+
+        if os.path.exists(clip_path):
+            print(f"Loading existing llava.clip from {clip_path}")
+            existing_clip, _ = load_model(clip_path)
+        else:
+            print(f"Creating new llava.clip at {clip_path}")
+            existing_clip = {}
+        # Update existing_clip with new tensors, avoid duplicates
+        for name in clip_tensors:
+            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
+            print(f"Adding {simple_name} to llava.clip")
+            if simple_name not in existing_clip:
+                existing_clip[simple_name] = checkpoint[name]
+
+        # Save the updated clip tensors back to llava.clip
+        save_model(existing_clip, clip_path, 'pytorch')
+
+        # Remove the tensors from the original checkpoint
+        for name in clip_tensors:
+            del checkpoint[name]
+
+        checkpoint_path = checkpoint_path
+        return True
+    return False
+
+def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
+    newline_checkpoint_path = None
+    projector_checkpoint_path = None
+
+    for path in checkpoint_paths:
+        checkpoint, _ = load_model(path)
+        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
+            newline_checkpoint_path = path
+        if projector(checkpoint):
+            projector_checkpoint_path = path
+
+    return newline_checkpoint_path, projector_checkpoint_path
+
+def newline_criteria(checkpoint):
+    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
+
+
+# Command-line interface setup
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
+ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
+args = ap.parse_args()
+
+if args.clean_vision_tower:
+    # Generalized to handle both PyTorch and SafeTensors models
+    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
+    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+    for projector_checkpoint_path in checkpoint_paths:
+        print(f"Cleaning {projector_checkpoint_path}")
+        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
+            print(f"No vision tower found in {projector_checkpoint_path}")
+            # we break once none is found, so far all models append them at the end
+            # break
+    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
+
+# Now we look for the projector in the last checkpoint
+model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+# last_checkpoint_path = checkpoint_paths[0]
+# first_checkpoint_path = checkpoint_paths[-1]
+newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
+
+print(f"Taking projector from {projector_checkpoint_path}")
+first_mm_tensors = []
+first_checkpoint = None
+if newline_checkpoint_path is not None:
+    print(f"Taking newline from {newline_checkpoint_path}")
+    first_checkpoint, file_type = load_model(newline_checkpoint_path)
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+
+# Load the checkpoint
+mm_tensors = []
+last_checkpoint = None
+if projector_checkpoint_path is not None:
+    last_checkpoint, file_type = load_model(projector_checkpoint_path)
+    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+
+if len(mm_tensors) == 0:
+    if last_checkpoint is not None:
+        for k, v in last_checkpoint.items():
+            print(k)
+    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.")
+    print("No tensors found. Is this a LLaVA model?")
+    exit()
+
+print(f"Found {len(mm_tensors)} tensors to extract.")
+print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
+# projector = {name: checkpoint.[name].float() for name in mm_tensors}
+projector = {}
+for name in mm_tensors:
+    projector[name] = last_checkpoint[name].float()
+for name in first_mm_tensors:
+    projector[name] = first_checkpoint[name].float()
+
+if len(projector) > 0:
+    save_model(projector, f"{args.model}/llava.projector", 'pytorch')
+
+print("Done!")
+print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava-surgery.py

@@ -19,19 +19,12 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_project
 projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
-# remove these tensors from the checkpoint and save it again
-for name in mm_tensors:
-    del checkpoint[name]
-
 # BakLLaVA models contain CLIP tensors in it
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
 if len(clip_tensors) > 0:
     clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
     torch.save(clip, f"{args.model}/llava.clip")
 
-    # remove these tensors
-    for name in clip_tensors:
-        del checkpoint[name]
 
     # added tokens should be removed to be able to convert Mistral models
     if os.path.exists(f"{args.model}/added_tokens.json"):
@@ -39,8 +32,7 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
-torch.save(checkpoint, path)
 
 print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava.cpp

@@ -2,32 +2,296 @@
 #include "common.h"
 #include "llama.h"
 #include "llava.h"
+#include "base64.hpp"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
+#include <numeric>
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+struct clip_image_grid_shape {
+    int first;
+    int second;
+};
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
+    int original_width  = original_size.first;
+    int original_height = original_size.second;
+
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width  = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+/**
+ * @brief Get the anyres image grid shape object
+ *
+ * @param image_size
+ * @param grid_pinpoints
+ * @param image_patch_size
+ * @return <int, int>
+ */
+static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
+    /**
+        Conversion from gguf flat array to vector:
+        std::vector<std::pair<int, int>> possible_resolutions;
+        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+        }
+     */
+    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
+    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
+}
+
+// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+    struct {
+        struct ggml_tensor * newline;
+        struct ggml_context * ctx;
+    } model;
+
+    const int32_t image_size = clip_image_size(ctx_clip);
+    const int32_t patch_size = clip_patch_size(ctx_clip);
+
+    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
+
+    int num_patches_width  = grid_shape.first;  // grid 1-4
+    int num_patches_height = grid_shape.second; // grid 1-4
+
+    const size_t num_images = num_patches_width * num_patches_height + 1;
+
+    // TODO: size calculation is not calculated - it's only tens of MB
+    size_t ctx_size = 0;
+
+    {
+        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+    }
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
+    };
+
+    // Python reference code for full unpad:
+    /*
+        base_image_feature = image_feature[0]
+        image_feature = image_feature[1:]
+        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+        image_feature = unpad_image(image_feature, image_sizes[image_idx])
+        image_feature = torch.cat((
+            image_feature,
+            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
+        ), dim=-1)
+        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+    */
+    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
+    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
+    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
+    // Once all images are processed to prepended the base_image_features without any changes.
+
+    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
+    /*
+        image_feature = image_feature.view(2, 2, 24, 24, 4096)
+        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+        image_feature = image_feature.view(2, 24, 2, 24, 4096)
+        image_feature = image_feature.flatten(0, 3)
+
+        // Reshape to 4D tensor by merging the last two dimensions
+        image_feature = image_feature.view(2, 2, 24, 24*4096)
+        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
+        image_feature = image_feature.view(-1, 4096)
+    */
+
+    model.ctx = ggml_init(params);
+
+    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
+    model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
+    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
+        if (newline_tmp->buffer == NULL) {
+            printf("newline_tmp tensor buffer is NULL\n");
+        }
+        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
+    } else {
+        model.newline->data = newline_tmp->data;
+        if (model.newline->data == NULL) {
+            printf("newline_tmp tensor data is NULL\n");
+        }
+    }
+
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
+    // fill it with the image embeddings, ignoring the base
+    for (size_t i = 1; i < num_images; i++) {
+        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
+        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
+    }
+
+    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
+    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
+
+    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
+                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                num_patches_per_side,
+                                                                num_patches_width,
+                                                                num_patches_height,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
+    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
+    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
+    /**
+     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
+         image_feature = torch.cat((
+        image_feature,
+        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+    ), dim=-1)
+     *
+     */
+
+    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
+    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
+    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
+    ggml_build_forward_expand(gf, flatten);
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+
+    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
+    // append without newline tokens (default behavior in llava_arch when not using unpad ):
+    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+
+    // Debug: Test single segments
+    // Current findings: sending base image, sending a segment embedding all works similar to python
+    // However, permuted embeddings do not work yet (stride issue?)
+    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
+    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
+    // *n_img_pos_out=576;
+
+    ggml_free(model.ctx);
+    return true;
+}
 
-#include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    clip_image_f32 * img_res = clip_image_f32_init();
-    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
+    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
+    clip_image_f32_batch img_res_v;
+    img_res_v.size = 0;
+    img_res_v.data = nullptr;
+    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
-        clip_image_f32_free(img_res);
+        delete[] img_res_v.data;
         return false;
     }
 
-    *n_img_pos = clip_n_patches(ctx_clip);
-
     const int64_t t_img_enc_start_us = ggml_time_us();
-    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
-    clip_image_f32_free(img_res);
-    if (!encoded) {
-        fprintf(stderr, "Unable to encode image\n");
 
-        return false;
+    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
+
+    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+        // flat / default llava-1.5 type embedding
+        *n_img_pos = clip_n_patches(ctx_clip);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
+        delete[] img_res_v.data;
+        if (!encoded) {
+            fprintf(stderr, "Unable to encode image\n");
+
+            return false;
+        }
+    } else {
+        // spatial_unpad llava-1.6 type embedding
+        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            if (!encoded) {
+                fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        const int32_t * image_grid = clip_image_grid(ctx_clip);
+
+        std::vector<std::pair<int, int>> grid_pinpoints;
+        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
+        }
+
+        // free all img_res_v - not needed anymore
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
+
+        const int32_t image_size = clip_image_size(ctx_clip);
+
+        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
+
+        int n_img_pos_out;
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        *n_img_pos = n_img_pos_out;
+
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+
+        // debug image/segment/normalization content:
+        // clip_image_u8 * tmp = clip_image_u8_init();
+        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
+        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
+    printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
@@ -47,11 +311,10 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
+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");
-        free(image_embd);
         return false;
     }
 
@@ -85,7 +348,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
     return true;
 }
 
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
@@ -142,7 +405,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     return true;
 }
 
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
     unsigned char* image_bytes;
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
@@ -151,13 +414,13 @@ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct
         return NULL;
     }
 
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
     free(image_bytes);
 
     return embed;
 }
 
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
+void llava_image_embed_free(struct llava_image_embed * embed) {
     free(embed->embed);
     free(embed);
 }

examples/llava/llava.h

@@ -3,7 +3,6 @@
 
 #include "ggml.h"
 
-
 #ifdef LLAMA_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
@@ -30,7 +29,9 @@ 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_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);
+
+LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct 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);
@@ -42,7 +43,6 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
 
-
 #ifdef __cplusplus
 }
 #endif

examples/llava/requirements.txt

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

examples/lookahead/lookahead.cpp

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

examples/lookup/lookup.cpp

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

examples/main/README.md

@@ -67,6 +67,7 @@ main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt
 In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:
 
 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
 -   `-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.
@@ -283,7 +284,11 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### NUMA support
 
--   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
+-   `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
+-   `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
+-   `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitraty core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
+
+ These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
 ### Memory Float 32
 

examples/main/main.cpp

@@ -98,7 +98,7 @@ static void write_logfile(
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
-        if (!is_interacting) {
+        if (!is_interacting && g_params->interactive) {
             is_interacting = true;
         } else {
             console::cleanup();
@@ -185,7 +185,8 @@ int main(int argc, char ** argv) {
     }
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -333,6 +334,8 @@ 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
@@ -352,12 +355,12 @@ int main(int argc, char ** argv) {
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
         params.interactive_first = true;
-        params.antiprompt.push_back("### Instruction:\n\n");
+        params.antiprompt.emplace_back("### Instruction:\n\n");
     }
     // similar for chatml mode
     else if (params.chatml) {
         params.interactive_first = true;
-        params.antiprompt.push_back("<|im_start|>user\n");
+        params.antiprompt.emplace_back("<|im_start|>user\n");
     }
 
     // enable interactive mode if interactive start is specified
@@ -382,8 +385,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > 0) {
-        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > add_bos) {
+            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());
             }
@@ -392,7 +395,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
     }
 
-    if (params.interactive) {
+    // ctrl+C handling
+    {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
         struct sigaction sigint_action;
         sigint_action.sa_handler = sigint_handler;
@@ -405,7 +409,9 @@ int main(int argc, char ** argv) {
         };
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
+    }
 
+    if (params.interactive) {
         LOG_TEE("%s: interactive mode on.\n", __func__);
 
         if (!params.antiprompt.empty()) {
@@ -505,6 +511,14 @@ 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) {
@@ -536,14 +550,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_left    = n_past - params.n_keep;
                     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);
 
-                    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);
+                    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);
 
                     n_past -= n_discard;
 
@@ -570,9 +584,9 @@ 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_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);
+                    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);
 
                     n_past -= bd;
 
@@ -763,6 +777,18 @@ 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());
                 }
@@ -852,6 +878,7 @@ 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

@@ -107,6 +107,9 @@ 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;
 
@@ -122,7 +125,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;
@@ -195,8 +199,8 @@ 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, 0, n_tokens_system);
+        for (int32_t i = 1; i <= n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
         }
 
         LOG_TEE("\n");
@@ -220,15 +224,17 @@ 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 }, true);
+            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
 
             client.n_decoded += 1;
         }
 
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
-            for (int i = 0; i < n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
+            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);
             }
 
             LOG_TEE("%s: clearing the KV cache\n", __func__);
@@ -254,7 +260,7 @@ 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 }, false);
+                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
                     }
 
                     // extract the logits only for the last token
@@ -365,7 +371,8 @@ 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, n_tokens_system, -1);
+                    llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
+                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -71,7 +71,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -125,7 +126,7 @@ int main(int argc, char ** argv) {
     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", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+    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);
 
     // print the prompt token-by-token
 
@@ -145,10 +146,11 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            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);
+            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);
 
-            n_past -= bd;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
 
         llama_batch_clear(batch);
@@ -178,10 +180,12 @@ int main(int argc, char ** argv) {
 
         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_shift(ctx, 0, n_keep + n_discard, n_ctx,  -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);
 
-        n_past -= n_discard;
+        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
 
         llama_batch_clear(batch);
 
@@ -207,10 +211,12 @@ int main(int argc, char ** argv) {
         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_shift(ctx, 0, n_keep + n_discard, n_ctx,  -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);
 
-            n_past -= n_discard;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -309,7 +309,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
 }
 
 static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
@@ -442,29 +442,28 @@ 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) {
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) {
     if (params.ppl_stride > 0) {
         return perplexity_v2(ctx, params);
     }
 
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    const int n_ctx = llama_n_ctx(ctx);
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
-        logits_stream.open(params.logits_file.c_str());
+        logits_stream.open(params.logits_file.c_str(), std::ios::binary);
         if (!logits_stream.is_open()) {
             fprintf(stderr, "%s: failed to open %s for writing\n", __func__, params.logits_file.c_str());
             return {};
         }
         fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str());
         logits_stream.write("_logits_", 8);
-        logits_stream.write((const char *)&n_ctx, sizeof(n_ctx));
+        logits_stream.write(reinterpret_cast<const char *>(&n_ctx), sizeof(n_ctx));
     }
 
     auto tim1 = std::chrono::high_resolution_clock::now();
@@ -499,13 +498,19 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     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);
 
     std::vector<float> logits;
     if (num_batches > 1) {
         logits.reserve((size_t)n_ctx * n_vocab);
     }
 
-    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
+    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);
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
@@ -518,10 +523,26 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         log_probs.resize(n_ctx * nv);
     }
 
-    for (int i = 0; i < n_chunk; ++i) {
+    // 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);
+
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
@@ -531,34 +552,50 @@ 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);
 
-            // save original token and restore it after eval
-            const auto token_org = tokens[batch_start];
+            batch.n_tokens = 0;
+            for (int seq = 0; seq < n_seq_batch; seq++) {
+                int seq_start = batch_start + seq*n_ctx;
 
-            // 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));
+                // 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;
             }
 
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            if (llama_decode(ctx, batch)) {
                 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);
+            int total_seconds = (int)(t_total*n_chunk/n_seq);
             if (total_seconds >= 60*60) {
                 fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
@@ -566,37 +603,31 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
-        // 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;
-        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, n_ctx - 1 - first,
-                    workers, log_probs, nll, nll2);
-        } else {
-            process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
-                    workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
-        }
-        count += n_ctx - first - 1;
+        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;
 
-        // 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*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 + 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);
+            }
         }
         fflush(stdout);
 
@@ -615,6 +646,8 @@ 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};
 }
 
@@ -809,7 +842,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 32;
-    const int max_seq = 4*max_tasks_per_batch;
+    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -881,7 +914,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             size_t li = hs_cur.common_prefix;
             for (int s = 0; s < 4; ++s) {
                 for (size_t j = hs_cur.common_prefix; j < hs_cur.seq_tokens[s].size() - 1; j++) {
-                    eval_pairs.push_back(std::make_pair(hs_cur.i_batch + li++, hs_cur.seq_tokens[s][j + 1]));
+                    eval_pairs.emplace_back(hs_cur.i_batch + li++, hs_cur.seq_tokens[s][j + 1]);
                 }
                 ++li;
             }
@@ -1086,7 +1119,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 128;
-    const int max_seq = 2*max_tasks_per_batch;
+    const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1159,13 +1192,13 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0;
             size_t li = n_base1 - 1;
             for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) {
-                eval_pairs.push_back(std::make_pair(task.i_batch + li++, task.seq_tokens[0][j+1]));
+                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[0][j+1]);
             }
             const auto& n_base2 = skip_choice ? task.n_base2 : task.common_prefix;
             const int last_2nd = task.seq_tokens[1].size() - n_base2 > 1 ? 1 : 0;
             li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - 1;
             for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) {
-                eval_pairs.push_back(std::make_pair(task.i_batch + li++, task.seq_tokens[1][j+1]));
+                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[1][j+1]);
             }
         }
         compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results);
@@ -1438,7 +1471,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 32;
-    const int max_seq = 4*max_tasks_per_batch;
+    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1524,7 +1557,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             size_t li = cur_task.common_prefix;
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
                 for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
-                    eval_pairs.push_back(std::make_pair(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]));
+                    eval_pairs.emplace_back(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]);
                 }
                 ++li;
             }
@@ -1623,7 +1656,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     uint32_t n_ctx;
     in.read((char *)&n_ctx, sizeof(n_ctx));
     if (n_ctx > llama_n_ctx(ctx)) {
-        fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
+        fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
                 __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
     }
 
@@ -1782,13 +1815,24 @@ 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;
-    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    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);
+    }
 
     if (params.ppl_stride > 0) {
         fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
@@ -1809,11 +1853,15 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
 
+    // 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) {
@@ -1843,7 +1891,7 @@ int main(int argc, char ** argv) {
     } else if (params.kl_divergence) {
         kl_divergence(ctx, params);
     } else {
-        results = perplexity(ctx, params);
+        results = perplexity(ctx, params, n_ctx);
     }
 
     llama_print_timings(ctx);

examples/quantize-stats/quantize-stats.cpp

@@ -257,13 +257,13 @@ int main(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            params.include_layers.push_back(argv[i]);
+            params.include_layers.emplace_back(argv[i]);
         } else if (arg == "-L" || arg == "--exclude-layer") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params.exclude_layers.push_back(argv[i]);
+            params.exclude_layers.emplace_back(argv[i]);
         } else if (arg == "-t" || arg == "--type") {
             if (++i >= argc) {
                 invalid_param = true;

examples/quantize/quantize.cpp

@@ -23,14 +23,21 @@ 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" },
-    { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
+    { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw 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", },
@@ -208,13 +215,13 @@ int main(int argc, char ** argv) {
             }
         } else if (strcmp(argv[arg_idx], "--include-weights") == 0) {
             if (arg_idx < argc-1) {
-                included_weights.push_back(argv[++arg_idx]);
+                included_weights.emplace_back(argv[++arg_idx]);
             } else {
                 usage(argv[0]);
             }
         } else if (strcmp(argv[arg_idx], "--exclude-weights") == 0) {
             if (arg_idx < argc-1) {
-                excluded_weights.push_back(argv[++arg_idx]);
+                excluded_weights.emplace_back(argv[++arg_idx]);
             } else {
                 usage(argv[0]);
             }
@@ -237,7 +244,7 @@ int main(int argc, char ** argv) {
         params.imatrix = &imatrix_data;
     }
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // parse command line arguments
     const std::string fname_inp = argv[arg_idx];
@@ -287,9 +294,11 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
+    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_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 IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
         fprintf(stderr, "===============================================================================================\n\n\n");
         return 1;
     }

examples/server-embd.py

@@ -0,0 +1,34 @@
+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/CMakeLists.txt

@@ -1,12 +1,18 @@
 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 oai.hpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp 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 llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${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(${TARGET} PRIVATE CPPHTTPLIB_OPENSSL_SUPPORT)
+endif()
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -1,12 +1,26 @@
-# llama.cpp/example/server
+# LLaMA.cpp HTTP Server
 
-This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
+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**.
 
-Command line options:
+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:**
 
 - `--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`).
+- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.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.
 - `-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.
@@ -16,13 +30,20 @@ Command line options:
 - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+- `--numa STRATEGY`: Attempt one of the below optimization strategies  that help on some NUMA systems
+- `--numa distribute`: Spread execution evenly over all nodes
+- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
+- `--numa numactl`: Use the CPU map provided by numactl
+if run without this previously, it is recommended to drop the system page cache before using this
+see https://github.com/ggerganov/llama.cpp/issues/1437
+
 - `--numa`: Attempt optimizations that help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 - `-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 examples/server/public)
+- `--path`: path from which to serve static files (default: disabled)
 - `--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.
@@ -32,6 +53,16 @@ Command line options:
 - `--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)
+- `--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
 
@@ -49,6 +80,28 @@ 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:
@@ -71,10 +124,10 @@ 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 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 ghcr.io/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 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 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
 ```
 
 ## Testing with CURL
@@ -88,6 +141,12 @@ 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.
@@ -125,18 +184,30 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - `{"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.
+  - 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.
 
 - **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, 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.
+    `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
 
     `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_exponent`: Dynamic temperature exponent (default: 1.0).
+
     `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
 
     `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
@@ -181,18 +252,22 @@ node index.js
 
     `ignore_eos`: Ignore end of stream token and continue generating (default: false).
 
-    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
+    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. (default: []).
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
+    `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum (default: 0)
+
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (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)
 
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. (default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values)
+
 ### Result JSON
 
 - Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
@@ -220,7 +295,7 @@ 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`
+- `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.).
 - `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
@@ -238,7 +313,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `content`: Set the text to tokenize.
 
-    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`.
+    Note that a special `BOS` token is never inserted.
 
 - **POST** `/detokenize`: Convert tokens to text.
 
@@ -264,9 +339,25 @@ Notice that each `probs` is an array of length `n_probs`.
 
     It also accepts all the options of `/completion` except `stream` and `prompt`.
 
-- **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
+- **GET** `/props`: Return current server settings.
 
-- **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.
+### Result JSON
+
+```json
+{
+  "assistant_name": "",
+  "user_name": "",
+  "default_generation_settings": { ... },
+  "total_slots": 1
+}
+```
+
+- `assistant_name` - the required assistant name to generate the prompt in case you have specified a system prompt for all slots.
+- `user_name` - the required anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
+- `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.
 
     *Options:*
 
@@ -350,6 +441,81 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
+- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+
+### Result JSON
+
+```json
+[
+    {
+        "dynatemp_exponent": 1.0,
+        "dynatemp_range": 0.0,
+        "frequency_penalty": 0.0,
+        "grammar": "",
+        "id": 0,
+        "ignore_eos": false,
+        "logit_bias": [],
+        "min_p": 0.05000000074505806,
+        "mirostat": 0,
+        "mirostat_eta": 0.10000000149011612,
+        "mirostat_tau": 5.0,
+        "model": "llama-2-7b-32k-instruct.Q2_K.gguf",
+        "n_ctx": 2048,
+        "n_keep": 0,
+        "n_predict": 100000,
+        "n_probs": 0,
+        "next_token": {
+            "has_next_token": true,
+            "n_remain": -1,
+            "n_decoded": 0,
+            "stopped_eos": false,
+            "stopped_limit": false,
+            "stopped_word": false,
+            "stopping_word": ""
+        },
+        "penalize_nl": true,
+        "penalty_prompt_tokens": [],
+        "presence_penalty": 0.0,
+        "prompt": "Say hello to llama.cpp",
+        "repeat_last_n": 64,
+        "repeat_penalty": 1.100000023841858,
+        "samplers": [
+            "top_k",
+            "tfs_z",
+            "typical_p",
+            "top_p",
+            "min_p",
+            "temperature"
+        ],
+        "seed": 42,
+        "state": 1,
+        "stop": [
+            "\n"
+        ],
+        "stream": false,
+        "task_id": 0,
+        "temperature": 0.0,
+        "tfs_z": 1.0,
+        "top_k": 40,
+        "top_p": 0.949999988079071,
+        "typical_p": 1.0,
+        "use_penalty_prompt_tokens": false
+    }
+]
+```
+
+- **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
@@ -391,26 +557,55 @@ Run with bash:
 bash chat.sh
 ```
 
-### API like OAI
+### OAI-like API
 
-API example using Python Flask: [api_like_OAI.py](api_like_OAI.py)
-This example must be used with server.cpp
+The HTTP server supports OAI-like API: https://github.com/openai/openai-openapi
 
-```sh
-python api_like_OAI.py
+### 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"
+    }
+}
 ```
 
-After running the API server, you can use it in Python by setting the API base URL.
+Apart from error types supported by OAI, we also have custom types that are specific to functionalities of llama.cpp:
 
-```python
-openai.api_base = "http://<Your api-server IP>:port"
+**When /metrics or /slots endpoint is disabled**
+
+```json
+{
+    "error": {
+        "code": 501,
+        "message": "This server does not support metrics endpoint.",
+        "type": "not_supported_error"
+    }
+}
 ```
 
-Then you can utilize llama.cpp as an OpenAI's **chat.completion** or **text_completion** API
+**When the server receives invalid grammar via */completions endpoint**
+
+```json
+{
+    "error": {
+        "code": 400,
+        "message": "Failed to parse grammar",
+        "type": "invalid_request_error"
+    }
+}
+```
 
 ### Extending or building alternative Web Front End
 
-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.
+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

@@ -1,228 +0,0 @@
-#!/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

@@ -0,0 +1,88 @@
+### 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

@@ -0,0 +1,120 @@
+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/chat.sh

@@ -48,6 +48,7 @@ chat_completion() {
         top_p: 0.9,
         n_keep: $n_keep,
         n_predict: 256,
+        cache_prompt: true,
         stop: ["\n### Human:"],
         stream: true
     }')"