wpm : portable unicode tolower (#6305)

Also use C locale for ispunct/isspace, and split unicode-data.cpp from unicode.cpp.

.clang-tidy

@@ -12,6 +12,7 @@ Checks: >
     -readability-implicit-bool-conversion,
     -readability-magic-numbers,
     -readability-uppercase-literal-suffix,
+    -readability-simplify-boolean-expr,
     clang-analyzer-*,
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,

.devops/full-cuda.Dockerfile

@@ -26,8 +26,8 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
+# Enable CUDA
+ENV LLAMA_CUDA=1
 
 RUN make
 

.devops/llama-cpp-cuda.srpm.spec

@@ -12,7 +12,7 @@
 # 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
 #    It is up to the user to install the correct vendor-specific support.
 
-Name:           llama.cpp-cublas
+Name:           llama.cpp-cuda
 Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
 Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
@@ -32,16 +32,16 @@ CPU inference for Meta's Lllama2 models using default options.
 %setup -n llama.cpp-master
 
 %build
-make -j LLAMA_CUBLAS=1
+make -j LLAMA_CUDA=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llamacppcublas
-cp -p server %{buildroot}%{_bindir}/llamacppcublasserver
-cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple
+cp -p main %{buildroot}%{_bindir}/llamacppcuda
+cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
-%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacublas.service
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacuda.service
 [Unit]
 Description=Llama.cpp server, CPU only (no GPU support in this build).
 After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
@@ -49,7 +49,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llamacppcublasserver $LLAMA_ARGS
+ExecStart=/usr/bin/llamacppcudaserver $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -67,10 +67,10 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llamacppcublas
-%{_bindir}/llamacppcublasserver
-%{_bindir}/llamacppcublassimple
-/usr/lib/systemd/system/llamacublas.service
+%{_bindir}/llamacppcuda
+%{_bindir}/llamacppcudaserver
+%{_bindir}/llamacppcudasimple
+/usr/lib/systemd/system/llamacuda.service
 %config /etc/sysconfig/llama
 
 %pre

.devops/main-cuda.Dockerfile

@@ -20,8 +20,8 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
+# Enable CUDA
+ENV LLAMA_CUDA=1
 
 RUN make
 

.devops/nix/package.nix

@@ -1,15 +1,17 @@
 {
   lib,
+  glibc,
   config,
   stdenv,
   mkShell,
+  runCommand,
   cmake,
   ninja,
   pkg-config,
   git,
   python3,
   mpi,
-  openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations
+  blas,
   cudaPackages,
   darwin,
   rocmPackages,
@@ -30,6 +32,12 @@
   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,
+  precompileMetalShaders ? false
 }@inputs:
 
 let
@@ -41,10 +49,7 @@ 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" ]
@@ -84,6 +89,11 @@ let
     ]
   );
 
+  xcrunHost = runCommand "xcrunHost" {} ''
+    mkdir -p $out/bin
+    ln -s /usr/bin/xcrun $out/bin
+  '';
+
   # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
   # separately
   darwinBuildInputs =
@@ -147,6 +157,8 @@ effectiveStdenv.mkDerivation (
     postPatch = ''
       substituteInPlace ./ggml-metal.m \
         --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+      substituteInPlace ./ggml-metal.m \
+        --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
 
       # TODO: Package up each Python script or service appropriately.
       # If we were to migrate to buildPythonPackage and prepare the `pyproject.toml`,
@@ -154,6 +166,14 @@ effectiveStdenv.mkDerivation (
       substituteInPlace ./*.py --replace "/usr/bin/env python" "${llama-python}/bin/python"
     '';
 
+    # With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
+    # `default.metallib` may be compiled with Metal compiler from XCode
+    # and we need to escape sandbox on MacOS to access Metal compiler.
+    # `xcrun` is used find the path of the Metal compiler, which is varible
+    # and not on $PATH
+    # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
+    __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
+
     nativeBuildInputs =
       [
         cmake
@@ -167,6 +187,11 @@ 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
+      ] ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [
+        xcrunHost
       ];
 
     buildInputs =
@@ -175,21 +200,23 @@ effectiveStdenv.mkDerivation (
       ++ optionals useMpi [ mpi ]
       ++ optionals useOpenCL [ clblast ]
       ++ optionals useRocm rocmBuildInputs
+      ++ optionals useBlas [ blas ]
       ++ 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)
-        (cmakeBool "LLAMA_CUBLAS" useCuda)
+        (cmakeBool "LLAMA_CUDA" useCuda)
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
+        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -209,8 +236,10 @@ effectiveStdenv.mkDerivation (
         # Should likely use `rocmPackages.clr.gpuTargets`.
         "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
       ]
-      ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
-      ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
+      ++ optionals useMetalKit [
+        (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
+        (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
+      ];
 
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.

.devops/nix/sif.nix

@@ -7,7 +7,7 @@
 }:
 
 let
-    optionalInt = cond: x: if cond then x else 0;
+  optionalInt = cond: x: if cond then x else 0;
 in
 singularity-tools.buildImage rec {
   inherit (llama-cpp) name;

.devops/server-cuda.Dockerfile

@@ -20,8 +20,8 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
+# Enable CUDA
+ENV LLAMA_CUDA=1
 
 RUN make
 

.github/workflows/build.yml

@@ -15,14 +15,133 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
 jobs:
+  macOS-latest-cmake-arm64:
+    runs-on: macos-14
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+
+  macOS-latest-cmake-x64:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+
   ubuntu-focal-make:
     runs-on: ubuntu-20.04
+    env:
+      LLAMA_NODE_AVAILABLE: true
+      LLAMA_PYTHON_AVAILABLE: true
 
     steps:
       - name: Clone
@@ -35,6 +154,14 @@ jobs:
           sudo apt-get update
           sudo apt-get install build-essential gcc-8
 
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "20"
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: "3.11"
+
       - name: Build
         id: make_build
         env:
@@ -48,6 +175,28 @@ 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
 
@@ -76,40 +225,51 @@ 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_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-
-      - name: Test
-        id: cmake_test
+      - name: Test llama2c conversion
+        id: llama2c_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
+#  ubuntu-latest-cmake-sanitizer:
+#    runs-on: ubuntu-latest
+#
+#    continue-on-error: true
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#        build_type: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v3
+#
+#      - name: Dependencies
+#        id: depends
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          mkdir build
+#          cd build
+#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ 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
@@ -145,6 +305,28 @@ 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
 
@@ -311,6 +493,7 @@ 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 \
@@ -339,6 +522,7 @@ 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 \
@@ -403,6 +587,8 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'arm64'
+            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -498,7 +684,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -542,13 +728,13 @@ jobs:
           path: |
             llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
 
-  windows-latest-cmake-cublas:
+  windows-latest-cmake-cuda:
     runs-on: windows-latest
 
     strategy:
       matrix:
         cuda: ['12.2.0', '11.7.1']
-        build: ['cublas']
+        build: ['cuda']
 
     steps:
       - name: Clone
@@ -569,7 +755,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
+          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUDA=ON -DBUILD_SHARED_LIBS=ON
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
@@ -614,6 +800,7 @@ jobs:
 
   windows-latest-cmake-sycl:
     runs-on: windows-latest
+
     defaults:
       run:
         shell: bash
@@ -622,7 +809,6 @@ jobs:
       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
@@ -637,6 +823,32 @@ jobs:
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
 
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+
   ios-xcode-build:
     runs-on: macos-latest
 
@@ -699,7 +911,9 @@ jobs:
       - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
-      - windows-latest-cmake-cublas
+      - windows-latest-cmake-cuda
+      - macOS-latest-cmake-arm64
+      - macOS-latest-cmake-x64
 
     steps:
       - name: Clone

.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: 10000
+          repo-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/code-coverage.yml

@@ -5,6 +5,10 @@ env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run:
     runs-on: ubuntu-20.04

.github/workflows/docker.yml

@@ -15,6 +15,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub

.github/workflows/editorconfig.yml

@@ -14,6 +14,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   editorconfig:
     runs-on: ubuntu-latest

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

@@ -17,6 +17,10 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/*.nix', 'flake.lock']
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -8,6 +8,10 @@ on:
   pull_request:
     types: [opened, synchronize, reopened]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   nix-eval:
     strategy:

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

@@ -3,17 +3,23 @@ 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'
       - 'requirements/*.txt'
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   python-check-requirements:
     runs-on: ubuntu-latest
@@ -26,4 +32,4 @@ 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

@@ -2,6 +2,10 @@ name: flake8 Lint
 
 on: [push, pull_request]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   flake8-lint:
     runs-on: ubuntu-latest

.github/workflows/server.yml

@@ -3,6 +3,11 @@ name: Server
 
 on:
   workflow_dispatch: # allows manual triggering
+    inputs:
+      slow_tests:
+        description: 'Run slow tests'
+        required: true
+        type: boolean
   push:
     branches:
       - master
@@ -10,6 +15,12 @@ on:
   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 * * *'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
 
 jobs:
   server:
@@ -17,18 +28,14 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        # TODO: temporary disabled due to linux kernel issues
+        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [UNDEFINED]
+        build_type: [Debug]
         include:
           - build_type: Release
             sanitizer: ""
-        exclude:
-          - build_type: Release
-            sanitizer: ADDRESS
-          - build_type: Release
-            sanitizer: THREAD
-          - build_type: Release
-            sanitizer: UNDEFINED
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     container:
       image: ubuntu:latest
@@ -40,6 +47,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
       - name: Dependencies
         id: depends
@@ -51,7 +60,8 @@ jobs:
             cmake \
             python3-pip \
             wget \
-            psmisc
+            language-pack-en \
+            libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -61,6 +71,7 @@ jobs:
           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
@@ -70,14 +81,74 @@ jobs:
         run: |
           pip install -r examples/server/tests/requirements.txt
 
-      - name: Download models
-        id: download_models
-        run: |
-          cd examples/server/tests
-          ../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf
-
       - name: Tests
-        id: server_integration_test
+        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

.github/workflows/zig-build.yml

@@ -6,6 +6,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   build:
     strategy:

.gitignore

@@ -11,7 +11,10 @@
 *.gcda
 *.dot
 *.bat
+*.tmp
 *.metallib
+*.etag
+*.lastModified
 .DS_Store
 .build/
 .cache/
@@ -25,6 +28,8 @@
 .vscode/
 .idea/
 
+ggml-metal-embed.metal
+
 lcov-report/
 gcovr-report/
 
@@ -45,6 +50,8 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/gguf-split
+/gritlm
 /imatrix
 /infill
 /libllama.so
@@ -52,6 +59,9 @@ models-mnt
 /llava-cli
 /lookahead
 /lookup
+/lookup-create
+/lookup-merge
+/lookup-stats
 /main
 /metal
 /passkey
@@ -67,6 +77,7 @@ models-mnt
 /batched-bench
 /export-lora
 /finetune
+/retrieval
 /speculative
 /parallel
 /train-text-from-scratch

CMakeLists.txt

@@ -89,8 +89,8 @@ endif()
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
-#option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
+option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
+option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
@@ -99,6 +99,8 @@ 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_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
+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)
@@ -116,7 +118,9 @@ option(LLAMA_MPI                             "llama: use MPI"
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
+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})
@@ -146,6 +150,8 @@ 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>)
@@ -196,63 +202,70 @@ 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_EMBED_LIBRARY)
         enable_language(ASM)
         add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-        set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
+        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
+        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+
         file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-        set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+
+        # merge ggml-common.h and ggml-metal.metal into a single file
+        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
 
         add_custom_command(
-            OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-            DEPENDS ${METALLIB_SOURCE}
+            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} ${EMBED_METALLIB_ASSEMBLY})
-    endif()
-
-    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)
+        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}
@@ -347,18 +360,25 @@ if (LLAMA_QKK_64)
 endif()
 
 if (LLAMA_CUBLAS)
+    message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
+    set(LLAMA_CUDA ON)
+endif()
+
+if (LLAMA_CUDA)
     cmake_minimum_required(VERSION 3.17)
 
     find_package(CUDAToolkit)
     if (CUDAToolkit_FOUND)
-        message(STATUS "cuBLAS found")
+        message(STATUS "CUDA found")
 
         enable_language(CUDA)
 
         set(GGML_HEADERS_CUDA ggml-cuda.h)
-        set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
-        add_compile_definitions(GGML_USE_CUBLAS)
+        file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu")
+        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
+
+        add_compile_definitions(GGML_USE_CUDA)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -375,6 +395,9 @@ if (LLAMA_CUBLAS)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
+        if (LLAMA_CUDA_NO_PEER_COPY)
+            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+        endif()
 
         if (LLAMA_STATIC)
             if (WIN32)
@@ -404,7 +427,7 @@ if (LLAMA_CUBLAS)
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
     else()
-        message(WARNING "cuBLAS not found")
+        message(WARNING "CUDA not found")
     endif()
 endif()
 
@@ -503,9 +526,11 @@ if (LLAMA_HIPBLAS)
     message(STATUS "HIP and hipBLAS found")
 
     set(GGML_HEADERS_ROCM ggml-cuda.h)
-    set(GGML_SOURCES_ROCM ggml-cuda.cu)
 
-    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+    file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
+    list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
+
+    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
 
     if (LLAMA_HIP_UMA)
         add_compile_definitions(GGML_HIP_UMA)
@@ -519,11 +544,15 @@ if (LLAMA_HIPBLAS)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
 
+    if (LLAMA_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    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)
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
 
     if (LLAMA_STATIC)
         message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
@@ -533,6 +562,10 @@ if (LLAMA_HIPBLAS)
 endif()
 
 if (LLAMA_SYCL)
+    if (NOT LLAMA_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$")
+        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA")
+    endif()
+
     if ( NOT DEFINED ENV{ONEAPI_ROOT})
         message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
     endif()
@@ -554,6 +587,9 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
+    if (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
+    endif()
 
     set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
@@ -561,7 +597,11 @@ if (LLAMA_SYCL)
     if (WIN32)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
     else()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+        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()
 
@@ -795,7 +835,7 @@ endif()
 
 set(CUDA_CXX_FLAGS "")
 
-if (LLAMA_CUBLAS)
+if (LLAMA_CUDA)
     set(CUDA_FLAGS -use_fast_math)
 
     if (LLAMA_FATAL_WARNINGS)
@@ -1020,7 +1060,7 @@ endif()
 add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
 add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
 
-if (LLAMA_CUBLAS)
+if (LLAMA_CUDA)
     list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
     list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
     if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
@@ -1128,6 +1168,9 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
+            unicode.h
+            unicode.cpp
+            unicode-data.cpp
             )
 
 target_include_directories(llama PUBLIC .)
@@ -1223,6 +1266,12 @@ if (LLAMA_METAL)
             GROUP_READ
             WORLD_READ
         DESTINATION ${CMAKE_INSTALL_BINDIR})
+    if (NOT LLAMA_METAL_EMBED_LIBRARY)
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
+    endif()
 endif()
 
 #

Makefile

@@ -1,15 +1,16 @@
 # Define the default target now so that it is always the first target
 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
+	simple batched batched-bench save-load-state server gguf gguf-split llama-bench libllava.a llava-cli baby-llama beam-search  \
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease                                 \
+	tests/test-json-schema-to-grammar
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -167,6 +168,10 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
+ifdef LLAMA_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
+endif
+
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
@@ -201,6 +206,10 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(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 ''
@@ -381,14 +390,20 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
+# LLAMA_CUBLAS is deprecated and will be removed in the future
+	LLAMA_CUDA := 1
+endif
+
+ifdef LLAMA_CUDA
 	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_CPPFLAGS  += -DGGML_USE_CUDA -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
+	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 	MK_NVCCFLAGS += -use_fast_math
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
@@ -443,19 +458,30 @@ ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-#ifdef LLAMA_CUDA_CUBLAS
-#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
-#endif # LLAMA_CUDA_CUBLAS
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+define NVCC_COMPILE
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -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 $@
+endef # NVCC_COMPILE
 else
+define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
 endif # JETSON_EOL_MODULE_DETECT
-endif # LLAMA_CUBLAS
+
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(NVCC_COMPILE)
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+	$(NVCC_COMPILE)
+
+endif # LLAMA_CUDA
 
 ifdef LLAMA_CLBLAST
 
@@ -501,7 +527,6 @@ ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
 endif # LLAMA_VULKAN
 
 ifdef LLAMA_HIPBLAS
-
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
 		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
@@ -513,7 +538,7 @@ ifdef LLAMA_HIPBLAS
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
-	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
 ifdef LLAMA_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
 endif # LLAMA_HIP_UMA
@@ -526,9 +551,17 @@ endif # LLAMA_HIP_UMA
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	HIPFLAGS 	+= -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	OBJS        += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
 endif # LLAMA_HIPBLAS
 
 ifdef LLAMA_METAL
@@ -545,19 +578,20 @@ 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-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:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@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
@@ -580,12 +614,17 @@ override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
 # identify CUDA host compiler
-ifdef LLAMA_CUBLAS
+ifdef LLAMA_CUDA
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
 CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif
 
+ifdef LLAMA_CURL
+override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
+override LDFLAGS  := $(LDFLAGS) -lcurl
+endif
+
 #
 # Print build information
 #
@@ -600,7 +639,7 @@ $(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))
-ifdef LLAMA_CUBLAS
+ifdef LLAMA_CUDA
 $(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)
@@ -610,9 +649,16 @@ $(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be exp
 endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # LLAMA_CUBLAS
+endif # LLAMA_CUDA
 $(info )
 
+ifdef LLAMA_CUBLAS
+$(info !!!!)
+$(info LLAMA_CUBLAS is deprecated and will be removed in the future. Use LLAMA_CUDA instead.)
+$(info !!!!)
+$(info )
+endif
+
 #
 # Build library
 #
@@ -626,12 +672,18 @@ 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
+unicode-data.o: unicode-data.cpp unicode-data.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.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
@@ -649,9 +701,15 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 train.o: common/train.cpp common/train.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
@@ -659,7 +717,8 @@ 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 *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf ggml-cuda/*.o
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -720,19 +779,26 @@ embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+gritlm: examples/gritlm/gritlm.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/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 examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp json-schema-to-grammar.o common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(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) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(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) -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) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -770,6 +836,10 @@ export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+retrieval: examples/retrieval/retrieval.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)
+
 speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -782,9 +852,15 @@ lookahead: examples/lookahead/lookahead.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)
 
-lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
 
 passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -841,6 +917,10 @@ tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server -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) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

Package.swift

@@ -31,6 +31,8 @@ let package = Package(
             sources: [
                 "ggml.c",
                 "llama.cpp",
+                "unicode.cpp",
+                "unicode-data.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",

README-sycl.md

@@ -1,6 +1,7 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
+- [News](#news)
 - [OS](#os)
 - [Intel GPU](#intel-gpu)
 - [Docker](#docker)
@@ -25,6 +26,23 @@ 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
+  - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
+  - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
+  - 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|
@@ -57,6 +75,29 @@ For iGPU, please make sure the shared memory from host memory is enough. For lla
 
 For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
 
+## Nvidia GPU
+
+### Verified
+
+|Intel GPU| Status | Verified Model|
+|-|-|-|
+|Ampere Series| Support| A100|
+
+### oneMKL for CUDA
+
+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:
@@ -75,7 +116,7 @@ You can choose between **F16** and **F32** build. F16 is faster for long-prompt
 # 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
+# Note: you can also use the ".devops/server-intel.Dockerfile", which compiles the "server" example
 ```
 
 ### Run
@@ -170,6 +211,9 @@ source /opt/intel/oneapi/setvars.sh
 # Or, for FP32:
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
+# For Nvidia GPUs
+cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
 # Build example/main only
 #cmake --build . --config Release --target main
 
@@ -212,16 +256,16 @@ Run without parameter:
 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
-
+found 6 SYCL devices:
+|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
+|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
+|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
+| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
+| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
+| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
+| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
+| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 ```
 
 |Attribute|Note|
@@ -229,12 +273,35 @@ found 4 SYCL devices:
 |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
+4. Device selection and execution of llama.cpp
 
-Set device ID = 0 by **GGML_SYCL_DEVICE=0**
+There are two device selection modes:
+
+- Single device: Use one device assigned by user.
+- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+
+|Device selection|Parameter|
+|-|-|
+|Single device|--split-mode none --main-gpu DEVICE_ID |
+|Multiple devices|--split-mode layer (default)|
+
+Examples:
+
+- Use 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
+ZES_ENABLE_SYSMAN=1 ./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 -sm none -mg 0
+```
+or run by script:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
+
+- Use multiple devices:
+
+```sh
+ZES_ENABLE_SYSMAN=1 ./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 -sm layer
 ```
 or run by script:
 
@@ -247,12 +314,18 @@ 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
+5. Verify the device ID in output
+
+Verify to see if the selected GPU is shown in the output, like:
 
-Like:
 ```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
+Or
+```
+use 1 SYCL GPUs: [0] with Max compute units:512
+```
+
 
 ## Windows
 
@@ -313,7 +386,7 @@ 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).
+- Download the 1.19.0 version of [w64devkit](https://github.com/skeeto/w64devkit/releases/download/v1.19.0/w64devkit-1.19.0.zip).
 
 - Extract `w64devkit` on your pc.
 
@@ -388,15 +461,16 @@ 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
+found 6 SYCL devices:
+|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
+|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
+|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
+| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
+| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
+| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
+| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
+| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 
 ```
 
@@ -405,13 +479,31 @@ found 4 SYCL devices:
 |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**
+4. Device selection and execution of llama.cpp
+
+There are two device selection modes:
+
+- Single device: Use one device assigned by user.
+- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+
+|Device selection|Parameter|
+|-|-|
+|Single device|--split-mode none --main-gpu DEVICE_ID |
+|Multiple devices|--split-mode layer (default)|
+
+Examples:
+
+- Use 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
+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 -sm none -mg 0
+```
+
+- Use multiple devices:
+
+```
+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 -sm layer
 ```
 or run by script:
 
@@ -424,11 +516,17 @@ 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:
+5. Verify the device ID in output
+
+Verify to see if the selected GPU is shown in the output, like:
+
 ```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+detect 1 SYCL GPUs: [0] with top Max compute units:512
+```
+Or
+```
+use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
 ## Environment Variable
@@ -447,8 +545,8 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
 |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
 
@@ -458,8 +556,15 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
   Solution: add **--no-mmap** or **--mmap 0**.
 
+- Split-mode: [row] is not supported
+
+  It's on developing.
+
 ## Q&A
 
+Note: please add prefix **[SYCL]** in issue title, so that we will check it as soon as possible.
+
+
 - 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.
@@ -491,4 +596,4 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
 ## Todo
 
-- Support multiple cards.
+- Support row layer split for multiple card runs.

README.md

@@ -8,12 +8,22 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
+### Recent API changes
+
+- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
+- [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
 
-- Support for chat templates: [Wiki (contributions welcome)](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631
-- Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590
-- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
+- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
+- 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
+- Support loading sharded model, using `gguf-split` CLI https://github.com/ggerganov/llama.cpp/pull/6187
 
 ----
 
@@ -104,10 +114,12 @@ Typically finetunes of the base models below are supported as well.
 - [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] [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)
@@ -125,6 +137,7 @@ Typically finetunes of the base models below are supported as well.
 - 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)
@@ -155,6 +168,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [RecurseChat](https://recurse.chat/) (proprietary)
 - [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)
@@ -435,30 +449,27 @@ Building the program with BLAS support may lead to some performance improvements
 
   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.
 
-- #### cuBLAS
+- #### CUDA
 
-  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+  This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
 
   For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
 
   - Using `make`:
     ```bash
-    make LLAMA_CUBLAS=1
+    make LLAMA_CUDA=1
     ```
   - Using `CMake`:
 
     ```bash
     mkdir build
     cd build
-    cmake .. -DLLAMA_CUBLAS=ON
+    cmake .. -DLLAMA_CUDA=ON
     cmake --build . --config Release
     ```
 
   The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
 
-<!---
-  | LLAMA_CUDA_CUBLAS       | Boolean                |   false | Use cuBLAS instead of custom CUDA kernels for prompt processing. Faster for all quantization formats except for q4_0 and q8_0, especially for k-quants. Increases VRAM usage (700 MiB for 7b, 970 MiB for 13b, 1430 MiB for 33b). |
---->
   | Option                         | Legal values           | Default | Description |
   |--------------------------------|------------------------|---------|-------------|
   | LLAMA_CUDA_FORCE_DMMV          | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
@@ -620,6 +631,15 @@ Building the program with BLAS support may lead to some performance improvements
 
 - #### Vulkan
 
+> [!WARNING]
+>
+> Vulkan support has been broken in https://github.com/ggerganov/llama.cpp/pull/6122
+> due to relying on `GGML_OP_GET_ROWS` which is not yet properly supported by the Vulkan backend,
+> but should be fixed relatively soon (possibly in https://github.com/ggerganov/llama.cpp/pull/6155
+> (ref: https://github.com/ggerganov/llama.cpp/pull/6122#issuecomment-2015327635)).
+>
+> Meanwhile, if you want to use the Vulkan backend, you should use the commit right before the breaking change, https://github.com/ggerganov/llama.cpp/commit/55c1b2a3bbd470e9e2a3a0618b92cf64a885f806
+
   **With docker**:
 
   You don't need to install Vulkan SDK. It will be installed inside the container.
@@ -785,7 +805,7 @@ 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
 
@@ -849,7 +869,7 @@ 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.
@@ -896,6 +916,9 @@ 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
@@ -904,7 +927,28 @@ $ 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
 

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

@@ -115,24 +115,27 @@ 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 unicode_data = make.obj("unicode-data", "unicode-data.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");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
+    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.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, unicode_data, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, 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, unicode_data, 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, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -40,12 +40,13 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUDA=1"
 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
 
@@ -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
@@ -411,8 +412,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 

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 "${GIT_DIR}/index")
+    else()
+        message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
+    endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
     set(GIT_INDEX "")
@@ -42,6 +47,8 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
+set(TARGET json-schema-to-grammar)
+add_library(${TARGET} OBJECT json-schema-to-grammar.cpp json-schema-to-grammar.h)
 
 set(TARGET common)
 
@@ -55,14 +62,28 @@ add_library(${TARGET} STATIC
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    json.hpp
     train.h
     train.cpp
+    ngram-cache.h
+    ngram-cache.cpp
     )
 
 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

@@ -37,13 +37,16 @@ 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;
@@ -51,13 +54,13 @@ struct gpt_params {
     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_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               = 8;     // number of tokens to draft during speculative decoding
+    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_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)
@@ -76,23 +79,31 @@ struct gpt_params {
     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
-    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-    ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
+
+    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_draft       = "";                              // draft model for speculative decoding
-    std::string model_alias       = "unknown"; // model alias
-    std::string prompt            = "";
-    std::string prompt_file       = "";  // store the external prompt file name
-    std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
-    std::string input_prefix      = "";  // string to prefix user inputs with
-    std::string input_suffix      = "";  // string to suffix user inputs with
+    std::string model                = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_draft          = "";  // draft model for speculative decoding
+    std::string model_alias          = "unknown"; // model alias
+    std::string model_url            = "";  // model url to download
+    std::string hf_repo              = "";  // HF repo
+    std::string hf_file              = "";  // HF file
+    std::string prompt               = "";
+    std::string prompt_file          = "";  // store the external prompt file name
+    std::string path_prompt_cache    = "";  // path to file for saving/loading prompt eval state
+    std::string input_prefix         = "";  // string to prefix user inputs with
+    std::string input_suffix         = "";  // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
-    std::string logdir            = "";  // directory in which to save YAML log files
-    std::string logits_file       = "";  // file for saving *all* logits
+    std::string logdir               = "";  // directory in which to save YAML log files
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
+    std::string logits_file          = "";  // file for saving *all* logits
 
     std::vector<llama_model_kv_override> kv_overrides;
 
@@ -100,6 +111,11 @@ 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)
@@ -115,7 +131,6 @@ 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
@@ -128,7 +143,7 @@ struct gpt_params {
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
-    bool cont_batching     = false; // insert new sequences for decoding on-the-fly
+    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
@@ -156,6 +171,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
 
+bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+
 std::string get_system_info(const gpt_params & params);
 
 std::string gpt_random_prompt(std::mt19937 & rng);
@@ -181,6 +198,9 @@ 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, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -259,3 +279,39 @@ 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);
+
+//
+// Split utils
+//
+static const char * const LLM_KV_SPLIT_NO            = "split.no";
+static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

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/json-schema-to-grammar.cpp

@@ -0,0 +1,721 @@
+#include "json-schema-to-grammar.h"
+#include <algorithm>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+const std::string SPACE_RULE = "\" \"?";
+
+std::unordered_map<std::string, std::string> PRIMITIVE_RULES = {
+    {"boolean", "(\"true\" | \"false\") space"},
+    {"number", "(\"-\"? ([0-9] | [1-9] [0-9]*)) (\".\" [0-9]+)? ([eE] [-+]? [0-9]+)? space"},
+    {"integer", "(\"-\"? ([0-9] | [1-9] [0-9]*)) space"},
+    {"value", "object | array | string | number | boolean"},
+    {"object", "\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space"},
+    {"array", "\"[\" space ( value (\",\" space value)* )? \"]\" space"},
+    {"uuid", "\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space"},
+    {"string", " \"\\\"\" (\n"
+               "        [^\"\\\\] |\n"
+               "        \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])\n"
+               "      )* \"\\\"\" space"},
+    {"null", "\"null\" space"}
+};
+std::vector<std::string> OBJECT_RULE_NAMES = {"object", "array", "string", "number", "boolean", "null", "value"};
+
+std::unordered_map<std::string, std::string> DATE_RULES = {
+    {"date", "[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )"},
+    {"time", "([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )"},
+    {"date-time", "date \"T\" time"},
+    {"date-string", "\"\\\"\" date \"\\\"\" space"},
+    {"time-string", "\"\\\"\" time \"\\\"\" space"},
+    {"date-time-string", "\"\\\"\" date-time \"\\\"\" space"}
+};
+
+static bool is_reserved_name(const std::string & name) {
+    static std::unordered_set<std::string> RESERVED_NAMES;
+    if (RESERVED_NAMES.empty()) {
+        RESERVED_NAMES.insert("root");
+        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
+        for (const auto &p : DATE_RULES) RESERVED_NAMES.insert(p.first);
+    }
+    return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
+}
+
+std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
+std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+};
+
+std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
+
+template <typename Iterator>
+std::string join(Iterator begin, Iterator end, const std::string & separator) {
+    std::ostringstream result;
+    if (begin != end) {
+        result << *begin;
+        for (Iterator it = begin + 1; it != end; ++it) {
+            result << separator << *it;
+        }
+    }
+    return result.str();
+}
+
+static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t start = 0;
+    size_t end = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        tokens.push_back(str.substr(start, end - start));
+        start = end + delimiter.length();
+        end = str.find(delimiter, start);
+    }
+
+    tokens.push_back(str.substr(start));
+
+    return tokens;
+}
+
+static std::string repeat(const std::string & str, size_t n) {
+    if (n == 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(str.length() * n);
+
+    for (size_t i = 0; i < n; ++i) {
+        result += str;
+    }
+
+    return result;
+}
+
+static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
+    std::smatch match;
+    std::string result;
+
+    std::string::const_iterator searchStart(input.cbegin());
+    std::string::const_iterator searchEnd(input.cend());
+
+    while (std::regex_search(searchStart, searchEnd, match, regex)) {
+        result.append(searchStart, searchStart + match.position());
+        result.append(replacement(match));
+        searchStart = match.suffix().first;
+    }
+
+    result.append(searchStart, searchEnd);
+
+    return result;
+}
+
+static std::string format_literal(const std::string & literal) {
+    std::string escaped = replacePattern(literal, GRAMMAR_LITERAL_ESCAPE_RE, [&](const std::smatch & match) {
+        char c = match.str()[0];
+        return GRAMMAR_LITERAL_ESCAPES.at(c);
+    });
+    return "\"" + escaped + "\"";
+}
+
+
+class SchemaConverter {
+private:
+    std::function<json(const std::string &)> _fetch_json;
+    bool _dotall;
+    std::map<std::string, std::string> _rules;
+    std::unordered_map<std::string, json> _refs;
+    std::unordered_set<std::string> _refs_being_resolved;
+    std::vector<std::string> _errors;
+    std::vector<std::string> _warnings;
+
+    std::string _add_rule(const std::string & name, const std::string & rule) {
+        std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-");
+        if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
+            _rules[esc_name] = rule;
+            return esc_name;
+        } else {
+            int i = 0;
+            while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) {
+                i++;
+            }
+            std::string key = esc_name + std::to_string(i);
+            _rules[key] = rule;
+            return key;
+        }
+    }
+
+    std::string _generate_union_rule(const std::string & name, const std::vector<json> & alt_schemas) {
+        std::vector<std::string> rules;
+        for (size_t i = 0; i < alt_schemas.size(); i++) {
+            rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
+        }
+        return join(rules.begin(), rules.end(), " | ");
+    }
+
+    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
+        if (!(pattern.front() == '^' && pattern.back() == '$')) {
+            _errors.push_back("Pattern must start with '^' and end with '$'");
+            return "";
+        }
+        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
+        std::unordered_map<std::string, std::string> sub_rule_ids;
+
+        size_t i = 0;
+        size_t length = sub_pattern.length();
+
+        using literal_or_rule = std::pair<std::string, bool>;
+        auto to_rule = [&](const literal_or_rule & ls) {
+            auto is_literal = ls.second;
+            auto s = ls.first;
+            return is_literal ? "\"" + s + "\"" : s;
+        };
+        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
+            size_t start = i;
+            std::vector<literal_or_rule> seq;
+
+            auto get_dot = [&]() {
+                std::string rule;
+                if (_dotall) {
+                    rule = "[\\U00000000-\\U0010FFFF]";
+                } else {
+                    rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]";
+                }
+                return _add_rule("dot", rule);
+            };
+
+            // Joins the sequence, merging consecutive literals together.
+            auto join_seq = [&]() {
+                std::vector<literal_or_rule> ret;
+
+                std::string literal;
+                auto flush_literal = [&]() {
+                    if (literal.empty()) {
+                        return false;
+                    }
+                    ret.push_back(std::make_pair(literal, true));
+                    literal.clear();
+                    return true;
+                };
+
+                for (const auto & item : seq) {
+                    auto is_literal = item.second;
+                    if (is_literal) {
+                        literal += item.first;
+                    } else {
+                        flush_literal();
+                        ret.push_back(item);
+                    }
+                }
+                flush_literal();
+
+                std::vector<std::string> results;
+                for (const auto & item : ret) {
+                    results.push_back(to_rule(item));
+                }
+                return std::make_pair(join(results.begin(), results.end(), " "), false);
+            };
+
+            while (i < length) {
+                char c = sub_pattern[i];
+                if (c == '.') {
+                    seq.push_back(std::make_pair(get_dot(), false));
+                    i++;
+                } else if (c == '(') {
+                    i++;
+                    if (i < length) {
+                        if (sub_pattern[i] == '?') {
+                            _warnings.push_back("Unsupported pattern syntax");
+                        }
+                    }
+                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
+                } else if (c == ')') {
+                    i++;
+                    if (start > 0 && sub_pattern[start - 1] != '(') {
+                        _errors.push_back("Unbalanced parentheses");
+                    }
+                    return join_seq();
+                } else if (c == '[') {
+                    std::string square_brackets = std::string(1, c);
+                    i++;
+                    while (i < length && sub_pattern[i] != ']') {
+                        if (sub_pattern[i] == '\\') {
+                            square_brackets += sub_pattern.substr(i, 2);
+                            i += 2;
+                        } else {
+                            square_brackets += sub_pattern[i];
+                            i++;
+                        }
+                    }
+                    if (i >= length) {
+                        _errors.push_back("Unbalanced square brackets");
+                    }
+                    square_brackets += ']';
+                    i++;
+                    seq.push_back(std::make_pair(square_brackets, false));
+                } else if (c == '|') {
+                    seq.push_back(std::make_pair("|", false));
+                    i++;
+                } else if (c == '*' || c == '+' || c == '?') {
+                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
+                    i++;
+                } else if (c == '{') {
+                    std::string curly_brackets = std::string(1, c);
+                    i++;
+                    while (i < length && sub_pattern[i] != '}') {
+                        curly_brackets += sub_pattern[i];
+                        i++;
+                    }
+                    if (i >= length) {
+                        _errors.push_back("Unbalanced curly brackets");
+                    }
+                    curly_brackets += '}';
+                    i++;
+                    auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
+                    int min_times = 0;
+                    int max_times = std::numeric_limits<int>::max();
+                    try {
+                        if (nums.size() == 1) {
+                            min_times = max_times = std::stoi(nums[0]);
+                        } else if (nums.size() != 2) {
+                            _errors.push_back("Wrong number of values in curly brackets");
+                        } else {
+                            if (!nums[0].empty()) {
+                                min_times = std::stoi(nums[0]);
+                            }
+                            if (!nums[1].empty()) {
+                                max_times = std::stoi(nums[1]);
+                            }
+                        }
+                    } catch (const std::invalid_argument & e) {
+                        _errors.push_back("Invalid number in curly brackets");
+                        return std::make_pair("", false);
+                    }
+                    auto &last = seq.back();
+                    auto &sub = last.first;
+                    auto sub_is_literal = last.second;
+
+                    if (min_times == 0 && max_times == std::numeric_limits<int>::max()) {
+                        sub += "*";
+                    } else if (min_times == 0 && max_times == 1) {
+                        sub += "?";
+                    } else if (min_times == 1 && max_times == std::numeric_limits<int>::max()) {
+                        sub += "+";
+                    } else {
+                        if (!sub_is_literal) {
+                            std::string & sub_id = sub_rule_ids[sub];
+                            if (sub_id.empty()) {
+                                sub_id = _add_rule(name + "-" + std::to_string(sub_rule_ids.size()), sub);
+                            }
+                            sub = sub_id;
+                        }
+                        std::string result;
+                        if (sub_is_literal && min_times > 0) {
+                            result = "\"" + repeat(sub.substr(1, sub.length() - 2), min_times) + "\"";
+                        } else {
+                            for (int j = 0; j < min_times; j++) {
+                                if (j > 0) {
+                                    result += " ";
+                                }
+                                result += sub;
+                            }
+                        }
+                        if (min_times > 0 && min_times < max_times) {
+                            result += " ";
+                        }
+                        if (max_times == std::numeric_limits<int>::max()) {
+                            result += sub + "*";
+                        } else {
+                            for (int j = min_times; j < max_times; j++) {
+                                if (j > min_times) {
+                                    result += " ";
+                                }
+                                result += sub + "?";
+                            }
+                        }
+                        seq.back().first = result;
+                        seq.back().second = false;
+                    }
+                } else {
+                    std::string literal;
+                    auto is_non_literal = [&](char c) {
+                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
+                    };
+                    while (i < length) {
+                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                            char next = sub_pattern[i + 1];
+                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
+                                i++;
+                                literal += sub_pattern[i];
+                                i++;
+                            } else {
+                                literal += sub_pattern.substr(i, 2);
+                                i += 2;
+                            }
+                        } else if (sub_pattern[i] == '"') {
+                            literal += "\\\"";
+                            i++;
+                        } else if (!is_non_literal(sub_pattern[i]) &&
+                                (i == length - 1 || literal.empty() || sub_pattern[i + 1] == '.' || !is_non_literal(sub_pattern[i + 1]))) {
+                            literal += sub_pattern[i];
+                            i++;
+                        } else {
+                            break;
+                        }
+                    }
+                    if (!literal.empty()) {
+                        seq.push_back(std::make_pair(literal, true));
+                    }
+                }
+            }
+            return join_seq();
+        };
+        return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
+    }
+
+    std::string _resolve_ref(const std::string & ref) {
+        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
+            _refs_being_resolved.insert(ref);
+            json resolved = _refs[ref];
+            ref_name = visit(resolved, ref_name);
+            _refs_being_resolved.erase(ref);
+        }
+        return ref_name;
+    }
+
+    std::string _build_object_rule(
+        const std::vector<std::pair<std::string, json>> & properties,
+        const std::unordered_set<std::string> & required,
+        const std::string & name,
+        const json & additional_properties)
+    {
+        std::vector<std::string> required_props;
+        std::vector<std::string> optional_props;
+        std::unordered_map<std::string, std::string> prop_kv_rule_names;
+        for (const auto & kv : properties) {
+            const auto &prop_name = kv.first;
+            const auto &prop_schema = kv.second;
+
+            std::string prop_rule_name = visit(prop_schema, name + (name.empty() ? "" : "-") + prop_name);
+            prop_kv_rule_names[prop_name] = _add_rule(
+                name + (name.empty() ? "" : "-") + prop_name + "-kv",
+                format_literal(json(prop_name).dump()) + " space \":\" space " + prop_rule_name
+            );
+            if (required.find(prop_name) != required.end()) {
+                required_props.push_back(prop_name);
+            } else {
+                optional_props.push_back(prop_name);
+            }
+        }
+        if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get<bool>())) {
+            std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
+            std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
+            std::string kv_rule = _add_rule(sub_name + "-kv", _add_rule("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            prop_kv_rule_names["*"] = kv_rule;
+            optional_props.push_back("*");
+        }
+
+        std::string rule = "\"{\" space ";
+        for (size_t i = 0; i < required_props.size(); i++) {
+            if (i > 0) {
+                rule += " \",\" space ";
+            }
+            rule += prop_kv_rule_names[required_props[i]];
+        }
+
+        if (!optional_props.empty()) {
+            rule += " (";
+            if (!required_props.empty()) {
+                rule += " \",\" space ( ";
+            }
+
+            std::function<std::string(const std::vector<std::string> &, bool)> get_recursive_refs = [&](const std::vector<std::string> & ks, bool first_is_optional) {
+                std::string res;
+                if (ks.empty()) {
+                    return res;
+                }
+                std::string k = ks[0];
+                std::string kv_rule_name = prop_kv_rule_names[k];
+                if (k == "*") {
+                    res = _add_rule(
+                        name + (name.empty() ? "" : "-") + "additional-kvs",
+                        kv_rule_name + " ( \",\" space " + kv_rule_name + " )*"
+                    );
+                } else if (first_is_optional) {
+                    res = "( \",\" space " + kv_rule_name + " )?";
+                } else {
+                    res = kv_rule_name;
+                }
+                if (ks.size() > 1) {
+                    res += " " + _add_rule(
+                        name + (name.empty() ? "" : "-") + k + "-rest",
+                        get_recursive_refs(std::vector<std::string>(ks.begin() + 1, ks.end()), true)
+                    );
+                }
+                return res;
+            };
+
+            for (size_t i = 0; i < optional_props.size(); i++) {
+                if (i > 0) {
+                    rule += " | ";
+                }
+                rule += get_recursive_refs(std::vector<std::string>(optional_props.begin() + i, optional_props.end()), false);
+            }
+            if (!required_props.empty()) {
+                rule += " )";
+            }
+            rule += " )?";
+        }
+
+        rule += " \"}\" space";
+
+        return rule;
+    }
+
+public:
+    SchemaConverter(
+        const std::function<json(const std::string &)> & fetch_json,
+        bool dotall)
+          : _fetch_json(fetch_json), _dotall(dotall)
+    {
+        _rules["space"] = SPACE_RULE;
+    }
+
+    void resolve_refs(json & schema, const std::string & url) {
+        /*
+        * Resolves all $ref fields in the given schema, fetching any remote schemas,
+        * replacing each $ref with absolute reference URL and populates _refs with the
+        * respective referenced (sub)schema dictionaries.
+        */
+        std::function<void(json &)> visit_refs = [&](json & n) {
+            if (n.is_array()) {
+                for (auto & x : n) {
+                    visit_refs(x);
+                }
+            } else if (n.is_object()) {
+                if (n.contains("$ref")) {
+                    std::string ref = n["$ref"];
+                    if (_refs.find(ref) == _refs.end()) {
+                        json target;
+                        if (ref.find("https://") == 0) {
+                            std::string base_url = ref.substr(0, ref.find('#'));
+                            auto it = _refs.find(base_url);
+                            if (it != _refs.end()) {
+                                target = it->second;
+                            } else {
+                                // Fetch the referenced schema and resolve its refs
+                                auto referenced = _fetch_json(ref);
+                                resolve_refs(referenced, base_url);
+                                _refs[base_url] = referenced;
+                            }
+                            if (ref.find('#') == std::string::npos || ref.substr(ref.find('#') + 1).empty()) {
+                                return;
+                            }
+                        } else if (ref.find("#/") == 0) {
+                            target = schema;
+                            n["$ref"] = url + ref;
+                            ref = url + ref;
+                        } else {
+                            _errors.push_back("Unsupported ref: " + ref);
+                            return;
+                        }
+                        std::string pointer = ref.substr(ref.find('#') + 1);
+                        std::vector<std::string> tokens = split(pointer, "/");
+                        for (size_t i = 1; i < tokens.size(); ++i) {
+                            std::string sel = tokens[i];
+                            if (target.is_null() || !target.contains(sel)) {
+                                _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                return;
+                            }
+                            target = target[sel];
+                        }
+                        _refs[ref] = target;
+                    }
+                } else {
+                    for (auto & kv : n.items()) {
+                        visit_refs(kv.value());
+                    }
+                }
+            }
+        };
+
+        visit_refs(schema);
+    }
+
+    std::string _generate_constant_rule(const json & value) {
+        return format_literal(value.dump());
+    }
+
+    std::string visit(const json & schema, const std::string & name) {
+        json schema_type = schema.contains("type") ? schema["type"] : json();
+        std::string schema_format = schema.contains("format") ? schema["format"].get<std::string>() : "";
+        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
+
+        if (schema.contains("$ref")) {
+            return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
+        } else if (schema.contains("oneOf") || schema.contains("anyOf")) {
+            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
+            return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
+        } else if (schema_type.is_array()) {
+            std::vector<json> schema_types;
+            for (const auto & t : schema_type) {
+                schema_types.push_back({{"type", t}});
+            }
+            return _add_rule(rule_name, _generate_union_rule(name, schema_types));
+        } else if (schema.contains("const")) {
+            return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
+        } else if (schema.contains("enum")) {
+            std::vector<std::string> enum_values;
+            for (const auto & v : schema["enum"]) {
+                enum_values.push_back(_generate_constant_rule(v));
+            }
+            return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | "));
+        } else if ((schema_type.is_null() || schema_type == "object")
+                && (schema.contains("properties") ||
+                    (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
+            std::unordered_set<std::string> required;
+            if (schema.contains("required") && schema["required"].is_array()) {
+                for (const auto & item : schema["required"]) {
+                    if (item.is_string()) {
+                        required.insert(item.get<std::string>());
+                    }
+                }
+            }
+            std::vector<std::pair<std::string, json>> properties;
+            if (schema.contains("properties")) {
+                for (const auto & prop : schema["properties"].items()) {
+                    properties.emplace_back(prop.key(), prop.value());
+                }
+            }
+            return _add_rule(rule_name,
+                _build_object_rule(
+                    properties, required, name,
+                    schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
+        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+            std::unordered_set<std::string> required;
+            std::vector<std::pair<std::string, json>> properties;
+            std::string hybrid_name = name;
+            std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
+                if (comp_schema.contains("$ref")) {
+                    add_component(_refs[comp_schema["$ref"]], is_required);
+                } else if (comp_schema.contains("properties")) {
+                    for (const auto & prop : comp_schema["properties"].items()) {
+                        properties.emplace_back(prop.key(), prop.value());
+                        if (is_required) {
+                            required.insert(prop.key());
+                        }
+                    }
+                } else {
+                  // todo warning
+                }
+            };
+            for (auto & t : schema["allOf"]) {
+                if (t.contains("anyOf")) {
+                    for (auto & tt : t["anyOf"]) {
+                        add_component(tt, false);
+                    }
+                } else {
+                    add_component(t, true);
+                }
+            }
+            return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
+        } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
+            json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
+            if (items.is_array()) {
+                std::string rule = "\"[\" space ";
+                for (size_t i = 0; i < items.size(); i++) {
+                    if (i > 0) {
+                        rule += " \",\" space ";
+                    }
+                    rule += visit(items[i], name + (name.empty() ? "" : "-") + "tuple-" + std::to_string(i));
+                }
+                rule += " \"]\" space";
+                return _add_rule(rule_name, rule);
+            } else {
+                std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
+                std::string list_item_operator = "( \",\" space " + item_rule_name + " )";
+                std::string successive_items;
+                int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
+                json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
+                int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : -1;
+                if (min_items > 0) {
+                    successive_items += repeat(list_item_operator, min_items - 1);
+                    min_items--;
+                }
+                if (max_items >= 0 && max_items > min_items) {
+                    successive_items += repeat(list_item_operator + "?", max_items - min_items - 1);
+                } else {
+                    successive_items += list_item_operator + "*";
+                }
+                std::string rule;
+                if (min_items == 0) {
+                    rule =  "\"[\" space ( " + item_rule_name + " " + successive_items + " )? \"]\" space";
+                } else {
+                    rule =  "\"[\" space " + item_rule_name + " " + successive_items + " \"]\" space";
+                }
+                return _add_rule(rule_name, rule);
+            }
+        } else if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
+            return _visit_pattern(schema["pattern"], rule_name);
+        } else if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
+            return _add_rule(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
+        } else if ((schema_type.is_null() || schema_type == "string") && DATE_RULES.find(schema_format) != DATE_RULES.end()) {
+            for (const auto & kv : DATE_RULES) {
+                _add_rule(kv.first, kv.second);
+            }
+            return schema_format + "-string";
+        } else if (schema.empty() || schema_type == "object") {
+            for (const auto & n : OBJECT_RULE_NAMES) {
+                _add_rule(n, PRIMITIVE_RULES.at(n));
+            }
+            return _add_rule(rule_name, "object");
+        } else {
+            if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
+                _errors.push_back("Unrecognized schema: " + schema.dump());
+                return "";
+            }
+            // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+            return _add_rule(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+        }
+    }
+
+    void check_errors() {
+        if (!_errors.empty()) {
+            throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
+        }
+        if (!_warnings.empty()) {
+            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
+        }
+    }
+
+    std::string format_grammar() {
+        std::stringstream ss;
+        for (const auto & kv : _rules) {
+            ss << kv.first << " ::= " << kv.second << std::endl;
+        }
+        return ss.str();
+    }
+};
+
+std::string json_schema_to_grammar(const json & schema) {
+    SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
+    auto copy = schema;
+    converter.resolve_refs(copy, "input");
+    converter.visit(copy, "");
+    converter.check_errors();
+    return converter.format_grammar();
+}

common/json-schema-to-grammar.h

@@ -0,0 +1,4 @@
+#pragma once
+#include "json.hpp"
+
+std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);

common/log.h

@@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER)
     #define LOG_IMPL(str, ...)                                                                                      \
     do {                                                                                                            \
         if (LOG_TARGET != nullptr)                                                                                  \
@@ -257,7 +257,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG_TEE() INSTEAD
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER)
     #define LOG_TEE_IMPL(str, ...)                                                                                                      \
     do {                                                                                                                                \
         if (LOG_TARGET != nullptr)                                                                                                      \
@@ -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
@@ -566,6 +566,7 @@ inline void log_print_usage()
     printf("  --log-new             Create a separate new log file on start. "
                                    "Each log file will have unique name: \"<name>.<ID>.log\"\n");
     printf("  --log-append          Don't truncate the old log file.\n");
+    printf("\n");
 }
 
 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)

common/ngram-cache.cpp

@@ -0,0 +1,282 @@
+#include "ngram-cache.h"
+#include "common.h"
+#include "log.h"
+
+#include <cstdint>
+#include <fstream>
+
+void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
+                              std::vector<llama_token> & inp, int nnew, bool print_progress) {
+    const int64_t t_start_ms = ggml_time_ms();
+    const int64_t inp_size = inp.size();
+
+    const int64_t n_todo = inp_size * (ngram_max - ngram_min + 1);
+    int64_t n_done = 0;
+
+    for (int64_t ngram_size = ngram_min; ngram_size <= ngram_max; ++ngram_size) {
+        const int64_t i_start = std::max(inp_size - nnew, ngram_size);
+        for (int64_t i = i_start; i < inp_size; ++i) {
+            const int64_t ngram_start = i - ngram_size;
+            llama_ngram ngram(&inp[ngram_start], ngram_size);
+            const llama_token token = inp[i];
+
+            llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
+            if (part_it == ngram_cache.end()) {
+                llama_ngram_cache_part part;
+                part.emplace(token, 1);
+                ngram_cache.emplace(ngram, part);
+            } else {
+                llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
+                if (token_count_it == part_it->second.end()) {
+                    part_it->second.emplace(token, 1);
+                } else {
+                    token_count_it->second++;
+                }
+            }
+            ++n_done;
+
+            if (print_progress && n_done % 10000000 == 0) {
+                const int64_t t_now_ms = ggml_time_ms();
+                const int64_t eta_ms   = (inp_size*(ngram_max-ngram_min+1) - n_done) * (t_now_ms - t_start_ms) / n_done;
+                const int64_t eta_min  = eta_ms / (60*1000);
+                const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+                fprintf(stderr, "%s: %" PRId64 "/%" PRId64 " done, ETA: %02" PRId64 ":%02" PRId64 "\n", __func__, n_done, n_todo, eta_min, eta_s);
+            }
+        }
+    }
+}
+
+// Helper function to get a token from the combined, speculative sequence of inp and draft.
+static llama_token get_token(const std::vector<llama_token> & inp, const std::vector<llama_token> & draft, const size_t i) {
+    return i < inp.size() ? inp[i] : draft[1 + i - inp.size()];
+}
+
+// If sample size or percentage are below these thresholds the draft is aborted early:
+constexpr int    draft_min_sample_size_lax[LLAMA_NGRAM_MAX] = { 2,  2,  1,  1};
+constexpr int        draft_min_percent_lax[LLAMA_NGRAM_MAX] = {66, 50, 50, 50};
+constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4,  3,  2,  2};
+constexpr int     draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
+
+// Helper function that tries to draft a token from only the static ngram cache:
+static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) {
+    llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+    if (part_static_it == nc_static.end()) {
+        return -1;
+    }
+    const llama_ngram_cache_part part_static = part_static_it->second;
+
+    int max_count_static  = 0;
+    int sum_count_static  = 0;
+    llama_token max_token = -1;
+
+    for (std::pair<llama_token, int> token_count_static : part_static) {
+        const llama_token token = token_count_static.first;
+        const int32_t count_static  = token_count_static.second;
+
+        if (count_static > max_count_static) {
+            max_token        = token;
+            max_count_static = count_static;
+        }
+        sum_count_static += count_static;
+    }
+
+    if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
+        return -1;
+    }
+    if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
+        return -1;
+    }
+    return max_token;
+}
+
+// Try to draft a token from primary cache (context/dynamic), validate with static cache:
+static llama_token try_draft(
+    llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static,
+    const int * min_sample_size, const int * min_percent) {
+
+    llama_token drafted_token = -1;
+
+    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
+        const llama_ngram ngram_primary = ngrams_primary[i];
+
+        llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
+        if (part_primary_it == nc_primary.end()) {
+            continue;
+        }
+        const llama_ngram_cache_part part_primary = part_primary_it->second;
+
+        int max_count_primary = 0;
+        int max_count_static  = 0;
+        int sum_count_primary = 0;
+        llama_token max_token = -1;
+
+        for (std::pair<llama_token, int> token_count_primary : part_primary) {
+            const llama_token token = token_count_primary.first;
+
+            llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
+
+            const int32_t count_primary = token_count_primary.second;
+            const int32_t count_static  = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
+
+            if (count_primary*count_static > max_count_primary*max_count_static) {
+                max_token         = token;
+                max_count_primary = count_primary;
+                max_count_static  = count_static;
+            }
+            sum_count_primary += count_primary;
+        }
+
+        if (sum_count_primary < min_sample_size[i]) {
+            continue;
+        }
+        if (100*max_count_primary < min_percent[i]*sum_count_primary) {
+            continue;;
+        }
+        drafted_token = max_token;
+    }
+
+    return drafted_token;
+}
+
+void llama_ngram_cache_draft(
+    std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
+    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static
+) {
+    GGML_ASSERT(draft.size() == 1);
+    const int inp_size = inp.size();
+
+    if (inp_size < LLAMA_NGRAM_STATIC) {
+        return;
+    }
+
+    while ((int) draft.size()-1 < n_draft) {
+        llama_token drafted_token = -1;
+
+        const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
+        llama_ngram ngram_static;
+        for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
+            ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
+        }
+        llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+        llama_ngram_cache_part part_static;
+        if (part_static_it != nc_static.end()) {
+            part_static = part_static_it->second;
+        }
+
+        // cd = context + dynamic
+        std::vector<llama_ngram> ngrams_cd;
+        for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
+            const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
+            llama_ngram ngram_cd;
+            for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
+                ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
+            }
+            ngrams_cd.push_back(ngram_cd);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_static, ngram_static);
+        }
+
+        if (drafted_token == -1) {
+            break;
+        }
+
+        LOG(" - draft candidate: token=%d\n", drafted_token);
+        draft.push_back(drafted_token);
+    }
+}
+
+void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) {
+    std::ofstream file_out(filename, std::ios::binary);
+    for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) {
+        const llama_ngram      ngram        = item.first;
+        llama_ngram_cache_part token_counts = item.second;
+        GGML_ASSERT(!token_counts.empty());
+        const int32_t ntokens = token_counts.size();
+        GGML_ASSERT(ntokens > 0);
+
+        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(llama_ngram));
+        file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
+        for (std::pair<llama_token, int32_t> item2 : token_counts) {
+            const llama_token token = item2.first;
+            const int32_t     count = item2.second;
+            GGML_ASSERT(count > 0);
+
+            file_out.write(reinterpret_cast<const char *>(&token), sizeof(llama_token));
+            file_out.write(reinterpret_cast<const char *>(&count), sizeof(int32_t));
+        }
+    }
+
+}
+
+llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
+    std::ifstream hashmap_file(filename, std::ios::binary);
+    if (!hashmap_file) {
+        throw std::ifstream::failure("Unable to open file " + filename);
+    }
+    llama_ngram_cache ngram_cache;
+
+    llama_ngram ngram;
+    int32_t     ntokens;
+    llama_token token;
+    int32_t     count;
+
+    char * ngramc   = reinterpret_cast<char*>(&ngram);
+    char * ntokensc = reinterpret_cast<char*>(&ntokens);
+    char * tokenc   = reinterpret_cast<char*>(&token);
+    char * countc   = reinterpret_cast<char*>(&count);
+    while(hashmap_file.read(ngramc, sizeof(llama_ngram))) {
+        GGML_ASSERT(!hashmap_file.eof());
+        GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
+        GGML_ASSERT(ntokens > 0);
+        llama_ngram_cache_part token_counts;
+
+        for (int i = 0; i < ntokens; ++i) {
+            GGML_ASSERT(!hashmap_file.eof());
+            GGML_ASSERT(hashmap_file.read(tokenc, sizeof(llama_token)));
+            GGML_ASSERT(!hashmap_file.eof());
+            GGML_ASSERT(hashmap_file.read(countc, sizeof(int32_t)));
+            GGML_ASSERT(count > 0);
+            token_counts.emplace(token, count);
+        }
+
+        ngram_cache.emplace(ngram, token_counts);
+    }
+    GGML_ASSERT(hashmap_file.eof());
+
+    return ngram_cache;
+}
+
+void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) {
+    for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) {
+        const llama_ngram      ngram = ngram_part.first;
+        llama_ngram_cache_part  part = ngram_part.second;
+
+        llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
+        if (part_merged_it == ngram_cache_target.end()) {
+            ngram_cache_target.emplace(ngram, part);
+            continue;
+        }
+
+        for (std::pair<llama_token, int32_t> token_count : part) {
+            const llama_token token = token_count.first;
+            const int32_t     count = token_count.second;
+            GGML_ASSERT(count > 0);
+
+            llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
+            if (token_count_merged_it == part_merged_it->second.end()) {
+                part_merged_it->second.emplace(token, count);
+                continue;
+            }
+
+            token_count_merged_it->second += count;
+        }
+    }
+}

common/ngram-cache.h

@@ -0,0 +1,94 @@
+#pragma once
+
+#include "llama.h"
+
+#include <unordered_map>
+#include <string>
+#include <vector>
+
+#define LLAMA_NGRAM_MIN    1
+#define LLAMA_NGRAM_MAX    4
+#define LLAMA_NGRAM_STATIC 2
+
+// Data structures to map n-grams to empirical token probabilities:
+
+struct llama_ngram {
+    llama_token tokens[LLAMA_NGRAM_MAX];
+
+    llama_ngram() {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            tokens[i] = -1;
+        }
+    }
+
+    llama_ngram(const llama_token * input, const int ngram_size) {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            tokens[i] = i < ngram_size ? input[i] : -1;
+        }
+    }
+
+    bool operator==(const llama_ngram & other) const {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            if (tokens[i] != other.tokens[i]) {
+                return false;
+            }
+        }
+        return true;
+    }
+};
+
+struct llama_ngram_hash_function {
+    size_t operator()(const llama_ngram & ngram) const {
+        size_t hash = 0;
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
+        }
+        return hash;
+    }
+};
+
+// token -> number of times token has been seen
+typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part;
+
+// n-gram -> empirical distribution of following tokens
+typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
+
+
+// Update an ngram cache with tokens.
+// ngram_cache:         the cache to modify.
+// ngram_min/ngram_max: the min/max size of the ngrams to extract from inp_data.
+// inp_data:            the token sequence with which to update ngram_cache.
+// nnew:                how many new tokens have been appended to inp_data since the last call to this function.
+// print_progress:      whether to print progress to stderr.
+//
+// In order to get correct results inp_data can ONLY BE APPENDED TO.
+// Changes in the middle need a complete rebuild.
+void llama_ngram_cache_update(
+    llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
+
+// Try to draft tokens from ngram caches.
+// inp:                the tokens generated so far.
+// draft:              the token sequence to draft. Expected to initially contain the previously sampled token.
+// n_draft:            maximum number of tokens to add to draft.
+// ngram_min/gram_max: the min/max size of the ngrams in nc_context and nc_dynamic.
+// nc_context:         ngram cache based on current context.
+// nc_dynamic:         ngram cache based on previous user generations.
+// nc_static:          ngram cache generated from a large text corpus, used for validation.
+void llama_ngram_cache_draft(
+    std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
+    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static);
+
+// Save an ngram cache to a file.
+// ngram_cache: the ngram cache to save.
+// filename:    the path under which to save the ngram cache.
+void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename);
+
+// Load an ngram cache saved with llama_ngram_cache_save.
+// filename: the path from which to load the ngram cache.
+// returns:  an ngram cache containing the information saved to filename.
+llama_ngram_cache llama_ngram_cache_load(std::string & filename);
+
+// Merge two ngram caches.
+// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
+// ngram_cache_add:    the ngram cache to add to ngram_cache_target.
+void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add);

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(
@@ -161,77 +168,20 @@ static llama_token llama_sampling_sample_impl(
                   bool is_resampling) {  // Add a parameter to indicate if we are resampling
     const llama_sampling_params & params = ctx_sampling->params;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
     const float   temp            = params.temp;
-    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
-    const float   penalty_repeat  = params.penalty_repeat;
-    const float   penalty_freq    = params.penalty_freq;
-    const float   penalty_present = params.penalty_present;
     const int     mirostat        = params.mirostat;
     const float   mirostat_tau    = params.mirostat_tau;
     const float   mirostat_eta    = params.mirostat_eta;
-    const bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
 
+    std::vector<float> original_logits;
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
+    if (!is_resampling) {
+        GGML_ASSERT(!original_logits.empty());
+    }
     llama_token id = 0;
-
     // 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;
-
-    if (!is_resampling) {
-        // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this.
-        original_logits = std::vector<float>(logits, logits + llama_n_vocab(llama_get_model(ctx_main)));
-    }
-
-    // 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;
-                }
-            }
-        }
-    }
-
-    // If we are in the resampling phase, apply grammar checks before sampling logic
-    if (is_resampling && ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    }
-
     if (temp < 0.0) {
         // greedy sampling, with probs
         llama_sample_softmax(ctx_main, &cur_p);
@@ -295,6 +245,81 @@ static llama_token llama_sampling_sample_impl(
     return id;
 }
 
+static llama_token_data_array llama_sampling_prepare_impl(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx,
+                  bool apply_grammar,
+                  std::vector<float> * original_logits) {
+    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);
+
+    if (apply_grammar && original_logits != NULL) {
+        // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
+        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
+    }
+
+    // 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 before sampling logic
+    if (apply_grammar && ctx_sampling->grammar != NULL) {
+        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
+    }
+
+    return cur_p;
+}
+
 llama_token llama_sampling_sample(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
@@ -304,6 +329,16 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
+llama_token_data_array llama_sampling_prepare(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx,
+                  bool apply_grammar,
+                  std::vector<float> * original_logits) {
+    return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits);
+}
+
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -32,13 +32,13 @@ 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
+    bool        penalize_nl           = false;     // consider newlines as a repeatable token
 
     std::vector<llama_sampler_type> samplers_sequence = {
         llama_sampler_type::TOP_K,
@@ -131,6 +131,15 @@ llama_token llama_sampling_sample(
         struct llama_context * ctx_cfg,
         int idx = 0);
 
+// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters.
+llama_token_data_array llama_sampling_prepare(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        struct llama_context * ctx_cfg,
+        int idx = 0,
+        bool apply_grammar = true,
+        std::vector<float> * original_logits = nullptr);
+
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

convert-hf-to-gguf.py

@@ -8,9 +8,10 @@ import json
 import os
 import re
 import sys
+from abc import ABC, abstractmethod
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
+from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterator, Sequence, TypeVar, cast
 
 import numpy as np
 import torch
@@ -36,7 +37,12 @@ class SentencePieceTokenTypes(IntEnum):
     BYTE = 6
 
 
-class Model:
+AnyModel = TypeVar("AnyModel", bound="type[Model]")
+
+
+class Model(ABC):
+    _model_classes: dict[str, type[Model]] = {}
+
     def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
         self.dir_model = dir_model
         self.ftype = ftype
@@ -47,10 +53,14 @@ class Model:
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
-        self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
 
+    @property
+    @abstractmethod
+    def model_arch(self) -> gguf.MODEL_ARCH:
+        pass
+
     def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
         key = next((k for k in keys if k in self.hparams), None)
         if key is not None:
@@ -83,29 +93,42 @@ class Model:
 
         if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
+            print(f"gguf: context length = {n_ctx}")
 
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         self.gguf_writer.add_embedding_length(n_embd)
+        print(f"gguf: embedding length = {n_embd}")
 
         if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
+            print(f"gguf: feed forward length = {n_ff}")
 
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         self.gguf_writer.add_head_count(n_head)
+        print(f"gguf: head count = {n_head}")
 
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
+            print(f"gguf: key-value head count = {n_head_kv}")
 
+        if (rope_theta := self.hparams.get("rope_theta")) is not None:
+            self.gguf_writer.add_rope_freq_base(rope_theta)
+            print(f"gguf: rope theta = {rope_theta}")
         if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
             self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
-        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
+            print(f"gguf: rms norm epsilon = {f_rms_eps}")
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+            print(f"gguf: layer norm epsilon = {f_norm_eps}")
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
+            print(f"gguf: expert count = {n_experts}")
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
+            print(f"gguf: experts used count = {n_experts_used}")
 
         self.gguf_writer.add_file_type(self.ftype)
+        print(f"gguf: file type = {self.ftype}")
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@@ -174,53 +197,22 @@ class Model:
         with open(dir_model / "config.json", "r", encoding="utf-8") as f:
             return json.load(f)
 
-    @staticmethod
-    def from_model_architecture(model_architecture):
-        if model_architecture == "GPTNeoXForCausalLM":
-            return GPTNeoXModel
-        if model_architecture == "BloomForCausalLM":
-            return BloomModel
-        if model_architecture == "MPTForCausalLM":
-            return MPTModel
-        if model_architecture in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
-            return BaichuanModel
-        if model_architecture in ("FalconForCausalLM", "RWForCausalLM"):
-            return FalconModel
-        if model_architecture == "GPTBigCodeForCausalLM":
-            return StarCoderModel
-        if model_architecture == "GPTRefactForCausalLM":
-            return RefactModel
-        if model_architecture == "PersimmonForCausalLM":
-            return PersimmonModel
-        if model_architecture in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
-            return StableLMModel
-        if model_architecture == "QWenLMHeadModel":
-            return QwenModel
-        if model_architecture == "Qwen2ForCausalLM":
-            return Model
-        if model_architecture == "MixtralForCausalLM":
-            return MixtralModel
-        if model_architecture == "GPT2LMHeadModel":
-            return GPT2Model
-        if model_architecture == "PhiForCausalLM":
-            return Phi2Model
-        if model_architecture == "PlamoForCausalLM":
-            return PlamoModel
-        if model_architecture == "CodeShellForCausalLM":
-            return CodeShellModel
-        if model_architecture == "OrionForCausalLM":
-            return OrionModel
-        if model_architecture == "InternLM2ForCausalLM":
-            return InternLM2Model
-        if model_architecture == "MiniCPMForCausalLM":
-            return MiniCPMModel
-        if model_architecture == "BertModel":
-            return BertModel
-        if model_architecture == "NomicBertModel":
-            return NomicBertModel
-        if model_architecture == "GemmaForCausalLM":
-            return GemmaModel
-        return Model
+    @classmethod
+    def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
+        assert names
+
+        def func(modelcls: type[Model]):
+            for name in names:
+                cls._model_classes[name] = modelcls
+            return modelcls
+        return func
+
+    @classmethod
+    def from_model_architecture(cls, arch):
+        try:
+            return cls._model_classes[arch]
+        except KeyError:
+            raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
 
     def _is_model_safetensors(self) -> bool:
         return Model.count_model_parts(self.dir_model, ".safetensors") > 0
@@ -235,55 +227,6 @@ class Model:
             return ("pytorch_model.bin",)
         return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
 
-    def _get_model_architecture(self) -> gguf.MODEL_ARCH:
-        arch = self.hparams["architectures"][0]
-        if arch == "GPTNeoXForCausalLM":
-            return gguf.MODEL_ARCH.GPTNEOX
-        if arch == "BloomForCausalLM":
-            return gguf.MODEL_ARCH.BLOOM
-        if arch == "MPTForCausalLM":
-            return gguf.MODEL_ARCH.MPT
-        if arch in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
-            return gguf.MODEL_ARCH.BAICHUAN
-        if arch in ("FalconForCausalLM", "RWForCausalLM"):
-            return gguf.MODEL_ARCH.FALCON
-        if arch == "GPTBigCodeForCausalLM":
-            return gguf.MODEL_ARCH.STARCODER
-        if arch == "GPTRefactForCausalLM":
-            return gguf.MODEL_ARCH.REFACT
-        if arch == "PersimmonForCausalLM":
-            return gguf.MODEL_ARCH.PERSIMMON
-        if arch in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
-            return gguf.MODEL_ARCH.STABLELM
-        if arch == "QWenLMHeadModel":
-            return gguf.MODEL_ARCH.QWEN
-        if arch == "Qwen2ForCausalLM":
-            return gguf.MODEL_ARCH.QWEN2
-        if arch == "MixtralForCausalLM":
-            return gguf.MODEL_ARCH.LLAMA
-        if arch == "GPT2LMHeadModel":
-            return gguf.MODEL_ARCH.GPT2
-        if arch == "PhiForCausalLM":
-            return gguf.MODEL_ARCH.PHI2
-        if arch == "PlamoForCausalLM":
-            return gguf.MODEL_ARCH.PLAMO
-        if arch == "CodeShellForCausalLM":
-            return gguf.MODEL_ARCH.CODESHELL
-        if arch == "OrionForCausalLM":
-            return gguf.MODEL_ARCH.ORION
-        if arch == "InternLM2ForCausalLM":
-            return gguf.MODEL_ARCH.INTERNLM2
-        if arch == "MiniCPMForCausalLM":
-            return gguf.MODEL_ARCH.MINICPM
-        if arch == "BertModel":
-            return gguf.MODEL_ARCH.BERT
-        if arch == "NomicBertModel":
-            return gguf.MODEL_ARCH.NOMIC_BERT
-        if arch == "GemmaForCausalLM":
-            return gguf.MODEL_ARCH.GEMMA
-
-        raise NotImplementedError(f'Architecture "{arch}" not supported!')
-
     def _set_vocab_gpt2(self):
         dir_model = self.dir_model
         hparams = self.hparams
@@ -388,7 +331,7 @@ class Model:
         tokenizer = SentencePieceProcessor(str(tokenizer_path))
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
-        for token_id in range(vocab_size):
+        for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.get_score(token_id)
@@ -413,9 +356,13 @@ class Model:
                 added_tokens_json = json.load(f)
 
                 for key in added_tokens_json:
-                    tokens.append(key.encode("utf-8"))
-                    scores.append(-1000.0)
-                    toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    key = key.encode("utf-8")
+                    if key not in tokens:
+                        tokens.append(key)
+                        scores.append(-1000.0)
+                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+
+        assert len(tokens) == vocab_size
 
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_token_list(tokens)
@@ -451,7 +398,10 @@ class Model:
         special_vocab.add_to_gguf(self.gguf_writer)
 
 
+@Model.register("GPTNeoXForCausalLM")
 class GPTNeoXModel(Model):
+    model_arch = gguf.MODEL_ARCH.GPTNEOX
+
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
 
@@ -468,7 +418,10 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
 
+@Model.register("BloomForCausalLM")
 class BloomModel(Model):
+    model_arch = gguf.MODEL_ARCH.BLOOM
+
     def set_gguf_parameters(self):
         self.gguf_writer.add_name("Bloom")
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
@@ -560,7 +513,10 @@ class BloomModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
+@Model.register("MPTForCausalLM")
 class MPTModel(Model):
+    model_arch = gguf.MODEL_ARCH.MPT
+
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
         self.gguf_writer.add_name(self.dir_model.name)
@@ -623,7 +579,10 @@ class MPTModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("OrionForCausalLM")
 class OrionModel(Model):
+    model_arch = gguf.MODEL_ARCH.ORION
+
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -702,7 +661,10 @@ class OrionModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
 class BaichuanModel(Model):
+    model_arch = gguf.MODEL_ARCH.BAICHUAN
+
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -817,7 +779,10 @@ class BaichuanModel(Model):
         return weights[r * n_part:r * n_part + r, ...]
 
 
+@Model.register("FalconForCausalLM", "RWForCausalLM")
 class FalconModel(Model):
+    model_arch = gguf.MODEL_ARCH.FALCON
+
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_hidden_layers")
         if block_count is None:
@@ -910,7 +875,10 @@ class FalconModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("GPTBigCodeForCausalLM")
 class StarCoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.STARCODER
+
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -925,7 +893,10 @@ class StarCoderModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
 
+@Model.register("GPTRefactForCausalLM")
 class RefactModel(Model):
+    model_arch = gguf.MODEL_ARCH.REFACT
+
     def set_gguf_parameters(self):
         hidden_dim = self.hparams["n_embd"]
         inner_dim = 4 * hidden_dim
@@ -1009,7 +980,10 @@ class RefactModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("PersimmonForCausalLM")
 class PersimmonModel(Model):
+    model_arch = gguf.MODEL_ARCH.PERSIMMON
+
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
         head_count = self.hparams["num_attention_heads"]
@@ -1057,7 +1031,10 @@ class PersimmonModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
 class StableLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.STABLELM
+
     def set_vocab(self):
         if (self.dir_model / "tokenizer.json").is_file():
             self._set_vocab_gpt2()
@@ -1081,12 +1058,33 @@ class StableLMModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
 
 
+@Model.register("MixtralForCausalLM")
 class MixtralModel(Model):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
 
+@Model.register("GrokForCausalLM")
+class GrokModel(Model):
+    model_arch = gguf.MODEL_ARCH.GROK
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_name("Grok")
+
+
+@Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
+    model_arch = gguf.MODEL_ARCH.MINICPM
+
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
         self.gguf_writer.add_name("MiniCPM")
@@ -1163,7 +1161,10 @@ class MiniCPMModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("QWenLMHeadModel")
 class QwenModel(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN
+
     @staticmethod
     def token_bytes_to_string(b):
         from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
@@ -1243,7 +1244,15 @@ class QwenModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("Qwen2ForCausalLM")
+class Qwen2Model(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN2
+
+
+@Model.register("GPT2LMHeadModel")
 class GPT2Model(Model):
+    model_arch = gguf.MODEL_ARCH.GPT2
+
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_block_count(self.hparams["n_layer"])
@@ -1305,7 +1314,10 @@ class GPT2Model(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
+@Model.register("PhiForCausalLM")
 class Phi2Model(Model):
+    model_arch = gguf.MODEL_ARCH.PHI2
+
     def set_gguf_parameters(self):
         block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
@@ -1327,7 +1339,10 @@ class Phi2Model(Model):
         self.gguf_writer.add_add_bos_token(False)
 
 
+@Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLAMO
+
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -1406,7 +1421,10 @@ class PlamoModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("CodeShellForCausalLM")
 class CodeShellModel(Model):
+    model_arch = gguf.MODEL_ARCH.CODESHELL
+
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -1471,7 +1489,10 @@ class CodeShellModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
+@Model.register("InternLM2ForCausalLM")
 class InternLM2Model(Model):
+    model_arch = gguf.MODEL_ARCH.INTERNLM2
+
     def set_vocab(self):
         # (TODO): Is there a better way?
         # Copy from _set_vocab_sentencepiece, The only difference is that we will treat the character
@@ -1643,7 +1664,10 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
+@Model.register("BertModel", "CamembertModel")
 class BertModel(Model):
+    model_arch = gguf.MODEL_ARCH.BERT
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.vocab_size = None
@@ -1653,16 +1677,17 @@ class BertModel(Model):
         self.gguf_writer.add_causal_attention(False)
 
         # get pooling path
-        with open(self.dir_model / "modules.json", encoding="utf-8") as f:
-            modules = json.load(f)
         pooling_path = None
-        for mod in modules:
-            if mod["type"] == "sentence_transformers.models.Pooling":
-                pooling_path = mod["path"]
-                break
+        module_path = self.dir_model / "modules.json"
+        if module_path.is_file():
+            with open(module_path, encoding="utf-8") as f:
+                modules = json.load(f)
+            for mod in modules:
+                if mod["type"] == "sentence_transformers.models.Pooling":
+                    pooling_path = mod["path"]
+                    break
 
         # get pooling type
-        pooling_type = gguf.PoolingType.NONE
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
@@ -1672,8 +1697,7 @@ class BertModel(Model):
                 pooling_type = gguf.PoolingType.CLS
             else:
                 raise NotImplementedError("Only MEAN and CLS pooling types supported")
-
-        self.gguf_writer.add_pooling_type(pooling_type.value)
+            self.gguf_writer.add_pooling_type(pooling_type)
 
     def set_vocab(self):
         path = self.dir_model
@@ -1749,7 +1773,10 @@ class BertModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("NomicBertModel")
 class NomicBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
 
@@ -1786,7 +1813,10 @@ class NomicBertModel(BertModel):
             yield name, data
 
 
+@Model.register("GemmaForCausalLM")
 class GemmaModel(Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA
+
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -1811,16 +1841,15 @@ class GemmaModel(Model):
         tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
 
         for name, data_torch in self.get_tensors():
-            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
-            if name.endswith("norm.weight"):
-                data_torch = data_torch + 1
-
             old_dtype = data_torch.dtype
 
             # convert any unsupported data types to float32
             if data_torch.dtype not in (torch.float16, torch.float32):
                 data_torch = data_torch.to(torch.float32)
 
+            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
+            if name.endswith("norm.weight"):
+                data_torch = data_torch + 1
             data = data_torch.squeeze().numpy()
 
             # map tensor names
@@ -1843,6 +1872,146 @@ class GemmaModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("Starcoder2ForCausalLM")
+class StarCoder2Model(Model):
+    model_arch = gguf.MODEL_ARCH.STARCODER2
+
+
+@Model.register("MambaForCausalLM", "MambaLMHeadModel")
+class MambaModel(Model):
+    model_arch = gguf.MODEL_ARCH.MAMBA
+
+    def set_vocab(self):
+        vocab_size = self.hparams["vocab_size"]
+        # Round vocab size to next multiple of 8
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
+        # pad using ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        if (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        else:
+            # Use the GPT-NeoX tokenizer when no tokenizer files are present
+            tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
+            print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
+            neox_reader = gguf.GGUFReader(tokenizer_path, "r")
+
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
+            self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
+            self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
+            self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
+            self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
+            self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
+            self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
+            self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size", "d_model"])
+        d_conv  = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
+        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+        d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
+        # ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
+        dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
+
+        # Fail early for models which don't have a block expansion factor of 2
+        assert d_inner == 2 * d_model
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
+        self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def write_tensors(self):
+        block_count = self.hparams["n_layer"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        tok_embd = None
+        tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight"
+        output_name   = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT]     + ".weight"
+
+        for name, data_torch in self.get_tensors():
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            if name.endswith(".A_log"):
+                print("A_log --> A ==> " + new_name)
+                data_torch = -torch.exp(data_torch)
+
+            # assuming token_embd.weight is seen before output.weight
+            if tok_embd is not None and new_name == output_name:
+                if torch.equal(tok_embd, data_torch):
+                    print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
+                    continue
+            if new_name == tok_embd_name:
+                tok_embd = data_torch
+
+            data = data_torch.squeeze().numpy()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert big float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+@Model.register("CohereForCausalLM")
+class CommandR2Model(Model):
+    model_arch = gguf.MODEL_ARCH.COMMAND_R
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # max_position_embeddings = 8192 in config.json but model was actually
+        # trained on 128k context length
+        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+
 ###### CONVERSION LOGIC ######
 
 

convert-llama-ggml-to-gguf.py

@@ -373,7 +373,7 @@ def handle_metadata(cfg, hp):
         raise ValueError('Unable to load metadata')
     vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
     vocab_factory = convert.VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, 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.py

@@ -332,6 +332,9 @@ 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())
         if isinstance(self.bpe_tokenizer.get('model'), dict):
@@ -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
@@ -553,7 +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"
 
 
 #
@@ -935,8 +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:
@@ -977,6 +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)
@@ -1013,21 +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 = []
@@ -1043,11 +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)
@@ -1074,6 +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()
 
@@ -1082,7 +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)
 
@@ -1120,8 +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():
@@ -1131,24 +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()
 
@@ -1156,9 +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
@@ -1282,38 +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,
@@ -1322,30 +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
@@ -1379,15 +1391,14 @@ 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 = 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",   choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
+    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)")
@@ -1397,18 +1408,10 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
-    if args.awq_path:
-        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
-        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
-        tmp_model_path = args.model / "weighted_model"
-        if tmp_model_path.is_dir():
-            print(f"{tmp_model_path} exists as a weighted model.")
-        else:
-            tmp_model_path.mkdir(parents=True, exist_ok=True)
-            print("Saving new weighted model ...")
-            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
-            print(f"Saved weighted model at {tmp_model_path}.")
-        args.model = tmp_model_path
+    if args.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)
@@ -1448,7 +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:
@@ -1459,7 +1462,7 @@ 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}")

docs/token_generation_performance_tips.md

@@ -1,7 +1,7 @@
 # Token generation performance troubleshooting
 
-## Verifying that the model is running on the GPU with cuBLAS
-Make sure you compiled llama with the correct env variables according to [this guide](../README.md#cublas), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
+## Verifying that the model is running on the GPU with CUDA
+Make sure you compiled llama with the correct env variables according to [this guide](../README.md#CUDA), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
 ```shell
 ./main -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
 ```

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)
@@ -32,6 +34,7 @@ else()
     add_subdirectory(perplexity)
     add_subdirectory(quantize)
     add_subdirectory(quantize-stats)
+    add_subdirectory(retrieval)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
     add_subdirectory(passkey)

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, };
@@ -65,19 +64,15 @@ 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
@@ -106,11 +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) {
@@ -141,6 +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;
@@ -159,7 +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 = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%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");
@@ -180,10 +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;
 
@@ -198,7 +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/batched.cpp

@@ -48,6 +48,8 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    process_escapes(params.prompt);
+
     // init LLM
 
     llama_backend_init();
@@ -78,8 +80,9 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed  = 1234;
-    ctx_params.n_ctx = n_kv_req;
+    ctx_params.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;
 
@@ -132,7 +135,7 @@ 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/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/README.md

@@ -21,6 +21,8 @@ An example command using a model from [karpathy/tinyllamas](https://huggingface.
 
 `$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
+Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
+
 Now you can use the model with a command like:
 
 `$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

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

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "log.h"
 
 #include <unordered_map>
 #include <vector>
@@ -78,111 +79,101 @@ typedef struct {
 
 struct TransformerWeights {
     // token embedding table
-    float* token_embedding_table;    // (vocab_size, dim)
+    std::vector<float> token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
-    float* rms_att_weight; // (layer, dim) rmsnorm weights
-    float* rms_ffn_weight; // (layer, dim)
+    std::vector<float> rms_att_weight; // (layer, dim) rmsnorm weights
+    std::vector<float> rms_ffn_weight; // (layer, dim)
     // weights for matmuls
-    float* wq; // (layer, dim, dim)
-    float* wk; // (layer, dim, dim)
-    float* wv; // (layer, dim, dim)
-    float* wo; // (layer, dim, dim)
+    std::vector<float> wq; // (layer, dim, dim)
+    std::vector<float> wk; // (layer, dim, dim)
+    std::vector<float> wv; // (layer, dim, dim)
+    std::vector<float> wo; // (layer, dim, dim)
     // weights for ffn
-    float* w1; // (layer, hidden_dim, dim)
-    float* w2; // (layer, dim, hidden_dim)
-    float* w3; // (layer, hidden_dim, dim)
+    std::vector<float> w1; // (layer, hidden_dim, dim)
+    std::vector<float> w2; // (layer, dim, hidden_dim)
+    std::vector<float> w3; // (layer, hidden_dim, dim)
     // final rmsnorm
-    float* rms_final_weight; // (dim,)
+    std::vector<float> rms_final_weight; // (dim,)
     // freq_cis for RoPE relatively positional embeddings
-    // float* freq_cis_real; // (seq_len, dim/2)
-    // float* freq_cis_imag; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_real; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    float* wcls;
-
-    ~TransformerWeights() {
-        delete[] token_embedding_table;
-        delete[] rms_att_weight;
-        delete[] rms_ffn_weight;
-        delete[] wq;
-        delete[] wk;
-        delete[] wv;
-        delete[] wo;
-        delete[] w1;
-        delete[] w2;
-        delete[] w3;
-        delete[] rms_final_weight;
-        delete[] wcls;
-    }
+    std::vector<float> wcls;
 };
 
-static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
-    // we calloc instead of malloc to keep valgrind happy
-    w->token_embedding_table = new float[p->vocab_size * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_weights) {
+    const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
+    try {
+        w->token_embedding_table.resize(p->vocab_size * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
 
-    w->rms_att_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+        w->rms_att_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
 
-    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+        w->rms_ffn_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
 
-    w->wq = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wq.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->wk = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wv = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wo = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wo.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->rms_final_weight = new float[p->dim]();
-    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        w->rms_final_weight.resize(p->dim);
+        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
-    if (shared_weights) {
-        w->wcls = NULL;
-    } else {
-        w->wcls = new float[p->vocab_size * p->dim]();
-        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        if (shared_weights) {
+            w->wcls = {};
+        } else {
+            w->wcls.resize(p->vocab_size * p->dim);
+            LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        }
+    }
+    catch (std::length_error &) {
+        die("Invalid configuration. Failed to allocate memory for weights");
     }
 }
 
-static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
-    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
-    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
-    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FILE * f, bool shared_weights) {
+    if (fread(w->token_embedding_table.data(), sizeof(float), w->token_embedding_table.size(), f) != w->token_embedding_table.size()) return 1;
+    if (fread(w->rms_att_weight.data(), sizeof(float), w->rms_att_weight.size(), f) != w->rms_att_weight.size()) return 1;
+    if (fread(w->wq.data(), sizeof(float), w->wq.size(), f) != w->wq.size()) return 1;
+    if (fread(w->wk.data(), sizeof(float), w->wk.size(), f) != w->wk.size()) return 1;
+    if (fread(w->wv.data(), sizeof(float), w->wv.size(), f) != w->wv.size()) return 1;
+    if (fread(w->wo.data(), sizeof(float), w->wo.size(), f) != w->wo.size()) return 1;
+    if (fread(w->rms_ffn_weight.data(), sizeof(float), w->rms_ffn_weight.size(), f) != w->rms_ffn_weight.size()) return 1;
+    if (fread(w->w1.data(), sizeof(float), w->w1.size(), f) != w->w1.size()) return 1;
+    if (fread(w->w2.data(), sizeof(float), w->w2.size(), f) != w->w2.size()) return 1;
+    if (fread(w->w3.data(), sizeof(float), w->w3.size(), f) != w->w3.size()) return 1;
+    if (fread(w->rms_final_weight.data(), sizeof(float), w->rms_final_weight.size(), f) != w->rms_final_weight.size()) return 1;
 
     // Skip freq_cis_real & freq_cis_imag
     int head_size = p->dim / p->n_heads;
     fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
 
-    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (!shared_weights && fread(w->wcls.data(), sizeof(float), w->wcls.size(), f) != w->wcls.size()) return 1;
 
     // Check we didn't forget to read anything
     auto curr = ftell(f);
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -190,20 +181,20 @@ static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bo
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    printf("----- Quick print of first of the weight vales of all the variables\n");
-    printf("%f\n", w->token_embedding_table[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    printf("%f\n", w->rms_ffn_weight[0]);
+    LOG("----- Quick print of first of the weight vales of all the variables\n");
+    LOG("%f\n", w->token_embedding_table[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    LOG("%f\n", w->rms_ffn_weight[0]);
 
-    printf("%f\n", w->wq[0]);
-    printf("%f\n", w->wk[0]);
-    printf("%f\n", w->wv[0]);
-    printf("%f\n", w->wo[0]);
-    printf("%f\n", w->w1[0]);
-    printf("%f\n", w->w2[0]);
-    printf("%f\n", w->w3[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    if (w->wcls) printf("%f\n", w->wcls[0]);
+    LOG("%f\n", w->wq[0]);
+    LOG("%f\n", w->wk[0]);
+    LOG("%f\n", w->wv[0]);
+    LOG("%f\n", w->wo[0]);
+    LOG("%f\n", w->w1[0]);
+    LOG("%f\n", w->w2[0]);
+    LOG("%f\n", w->w3[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -225,14 +216,16 @@ struct llama_vocab {
 };
 
 struct my_llama_hparams {
-    uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
-    uint32_t n_embd  = 4096;
-    uint32_t n_ff    = 11008;
-    uint32_t n_mult  = 4;
-    uint32_t n_head  = 32;
-    uint32_t n_layer = 32;
-    uint32_t n_rot   = 64;
+    uint32_t n_vocab   = 32000;
+    uint32_t n_ctx     = 512;   // this is provided as user input?
+    uint32_t n_embd    = 4096;
+    uint32_t n_ff      = 11008;
+    uint32_t n_mult    = 4;
+    uint32_t n_head    = 32;
+    uint32_t n_head_kv = 32;
+    uint32_t n_layer   = 32;
+    uint32_t n_rot     = 64;
+
     bool operator!=(const my_llama_hparams& other) const {
         return memcmp(this, &other, sizeof(my_llama_hparams));
     }
@@ -325,14 +318,30 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
-    printf("%s: 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);
+    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
+}
+
+static void print_tensor_info(const struct ggml_context * ctx) {
+    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        LOG("%s: Allocating ", __func__);
+        int64_t total = 1;
+        int i = 0;
+        for (; i < ggml_n_dims(t); ++i) {
+            if (i > 0) LOG("x ");
+            LOG("[%" PRId64 "] ", t->ne[i]);
+            total *= t->ne[i];
+        }
+        if (i > 1) LOG("= [%" PRId64 "] ", total);
+        LOG("float space for %s\n", ggml_get_name(t));
+    }
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -342,6 +351,8 @@ static void init_model(struct my_llama_model * model) {
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
+    const uint32_t n_multiqueries = hparams.n_head_kv <= 0 || hparams.n_head_kv >= hparams.n_head ? 1 : hparams.n_head / hparams.n_head_kv;
+
     const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
@@ -350,25 +361,8 @@ 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 [%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 [%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 [%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 [%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 [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, 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");
@@ -383,8 +377,8 @@ static void init_model(struct my_llama_model * model) {
         layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
         layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
         layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
@@ -406,6 +400,8 @@ static void init_model(struct my_llama_model * model) {
         ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
         ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
     }
+
+    print_tensor_info(ctx);
 }
 
 static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
@@ -421,9 +417,9 @@ static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
 static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = get_f32_2d(probs, k, i);
-        printf(" %f", p);
+        LOG(" %f", p);
     }
-    printf("\n");
+    LOG("\n");
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
@@ -431,33 +427,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
-            printf(" %.2f", p);
+            LOG(" %.2f", p);
         }
-        printf("\n");
+        LOG("\n");
     }
 }
 
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
@@ -549,8 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
+        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -578,6 +554,9 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
 
         const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+        if (n_vocab != static_cast<uint32_t>(config->vocab_size)) {
+            die_fmt("vocab size mismatch: (gguf) %u != (llama2c) %d", n_vocab, config->vocab_size);
+        }
 
         vocab->id_to_token.resize(n_vocab);
 
@@ -595,7 +574,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
+        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -638,38 +617,15 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 }
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
-    int ct;
-    switch (ggml_n_dims(gg_weights)) {
-        case 1:
-            ct = 0;
-            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
-                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
-                *ptr = karpathy_weights[ct];
-                ct++;
-            }
-            break;
-        case 2:
-            ct = 0;
-            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
-                    *ptr = karpathy_weights[ct];
-                    ct++;
-                }
-            }
-            break;
-        case 3:
-            ct = 0;
-            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
-                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
-                        *ptr = karpathy_weights[ct];
-                        ct++;
-                    }
-                }
-            }
-            break;
+    int size = 1;
+    for (int dim = 0; dim < ggml_n_dims(gg_weights); ++dim) {
+        size *= gg_weights->ne[dim];
+    }
+    for (int ct = 0; ct < size; ++ct) {
+        int64_t i0 = 0; int64_t i1 = 0;
+        int64_t i2 = 0; int64_t i3 = 0;
+        ggml_unravel_index(gg_weights, ct, &i0, &i1, &i2, &i3);
+        ggml_set_f32_nd(gg_weights, i0, i1, i2, i3, karpathy_weights[ct]);
     }
 }
 
@@ -679,16 +635,18 @@ static void save_as_llama_model(
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
-    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
-    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table.data());
+    convert_weights_ak_to_gg(model->output, !w->wcls.empty() ? w->wcls.data() : w->token_embedding_table.data());
 
-    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight.data());
     //print_row(model->norm, 0);
 
     // for rms-att-weight
     int row_length = model->hparams.n_embd;
     int n_ff = model->hparams.n_ff;
 
+    const uint32_t n_multiqueries = model->hparams.n_head_kv <= 0 || model->hparams.n_head_kv >= model->hparams.n_head ? 1 : model->hparams.n_head / model->hparams.n_head_kv;
+
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
         // 1d
@@ -697,9 +655,10 @@ static void save_as_llama_model(
 
         // from 3d matrix layer x dim x dim to 2d matrix dim x dim
         convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
         convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim / n_multiqueries
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length/n_multiqueries]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length/n_multiqueries]);
 
         convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
         convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
@@ -736,8 +695,8 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
     gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
     gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
-    // n_head_kv is optional, default to n_head
-    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, model->hparams.n_head_kv);
     gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
     gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
     gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
@@ -789,12 +748,12 @@ static void save_as_llama_model(
 
 static struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
+    params.fn_vocab_model          = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
-    params.fn_train_data     = "shakespeare.txt";
-    params.fn_checkpoint_in  = "checkpoint.bin";
-    params.fn_checkpoint_out = "checkpoint.bin";
-    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+    params.fn_train_data           = "shakespeare.txt";
+    params.fn_checkpoint_in        = "checkpoint.bin";
+    params.fn_checkpoint_out       = "checkpoint.bin";
+    params.fn_model_out            = "ggml-checkpoint-f32.bin";
 
     params.seed       =   -1;
 
@@ -829,8 +788,8 @@ static struct train_params get_default_train_params() {
     params.adam_alpha        = 1e-3f;
     params.adam_decay        = 1e-3f;
 
-    params.mem_model_gb   = 2;
-    params.mem_compute_gb = 24;
+    params.mem_model_gb    = 2;
+    params.mem_compute_gb  = 24;
     params.mem_compute0_gb = 8;
     params.mem_compute1_gb = 2;
 
@@ -916,19 +875,30 @@ int main(int argc, char ** argv) {
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
+    log_set_target(stdout);
     Config config;
     TransformerWeights weights = {};
     {
-        FILE *file = fopen(params.fn_llama2c_model, "rb");
-        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
+        FILE *file = fopen(params.fn_llama2c_model, "r");
+        if (!file) {
+            LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
+            return 1;
+        }
         // read in the config header
-        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        if (fread(&config, sizeof(Config), 1, file) != 1) {
+            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         auto shared_weights = config.vocab_size > 0;
         config.vocab_size = abs(config.vocab_size);
 
         // read in the Transformer weights
-        malloc_weights(&weights, &config, shared_weights);
-        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
+        alloc_weights(&weights, &config, shared_weights);
+        if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
+            LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         fclose(file);
     }
 
@@ -936,15 +906,18 @@ int main(int argc, char ** argv) {
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
-    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
-    model.hparams.n_embd  = config.dim; //params.n_embd;
-    model.hparams.n_ff    = config.hidden_dim;
-    model.hparams.n_mult  = 32;//params.n_mult;
-    model.hparams.n_head  = config.n_heads; //params.n_head;
-    model.hparams.n_layer = config.n_layers; //params.n_layer;
-    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    model.hparams.n_vocab   = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx     = params.n_ctx;
+    model.hparams.n_embd    = config.dim; //params.n_embd;
+    model.hparams.n_ff      = config.hidden_dim;
+    model.hparams.n_mult    = 32;//params.n_mult;
+    model.hparams.n_head    = config.n_heads; //params.n_head;
+    model.hparams.n_head_kv = config.n_kv_heads;
+    model.hparams.n_layer   = config.n_layers; //params.n_layer;
+    model.hparams.n_rot     = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+
     print_params(&model.hparams);
+
     struct ggml_init_params lcparams;
     lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
     lcparams.mem_buffer = NULL;
@@ -956,7 +929,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

examples/embedding/embedding.cpp

@@ -19,18 +19,7 @@ static std::vector<std::string> split_lines(const std::string & s) {
 
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
     for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
-    }
-}
-
-static void normalize(float * vec, float * out, int n) {
-    float norm = 0;
-    for (int i = 0; i < n; i++) {
-        norm += vec[i] * vec[i];
-    }
-    norm = sqrt(norm);
-    for (int i = 0; i < n; i++) {
-        out[i] = vec[i] / norm;
+        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
     }
 }
 
@@ -44,11 +33,23 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
-    // normalize on copy
-    for (int k = 0; k < n_seq; k++) {
-        float * emb = llama_get_embeddings_ith(ctx, k);
-        float * out = output + k * n_embd;
-        normalize(emb, out, n_embd);
+    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);
     }
 }
 
@@ -60,6 +61,8 @@ int main(int argc, char ** argv) {
     }
 
     params.embedding = true;
+    // For non-causal models, batch size must be equal to ubatch size
+    params.n_ubatch = params.n_batch;
 
     print_build_info();
 
@@ -106,18 +109,27 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch == params.n_ctx);
+    GGML_ASSERT(params.n_batch >= params.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);
+        auto inp = ::llama_tokenize(ctx, prompt, true, false);
         if (inp.size() > n_batch) {
-            inp.resize(n_batch);
+            fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
+                    __func__, (long long int) inp.size(), (long long int) n_batch);
+            return 1;
         }
         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) {
         for (int i = 0; i < (int) inputs.size(); i++) {
@@ -132,7 +144,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
     // allocate output
     const int n_embd = llama_n_embd(model);
@@ -145,6 +157,7 @@ int main(int argc, char ** argv) {
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
         auto & inp = inputs[k];
+
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity
@@ -165,15 +178,26 @@ int main(int argc, char ** argv) {
     float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd);
 
-    // print first 3 embeddings
-    for (int j = 0; j < std::min(3, n_prompts); j++) {
-        fprintf(stderr, "embedding %d: ", j);
-        for (int i = 0; i < n_embd; i++) {
-            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+    // 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(stderr, "\n\n");
+        fprintf(stdout, "\n");
+    }
+
+    // print cosine similarity matrix
+    fprintf(stdout, "\n");
+    printf("cosine similarity matrix:\n\n");
+    for (int i = 0; i < n_prompts; i++) {
+        for (int j = 0; j < n_prompts; j++) {
+            float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+            fprintf(stdout, "%6.2f ", sim);
+        }
+        fprintf(stdout, "\n");
     }
-    fprintf(stderr, "\n");
 
     // clean up
     llama_print_timings(ctx);

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 +1,468 @@
+#include "llama.h"
+#include "common.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include <stdio.h>
+#include <string.h>
+#include <climits>
+#include <stdexcept>
+
+#if defined(_WIN32)
+    #include <windows.h>
+    #ifndef PATH_MAX
+        #define PATH_MAX MAX_PATH
+    #endif
+    #include <io.h>
+#endif
+
+enum split_operation : uint8_t {
+    SPLIT_OP_SPLIT,
+    SPLIT_OP_MERGE,
+};
+
+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(EXIT_FAILURE);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        split_print_usage(argv[0]);
+        exit(EXIT_FAILURE);
+    }
+    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);
+    }
+}
+
+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_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
+        gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, n_split);
+        gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
+
+        // 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);
+        }
+
+        char split_path[PATH_MAX] = {0};
+        llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
+
+        fprintf(stderr, "%s: %s ...", __func__, split_path);
+        fout = std::ofstream(split_path, 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(EXIT_FAILURE);
+    }
+
+    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(EXIT_FAILURE);
+    }
+
+    split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+
+    char first_split_path[PATH_MAX] = {0};
+    llama_split_path(first_split_path, sizeof(first_split_path),
+                     split_params.output.c_str(), strategy.i_split, strategy.n_split);
+    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
+            __func__, split_params.input.c_str(),
+            first_split_path,
+            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;
+
+    char split_path[PATH_MAX] = {0};
+    strncpy(split_path, split_params.input.c_str(), sizeof(split_path) - 1);
+    char split_prefix[PATH_MAX] = {0};
+
+    // 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,
+        };
+
+        if (i_split > 0) {
+            llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        }
+        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_path);
+
+        auto * ctx_gguf = gguf_init_from_file(split_path, params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+            exit(EXIT_FAILURE);
+        }
+        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_SPLIT_COUNT);
+            if (key_n_split < 0) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain %s metadata\n",
+                        __func__,
+                        LLM_KV_SPLIT_COUNT);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
+            }
+
+            n_split = gguf_get_val_u16(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);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
+            }
+
+            // Verify the file naming and extract split_prefix
+            if (!llama_split_prefix(split_prefix, sizeof (split_prefix), split_path, i_split, n_split)) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s"
+                                " i_split=%d"
+                                " n_split=%d\n", __func__,
+                        split_path, i_split, n_split);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
+            }
+
+            // Do not trigger merge if we try to merge again the output
+            gguf_set_val_u16(ctx_gguf, LLM_KV_SPLIT_COUNT, 0);
+
+            // Set metadata from the first split
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+
+        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++) {
+        llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        std::ifstream f_input(split_path, std::ios::binary);
+        if (!f_input.is_open()) {
+            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_path);
+            for (uint32_t i = 0; i < ctx_ggufs.size(); i++) {
+                gguf_free(ctx_ggufs[i]);
+                ggml_free(ctx_metas[i]);
+            }
+            gguf_free(ctx_out);
+            fout.close();
+            exit(EXIT_FAILURE);
+        }
+        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
+
+        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(EXIT_FAILURE);
+    }
+
+    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/README.md

@@ -22,7 +22,7 @@ For faster computation, make sure to use GPU offloading via the `-ngl` argument
 ## Example
 
 ```bash
-LLAMA_CUBLAS=1 make -j
+LLAMA_CUDA=1 make -j
 
 # generate importance matrix (imatrix.dat)
 ./imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99

examples/imatrix/imatrix.cpp

@@ -50,11 +50,31 @@ private:
     void keep_imatrix(int ncall) const;
 };
 
+// remove any prefix and suffixes from the name
+// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+static std::string filter_tensor_name(const char * name) {
+    std::string wname;
+    const char * p = strchr(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 = name;
+    }
+    return wname;
+}
+
 bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
     GGML_UNUSED(user_data);
 
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
+    std::string wname = filter_tensor_name(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
@@ -62,7 +82,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         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;
     }
 
@@ -94,12 +114,13 @@ 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];
+            wname = filter_tensor_name(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
@@ -107,7 +128,7 @@ 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];
@@ -129,17 +150,17 @@ 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];
@@ -403,6 +424,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
+            // TODO: use batch.logits to save computations instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return false;

examples/infill/infill.cpp

@@ -378,10 +378,10 @@ int main(int argc, char ** argv) {
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
-            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
+            control_message = " - To return control to LLaMA, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
-            control_message = " - Press Return to return control to LLaMa.\n"
+            control_message = " - Press Return to return control to LLaMA.\n"
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }

examples/json-schema-pydantic-example.py

@@ -0,0 +1,74 @@
+# Usage:
+#! ./server -m some-model.gguf &
+#! pip install pydantic
+#! python json-schema-pydantic-example.py
+
+from pydantic import BaseModel, TypeAdapter
+from annotated_types import MinLen
+from typing import Annotated, List, Optional
+import json, requests
+
+if True:
+
+    def create_completion(*, response_model=None, endpoint="http://localhost:8080/v1/chat/completions", messages, **kwargs):
+        '''
+        Creates a chat completion using an OpenAI-compatible endpoint w/ JSON schema support
+        (llama.cpp server, llama-cpp-python, Anyscale / Together...)
+
+        The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below)
+        '''
+        if response_model:
+            type_adapter = TypeAdapter(response_model)
+            schema = type_adapter.json_schema()
+            messages = [{
+                "role": "system",
+                "content": f"You respond in JSON format with the following schema: {json.dumps(schema, indent=2)}"
+            }] + messages
+            response_format={"type": "json_object", "schema": schema}
+
+        data = requests.post(endpoint, headers={"Content-Type": "application/json"},
+                             json=dict(messages=messages, response_format=response_format, **kwargs)).json()
+        if 'error' in data:
+            raise Exception(data['error']['message'])
+
+        content = data["choices"][0]["message"]["content"]
+        return type_adapter.validate_json(content) if type_adapter else content
+
+else:
+
+    # This alternative branch uses Instructor + OpenAI client lib.
+    # Instructor support streamed iterable responses, retry & more.
+    # (see https://python.useinstructor.com/)
+    #! pip install instructor openai
+    import instructor, openai
+    client = instructor.patch(
+        openai.OpenAI(api_key="123", base_url="http://localhost:8080"),
+        mode=instructor.Mode.JSON_SCHEMA)
+    create_completion = client.chat.completions.create
+
+
+if __name__ == '__main__':
+
+    class QAPair(BaseModel):
+        question: str
+        concise_answer: str
+        justification: str
+
+    class PyramidalSummary(BaseModel):
+        title: str
+        summary: str
+        question_answers: Annotated[List[QAPair], MinLen(2)]
+        sub_sections: Optional[Annotated[List['PyramidalSummary'], MinLen(2)]]
+
+    print("# Summary\n", create_completion(
+        model="...",
+        response_model=PyramidalSummary,
+        messages=[{
+            "role": "user",
+            "content": f"""
+                You are a highly efficient corporate document summarizer.
+                Create a pyramidal summary of an imaginary internal document about our company processes
+                (starting high-level, going down to each sub sections).
+                Keep questions short, and answers even shorter (trivia / quizz style).
+            """
+        }]))

examples/json-schema-to-grammar.py

@@ -1,8 +1,10 @@
 #!/usr/bin/env python3
 import argparse
+import itertools
 import json
 import re
 import sys
+from typing import Any, Dict, List, Set, Tuple, Union
 
 # whitespace is constrained to a single space char to prevent model "running away" in
 # whitespace. Also maybe improves generation quality?
@@ -12,26 +14,54 @@ PRIMITIVE_RULES = {
     'boolean': '("true" | "false") space',
     'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
     'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
+    'value'  : 'object | array | string | number | boolean',
+    'object' : '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space',
+    'array'  : '"[" space ( value ("," space value)* )? "]" space',
+    'uuid'   : '"\\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + ' "\\"" space',
     'string': r''' "\"" (
         [^"\\] |
         "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-      )* "\"" space ''',
+      )* "\"" space''',
     'null': '"null" space',
 }
+OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value']
+
+# TODO: support "uri", "email" string formats
+DATE_RULES = {
+    'date'   : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )',
+    'time'   : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )',
+    'date-time': 'date "T" time',
+    'date-string': '"\\"" date "\\"" space',
+    'time-string': '"\\"" time "\\"" space',
+    'date-time-string': '"\\"" date-time "\\"" space',
+}
+
+RESERVED_NAMES = set(["root", *PRIMITIVE_RULES.keys(), *DATE_RULES.keys()])
 
 INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
 GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'}
+GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
 
+NON_LITERAL_SET = set('|.()[]{}*+?')
+ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?')
+
+DATE_PATTERN = '[0-9]{4}-(0[1-9]|1[0-2])-([0-2][0-9]|3[0-1])'
+TIME_PATTERN = '([01][0-9]|2[0-3])(:[0-5][0-9]){2}(\\.[0-9]{1,3})?(Z|[+-](([01][0-9]|2[0-3]):[0-5][0-9]))' # Cap millisecond precision w/ 3 digits
 
 class SchemaConverter:
-    def __init__(self, prop_order):
+    def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
         self._prop_order = prop_order
+        self._allow_fetch = allow_fetch
+        self._dotall = dotall
+        self._raw_pattern = raw_pattern
         self._rules = {'space': SPACE_RULE}
+        self._refs = {}
+        self._refs_being_resolved = set()
 
     def _format_literal(self, literal):
         escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
-            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal)
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal
         )
         return f'"{escaped}"'
 
@@ -41,78 +71,420 @@ class SchemaConverter:
             key = esc_name
         else:
             i = 0
-            while f'{esc_name}{i}' in self._rules:
+            while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule:
                 i += 1
             key = f'{esc_name}{i}'
         self._rules[key] = rule
         return key
 
+    def resolve_refs(self, schema: dict, url: str):
+        '''
+            Resolves all $ref fields in the given schema, fetching any remote schemas,
+            replacing $ref with absolute reference URL and populating self._refs with the
+            respective referenced (sub)schema dictionaries.
+        '''
+        def visit(n: dict):
+            if isinstance(n, list):
+                return [visit(x) for x in n]
+            elif isinstance(n, dict):
+                ref = n.get('$ref')
+                if ref is not None and ref not in self._refs:
+                    if ref.startswith('https://'):
+                        assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)'
+                        import requests
+
+                        frag_split = ref.split('#')
+                        base_url = frag_split[0]
+
+                        target = self._refs.get(base_url)
+                        if target is None:
+                            target = self.resolve_refs(requests.get(ref).json(), base_url)
+                            self._refs[base_url] = target
+
+                        if len(frag_split) == 1 or frag_split[-1] == '':
+                            return target
+                    elif ref.startswith('#/'):
+                        target = schema
+                        ref = f'{url}{ref}'
+                        n['$ref'] = ref
+                    else:
+                        raise ValueError(f'Unsupported ref {ref}')
+
+                    for sel in ref.split('#')[-1].split('/')[1:]:
+                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                        target = target[sel]
+
+                    self._refs[ref] = target
+                else:
+                    for v in n.values():
+                        visit(v)
+
+            return n
+        return visit(schema)
+
+    def _generate_union_rule(self, name, alt_schemas):
+        return ' | '.join((
+            self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}')
+            for i, alt_schema in enumerate(alt_schemas)
+        ))
+
+    def _visit_pattern(self, pattern, name):
+        '''
+            Transforms a regular expression pattern into a GBNF rule.
+
+            Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
+            Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md
+
+            Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
+
+            Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which
+            we define sub-rules to keep the output lean.
+        '''
+
+        assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"'
+        pattern = pattern[1:-1]
+        sub_rule_ids = {}
+
+        i = 0
+        length = len(pattern)
+
+        def to_rule(s: Tuple[str, bool]) -> str:
+            (txt, is_literal) = s
+            return "\"" + txt + "\"" if is_literal else txt
+
+        def transform() -> Tuple[str, bool]:
+            '''
+                Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
+            '''
+            nonlocal i
+            nonlocal pattern
+            nonlocal sub_rule_ids
+
+            start = i
+            # For each component of this sequence, store its string representation and whether it's a literal.
+            # We only need a flat structure here to apply repetition operators to the last item, and
+            # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
+            # (GBNF's syntax is luckily very close to regular expressions!)
+            seq: list[Tuple[str, bool]] = []
+
+            def get_dot():
+                if self._dotall:
+                    rule = '[\\U00000000-\\U0010FFFF]'
+                else:
+                    # Accept any character... except \n and \r line break chars (\x0A and \xOD)
+                    rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'
+                return self._add_rule(f'dot', rule)
+
+            def join_seq():
+                nonlocal seq
+                ret = []
+                for is_literal, g in itertools.groupby(seq, lambda x: x[1]):
+                    if is_literal:
+                        ret.append((''.join(x[0] for x in g), True))
+                    else:
+                        ret.extend(g)
+                if len(ret) == 1:
+                    return ret[0]
+                return (' '.join(to_rule(x) for x in seq), False)
+
+            while i < length:
+                c = pattern[i]
+                if c == '.':
+                    seq.append((get_dot(), False))
+                    i += 1
+                elif c == '(':
+                    i += 1
+                    if i < length:
+                        assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/'
+                    seq.append((f'({to_rule(transform())})', False))
+                elif c == ')':
+                    i += 1
+                    assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}'
+                    return join_seq()
+                elif c == '[':
+                    square_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != ']':
+                        if pattern[i] == '\\':
+                            square_brackets += pattern[i:i+2]
+                            i += 2
+                        else:
+                            square_brackets += pattern[i]
+                            i += 1
+                    assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    square_brackets += ']'
+                    i += 1
+                    seq.append((square_brackets, False))
+                elif c == '|':
+                    seq.append(('|', False))
+                    i += 1
+                elif c in ('*', '+', '?'):
+                    seq[-1] = (to_rule(seq[-1]) + c, False)
+                    i += 1
+                elif c == '{':
+                    curly_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != '}':
+                        curly_brackets += pattern[i]
+                        i += 1
+                    assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    curly_brackets += '}'
+                    i += 1
+                    nums = [s.strip() for s in curly_brackets[1:-1].split(',')]
+                    min_times = 0
+                    max_times = None
+                    try:
+                        if len(nums) == 1:
+                            min_times = int(nums[0])
+                            max_times = min_times
+                        else:
+                            assert len(nums) == 2
+                            min_times = int(nums[0]) if nums[0] else 0
+                            max_times = int(nums[1]) if nums[1] else None
+                    except ValueError:
+                        raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/')
+
+                    (sub, sub_is_literal) = seq[-1]
+
+                    if min_times == 0 and max_times is None:
+                        seq[-1] = (f'{sub}*', False)
+                    elif min_times == 0 and max_times == 1:
+                        seq[-1] = (f'{sub}?', False)
+                    elif min_times == 1 and max_times is None:
+                        seq[-1] = (f'{sub}+', False)
+                    else:
+                        if not sub_is_literal:
+                            id = sub_rule_ids.get(sub)
+                            if id is None:
+                                id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
+                                sub_rule_ids[sub] = id
+                            sub = id
+
+                        seq[-1] = (
+                            ' '.join(
+                                ([f'"{sub[1:-1] * min_times}"'] if sub_is_literal else [sub] * min_times) +
+                                ([f'{sub}?'] * (max_times - min_times) if max_times is not None else [f'{sub}*'])),
+                            False
+                        )
+                else:
+                    literal = ''
+                    while i < length:
+                        if pattern[i] == '\\' and i < length - 1:
+                            next = pattern[i + 1]
+                            if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS:
+                                i += 1
+                                literal += pattern[i]
+                                i += 1
+                            else:
+                                literal += pattern[i:i+2]
+                                i += 2
+                        elif pattern[i] == '"' and not self._raw_pattern:
+                            literal += '\\"'
+                            i += 1
+                        elif pattern[i] not in NON_LITERAL_SET and \
+                                (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET):
+                            literal += pattern[i]
+                            i += 1
+                        else:
+                            break
+                    if literal:
+                        seq.append((literal, True))
+
+            return join_seq()
+
+        return self._add_rule(
+            name,
+            to_rule(transform()) if self._raw_pattern \
+                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
+
+
+    def _resolve_ref(self, ref):
+        ref_name = ref.split('/')[-1]
+        if ref_name not in self._rules and ref not in self._refs_being_resolved:
+            self._refs_being_resolved.add(ref)
+            resolved = self._refs[ref]
+            ref_name = self.visit(resolved, ref_name)
+            self._refs_being_resolved.remove(ref)
+        return ref_name
+
+    def _generate_constant_rule(self, value):
+        return self._format_literal(json.dumps(value))
+
     def visit(self, schema, name):
         schema_type = schema.get('type')
-        rule_name = name or 'root'
+        schema_format = schema.get('format')
+        rule_name = name + '-' if name in RESERVED_NAMES else name or 'root'
 
-        if 'oneOf' in schema or 'anyOf' in schema:
-            rule = ' | '.join((
-                self.visit(alt_schema, f'{name}{"-" if name else ""}{i}')
-                for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf'])
-            ))
-            return self._add_rule(rule_name, rule)
+        if (ref := schema.get('$ref')) is not None:
+            return self._add_rule(rule_name, self._resolve_ref(ref))
+
+        elif 'oneOf' in schema or 'anyOf' in schema:
+            return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
+
+        elif isinstance(schema_type, list):
+            return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type]))
 
         elif 'const' in schema:
-            return self._add_rule(rule_name, self._format_literal(schema['const']))
+            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']))
 
         elif 'enum' in schema:
-            rule = ' | '.join((self._format_literal(v) for v in schema['enum']))
+            rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum']))
             return self._add_rule(rule_name, rule)
 
-        elif schema_type == 'object' and 'properties' in schema:
-            # TODO: `required` keyword
-            prop_order = self._prop_order
-            prop_pairs = sorted(
-                schema['properties'].items(),
-                # sort by position in prop_order (if specified) then by key
-                key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]),
+        elif schema_type in (None, 'object') and \
+             ('properties' in schema or \
+              ('additionalProperties' in schema and schema['additionalProperties'] is not True)):
+            required = set(schema.get('required', []))
+            properties = list(schema.get('properties', {}).items())
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
+
+        elif schema_type in (None, 'object') and 'allOf' in schema:
+            required = set()
+            properties = []
+            hybrid_name = name
+            def add_component(comp_schema, is_required):
+                if (ref := comp_schema.get('$ref')) is not None:
+                    comp_schema = self._refs[ref]
+
+                if 'properties' in comp_schema:
+                    for prop_name, prop_schema in comp_schema['properties'].items():
+                        properties.append((prop_name, prop_schema))
+                        if is_required:
+                            required.add(prop_name)
+
+            for t in schema['allOf']:
+                if 'anyOf' in t:
+                    for tt in t['anyOf']:
+                        add_component(tt, is_required=False)
+                else:
+                    add_component(t, is_required=True)
+
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[]))
+
+        elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
+            items = schema.get('items') or schema['prefixItems']
+            if isinstance(items, list):
+                return self._add_rule(
+                    rule_name,
+                    '"[" space ' +
+                    ' "," space '.join(
+                        self.visit(item, f'{name}{"-" if name else ""}tuple-{i}')
+                        for i, item in enumerate(items)) +
+                    ' "]" space')
+            else:
+                item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
+                list_item_operator = f'( "," space {item_rule_name} )'
+                successive_items = ""
+                min_items = schema.get("minItems", 0)
+                max_items = schema.get("maxItems")
+                if min_items > 0:
+                    successive_items = list_item_operator * (min_items - 1)
+                    min_items -= 1
+                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 + "*"
+                if min_items == 0:
+                    rule = f'"[" space ( {item_rule_name} {successive_items} )? "]" space'
+                else:
+                    rule = f'"[" space {item_rule_name} {successive_items} "]" space'
+                return self._add_rule(rule_name, rule)
+
+        elif schema_type in (None, 'string') and 'pattern' in schema:
+            return self._visit_pattern(schema['pattern'], rule_name)
+
+        elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''):
+            return self._add_rule(
+                'root' if rule_name == 'root' else schema_format,
+                PRIMITIVE_RULES['uuid']
             )
 
-            rule = '"{" space'
-            for i, (prop_name, prop_schema) in enumerate(prop_pairs):
-                prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
-                if i > 0:
-                    rule += ' "," space'
-                rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}'
-            rule += ' "}" space'
+        elif schema_type in (None, 'string') and schema_format in DATE_RULES:
+            for t, r in DATE_RULES.items():
+                self._add_rule(t, r)
+            return schema_format + '-string'
 
-            return self._add_rule(rule_name, rule)
-
-        elif schema_type == 'array' and 'items' in schema:
-            # TODO `prefixItems` keyword
-            item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
-            list_item_operator = f'("," space {item_rule_name})'
-            successive_items = ""
-            min_items = schema.get("minItems", 0)
-            if min_items > 0:
-               first_item = f"({item_rule_name})"
-               successive_items = list_item_operator * (min_items - 1)
-               min_items -= 1
-            else:
-               first_item = f"({item_rule_name})?"
-            max_items = schema.get("maxItems")
-            if max_items is not None and max_items > min_items:
-                successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-            else:
-                successive_items += list_item_operator + "*"
-            rule = f'"[" space {first_item} {successive_items} "]" space'
-            return self._add_rule(rule_name, rule)
+        elif (schema_type == 'object') or (len(schema) == 0):
+            for n in OBJECT_RULE_NAMES:
+                self._add_rule(n, PRIMITIVE_RULES[n])
+            return self._add_rule(rule_name, 'object')
 
         else:
             assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
+            # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
             return self._add_rule(
                 'root' if rule_name == 'root' else schema_type,
                 PRIMITIVE_RULES[schema_type]
             )
 
+    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]):
+        prop_order = self._prop_order
+        # sort by position in prop_order (if specified) then by original order
+        sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
+
+        prop_kv_rule_names = {}
+        for prop_name, prop_schema in properties:
+            prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+            prop_kv_rule_names[prop_name] = self._add_rule(
+                f'{name}{"-" if name else ""}{prop_name}-kv',
+                fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}'
+            )
+        required_props = [k for k in sorted_props if k in required]
+        optional_props = [k for k in sorted_props if k not in required]
+
+        if additional_properties == True or isinstance(additional_properties, dict):
+            sub_name = f'{name}{"-" if name else ""}additional'
+            value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value')
+            prop_kv_rule_names["*"] = self._add_rule(
+                f'{sub_name}-kv',
+                self._add_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+            )
+            optional_props.append("*")
+
+        rule = '"{" space '
+        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)
+
+        if optional_props:
+            rule += ' ('
+            if required_props:
+                rule += ' "," space ( '
+
+            def get_recursive_refs(ks, first_is_optional):
+                [k, *rest] = ks
+                kv_rule_name = prop_kv_rule_names[k]
+                if k == '*':
+                    res = self._add_rule(
+                        f'{name}{"-" if name else ""}additional-kvs',
+                        f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*'
+                    )
+                elif first_is_optional:
+                    res = f'( "," space {kv_rule_name} )?'
+                else:
+                    res = kv_rule_name
+                if len(rest) > 0:
+                    res += ' ' + self._add_rule(
+                        f'{name}{"-" if name else ""}{k}-rest',
+                        get_recursive_refs(rest, first_is_optional=True)
+                    )
+                return res
+
+            rule += ' | '.join(
+                get_recursive_refs(optional_props[i:], first_is_optional=False)
+                for i in range(len(optional_props))
+            )
+            if required_props:
+                rule += ' )'
+            rule += ' )?'
+
+        rule += ' "}" space'
+
+        return rule
+
     def format_grammar(self):
-        return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items()))
+        return '\n'.join(
+            f'{name} ::= {rule}'
+            for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0])
+        )
 
 
 def main(args_in = None):
@@ -129,16 +501,47 @@ def main(args_in = None):
         type=lambda s: s.split(','),
         help='''
             comma-separated property names defining the order of precedence for object properties;
-            properties not specified here are given lower precedence than those that are, and are
-            sorted alphabetically
+            properties not specified here are given lower precedence than those that are, and
+            are kept in their original order from the schema. Required properties are always
+            given precedence over optional properties.
         '''
     )
+    parser.add_argument(
+        '--allow-fetch',
+        action='store_true',
+        default=False,
+        help='Whether to allow fetching referenced schemas over HTTPS')
+    parser.add_argument(
+        '--dotall',
+        action='store_true',
+        default=False,
+        help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns')
+    parser.add_argument(
+        '--raw-pattern',
+        action='store_true',
+        default=False,
+        help='Treats string patterns as raw patterns w/o quotes (or quote escapes)')
+
     parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
     args = parser.parse_args(args_in)
 
-    schema = json.load(sys.stdin if args.schema == '-' else open(args.schema))
-    prop_order = {name: idx for idx, name in enumerate(args.prop_order)}
-    converter = SchemaConverter(prop_order)
+    if args.schema.startswith('https://'):
+        url = args.schema
+        import requests
+        schema = requests.get(url).json()
+    elif args.schema == '-':
+        url = 'stdin'
+        schema = json.load(sys.stdin)
+    else:
+        url = f'file://{args.schema}'
+        with open(args.schema) as f:
+            schema = json.load(f)
+    converter = SchemaConverter(
+        prop_order={name: idx for idx, name in enumerate(args.prop_order)},
+        allow_fetch=args.allow_fetch,
+        dotall=args.dotall,
+        raw_pattern=args.raw_pattern)
+    schema = converter.resolve_refs(schema, url)
     converter.visit(schema, '')
     print(converter.format_grammar())
 

examples/llama-bench/README.md

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

examples/llama-bench/llama-bench.cpp

@@ -8,6 +8,7 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <cstdlib>
 #include <iterator>
 #include <map>
 #include <numeric>
@@ -103,6 +104,7 @@ static std::string get_cpu_info() {
                 }
             }
         }
+        fclose(f);
     }
 #endif
     // TODO: other platforms
@@ -111,11 +113,11 @@ 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();
+#ifdef GGML_USE_CUDA
+    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 += "/";
@@ -123,20 +125,15 @@ static std::string get_gpu_info() {
     }
 #endif
 #ifdef GGML_USE_SYCL
-    int device_list[GGML_SYCL_MAX_DEVICES];
-    ggml_sycl_get_gpu_list(device_list, GGML_SYCL_MAX_DEVICES);
-
-    for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
-        if (device_list[i] >0 ){
-            char buf[128];
-            ggml_sycl_get_device_description(i, buf, sizeof(buf));
-            id += buf;
+    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 += "/";
         }
     }
-    if (id.length() >2 ) {
-        id.pop_back();
-    }
 #endif
     // TODO: other backends
     return id;
@@ -169,6 +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;
@@ -176,9 +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::vector<float>> tensor_split;
     std::vector<bool> use_mmap;
+    std::vector<bool> embeddings;
     int reps;
     bool verbose;
     output_formats output_format;
@@ -188,7 +186,8 @@ 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()},
@@ -196,9 +195,9 @@ static const cmd_params cmd_params_defaults = {
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
-    /* mul_mat_q     */ {true},
     /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
     /* use_mmap      */ {true},
+    /* embeddings    */ {false},
     /* reps          */ 5,
     /* verbose       */ false,
     /* output_format */ MARKDOWN
@@ -213,6 +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());
@@ -221,8 +221,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
-    printf("  -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("  -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");
@@ -249,6 +249,9 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "q5_1") {
         return GGML_TYPE_Q5_1;
     }
+    if (s == "iq4_nl") {
+        return GGML_TYPE_IQ4_NL;
+    }
 
     return GGML_TYPE_COUNT;
 }
@@ -302,6 +305,13 @@ 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;
@@ -383,13 +393,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
-        } else if (arg == "-mmq" || arg == "--mul-mat-q") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.mul_mat_q.insert(params.mul_mat_q.end(), p.begin(), p.end());
         } else if (arg == "-mmp" || arg == "--mmap") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -397,6 +400,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
+        } else if (arg == "-embd" || arg == "--embeddings") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -460,15 +470,16 @@ 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;
@@ -479,6 +490,7 @@ 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;
@@ -486,9 +498,9 @@ struct cmd_params_instance {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
-    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
+    bool embeddings;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -516,10 +528,11 @@ 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;
     }
@@ -535,10 +548,11 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & mmp : params.use_mmap)
+    for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
+    for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
-    for (const auto & mmq : params.mul_mat_q)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -550,6 +564,7 @@ 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,
@@ -557,9 +572,9 @@ 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);
         }
@@ -573,6 +588,7 @@ 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,
@@ -580,9 +596,9 @@ 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);
         }
@@ -609,6 +625,7 @@ 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;
@@ -616,9 +633,9 @@ struct test {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
-    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
+    bool embeddings;
     int n_prompt;
     int n_gen;
     std::string test_time;
@@ -632,6 +649,7 @@ 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;
@@ -639,9 +657,9 @@ 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
@@ -710,10 +728,11 @@ struct test {
             "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_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", "use_mmap",
+            "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -724,7 +743,8 @@ 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" ||
@@ -733,7 +753,7 @@ struct test {
         }
         if (field == "cuda" || field == "opencl"  || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
-            field == "mul_mat_q" || field == "use_mmap") {
+            field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -764,10 +784,11 @@ struct test {
             std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_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, std::to_string(use_mmap),
+            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())
@@ -787,7 +808,7 @@ struct test {
 
 const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
-const bool        test::cuda         = !!ggml_cpu_has_cublas();
+const bool        test::cuda         = !!ggml_cpu_has_cuda();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::kompute      = !!ggml_cpu_has_kompute();
@@ -931,15 +952,15 @@ 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";
         }
@@ -962,6 +983,9 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
+        if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) {
+            fields.emplace_back("n_ubatch");
+        }
         if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
             fields.emplace_back("type_k");
         }
@@ -974,9 +998,6 @@ struct markdown_printer : public printer {
         if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
             fields.emplace_back("split_mode");
         }
-        if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
-            fields.emplace_back("mul_mat_q");
-        }
         if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
             fields.emplace_back("no_kv_offload");
         }
@@ -986,6 +1007,9 @@ struct markdown_printer : public printer {
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }
+        if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
+            fields.emplace_back("embeddings");
+        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");
 
@@ -1101,25 +1125,40 @@ 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;
     }
 }
 
@@ -1208,7 +1247,8 @@ 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);
@@ -1224,6 +1264,7 @@ 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/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);
@@ -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

@@ -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

@@ -1,11 +1,13 @@
 # MobileVLM
 
-Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants.
+Currently this implementation supports [MobileVLM-1.7B](https://huggingface.co/mtgv/MobileVLM-1.7B) / [MobileVLM_V2-1.7B](https://huggingface.co/mtgv/MobileVLM_V2-1.7B) variants.
 
 for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM)
 
 The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava.
 
+Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion  is a little different. Therefore, using MobiVLM as an example, the different conversion step will be shown.
+
 ## Usage
 Build with cmake or run `make llava-cli` to build it.
 
@@ -34,7 +36,7 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF:
+3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** the arg is `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
 python ./examples/llava/convert-image-encoder-to-gguf \
@@ -44,6 +46,14 @@ python ./examples/llava/convert-image-encoder-to-gguf \
     --projector-type ldp
 ```
 
+```sh
+python ./examples/llava/convert-image-encoder-to-gguf \
+    -m path/to/clip-vit-large-patch14-336 \
+    --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
+    --output-dir path/to/MobileVLM-1.7B_V2 \
+    --projector-type ldpv2
+```
+
 4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
@@ -114,7 +124,7 @@ 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
+make LLAMA_CUDA=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
 ```
 
 ### run on Orin

examples/llava/README.md

@@ -63,12 +63,20 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 ```console
 git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 ```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+
+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
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+
+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
@@ -76,18 +84,18 @@ 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
 ```
 
-4) Create the visual gguf model:
+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
 
-5) Then convert the model to gguf format:
+6) Then convert the model to gguf format:
 ```console
 python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
-6) And finally we can run the llava-cli using the 1.6 model version:
+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
 ```

examples/llava/clip.cpp

@@ -7,7 +7,7 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
@@ -119,6 +119,7 @@ static std::string format(const char * fmt, ...) {
 #define TN_LLAVA_PROJ      "mm.%d.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 
 
@@ -126,12 +127,14 @@ enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
 static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP, "mlp" },
     { PROJECTOR_TYPE_LDP, "ldp" },
+    { PROJECTOR_TYPE_LDPV2, "ldpv2"},
 };
 
 
@@ -475,6 +478,14 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_block_2_block_2_0_w;
     struct ggml_tensor * mm_model_block_2_block_2_1_w;
     struct ggml_tensor * mm_model_block_2_block_2_1_b;
+
+    // MobileVLM_V2 projection
+    struct ggml_tensor * mm_model_mlp_0_w;
+    struct ggml_tensor * mm_model_mlp_0_b;
+    struct ggml_tensor * mm_model_mlp_2_w;
+    struct ggml_tensor * mm_model_mlp_2_b;
+    struct ggml_tensor * mm_model_peg_0_w;
+    struct ggml_tensor * mm_model_peg_0_b;
 };
 
 struct clip_ctx {
@@ -497,7 +508,6 @@ struct clip_ctx {
 
     // memory buffers to evaluate the model
     ggml_backend_buffer_t params_buffer  = NULL;
-    ggml_backend_buffer_t compute_buffer = NULL;
 
     ggml_backend_t backend       = NULL;
     ggml_gallocr_t compute_alloc = NULL;
@@ -808,6 +818,29 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             }
             embeddings = block_1;
         }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
         else {
             GGML_ASSERT(false);
         }
@@ -935,7 +968,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
     new_clip->backend = ggml_backend_cuda_init(0);
     printf("%s: CLIP using CUDA backend\n", __func__);
 #endif
@@ -995,6 +1028,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         if (!new_clip->ctx_data) {
             fprintf(stderr, "%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
+            gguf_free(ctx);
             return nullptr;
         }
 
@@ -1002,6 +1036,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         if (!fin) {
             printf("cannot open model file for loading tensors\n");
             clip_free(new_clip);
+            gguf_free(ctx);
             return nullptr;
         }
 
@@ -1023,6 +1058,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             if (!fin) {
                 printf("%s: failed to seek for tensor %s\n", __func__, name);
                 clip_free(new_clip);
+                gguf_free(ctx);
                 return nullptr;
             }
             int num_bytes = ggml_nbytes(cur);
@@ -1175,7 +1211,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_block_2_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
             vision_model.mm_model_block_2_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
             vision_model.mm_model_block_2_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
-        } else {
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            // MobilVLM_V2 projection
+            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight"));
+            vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias"));
+            vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight"));
+            vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias"));
+            vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
+            vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
+        }
+        else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
         }
@@ -1232,16 +1279,16 @@ struct clip_image_f32 * clip_image_f32_init() {
 
 void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
 void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
-void clip_image_u8_batch_free(struct clip_image_u8_batch  & batch) {
-    if (batch.size > 0) {
-        delete[] batch.data;
-        batch.size = 0;
+void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
     }
 }
-void clip_image_f32_batch_free(struct clip_image_f32_batch  & batch) {
-    if (batch.size > 0) {
-        delete[] batch.data;
-        batch.size = 0;
+void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
     }
 }
 
@@ -1494,7 +1541,7 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
 
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
-bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs) {
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
@@ -1506,11 +1553,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         pad_to_square = false;
     }
     // free the previous res_imgs if any set
-    if (res_imgs.size > 0) {
+    if (res_imgs->size > 0) {
         clip_image_f32_batch_free(res_imgs);
     }
-    res_imgs.data = nullptr;
-    res_imgs.size = 0;
+    res_imgs->data = nullptr;
+    res_imgs->size = 0;
 
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
     // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
@@ -1565,11 +1612,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
             bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
             patches.insert(patches.begin(), image_original_resize);
             // clip_image_f32_batch_init(patches.size());
-            res_imgs.size = patches.size();
-            res_imgs.data = new clip_image_f32[res_imgs.size];
+            res_imgs->size = patches.size();
+            res_imgs->data = new clip_image_f32[res_imgs->size];
             int num=0;
             for (auto& patch : patches) {
-                normalize_image_u8_to_f32(patch, &res_imgs.data[num], ctx->image_mean, ctx->image_std);
+                normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std);
                 num++;
             }
 
@@ -1657,9 +1704,9 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
     // }
     // res_imgs.push_back(res);
 
-    res_imgs.size = 1;
-    res_imgs.data = new clip_image_f32[res_imgs.size];
-    res_imgs.data[0] = *res;
+    res_imgs->size = 1;
+    res_imgs->data = new clip_image_f32[res_imgs->size];
+    res_imgs->data[0] = *res;
     clip_image_f32_free(res);
 
     return true;
@@ -1673,6 +1720,9 @@ void clip_free(clip_ctx * ctx) {
     ggml_free(ctx->ctx_data);
     gguf_free(ctx->ctx_gguf);
 
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
     delete ctx;
 }
 
@@ -1862,7 +1912,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
     std::vector<uint8_t> work(512);
     std::vector<float> conv_buf(512);
-    std::vector<int64_t> hist_all(1 << 4, 0);
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 
@@ -1909,6 +1958,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 break;
             default:
                 printf("Please use an input file in f32 or f16\n");
+                gguf_free(ctx_out);
                 return false;
             }
 
@@ -1917,48 +1967,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             }
             new_data = work.data();
 
-            std::vector<int64_t> hist_cur(1 << 4, 0);
-
-            switch (new_type) {
-                case GGML_TYPE_Q4_0: {
-                    new_size = ggml_quantize_q4_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q4_1: {
-                    new_size = ggml_quantize_q4_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q5_0: {
-                    new_size = ggml_quantize_q5_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q5_1: {
-                    new_size = ggml_quantize_q5_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q8_0: {
-                    new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q2_K: {
-                    new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q3_K: {
-                    new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q4_K: {
-                    new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q5_K: {
-                    new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                case GGML_TYPE_Q6_K: {
-                    new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
-                } break;
-                default: {
-                    fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
-                    return false;
-                }
-            }
-
-            for (size_t j = 0; j < hist_cur.size(); ++j) {
-                hist_all[j] += hist_cur[j];
-            }
+            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr);
         } else {
             new_type = cur->type;
             new_data = cur->data;
@@ -1993,17 +2002,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     {
         printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
         printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
-
-        int64_t sum_all = 0;
-        for (size_t i = 0; i < hist_all.size(); ++i) {
-            sum_all += hist_all[i];
-        }
-
-        printf("%s: hist: ", __func__);
-        for (size_t i = 0; i < hist_all.size(); ++i) {
-            printf("%5.3f ", hist_all[i] / (float)sum_all);
-        }
-        printf("\n");
     }
 
     return true;
@@ -2013,6 +2011,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
         return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        return ctx->vision_model.mm_model_peg_0_b->ne[0];
+    }
     if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
         return ctx->vision_model.mm_2_b->ne[0];
     }

examples/llava/clip.h

@@ -60,8 +60,8 @@ 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_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 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);
 
@@ -69,7 +69,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
 /** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
-CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
+CLIP_API 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);
 

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

@@ -1,6 +1,7 @@
 import argparse
 import os
 import json
+import re
 
 import torch
 import numpy as np
@@ -38,9 +39,11 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b
 def get_tensor_name(name: str) -> str:
     if "projection" in name:
         return name
-
     if "mm_projector" in name:
-        return name.replace("model.mm_projector", "mm")
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
 
     return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
 
@@ -83,7 +86,7 @@ ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
 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("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
 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

examples/llava/llava.cpp

@@ -223,7 +223,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     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)) {
+    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
         delete[] img_res_v.data;
         return false;

examples/llava/llava.h

@@ -29,9 +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(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
+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);

examples/lookup/CMakeLists.txt

@@ -3,3 +3,21 @@ add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-create)
+add_executable(${TARGET} lookup-create.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-merge)
+add_executable(${TARGET} lookup-merge.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-stats)
+add_executable(${TARGET} lookup-stats.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/lookup/lookup-create.cpp

@@ -0,0 +1,43 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    GGML_ASSERT(model != nullptr);
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+    fprintf(stderr, "%s: tokenization done\n", __func__);
+
+
+    llama_ngram_cache ngram_cache;
+    llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
+
+    llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+}

examples/lookup/lookup-merge.cpp

@@ -0,0 +1,47 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+static void print_usage() {
+    fprintf(stderr, "Merges multiple lookup cache files into a single one.\n");
+    fprintf(stderr, "Usage: lookup-merge [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n");
+}
+
+int main(int argc, char ** argv){
+    if (argc < 3) {
+        print_usage();
+        exit(1);
+    }
+
+    std::vector<std::string> args;
+    args.resize(argc-1);
+    for (int i = 0; i < argc-1; ++i) {
+        args[i] = argv[i+1];
+        if (args[i] == "-h" || args[i] == "--help") {
+            print_usage();
+            exit(0);
+        }
+    }
+
+    fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
+    llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]);
+
+    for (size_t i = 1; i < args.size()-1; ++i) {
+        fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
+        llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]);
+
+        llama_ngram_cache_merge(ngram_cache_merged, ngram_cache);
+    }
+
+    fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
+    llama_ngram_cache_save(ngram_cache_merged, args.back());
+}

examples/lookup/lookup-stats.cpp

@@ -0,0 +1,163 @@
+#include "ggml.h"
+#include "common.h"
+#include "llama.h"
+#include "log.h"
+#include "ngram-cache.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    const int n_draft = params.n_draft;
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
+    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+    LOG("add_bos tgt: %d\n", add_bos);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        const int64_t t_start_draft_us = ggml_time_us();
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
+    const int n_input = inp.size();
+    const int n_ctx = params.n_ctx;
+
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    const int64_t t_start_ms = ggml_time_ms();
+
+    // Iterate over input tokens in chunks of size n_ctx.
+    // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
+    for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) {
+        const std::vector<llama_token> inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx);
+        std::vector<llama_token> pseudo_output;
+        pseudo_output.push_back(inp_slice[0]);
+
+        while ((int) pseudo_output.size() < n_ctx) {
+            // Simulate drafting and decoding from draft:
+            std::vector<llama_token> draft;
+            draft.push_back(pseudo_output.back());
+
+            {
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
+
+            n_drafted += draft.size() - 1;
+
+            for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) {
+                const llama_token ground_truth = inp_slice[pseudo_output.size()];
+                const llama_token drafted = draft[j];
+
+                if (ground_truth != drafted) {
+                    break;
+                }
+
+                ++n_accept;
+                pseudo_output.push_back(ground_truth);
+
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            // After each simulated batch decoding simulate the sampling of a single token:
+            if ((int) pseudo_output.size() < n_ctx) {
+                pseudo_output.push_back(inp_slice[pseudo_output.size()]);
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            draft.erase(draft.begin());
+
+        }
+        if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
+            const int64_t t_now_ms = ggml_time_ms();
+            const int64_t eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
+            const int64_t eta_min  = eta_ms / (60*1000);
+            const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+            LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
+        }
+
+        // After each chunk, update the dynamic ngram cache with the context ngram cache:
+        llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+        ngram_cache_context.clear();
+    }
+
+    LOG_TEE("\n");
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_input - n_input % n_ctx);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    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);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/lookup/lookup.cpp

@@ -1,12 +1,15 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
 
 #include <cmath>
 #include <cstdint>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>
+#include <unordered_map>
 
 int main(int argc, char ** argv){
     gpt_params params;
@@ -15,11 +18,7 @@ int main(int argc, char ** argv){
         return 1;
     }
 
-    // max/min n-grams size to search for in prompt
-    const int ngram_max = 4;
-    const int ngram_min = 1;
-
-    // length of the candidate / draft sequence, if match is found
+    // max. number of additional tokens to draft if match is found
     const int n_draft = params.n_draft;
 
     const bool dump_kv_cache = params.dump_kv_cache;
@@ -39,6 +38,8 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
+    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
     const bool add_bos = llama_should_add_bos_token(model);
@@ -47,6 +48,35 @@ int main(int argc, char ** argv){
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
 
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        // Fill up context ngram cache with tokens from user input:
+        const int64_t t_start_draft_us = ggml_time_us();
+        llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
     const int max_context_size     = llama_n_ctx(ctx);
     const int max_tokens_list_size = max_context_size - 4;
 
@@ -76,8 +106,6 @@ 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;
@@ -129,6 +157,12 @@ int main(int argc, char ** argv){
                 ++n_past;
                 ++i_dft;
                 inp.push_back(id);
+                {
+                    // Update context ngram cache with the newly accepted token:
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
 
                 if (params.use_color) {
                     // color accepted draft token
@@ -149,6 +183,12 @@ int main(int argc, char ** argv){
             draft.clear();
             draft.push_back(id);
             inp.push_back(id);
+            {
+                // Update context ngram cache with the newly accepted token:
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
             break;
         }
 
@@ -163,44 +203,19 @@ int main(int argc, char ** argv){
         llama_batch_clear(batch_tgt);
         llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
 
-        // generate n_pred tokens through prompt lookup
-        auto prompt_lookup = [&]() -> void {
-            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];
-
-                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
-                    bool match = true;
-                    for (int j = 0; j < ngram_size; ++j) {
-                        if (inp[i + j] != ngram[j]) {
-                            match = false;
-                            break;
-                        }
-                    }
-
-                    if (match) {
-                        const int startIdx = i + ngram_size;
-                        const int endIdx = startIdx + n_draft;
-                        if (endIdx < inp_size) {
-                            for (int j = startIdx; j < endIdx; ++j) {
-                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
-                                draft.push_back(inp[j]);
-                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
-                                ++n_drafted;
-                            }
-                            return;
-                        }
-                    }
-                }
-            }
-            return;
-        };
-
+        // Draft already contains a single token sampled from the model:
+        GGML_ASSERT(draft.size() == 1);
+        GGML_ASSERT(draft[0] == inp.back());
         const int64_t t_start_draft_us = ggml_time_us();
 
-        prompt_lookup();
+        llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+
+        for (size_t i = 1; i < draft.size(); ++i) {
+            llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+        }
 
         t_draft_us += ggml_time_us() - t_start_draft_us;
+        n_drafted += draft.size() - 1;
 
         llama_decode(ctx, batch_tgt);
         ++n_past;
@@ -210,19 +225,24 @@ int main(int argc, char ** argv){
 
     auto t_dec_end = ggml_time_us();
 
+    // Update dynamic ngram cache with context ngram cache and save it to disk:
+    llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+    llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+
     LOG_TEE("\n\n");
 
     LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
     LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
     LOG_TEE("\n");
-    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",
+    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_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    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);
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_TEE("\ntarget:\n");
     llama_print_timings(ctx);

examples/main-cmake-pkg/README.md

@@ -8,7 +8,7 @@ Because this example is "outside of the source tree", it is important to first b
 
 ### Considerations
 
-When hardware acceleration libraries are used (e.g. CUBlas, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
+When hardware acceleration libraries are used (e.g. CUDA, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
 
 ### Build llama.cpp and install to C:\LlamaCPP directory
 

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.
@@ -315,8 +316,8 @@ These options provide extra functionality and customization when running the LLa
 
 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
--   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
--   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
--   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -511,6 +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) {
@@ -769,6 +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());
                 }
@@ -858,6 +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;
 
@@ -129,7 +132,6 @@ int main(int argc, char ** argv) {
     llama_context * ctx = NULL;
 
     // load the target model
-    params.logits_all = true;
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // load the prompts from an external file if there are any
@@ -196,8 +198,8 @@ 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");
@@ -221,15 +223,17 @@ 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__);
@@ -255,7 +259,7 @@ 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
@@ -366,7 +370,8 @@ 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/perplexity/perplexity.cpp

@@ -380,6 +380,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
             const int batch_size  = std::min(end - batch_start, n_batch);
 
             //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
+            // TODO: use llama_batch.logits instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 //fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
@@ -442,7 +443,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     return {tokens, std::exp(nll / count), logit_history, prob_history};
 }
 
-static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
+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);
     }
@@ -453,7 +454,6 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // BOS tokens will be added for each chunk before eval
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    const int n_ctx = llama_n_ctx(ctx);
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -499,13 +499,19 @@ 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 +524,26 @@ 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 +553,54 @@ 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];
+            int n_outputs = 0;
 
-            // 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));
+            batch.n_tokens = 0;
+            for (int seq = 0; seq < n_seq_batch; seq++) {
+                int seq_start = batch_start + seq*n_ctx;
+
+                // save original token and restore it after eval
+                const auto token_org = tokens[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;
+
+                    n_outputs += batch.logits[idx] != 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) {
+            if (num_batches > 1 && n_outputs > 0) {
                 const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+                logits.insert(logits.end(), batch_logits, batch_logits + n_outputs * 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 +608,32 @@ 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 + first);
 
-        // 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);
+            llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;
+            if (!params.logits_file.empty()) {
+                process_logits(logits_stream, n_vocab, all_logits,
+                        tokens_data, n_ctx - 1 - first,
+                        workers, log_probs, nll, nll2);
+            } else {
+                process_logits(n_vocab, all_logits,
+                        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 + 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);
 
@@ -615,10 +652,13 @@ 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};
 }
 
 static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int32_t n_batch, int32_t n_vocab) {
+    int prev_outputs = 0;
     for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
         const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
 
@@ -639,7 +679,14 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             return false;
         }
 
-        memcpy(batch_logits.data() + i*n_vocab, llama_get_logits(ctx), n_tokens*n_vocab*sizeof(float));
+        int n_outputs = 0;
+        for (int i = 0; i < n_tokens; ++i) {
+            n_outputs += batch_view.logits[i] != 0;
+        }
+
+        memcpy(batch_logits.data() + prev_outputs*n_vocab, llama_get_logits(ctx), n_outputs*n_vocab*sizeof(float));
+
+        prev_outputs += n_outputs;
     }
 
     return true;
@@ -746,7 +793,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         size_t ending_logprob_count[4];
         double ending_logprob[4];
 
-        size_t i_batch;         // starting index in the llama_batch
+        size_t i_logits;        // starting index of logits in the llama_batch
         size_t common_prefix;   // max number of initial tokens that are the same in all sentences
         size_t required_tokens; // needed number of tokens to evaluate all 4 endings
         std::vector<llama_token> seq_tokens[4];
@@ -809,11 +856,12 @@ 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);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 4);
 
     std::vector<float> tok_logits(n_vocab);
+    // TODO: this could be made smaller; it's currently the worst-case size
     std::vector<float> batch_logits(n_vocab*n_ctx);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
@@ -824,16 +872,17 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+        size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
         llama_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
-        // each task has 4 unique seuqnce ids - one for each ending
+        // each task has 4 unique sequence ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         while (n_cur + (int) hs_data[i1].required_tokens <= n_ctx) {
             auto & hs_cur = hs_data[i1];
+            int n_logits = 0;
 
             const int s0 = 4*(i1 - i0);
             if (s0 + 4 > max_seq) {
@@ -841,18 +890,23 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             }
 
             for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
-                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
+                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            n_logits += 1;
 
             for (int s = 0; s < 4; ++s) {
-                for (size_t i = hs_cur.common_prefix; i < hs_cur.seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, true);
+                const size_t seq_tokens_size = hs_cur.seq_tokens[s].size();
+                // TODO: don't evaluate the last token of each sequence
+                for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
+                    const bool needs_logits = i < seq_tokens_size - 1;
+                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    n_logits += needs_logits;
                 }
             }
 
-            hs_cur.i_batch = i_batch;
-            i_batch += hs_cur.required_tokens;
+            hs_cur.i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += hs_data[i1].required_tokens;
             if (++i1 == hs_task_count) {
@@ -878,12 +932,11 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         eval_pairs.clear();
         for (size_t i = i0; i < i1; ++i) {
             auto & hs_cur = hs_data[i];
-            size_t li = hs_cur.common_prefix;
+            size_t li = 1; // skip the last logit of the common prefix (computed separately below)
             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.emplace_back(hs_cur.i_batch + li++, hs_cur.seq_tokens[s][j + 1]);
+                    eval_pairs.emplace_back(hs_cur.i_logits + li++, hs_cur.seq_tokens[s][j + 1]);
                 }
-                ++li;
             }
         }
         // Then we do the actual calculation
@@ -895,7 +948,8 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t i = i0; i < i1; ++i) {
             auto & hs_cur = hs_data[i];
 
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(hs_cur.i_batch + hs_cur.common_prefix - 1), n_vocab*sizeof(float));
+            // get the logits of the last token of the common prefix
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*hs_cur.i_logits, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -945,7 +999,7 @@ struct winogrande_entry {
     std::array<std::string, 2> choices;
     int answer;
 
-    size_t i_batch;
+    size_t i_logits;
     size_t common_prefix;
     size_t required_tokens;
     size_t n_base1; // number of tokens for context + choice 1
@@ -1071,6 +1125,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             task.common_prefix++;
         }
 
+        // TODO: the last token of each of the sequences don't need to be evaluated
         task.required_tokens = task.common_prefix +
             task.seq_tokens[0].size() - task.common_prefix +
             task.seq_tokens[1].size() - task.common_prefix;
@@ -1086,11 +1141,12 @@ 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);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 2);
 
     std::vector<float> tok_logits(n_vocab);
+    // TODO: this could be made smaller; it's currently the worst-case size
     std::vector<float> batch_logits(n_vocab*n_ctx);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
@@ -1104,29 +1160,33 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0;
+        size_t i_logits = 0;
 
         llama_batch_clear(batch);
 
         while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
+            int n_logits = 0;
             const int s0 = 2*(i1 - i0);
             if (s0 + 2 > max_seq) {
                 break;
             }
 
             for (size_t i = 0; i < data[i1].common_prefix; ++i) {
-                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1}, false);
+                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true;
+            n_logits += 1;
 
             for (int s = 0; s < 2; ++s) {
+                // TODO: end before the last token, no need to predict past the end of the sequences
                 for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
                     llama_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
+                    n_logits += 1;
                 }
             }
 
-            data[i1].i_batch = i_batch;
-            i_batch += data[i1].required_tokens;
+            data[i1].i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += data[i1].required_tokens;
             if (++i1 == data.size()) {
@@ -1157,15 +1217,16 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
 
             const auto& n_base1 = skip_choice ? task.n_base1 : task.common_prefix;
             const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0;
-            size_t li = n_base1 - 1;
+            size_t li = n_base1 - task.common_prefix;
             for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) {
-                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[0][j+1]);
+                eval_pairs.emplace_back(task.i_logits + 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;
+            // FIXME: this uses the wrong first logits when not skipping the choice word
+            li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - task.common_prefix;
             for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) {
-                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[1][j+1]);
+                eval_pairs.emplace_back(task.i_logits + li++, task.seq_tokens[1][j+1]);
             }
         }
         compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results);
@@ -1254,7 +1315,7 @@ struct multiple_choice_task {
     }
 
     // For evaluation
-    size_t i_batch;         // starting index in the llama_batch
+    size_t i_logits;        // starting index of logits in the llama_batch
     size_t common_prefix;   // max number of initial tokens that are the same in all sentences
     size_t required_tokens; // needed number of tokens to evaluate all answers
     std::vector<std::vector<llama_token>> seq_tokens;
@@ -1333,7 +1394,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     std::vector<uint32_t> task_pos(n_task);
     strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t));
     if (strstream.fail()) {
-        printf("%s: failed to raad task positions from prompt\n", __func__);
+        printf("%s: failed to read task positions from prompt\n", __func__);
         return;
     }
 
@@ -1414,7 +1475,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             return;
         }
     } else {
-        int n_dot = n_task/100;
+        int n_dot = std::max((int) n_task/100, 1);
         int i_task = 0;
         for (auto& task : tasks) {
             ++i_task;
@@ -1438,7 +1499,7 @@ 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);
 
@@ -1458,17 +1519,18 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+        size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
         llama_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
-        // each task has 4 unique seuqnce ids - one for each ending
+        // each task has 4 unique sequence ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         int s0 = 0;
         while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) {
             auto& cur_task = tasks[i1];
+            int n_logits = 0;
 
             int num_answers = cur_task.seq_tokens.size();
             if (s0 + num_answers > max_seq) {
@@ -1485,17 +1547,22 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
                 llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            n_logits += 1;
 
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
-                for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true);
+                const size_t seq_tokens_size = cur_task.seq_tokens[s].size();
+                // TODO: don't evaluate the last token of each sequence
+                for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
+                    const bool needs_logits = i < seq_tokens_size - 1;
+                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    n_logits += needs_logits;
                 }
             }
 
             s0 += num_answers;
 
-            cur_task.i_batch = i_batch;
-            i_batch += cur_task.required_tokens;
+            cur_task.i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += cur_task.required_tokens;
             if (++i1 == tasks.size()) {
@@ -1521,12 +1588,11 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         eval_pairs.clear();
         for (size_t i = i0; i < i1; ++i) {
             auto& cur_task = tasks[i];
-            size_t li = cur_task.common_prefix;
+            size_t li = 1; // skip the last logit of the common prefix (computed separately below)
             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.emplace_back(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]);
+                    eval_pairs.emplace_back(cur_task.i_logits + li++, cur_task.seq_tokens[s][j + 1]);
                 }
-                ++li;
             }
         }
         // Then we do the actual calculation
@@ -1545,7 +1611,8 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             //}
             //printf("\n    common_prefix: %zu\n", cur_task.common_prefix);
 
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float));
+            // get the logits of the last token of the common prefix
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*cur_task.i_logits, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -1697,6 +1764,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
+            // TODO: use llama_batch.logits instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;
@@ -1782,13 +1850,24 @@ 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",
@@ -1815,6 +1894,9 @@ int main(int argc, char ** argv) {
     llama_model * model;
     llama_context * ctx;
 
+    // ensure there's at least enough seq_ids for HellaSwag
+    params.n_parallel = std::max(4, params.n_parallel);
+
     // load the model and apply lora adapter, if any
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == NULL) {
@@ -1844,7 +1926,7 @@ 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/quantize.cpp

@@ -26,6 +26,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "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",            },
+    { "IQ1_M",  LLAMA_FTYPE_MOSTLY_IQ1_M,  " 1.75 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",            },
@@ -36,7 +37,8 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "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, " 4.25 bpw non-linear quantization", },
+    { "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", },
@@ -86,13 +88,17 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("  --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n");
     printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
+    printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
@@ -106,14 +112,14 @@ static void usage(const char * executable) {
     exit(1);
 }
 
-static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+static void load_imatrix(const std::string & imatrix_file, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     std::ifstream in(imatrix_file.c_str(), std::ios::binary);
     if (!in) {
-        printf("%s: failed to open %s\n",__func__,imatrix_file.c_str());
+        printf("%s: failed to open %s\n",__func__, imatrix_file.c_str());
         return;
     }
     int n_entries;
-    in.read((char*)&n_entries, sizeof(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.c_str());
         return;
@@ -123,25 +129,25 @@ static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std
         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.c_str());
+            printf("%s: failed reading name for entry %d from %s\n", __func__, i+1, imatrix_file.c_str());
             return;
         }
         name_as_vec[len] = 0;
         std::string name{name_as_vec.data()};
-        auto& e = imatrix_data[std::move(name)];
+        auto & e = imatrix_data[std::move(name)];
         int ncall;
-        in.read((char*)&ncall, sizeof(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);
+            printf("%s: failed reading number of values for entry %d\n", __func__, i);
             imatrix_data = {};
             return;
         }
         e.resize(nval);
-        in.read((char*)e.data(), nval*sizeof(float));
+        in.read((char *)e.data(), nval*sizeof(float));
         if (in.fail()) {
-            printf("%s: failed reading data for entry %d\n",__func__,i);
+            printf("%s: failed reading data for entry %d\n", __func__, i);
             imatrix_data = {};
             return;
         }
@@ -149,13 +155,13 @@ static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std
             for (auto& v : e) v /= ncall;
         }
     }
-    printf("%s: loaded %d importance matrix entries from %s\n",__func__,int(imatrix_data.size()),imatrix_file.c_str());
+    printf("%s: loaded %d importance matrix entries from %s\n", __func__, int(imatrix_data.size()), imatrix_file.c_str());
 }
 
-static void prepare_imatrix(const std::string& imatrix_file,
-        const std::vector<std::string>& included_weights,
-        const std::vector<std::string>& excluded_weights,
-        std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+static void prepare_imatrix(const std::string & imatrix_file,
+        const std::vector<std::string> & included_weights,
+        const std::vector<std::string> & excluded_weights,
+        std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     if (!imatrix_file.empty()) {
         load_imatrix(imatrix_file, imatrix_data);
     }
@@ -188,6 +194,55 @@ static void prepare_imatrix(const std::string& imatrix_file,
     }
 }
 
+static ggml_type parse_ggml_type(const char * arg) {
+    ggml_type result = GGML_TYPE_COUNT;
+    for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
+        auto type = ggml_type(j);
+        const auto * name = ggml_type_name(type);
+        if (name && strcmp(arg, name) == 0) {
+            result = type; break;
+        }
+    }
+    return result;
+}
+
+static bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides) {
+    const char* sep = strchr(data, '=');
+    if (sep == nullptr || sep - data >= 128) {
+        fprintf(stderr, "%s: malformed KV override '%s'\n", __func__, data);
+        return false;
+    }
+    llama_model_kv_override kvo;
+    std::strncpy(kvo.key, data, sep - data);
+    kvo.key[sep - data] = 0;
+    sep++;
+    if (strncmp(sep, "int:", 4) == 0) {
+        sep += 4;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+        kvo.int_value = std::atol(sep);
+    } else if (strncmp(sep, "float:", 6) == 0) {
+        sep += 6;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
+        kvo.float_value = std::atof(sep);
+    } else if (strncmp(sep, "bool:", 5) == 0) {
+        sep += 5;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
+        if (std::strcmp(sep, "true") == 0) {
+            kvo.bool_value = true;
+        } else if (std::strcmp(sep, "false") == 0) {
+            kvo.bool_value = false;
+        } else {
+            fprintf(stderr, "%s: invalid boolean value for KV override '%s'\n", __func__, data);
+            return false;
+        }
+    } else {
+        fprintf(stderr, "%s: invalid type for KV override '%s'\n", __func__, data);
+        return false;
+    }
+    overrides.emplace_back(std::move(kvo));
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -198,10 +253,27 @@ int main(int argc, char ** argv) {
     int arg_idx = 1;
     std::string imatrix_file;
     std::vector<std::string> included_weights, excluded_weights;
+    std::vector<llama_model_kv_override> kv_overrides;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
             params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
+            if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {
@@ -242,6 +314,11 @@ int main(int argc, char ** argv) {
     if (!imatrix_data.empty()) {
         params.imatrix = &imatrix_data;
     }
+    if (!kv_overrides.empty()) {
+        kv_overrides.emplace_back();
+        kv_overrides.back().key[0] = 0;
+        params.kv_overrides = &kv_overrides;
+    }
 
     llama_backend_init();
 
@@ -263,8 +340,7 @@ int main(int argc, char ** argv) {
         if (ftype_str == "COPY") {
             params.only_copy = true;
         }
-    }
-    else {
+    } else {
         fname_out = argv[arg_idx];
         arg_idx++;
 
@@ -295,10 +371,12 @@ int main(int argc, char ** argv) {
 
     if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
          params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
-        fprintf(stderr, "\n===============================================================================================\n");
-        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
-        fprintf(stderr, "===============================================================================================\n\n\n");
+         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S  ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) && imatrix_data.empty()) {
+        fprintf(stderr, "\n==========================================================================================================\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ1_M, IQ2_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "==========================================================================================================\n\n\n");
         return 1;
     }
 

examples/regex-to-grammar.py

@@ -0,0 +1,20 @@
+import json, subprocess, sys, os
+
+assert len(sys.argv) >= 2
+[_, pattern, *rest] = sys.argv
+
+print(subprocess.check_output(
+    [
+        "python",
+        os.path.join(
+        os.path.dirname(os.path.realpath(__file__)),
+        "json-schema-to-grammar.py"),
+        *rest,
+        "-",
+        "--raw-pattern",
+    ],
+    text=True,
+    input=json.dumps({
+        "type": "string",
+        "pattern": pattern,
+    }, indent=2)))

examples/retrieval/CMakeLists.txt

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

@@ -0,0 +1,69 @@
+# llama.cpp/examples/retrieval
+
+Demonstration of simple retrieval technique based on cosine similarity
+
+More info:
+https://github.com/ggerganov/llama.cpp/pull/6193
+
+### How to use
+
+`retieval.cpp` has parameters of its own:
+- `--context-file`: file to be embedded - state this option multiple times to embed multiple files
+- `--chunk-size`: minimum size of each text chunk to be embedded
+- `--chunk-separator`: STRING to divide chunks by. newline by default
+
+`retrieval` example can be tested as follows:
+
+```bash
+make -j && ./retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
+```
+
+This chunks and embeds all given files and starts a loop requesting query inputs:
+
+```
+Enter query:
+```
+
+On each query input, top k chunks are shown along with file name, chunk position within file and original text:
+
+```
+Enter query: describe the mit license
+batch_decode: n_tokens = 6, n_seq = 1
+Top 3 similar chunks:
+filename: README.md
+filepos: 119
+similarity: 0.762334
+textdata:
+png)
+
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+
+[Roadmap](https://github.
+--------------------
+filename: License
+filepos: 0
+similarity: 0.725146
+textdata:
+MIT License
+
+Copyright (c) 2023 Georgi Gerganov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+--------------------
+filename: README.md
+filepos: 9178
+similarity: 0.621722
+textdata:
+com/cztomsik/ava) (MIT)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [pythops/tenere](https://github.
+--------------------
+```

examples/retrieval/retrieval.cpp

@@ -0,0 +1,350 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <fstream>
+
+struct retrieval_params {
+    std::vector<std::string> context_files; // context files to embed
+    int32_t chunk_size            = 64;     // chunk size for context embedding
+    std::string chunk_separator   = "\n";   // chunk separator for context embedding
+};
+
+static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
+    gpt_print_usage(argc, argv, gpt_params);
+    printf("retrieval options:\n");
+    printf("  --context-file FNAME  file containing context to embed.\n");
+    printf("                        specify multiple files by providing --context-file option multiple times.\n");
+    printf("  --chunk-size N        minimum length of embedded text chunk (default:%d)\n", params.chunk_size);
+    printf("  --chunk-separator STRING\n");
+    printf("                        string to separate chunks (default: \"\\n\")\n");
+    printf("\n");
+}
+
+static void retrieval_params_parse(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & retrieval_params) {
+    int i = 1;
+    std::string arg;
+    while (i < argc) {
+        arg = argv[i];
+        bool invalid_gpt_param = false;
+        if(gpt_params_find_arg(argc, argv, argv[i], gpt_params, i, invalid_gpt_param)) {
+            if (invalid_gpt_param) {
+                fprintf(stderr, "error: invalid argument: %s\n", arg.c_str());
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            // option was parsed by gpt_params_find_arg
+        } else if (arg == "--context-file") {
+            if (++i >= argc) {
+                fprintf(stderr, "error: missing argument for --context-file\n");
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            // store the external file name in params
+            retrieval_params.context_files.push_back(argv[i]);
+        } else if (arg == "--chunk-size") {
+            if (++i >= argc) {
+                fprintf(stderr, "error: missing argument for --chunk-size\n");
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            retrieval_params.chunk_size = std::stoi(argv[i]);
+        } else if (arg == "--chunk-separator") {
+            if (++i >= argc) {
+                fprintf(stderr, "error: missing argument for --chunk-separator\n");
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            retrieval_params.chunk_separator = argv[i];
+        } else {
+            // unknown argument
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+            exit(1);
+        }
+        i++;
+    }
+}
+
+struct chunk {
+    // filename
+    std::string filename;
+    // original file position
+    size_t filepos;
+    // original text data
+    std::string textdata = "";
+    // tokenized text data
+    std::vector<llama_token> tokens;
+    // embedding
+    std::vector<float> embedding;
+};
+
+// chunk file data to chunks of size >= chunk_size
+// chunk_separator is the separator between chunks
+static std::vector<chunk> chunk_file(const std::string & filename, int chunk_size, const std::string & chunk_separator) {
+    std::vector<chunk> chunks;
+    std::ifstream f(filename.c_str());
+
+    if (!f.is_open()) {
+        fprintf(stderr, "Error: could not open file %s\n", filename.c_str());
+        return chunks;
+    }
+
+    chunk current_chunk;
+    char buffer[1024];
+    int64_t filepos = 0;
+    std::string current = "";
+    while (f.read(buffer, 1024)) {
+        current += std::string(buffer, f.gcount());
+        size_t pos;
+        while ((pos = current.find(chunk_separator)) != std::string::npos) {
+            current_chunk.textdata += current.substr(0, pos + chunk_separator.size());
+            if ((int) current_chunk.textdata.size() > chunk_size) {
+                // save chunk
+                current_chunk.filepos = filepos;
+                current_chunk.filename = filename;
+                chunks.push_back(current_chunk);
+                // update filepos
+                filepos += (int) current_chunk.textdata.size();
+                // reset current_chunk
+                current_chunk = chunk();
+            }
+            current = current.substr(pos + chunk_separator.size());
+        }
+
+    }
+    // add leftover data to last chunk
+    if (current_chunk.textdata.size() > 0) {
+        if (chunks.empty()) {
+            current_chunk.filepos = filepos;
+            current_chunk.filename = filename;
+            chunks.push_back(current_chunk);
+        } else {
+            chunks.back().textdata += current_chunk.textdata;
+        }
+    }
+    f.close();
+    return chunks;
+}
+
+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;
+    retrieval_params retrieval_params;
+
+    retrieval_params_parse(argc, argv, params, retrieval_params);
+
+    // For BERT models, batch size must be equal to ubatch size
+    params.n_ubatch = params.n_batch;
+
+    if (retrieval_params.chunk_size <= 0) {
+        fprintf(stderr, "chunk_size must be positive\n");
+        return 1;
+    }
+    if (retrieval_params.context_files.empty()) {
+        fprintf(stderr, "context_files must be specified\n");
+        return 1;
+    }
+    params.embedding = true;
+
+    print_build_info();
+
+    printf("processing files:\n");
+    for (auto & context_file : retrieval_params.context_files) {
+        printf("%s\n", context_file.c_str());
+    }
+
+    std::vector<chunk> chunks;
+    for (auto & context_file : retrieval_params.context_files) {
+        std::vector<chunk> file_chunk = chunk_file(context_file, retrieval_params.chunk_size, retrieval_params.chunk_separator);
+        chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
+    }
+    printf("Number of chunks: %ld\n", chunks.size());
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch >= params.n_ctx);
+
+    // tokenize the prompts and trim
+    for (auto & chunk : chunks) {
+        auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false);
+        if (inp.size() > n_batch) {
+            fprintf(stderr, "%s: error: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
+                    __func__, (long long int) inp.size(), (long long int) n_batch);
+            return 1;
+        }
+        // add eos if not present
+        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+            inp.push_back(llama_token_eos(model));
+        }
+        chunk.tokens = inp;
+    }
+
+    // tokenization stats
+    if (params.verbose_prompt) {
+        for (int i = 0; i < (int) chunks.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
+            for (int j = 0; j < (int) chunks[i].tokens.size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
+        }
+    }
+
+    // initialize batch
+    const int n_chunks = chunks.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+
+    // allocate output
+    const int n_embd = llama_n_embd(model);
+    std::vector<float> embeddings(n_chunks * n_embd, 0);
+    float * emb = embeddings.data();
+
+    // 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_chunks; k++) {
+        // clamp to n_batch tokens
+        auto & inp = chunks[k].tokens;
+
+        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;
+    }
+
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
+
+    // save embeddings to chunks
+    for (int i = 0; i < n_chunks; i++) {
+        chunks[i].embedding = std::vector<float>(emb + i * n_embd, emb + (i + 1) * n_embd);
+        // clear tokens as they are no longer needed
+        chunks[i].tokens.clear();
+    }
+
+    // start loop, receive query and return top k similar chunks based on cosine similarity
+    std::string query;
+    while (true) {
+        printf("Enter query: ");
+        std::getline(std::cin, query);
+        std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);
+
+        struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
+        batch_add_seq(query_batch, query_tokens, 0);
+
+        std::vector<float> query_emb(n_embd, 0);
+        batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
+
+        llama_batch_clear(query_batch);
+
+        // compute cosine similarities
+        {
+            std::vector<std::pair<int, float>> similarities;
+            for (int i = 0; i < n_chunks; i++) {
+                float sim = llama_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
+                similarities.push_back(std::make_pair(i, sim));
+            }
+
+            // sort similarities
+            std::sort(similarities.begin(), similarities.end(), [](const std::pair<int, float> & a, const std::pair<int, float> & b) {
+                return a.second > b.second;
+            });
+
+            printf("Top %d similar chunks:\n", params.sparams.top_k);
+            for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
+                printf("filename: %s\n", chunks[similarities[i].first].filename.c_str());
+                printf("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
+                printf("similarity: %f\n", similarities[i].second);
+                printf("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str());
+                printf("--------------------\n");
+            }
+        }
+    }
+
+    // clean up
+    llama_print_timings(ctx);
+    llama_free(ctx);
+    llama_free_model(model);
+    llama_backend_free();
+}

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 +1,22 @@
 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
+    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 json-schema-to-grammar ${CMAKE_THREAD_LIBS_INIT})
+if (LLAMA_SERVER_SSL)
+    find_package(OpenSSL REQUIRED)
+    target_link_libraries(${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

@@ -16,15 +16,20 @@ The project is under active development, and we are [looking for feedback and co
 
 **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 N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching, this parameter is used only if one token is to be processed on CPU backend.
+- `-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. Not used if model layers are offloaded to GPU.
+- `--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 (default: unused).
+- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository (default: unused).
+- `-hff FILE, --hf-file FILE`: Hugging Face model file (default: unused).
 - `-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.
-- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
-- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
-- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
+- `-ngl N`, `--n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
+- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
+- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048`.
+- `-ub N`, `--ubatch-size N`: physical maximum batch size. Default: `512`.
 - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
@@ -41,7 +46,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
 - `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 - `--port`: Set the port to listen. Default: `8080`.
-- `--path`: path from which to serve static files (default 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.
@@ -55,9 +60,13 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--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-disable`: Output logs to stdout only, not to `llama.log`. 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
 
 server is build alongside everything else from the root of the project
@@ -74,6 +83,28 @@ 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:
@@ -96,10 +127,10 @@ 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
@@ -168,7 +199,11 @@ node index.js
 
     *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).
 
@@ -228,7 +263,7 @@ node index.js
 
     `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)
+    `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
     `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
 
@@ -281,7 +316,7 @@ 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.
 
@@ -325,7 +360,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `default_generation_settings` - the default generation settings for the `/completion` endpoint, has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
+- **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 model with [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, ChatML template will be used.
 
     *Options:*
 
@@ -435,7 +470,7 @@ Notice that each `probs` is an array of length `n_probs`.
         "next_token": {
             "has_next_token": true,
             "n_remain": -1,
-            "num_tokens_predicted": 0,
+            "n_decoded": 0,
             "stopped_eos": false,
             "stopped_limit": false,
             "stopped_word": false,
@@ -525,26 +560,55 @@ 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.mjs

@@ -26,8 +26,9 @@ const propOrder = grammarJsonSchemaPropOrder
 
 let grammar = null
 if (grammarJsonSchemaFile) {
-    const schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8'))
-    const converter = new SchemaConverter(propOrder)
+    let schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8'))
+    const converter = new SchemaConverter({prop_order: propOrder, allow_fetch: true})
+    schema = await converter.resolveRefs(schema, grammarJsonSchemaFile)
     converter.visit(schema, '')
     grammar = converter.formatGrammar()
 }