ggml-alloc : fix assert in debug builds

* refact : fix convert script + zero out KV cache to avoid nans

* ggml : silu(-inf) should never happen

* metal : assert various kernel requirements

.clang-tidy

@@ -3,6 +3,7 @@ Checks: >
     bugprone-*,
     -bugprone-easily-swappable-parameters,
     -bugprone-implicit-widening-of-multiplication-result,
+    -bugprone-misplaced-widening-cast,
     -bugprone-narrowing-conversions,
     readability-*,
     -readability-avoid-unconditional-preprocessor-if,
@@ -15,4 +16,8 @@ Checks: >
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,
     portability-*,
+    misc-*,
+    -misc-const-correctness,
+    -misc-non-private-member-variables-in-classes,
+    -misc-no-recursion,
 FormatStyle: none

.devops/cloud-v-pipeline

@@ -0,0 +1,22 @@
+node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
+    stage('Cleanup'){
+        cleanWs()               // Cleaning previous CI build in workspace
+    }
+    stage('checkout repo'){
+        retry(5){               // Retry if the cloning fails due to some reason
+            checkout scm        // Clone the repo on Runner
+        }
+    }
+    stage('Compiling llama.cpp'){
+        sh'''#!/bin/bash
+            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
+        '''
+    }
+    stage('Running llama.cpp'){
+        sh'''#!/bin/bash
+            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            cat llama_log.txt                   # Printing results
+        '''
+    }
+}

.devops/full-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip
+    apt-get install -y build-essential python3 python3-pip git
 
 COPY requirements.txt requirements.txt
 

.devops/main-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential
+    apt-get install -y build-essential git
 
 WORKDIR /app
 

.devops/tools.sh

@@ -7,15 +7,12 @@ arg1="$1"
 # Shift the arguments to remove the first one
 shift
 
-# Join the remaining arguments into a single string
-arg2="$@"
-
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert.py "$arg2"
+    python3 ./convert.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./quantize "$arg2"
+    ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./main "$arg2"
+    ./main "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -27,7 +24,7 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./server "$arg2"
+    ./server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "

.dockerignore

@@ -1,6 +1,9 @@
 *.o
 *.a
 .cache/
+.git/
+.github/
+.gitignore
 .vs/
 .vscode/
 .DS_Store

.editorconfig

@@ -17,3 +17,6 @@ indent_style = tab
 
 [prompts/*.txt]
 insert_final_newline = unset
+
+[examples/server/public/*]
+indent_size = 2

.github/workflows/build.yml

@@ -10,15 +10,14 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
-  GGML_NITER: 1
   GGML_N_THREADS: 1
 
 jobs:
@@ -28,7 +27,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -39,7 +38,13 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          CC=gcc-8 make
+          CC=gcc-8 make -j $(nproc)
+
+      - name: Test
+        id: make_test
+        run: |
+          CC=gcc-8 make tests -j $(nproc)
+          make test -j $(nproc)
 
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
@@ -47,7 +52,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -61,7 +66,7 @@ jobs:
           mkdir build
           cd build
           cmake ..
-          cmake --build . --config Release
+          cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -82,7 +87,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -96,7 +101,7 @@ jobs:
           mkdir build
           cd build
           cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -116,7 +121,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -130,7 +135,7 @@ jobs:
           mkdir build
           cd build
           cmake -DLLAMA_MPI=ON ..
-          cmake --build . --config Release
+          cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -144,7 +149,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -155,11 +160,46 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          make
+          make -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: make_test
+        run: |
+          make tests -j $(sysctl -n hw.logicalcpu)
+          make test -j $(sysctl -n hw.logicalcpu)
 
   macOS-latest-cmake:
     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 ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest --verbose --timeout 900
+
+  macOS-latest-cmake-ios:
+    runs-on: macos-latest
+
     steps:
       - name: Clone
         id: checkout
@@ -177,14 +217,64 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF ..
-          cmake --build . --config Release
+          cmake -G Xcode .. \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: Test
-        id: cmake_test
+  macOS-latest-cmake-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
         run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
           cd build
-          ctest --verbose --timeout 900
+          cmake -G Xcode .. \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+  macOS-latest-swift:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
 
   windows-latest-cmake:
     runs-on: windows-latest
@@ -198,22 +288,24 @@ jobs:
       matrix:
         include:
           - build: 'noavx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2'
-            defines: '-DLLAMA_BUILD_SERVER=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'clblast'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
       - name: Download OpenCL SDK
         id: get_opencl
@@ -255,7 +347,7 @@ jobs:
           mkdir build
           cd build
           cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Add clblast.dll
         id: add_clblast_dll
@@ -323,28 +415,30 @@ jobs:
 
     strategy:
       matrix:
-        cuda: ['12.1.0', '11.7.1']
+        cuda: ['12.2.0', '11.7.1']
         build: ['cublas']
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
-      - uses: Jimver/cuda-toolkit@v0.2.10
+      - uses: Jimver/cuda-toolkit@v0.2.11
         id: cuda-toolkit
         with:
           cuda: ${{ matrix.cuda }}
-          # TODO(green-sky): _dev seems to fail, and non dev are not enought
-          #sub-packages: '["nvcc", "cudart", "cublas", "cudart_dev", "cublas_dev"]'
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON
-          cmake --build . --config Release
+          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
         id: tag
@@ -373,27 +467,11 @@ jobs:
             llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
-        if: ${{ matrix.cuda == '12.1.0' }}
-        # TODO(green-sky): paths are cuda 12 specific
         run: |
           echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
-          mkdir '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_12.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_12.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_12.dll" '.\build\bin\cudart\'
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
-
-      - name: Copy and pack Cuda runtime
-        if: ${{ matrix.cuda == '11.7.1' }}
-        # TODO(green-sky): paths are cuda 11 specific
-        run: |
-          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
-          mkdir '.\build\bin\cudart\'
-          ls "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin"
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_110.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_11.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_11.dll" '.\build\bin\cudart\'
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
+          $dst='.\build\bin\cudart\'
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -402,6 +480,23 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
+#  freeBSD-latest:
+#    runs-on: macos-12
+#    steps:
+#    - name: Clone
+#      uses: actions/checkout@v3
+#
+#    - name: Build
+#      uses: cross-platform-actions/action@v0.19.0
+#      with:
+#        operating_system: freebsd
+#        version: '13.2'
+#        hypervisor: 'qemu'
+#        run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 clinfo clover opencl clblast openblas
+#            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -418,7 +513,9 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
       - name: Determine tag name
         id: tag
@@ -476,7 +573,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -500,7 +597,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -524,7 +621,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -554,7 +651,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -593,7 +690,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -639,7 +736,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |

.github/workflows/code-coverage.yml

@@ -0,0 +1,36 @@
+name: Code Coverage
+on: [push, pull_request]
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+
+jobs:
+  run:
+    runs-on: ubuntu-20.04
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8 lcov
+
+      - name: Build
+        run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
+
+      - name: Run tests
+        run: CC=gcc-8 make test
+
+      - name: Generate coverage report
+        run: |
+          make coverage
+          make lcov-report
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v3
+        env:
+           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
+        with:
+          files: lcov-report/coverage.info

.github/workflows/docker.yml

@@ -26,8 +26,15 @@ jobs:
     strategy:
       matrix:
         config:
-          - { tag: "light", dockerfile: ".devops/main.Dockerfile" }
-          - { tag: "full", dockerfile: ".devops/full.Dockerfile" }
+          - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
+          #                     have disabled them for now until the reason why
+          #                     is understood.
+          - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3
@@ -51,7 +58,7 @@ jobs:
         with:
           context: .
           push: true
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.config.platforms }}
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
@@ -60,6 +67,6 @@ jobs:
         with:
           context: .
           push: ${{ github.event_name == 'push' }}
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.config.platforms }}
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}

.github/workflows/gguf-publish.yml

@@ -0,0 +1,44 @@
+# This workflow will upload a Python Package using Twine when a GGUF release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# See `gguf-py/README.md` for how to make a release.
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  workflow_dispatch:
+  push:
+    # Pattern matched against refs/tags
+    tags:
+      - 'gguf-v*'           # Push events to every version tag
+
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v3
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.9.x'
+    - name: Install dependencies
+      run: |
+        cd gguf-py
+        python -m pip install poetry
+        poetry install
+
+    - name: Build package
+      run: cd gguf-py && poetry build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        password: ${{ secrets.PYPI_API_TOKEN }}
+        packages-dir: gguf-py/dist

.github/workflows/zig-build.yml

@@ -0,0 +1,25 @@
+name: Zig CI
+
+on:
+  pull_request:
+  push:
+    branches:
+      - master
+
+jobs:
+  build:
+    strategy:
+      fail-fast: false
+      matrix:
+        runs-on: [ubuntu-latest, macos-latest, windows-latest]
+    runs-on: ${{ matrix.runs-on }}
+    steps:
+      - uses: actions/checkout@v3
+        with:
+          submodules: recursive
+          fetch-depth: 0
+      - uses: goto-bus-stop/setup-zig@v2
+        with:
+          version: 0.11.0
+      - name: Build Summary
+        run: zig build --summary all -freference-trace

.gitignore

@@ -5,6 +5,12 @@
 *.bin
 *.exe
 *.dll
+*.log
+*.gcov
+*.gcno
+*.gcda
+*.dot
+*.metallib
 .DS_Store
 .build/
 .cache/
@@ -16,6 +22,9 @@
 .vs/
 .vscode/
 
+lcov-report/
+gcovr-report/
+
 build*/
 out/
 tmp/
@@ -23,24 +32,35 @@ tmp/
 models/*
 models-mnt
 
+/Pipfile
+/baby-llama
+/beam-search
+/benchmark-matmult
+/convert-llama2c-to-ggml
+/embd-input-test
+/embedding
+/gguf
+/gguf-llama-simple
+/infill
+/libllama.so
+/llama-bench
 /main
+/metal
+/perplexity
+/q8dot
 /quantize
 /quantize-stats
 /result
-/perplexity
-/embedding
-/train-text-from-scratch
-/convert-llama2c-to-ggml
-/simple
-/benchmark-matmult
-/vdot
+/save-load-state
 /server
-/Pipfile
-/embd-input-test
-/gguf
-/gguf-llama-simple
-/libllama.so
-/llama-bench
+/simple
+/batched
+/export-lora
+/finetune
+/speculative
+/parallel
+/train-text-from-scratch
+/vdot
 build-info.h
 arm_neon.h
 compile_commands.json
@@ -72,4 +92,5 @@ tests/test-quantize-perf
 tests/test-sampling
 tests/test-tokenizer-0-llama
 tests/test-tokenizer-0-falcon
-tests/test-tokenizer-1
+tests/test-tokenizer-1-llama
+tests/test-tokenizer-1-bpe

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.12) # Don't bump this version for no reason
+cmake_minimum_required(VERSION 3.13)  # for add_link_options
 project("llama.cpp" C CXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -36,9 +36,15 @@ endif()
 # Option list
 #
 
+if (APPLE)
+    set(LLAMA_METAL_DEFAULT ON)
+else()
+    set(LLAMA_METAL_DEFAULT OFF)
+endif()
+
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
@@ -52,15 +58,21 @@ option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"
 option(LLAMA_SANITIZE_UNDEFINED         "llama: enable undefined sanitizer"                     OFF)
 
 # instruction set specific
-option(LLAMA_AVX                        "llama: enable AVX"                                     ON)
-option(LLAMA_AVX2                       "llama: enable AVX2"                                    ON)
-option(LLAMA_AVX512                     "llama: enable AVX512"                                  OFF)
-option(LLAMA_AVX512_VBMI                "llama: enable AVX512-VBMI"                             OFF)
-option(LLAMA_AVX512_VNNI                "llama: enable AVX512-VNNI"                             OFF)
-option(LLAMA_FMA                        "llama: enable FMA"                                     ON)
+if (LLAMA_NATIVE)
+    set(INS_ENB OFF)
+else()
+    set(INS_ENB ON)
+endif()
+
+option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
+option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
+option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
+option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
+option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
+option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
-    option(LLAMA_F16C                   "llama: enable F16C"                                    ON)
+    option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
 endif()
 
 # 3rd party libs
@@ -74,9 +86,12 @@ set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kern
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
 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_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
-option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
+option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
+option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
@@ -109,7 +124,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
     add_custom_command(
         OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
         COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
+        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
         DEPENDS "${GIT_DIR}/index"
         VERBATIM
@@ -128,6 +143,7 @@ set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
+include(CheckCXXCompilerFlag)
 
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
@@ -152,12 +168,40 @@ if (APPLE AND LLAMA_ACCELERATE)
         message(STATUS "Accelerate framework found")
 
         add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
     else()
         message(WARNING "Accelerate framework not found")
     endif()
 endif()
 
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+    message(STATUS "Metal framework found")
+    set(GGML_HEADERS_METAL ggml-metal.h)
+    set(GGML_SOURCES_METAL ggml-metal.m)
+
+    add_compile_definitions(GGML_USE_METAL)
+    if (LLAMA_METAL_NDEBUG)
+        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
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        )
+endif()
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -234,7 +278,8 @@ if (LLAMA_BLAS)
 endif()
 
 if (LLAMA_K_QUANTS)
-    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
+    set(GGML_HEADERS_EXTRA k_quants.h)
+    set(GGML_SOURCES_EXTRA k_quants.c)
     add_compile_definitions(GGML_USE_K_QUANTS)
     if (LLAMA_QKK_64)
         add_compile_definitions(GGML_QKK_64)
@@ -250,7 +295,8 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+        set(GGML_HEADERS_CUDA ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
         add_compile_definitions(GGML_USE_CUBLAS)
 #        if (LLAMA_CUDA_CUBLAS)
@@ -268,6 +314,7 @@ if (LLAMA_CUBLAS)
             add_compile_definitions(GGML_CUDA_F16)
         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_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -293,39 +340,18 @@ if (LLAMA_CUBLAS)
     endif()
 endif()
 
-if (LLAMA_METAL)
-    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-
-    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
-
-    add_compile_definitions(GGML_USE_METAL)
-    add_compile_definitions(GGML_METAL_NDEBUG)
-
-    # get full path to the file
-    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
-    # copy ggml-metal.metal to bin directory
-    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
-        ${FOUNDATION_LIBRARY}
-        ${METAL_FRAMEWORK}
-        ${METALKIT_FRAMEWORK}
-        )
-endif()
-
 if (LLAMA_MPI)
     cmake_minimum_required(VERSION 3.10)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
+        set(GGML_HEADERS_MPI ggml-mpi.h)
         set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-        set(cxx_flags ${cxx_flags} -Wno-cast-qual)
-        set(c_flags   ${c_flags}   -Wno-cast-qual)
+        if (NOT MSVC)
+            add_compile_options(-Wno-cast-qual)
+        endif()
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
         # Even if you're only using the C header, C++ programs may bring in MPI
@@ -343,7 +369,8 @@ if (LLAMA_CLBLAST)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_HEADERS_OPENCL ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
@@ -371,13 +398,15 @@ if (LLAMA_HIPBLAS)
         message(STATUS "HIP and hipBLAS found")
         add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
         add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
         if (LLAMA_CUDA_FORCE_DMMV)
             target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
         endif()
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-        target_compile_definitions(ggml-rocm PRIVATE CC_TURING=1000000000)
         set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
         target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 
@@ -392,37 +421,57 @@ endif()
 
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        set(c_flags
-            -Wall
-            -Wextra
-            -Wpedantic
-            -Wcast-qual
-            -Wdouble-promotion
-            -Wshadow
-            -Wstrict-prototypes
-            -Wpointer-arith
-            -Wmissing-prototypes
-        )
-        set(cxx_flags
-            -Wall
-            -Wextra
-            -Wpedantic
-            -Wcast-qual
-            -Wno-unused-function
-            -Wno-multichar
-        )
+        set(warning_flags -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int
+            -Werror=implicit-function-declaration)
+        set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
+        set(host_cxx_flags "")
+
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
+            set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
+
+            if (
+                (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
+                (CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
+            )
+                set(c_flags ${c_flags} -Wdouble-promotion)
+            endif()
+        elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+            set(c_flags ${c_flags} -Wdouble-promotion)
+            set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
+
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
+            endif()
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
+            endif()
+        endif()
     else()
         # todo : msvc
     endif()
 
-    add_compile_options(
-            "$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
-            "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
-    )
+    set(c_flags   ${c_flags}   ${warning_flags})
+    set(cxx_flags ${cxx_flags} ${warning_flags})
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags} ${host_cxx_flags}>")
 
 endif()
 
-if (MSVC)
+if (NOT MSVC)
+    set(cuda_flags -Wno-pedantic)
+endif()
+set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+
+list(JOIN host_cxx_flags " " cuda_host_flags)  # pass host compiler flags as a single argument
+if (NOT cuda_host_flags STREQUAL "")
+    set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
+endif()
+
+add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
+
+if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 
     if (BUILD_SHARED_LIBS)
@@ -444,6 +493,13 @@ endif()
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+if (MSVC)
+  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+else ()
+  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+endif ()
+
 if (NOT MSVC)
     if (LLAMA_STATIC)
         add_link_options(-static)
@@ -454,30 +510,35 @@ if (NOT MSVC)
     if (LLAMA_GPROF)
         add_compile_options(-pg)
     endif()
-    if (LLAMA_NATIVE)
-        add_compile_options(-march=native)
-    endif()
 endif()
 
-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
     message(STATUS "ARM detected")
     if (MSVC)
-        # TODO: arm msvc?
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+        add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
     else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            add_compile_options(-mfp16-format=ieee)
+        endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
             # Raspberry Pi 1, Zero
-            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
             # Raspberry Pi 2
-            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            add_compile_options(-mfp16-format=ieee -mno-unaligned-access)
+            add_compile_options(-mno-unaligned-access)
         endif()
     endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
     message(STATUS "x86 detected")
     if (MSVC)
         if (LLAMA_AVX512)
@@ -503,6 +564,9 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
             add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
+        if (LLAMA_NATIVE)
+            add_compile_options(-march=native)
+        endif()
         if (LLAMA_F16C)
             add_compile_options(-mf16c)
         endif()
@@ -534,27 +598,86 @@ else()
     message(STATUS "Unknown architecture")
 endif()
 
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (
+    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
+    CMAKE_SYSTEM_NAME MATCHES "iOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "tvOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
+)
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
 #
 # libraries
 #
 
 # ggml
 
+if (GGML_USE_CPU_HBM)
+    add_definitions(-DGGML_USE_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+endif()
+
 add_library(ggml OBJECT
             ggml.c
             ggml.h
             ggml-alloc.c
             ggml-alloc.h
-            ${GGML_SOURCES_CUDA}
-            ${GGML_SOURCES_OPENCL}
-            ${GGML_SOURCES_METAL}
-            ${GGML_SOURCES_MPI}
-            ${GGML_SOURCES_EXTRA}
+            ggml-backend.c
+            ggml-backend.h
+            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+if (GGML_USE_CPU_HBM)
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
 if (BUILD_SHARED_LIBS)
@@ -584,14 +707,54 @@ if (BUILD_SHARED_LIBS)
     if (LLAMA_METAL)
         set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
     endif()
-    install(TARGETS llama LIBRARY)
 endif()
 
+
 #
 # install
 #
 
 include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR}
+    CACHE PATH "Location of header files")
+set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR}
+    CACHE PATH "Location of library files")
+set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR}
+    CACHE PATH "Location of binary files")
+set(LLAMA_BUILD_NUMBER ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/scripts/LlamaConfig.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama
+    PATH_VARS LLAMA_INCLUDE_INSTALL_DIR
+              LLAMA_LIB_INSTALL_DIR
+              LLAMA_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+    VERSION ${LLAMA_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
+
+set(GGML_PUBLIC_HEADERS "ggml.h"
+        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
+
+set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+install(TARGETS ggml PUBLIC_HEADER)
+
+set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/llama.h)
+install(TARGETS llama LIBRARY PUBLIC_HEADER)
+
 install(
     FILES convert.py
     PERMISSIONS

Makefile

@@ -1,10 +1,11 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple server embd-input-test gguf llama-bench
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative infill benchmark-matmult parallel finetune export-lora 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
+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
 
-default: $(BUILD_TARGETS)
+# Code coverage output files
+COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -18,12 +19,27 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
-CXXV := $(shell $(CXX) --version | head -n 1)
+ifeq '' '$(findstring clang,$(shell $(CC) --version))'
+	CC_IS_GCC=1
+	CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+else
+	CC_IS_CLANG=1
+	ifeq '' '$(findstring Apple LLVM,$(shell $(CC) --version))'
+		CC_IS_LLVM_CLANG=1
+	else
+		CC_IS_APPLE_CLANG=1
+	endif
+	CC_VER := $(shell $(CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
+				| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+endif
 
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
+	ifndef LLAMA_NO_METAL
+		LLAMA_METAL := 1
+	endif
+
 	ifneq ($(UNAME_P),arm)
 		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
 		ifeq ($(SYSCTL_M),1)
@@ -34,64 +50,176 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
+ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
+BUILD_TARGETS += metal
+endif
+
+default: $(BUILD_TARGETS)
+
+test: $(TEST_TARGETS)
+	@failures=0; \
+	for test_target in $(TEST_TARGETS); do \
+		if [ "$$test_target" = "tests/test-tokenizer-0-llama" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \
+		elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; then \
+			continue; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \
+			continue; \
+		else \
+			echo "Running test $$test_target..."; \
+			./$$test_target; \
+		fi; \
+		if [ $$? -ne 0 ]; then \
+			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
+			failures=$$(( failures + 1 )); \
+		else \
+			printf 'Test %s passed.\n\n' $$test_target; \
+		fi; \
+	done; \
+	if [ $$failures -gt 0 ]; then \
+		printf '\n%s tests failed.\n' $$failures; \
+		exit 1; \
+	fi
+	@echo 'All tests passed.'
+
+all: $(BUILD_TARGETS) $(TEST_TARGETS)
+
+coverage: ## Run code coverage
+	gcov -pb tests/*.cpp
+
+lcov-report: coverage ## Generate lcov report
+	mkdir -p lcov-report
+	lcov --capture --directory . --output-file lcov-report/coverage.info
+	genhtml lcov-report/coverage.info --output-directory lcov-report
+
+gcovr-report: coverage ## Generate gcovr report
+	mkdir -p gcovr-report
+	gcovr --root . --html --html-details --output gcovr-report/coverage.html
+
+ifdef RISCV_CROSS_COMPILE
+CC	:= riscv64-unknown-linux-gnu-gcc
+CXX	:= riscv64-unknown-linux-gnu-g++
+endif
+
 #
 # Compile flags
 #
 
 # keep standard at C11 and C++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
+
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-OPT = -Ofast
+MK_CFLAGS        += -Ofast
+MK_HOST_CXXFLAGS += -Ofast
+MK_CUDA_CXXFLAGS += -O3
 else
-OPT = -O3
+MK_CFLAGS        += -O3
+MK_CXXFLAGS      += -O3
+endif
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+MK_CPPFLAGS += -D_XOPEN_SOURCE=600
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+ifeq ($(UNAME_S),OpenBSD)
+	MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+endif
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+ifeq ($(UNAME_S),Linux)
+	MK_CPPFLAGS += -D_GNU_SOURCE
+endif
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+ifeq ($(UNAME_S),Darwin)
+	MK_CPPFLAGS += -D_DARWIN_C_SOURCE
+endif
+ifeq ($(UNAME_S),DragonFly)
+	MK_CPPFLAGS += -D__BSD_VISIBLE
+endif
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+ifeq ($(UNAME_S),FreeBSD)
+	MK_CPPFLAGS += -D__BSD_VISIBLE
+endif
+ifeq ($(UNAME_S),NetBSD)
+	MK_CPPFLAGS += -D_NETBSD_SOURCE
+endif
+ifeq ($(UNAME_S),OpenBSD)
+	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
-CFLAGS   = -I.            $(OPT) -std=c11   -fPIC
-CXXFLAGS = -I. -I./common $(OPT) -std=c++11 -fPIC
-LDFLAGS  =
 
 ifdef LLAMA_DEBUG
-	CFLAGS   += -O0 -g
-	CXXFLAGS += -O0 -g
-	LDFLAGS  += -g
+	MK_CFLAGS   += -O0 -g
+	MK_CXXFLAGS += -O0 -g
+	MK_LDFLAGS  += -g
 else
-	CFLAGS   += -DNDEBUG
-	CXXFLAGS += -DNDEBUG
+	MK_CPPFLAGS += -DNDEBUG
 endif
 
 ifdef LLAMA_SERVER_VERBOSE
-	CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
+	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
+
+ifdef LLAMA_CODE_COVERAGE
+	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
+endif
+
+ifdef LLAMA_DISABLE_LOGS
+	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
+endif # LLAMA_DISABLE_LOGS
+
 # warnings
-CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
-			-Wmissing-prototypes
-CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
+WARN_FLAGS    = -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int \
+				-Werror=implicit-function-declaration
+MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
+
+ifeq ($(CC_IS_CLANG), 1)
+	# clang options
+	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
+	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+
+	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+else
+	# gcc options
+	MK_CFLAGS        += -Wdouble-promotion
+	MK_HOST_CXXFLAGS += -Wno-array-bounds
+
+	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
+		MK_HOST_CXXFLAGS += -Wno-format-truncation
+	endif
+	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
+		MK_HOST_CXXFLAGS += -Wextra-semi
+	endif
+endif
 
 # OS specific
 # TODO: support Windows
-ifeq ($(UNAME_S),Linux)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Darwin)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),FreeBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),NetBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),OpenBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Haiku)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
+ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)'
+	MK_CFLAGS   += -pthread
+	MK_CXXFLAGS += -pthread
 endif
 
 # detect Windows
@@ -117,104 +245,119 @@ ifeq ($(_WIN32),1)
 endif
 
 ifdef LLAMA_GPROF
-	CFLAGS   += -pg
-	CXXFLAGS += -pg
+	MK_CFLAGS   += -pg
+	MK_CXXFLAGS += -pg
 endif
 ifdef LLAMA_PERF
-	CFLAGS   += -DGGML_PERF
-	CXXFLAGS += -DGGML_PERF
+	MK_CPPFLAGS += -DGGML_PERF
 endif
 
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
+
+ifndef RISCV
+
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
-	CFLAGS   += -march=native -mtune=native
-	CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS   += -march=native -mtune=native
+	MK_HOST_CXXFLAGS += -march=native -mtune=native
 
 	# Usage AVX-only
-	#CFLAGS   += -mfma -mf16c -mavx
-	#CXXFLAGS += -mfma -mf16c -mavx
+	#MK_CFLAGS   += -mfma -mf16c -mavx
+	#MK_CXXFLAGS += -mfma -mf16c -mavx
 
 	# Usage SSSE3-only (Not is SSE3!)
-	#CFLAGS   += -mssse3
-	#CXXFLAGS += -mssse3
+	#MK_CFLAGS   += -mssse3
+	#MK_CXXFLAGS += -mssse3
+endif
+
+# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+# https://github.com/ggerganov/llama.cpp/issues/2922
+ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
+	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
+	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
-	CFLAGS   += -mcpu=native
-	CXXFLAGS += -mcpu=native
+	MK_CFLAGS   += -mcpu=native
+	MK_CXXFLAGS += -mcpu=native
 endif
 
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, Zero
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
 
 ifneq ($(filter armv7%,$(UNAME_M)),)
 	# Raspberry Pi 2
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 3, 4, Zero 2 (32-bit)
-	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+	MK_CFLAGS   += -mfp16-format=ieee -mno-unaligned-access
+	MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		CFLAGS   += -mcpu=power9
-		CXXFLAGS += -mcpu=power9
-	endif
-	# Require c++23's std::byteswap for big-endian support.
-	ifeq ($(UNAME_M),ppc64)
-		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
+		MK_CFLAGS   += -mcpu=power9
+		MK_CXXFLAGS += -mcpu=power9
 	endif
 endif
 
+else
+	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
+	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
+endif
+
 ifndef LLAMA_NO_K_QUANTS
-	CFLAGS   += -DGGML_USE_K_QUANTS
-	CXXFLAGS += -DGGML_USE_K_QUANTS
+	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
 	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
-	CFLAGS   += -DGGML_QKK_64
-	CXXFLAGS += -DGGML_QKK_64
+	MK_CPPFLAGS += -DGGML_QKK_64
 endif
 endif
 
 ifndef LLAMA_NO_ACCELERATE
-	# Mac M1 - include Accelerate framework.
-	# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
+	# Mac OS - include Accelerate framework.
+	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		CFLAGS  += -DGGML_USE_ACCELERATE
-		LDFLAGS += -framework Accelerate
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE
+		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
+		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
+		MK_LDFLAGS  += -framework Accelerate
 	endif
 endif # LLAMA_NO_ACCELERATE
 
 ifdef LLAMA_MPI
-	CFLAGS += -DGGML_USE_MPI -Wno-cast-qual
-	CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual
+	MK_CPPFLAGS += -DGGML_USE_MPI
+	MK_CFLAGS   += -Wno-cast-qual
+	MK_CXXFLAGS += -Wno-cast-qual
 	OBJS     += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags openblas)
-	LDFLAGS += $(shell pkg-config --libs openblas)
+	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
 ifdef LLAMA_BLIS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
-	LDFLAGS += -lblis -L/usr/local/lib
+	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS  += -lblis -L/usr/local/lib
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
+	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
@@ -253,6 +396,11 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+else
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
 #	NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
@@ -260,19 +408,20 @@ ifdef LLAMA_CUDA_CCBIN
 	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
 
-	CFLAGS   += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
-	CXXFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
+	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
+	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
 
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
-		LDFLAGS += -lclblast -framework OpenCL
+		MK_LDFLAGS += -lclblast -framework OpenCL
 	else
-		LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
+		MK_LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
 	endif
 	OBJS    += ggml-opencl.o
 
@@ -287,15 +436,13 @@ ifdef LLAMA_HIPBLAS
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
-	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
-	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
-	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
-	LDFLAGS		+= -lhipblas -lamdhip64 -lrocblas
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
 	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
-	HIPFLAGS    += -DCC_TURING=1000000000
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
@@ -305,10 +452,12 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 endif # LLAMA_HIPBLAS
 
 ifdef LLAMA_METAL
-	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
-	CXXFLAGS += -DGGML_USE_METAL
-	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJS     += ggml-metal.o
+	MK_CPPFLAGS += -DGGML_USE_METAL
+	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	OBJS		+= ggml-metal.o
+ifdef LLAMA_METAL_NDEBUG
+	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
+endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
@@ -321,24 +470,36 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifdef LLAMA_NO_K_QUANTS
+ifndef LLAMA_NO_K_QUANTS
 k_quants.o: k_quants.c k_quants.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_NO_K_QUANTS
 
+# combine build flags with cmdline overrides
+override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
+override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
+override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
+override LDFLAGS       := $(MK_LDFLAGS) $(LDFLAGS)
+
+# save CXXFLAGS before we add host-only options
+NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
+override CXXFLAGS += $(HOST_CXXFLAGS)
+
 #
 # Print build information
 #
 
 $(info I llama.cpp build info: )
-$(info I UNAME_S:  $(UNAME_S))
-$(info I UNAME_P:  $(UNAME_P))
-$(info I UNAME_M:  $(UNAME_M))
-$(info I CFLAGS:   $(CFLAGS))
-$(info I CXXFLAGS: $(CXXFLAGS))
-$(info I LDFLAGS:  $(LDFLAGS))
-$(info I CC:       $(CCV))
-$(info I CXX:      $(CXXV))
+$(info I UNAME_S:   $(UNAME_S))
+$(info I UNAME_P:   $(UNAME_P))
+$(info I UNAME_M:   $(UNAME_M))
+$(info I CFLAGS:    $(CFLAGS))
+$(info I CXXFLAGS:  $(CXXFLAGS))
+$(info I NVCCFLAGS: $(NVCCFLAGS))
+$(info I LDFLAGS:   $(LDFLAGS))
+$(info I CC:        $(shell $(CC) --version | head -n 1))
+$(info I CXX:       $(shell $(CXX) --version | head -n 1))
 $(info )
 
 #
@@ -351,12 +512,15 @@ ggml.o: ggml.c ggml.h ggml-cuda.h
 ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-OBJS += ggml-alloc.o
+ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h
+OBJS += ggml-alloc.o ggml-backend.o
+
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common.o: common/common.cpp common/common.h
+common.o: common/common.cpp common/common.h build-info.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 console.o: common/console.cpp common/console.h
@@ -365,11 +529,14 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+train.o: common/train.cpp common/train.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch convert-llama2c-to-ggml embd-input-test gguf llama-bench build-info.h $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
@@ -381,9 +548,15 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o common.o console.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -409,20 +582,43 @@ $(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-in
 embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
-gguf: examples/gguf/gguf.cpp                                  build-info.h ggml.o llama.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o common.o $(OBJS)
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp    build-info.h ggml.o llama.o $(OBJS)
+convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o train.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o common.o train.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+ifdef LLAMA_METAL
+metal: examples/metal/metal.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
-	@sh scripts/build-info.sh > $@.tmp
+	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
 	else \
@@ -437,40 +633,53 @@ tests: $(TEST_TARGETS)
 
 benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+run-benchmark-matmult: benchmark-matmult
 	./$@
 
+.PHONY: run-benchmark-matmult
+
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
+q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+
 tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o common.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-c.o: tests/test-c.c llama.h
+	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@

Package.swift

@@ -2,8 +2,33 @@
 
 import PackageDescription
 
+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v11),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let resources: [Resource] = [
+    .process("ggml-metal.metal")
+]
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let resources: [Resource] = []
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
 let package = Package(
     name: "llama",
+    platforms: platforms,
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -11,14 +36,29 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: ["ggml-metal.metal"],
-            sources: ["ggml.c", "llama.cpp"],
+            exclude: exclude,
+            sources: [
+                "ggml.c",
+                "llama.cpp",
+                "ggml-alloc.c",
+                "k_quants.c",
+            ] + additionalSources,
+            resources: resources,
             publicHeadersPath: "spm-headers",
-            cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],
+            cSettings: [
+                .unsafeFlags(["-Wno-shorten-64-to-32"]),
+                .define("GGML_USE_K_QUANTS"),
+                .define("GGML_USE_ACCELERATE")
+                // NOTE: NEW_LAPACK will required iOS version 16.4+
+                // We should consider add this in the future when we drop support for iOS 14
+                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+                // .define("ACCELERATE_NEW_LAPACK"),
+                // .define("ACCELERATE_LAPACK_ILP64")
+            ] + additionalSettings,
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]
-        ),
+        )
     ],
     cxxLanguageStandard: .cxx11
 )

README.md

@@ -5,27 +5,18 @@
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- #### IMPORTANT: Tokenizer fixes and API change (developers and projects using `llama.cpp` built-in tokenization must read): https://github.com/ggerganov/llama.cpp/pull/2810
+- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
+- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
+  **Devs should become familiar with the new API**
+- Local Falcon 180B inference on Mac Studio
 
-- GGUFv2 adds support for 64-bit sizes + backwards compatible: https://github.com/ggerganov/llama.cpp/pull/2821
-
-- Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717
-
-- A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398)
-
-  Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa)
-
-  ### Current `master` should be considered in Beta - expect some issues for a few days!
-
-  ### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up!
-
-  ### Issues with non-GGUF models will be considered with low priority!
+  https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
 
 ----
 
@@ -102,22 +93,28 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
 - [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
+- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
+- [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
+- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
 
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
+- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
+- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
 
 **UI:**
 
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
+- [withcatai/catai](https://github.com/withcatai/catai)
 
 ---
 
@@ -277,29 +274,11 @@ In order to build llama.cpp you have three different options.
 
 ### Metal Build
 
-Using Metal allows the computation to be executed on the GPU for Apple devices:
+On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
+To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or the `LLAMA_METAL=OFF` cmake option.
 
-- Using `make`:
-
-  ```bash
-  LLAMA_METAL=1 make
-  ```
-
-- Using `CMake`:
-
-    ```bash
-    mkdir build-metal
-    cd build-metal
-    cmake -DLLAMA_METAL=ON ..
-    cmake --build . --config Release
-    ```
-
-When built with Metal support, you can enable GPU inference with the `--gpu-layers|-ngl` command-line argument.
-Any value larger than 0 will offload the computation to the GPU. For example:
-
-```bash
-./main -m ./models/7B/ggml-model-q4_0.gguf -n 128 -ngl 1
-```
+When built with Metal support, you can explicitly disable GPU inference with the `--gpu-layers|-ngl 0` command-line
+argument.
 
 ### MPI Build
 
@@ -399,7 +378,7 @@ Building the program with BLAS support may lead to some performance improvements
 
 - #### cuBLAS
 
-  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+  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).
   - Using `make`:
     ```bash
     make LLAMA_CUBLAS=1
@@ -418,17 +397,18 @@ Building the program with BLAS support may lead to some performance improvements
 <!---
   | 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. |
-  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
-  | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | 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. |
+  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
+  | LLAMA_CUDA_F16                 | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       |     128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
 
 - #### hipBLAS
 
-  This provide BLAS acceleation on HIP supported GPU like AMD GPU.
+  This provides BLAS acceleration on HIP-supported AMD GPUs.
   Make sure to have ROCm installed.
   You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
   Windows support is coming soon...
@@ -462,6 +442,8 @@ Building the program with BLAS support may lead to some performance improvements
   You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
     - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed.
 
+    - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page.
+
     - <details>
         <summary>Installing the OpenCL SDK from source</summary>
 
@@ -479,10 +461,27 @@ Building the program with BLAS support may lead to some performance improvements
         ```
       </details>
 
-  Installing CLBlast: it may be found in your operating system's packages.
+  ##### Installing CLBlast
+
+  Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages.
+
+  Alternatively, they may be built from source.
 
   - <details>
-    <summary>If not, then installing from source:</summary>
+    <summary>Windows:</summary>
+
+      ```cmd
+      set OPENCL_SDK_ROOT="C:/OpenCL-SDK-v2023.04.17-Win-x64"
+      git clone https://github.com/CNugteren/CLBlast.git
+      mkdir CLBlast\build
+      cd CLBlast\build
+      cmake .. -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64
+      cmake --build . --config Release
+      cmake --install . --prefix C:/CLBlast
+      ```
+
+  - <details>
+    <summary>Unix:</summary>
 
       ```sh
       git clone https://github.com/CNugteren/CLBlast.git
@@ -496,21 +495,32 @@ Building the program with BLAS support may lead to some performance improvements
       Where `/some/path` is where the built library will be installed (default is `/usr/local`).
     </details>
 
-  Building:
+  ##### Building Llama with CLBlast
 
   - Build with make:
     ```sh
     make LLAMA_CLBLAST=1
     ```
-  - CMake:
+  - CMake (Unix):
     ```sh
     mkdir build
     cd build
-    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_dir=/some/path
+    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
     cmake --build . --config Release
     ```
+  - CMake (Windows):
+    ```cmd
+    set CL_BLAST_CMAKE_PKG="C:/CLBlast/lib/cmake/CLBlast"
+    git clone https://github.com/ggerganov/llama.cpp
+    cd llama.cpp
+    mkdir build
+    cd build
+    cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64
+    cmake --build . --config Release
+    cmake --install . --prefix C:/LlamaCPP
+    ```
 
-  Running:
+  ##### Running Llama with CLBlast
 
   The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
 
@@ -551,6 +561,10 @@ python3 convert.py models/7B/
 # quantize the model to 4-bits (using q4_0 method)
 ./quantize ./models/7B/ggml-model-f16.gguf ./models/7B/ggml-model-q4_0.gguf q4_0
 
+# update the gguf filetype to current if older version is unsupported by another application
+./quantize ./models/7B/ggml-model-q4_0.gguf ./models/7B/ggml-model-q4_0-v2.gguf COPY
+
+
 # run the inference
 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
 ```
@@ -587,6 +601,11 @@ Several quantization methods are supported. They differ in the resulting model d
 |   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
 |   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
 
+- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
+- recent k-quants improvements
+  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+
 ### Perplexity (measuring model quality)
 
 You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
@@ -595,6 +614,18 @@ For more information, see [https://huggingface.co/docs/transformers/perplexity](
 The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
 The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
 
+#### How to run
+
+1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
+3. Output:
+```
+perplexity : calculating perplexity over 655 chunks
+24.43 seconds per pass - ETA 4.45 hours
+[1]4.5970,[2]5.1807,[3]6.0382,...
+```
+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.
@@ -646,6 +677,8 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
 
 The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
 
+For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder
@@ -722,14 +755,12 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
 
 - Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
 - Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
-  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGML)
-  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGML)
-  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGML)
-  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML)
-  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML)
-  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML)
-- Specify `-eps 1e-5` for best generation quality
-- Specify `-gqa 8` for 70B models to work
+  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF)
+  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF)
+  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF)
+  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF)
+  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
+  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
 
 ### Verifying the model files
 
@@ -757,18 +788,6 @@ If your issue is with model generation quality, then please at least scan the fo
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
-#### How to run
-
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
-2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
-3. Output:
-```
-perplexity : calculating perplexity over 655 chunks
-24.43 seconds per pass - ETA 4.45 hours
-[1]4.5970,[2]5.1807,[3]6.0382,...
-```
-And after 4.45 hours, you will have the final perplexity.
-
 ### Android
 
 #### Building the Project using Android NDK
@@ -843,8 +862,17 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 #### Images
 We have two Docker images available for this project:
 
-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file.
+1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+
+Additionally, there the following images, similar to the above:
+
+- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
 #### Usage
 

build.zig

@@ -36,17 +36,23 @@ const Maker = struct {
     }
 
     fn init(builder: *std.build.Builder) !Maker {
-        const commit_hash = @embedFile(".git/refs/heads/master");
+        const target = builder.standardTargetOptions(.{});
+        const zig_version = @import("builtin").zig_version_string;
+        const commit_hash = try std.ChildProcess.exec(
+            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
+        );
         const config_header = builder.addConfigHeader(
             .{ .style = .blank, .include_path = "build-info.h" },
             .{
                 .BUILD_NUMBER = 0,
-                .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
+                .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline
+                .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}),
+                .BUILD_TARGET = try target.allocDescription(builder.allocator),
             },
         );
         var m = Maker{
             .builder = builder,
-            .target = builder.standardTargetOptions(.{}),
+            .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
             .config_header = config_header,
             .enable_lto = false,
@@ -58,19 +64,28 @@ const Maker = struct {
         try m.addCFlag("-std=c11");
         try m.addCxxFlag("-std=c++11");
         try m.addProjectInclude(&.{});
-        try m.addProjectInclude(&.{"examples"});
+        try m.addProjectInclude(&.{"common"});
         return m;
     }
 
     fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
+        if (o.target.getAbi() != .msvc)
+            o.defineCMacro("_GNU_SOURCE", null);
+        o.addConfigHeader(m.config_header);
         if (std.mem.endsWith(u8, src, ".c")) {
             o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
         } else {
             o.addCSourceFiles(&.{src}, m.cxxflags.items);
-            o.linkLibCpp();
+            if (o.target.getAbi() == .msvc) {
+                o.linkLibC(); // need winsdk + crt
+            } else {
+                // linkLibCpp already add (libc++ + libunwind + libc)
+                o.linkLibCpp();
+            }
         }
+        o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
         o.want_lto = m.enable_lto;
         return o;
@@ -82,8 +97,14 @@ const Maker = struct {
         for (deps) |d| e.addObject(d);
         for (m.objs.items) |o| e.addObject(o);
         for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
-        e.linkLibC();
-        e.linkLibCpp();
+
+        // https://github.com/ziglang/zig/issues/15448
+        if (e.target.getAbi() == .msvc) {
+            e.linkLibC(); // need winsdk + crt
+        } else {
+            // linkLibCpp already add (libc++ + libunwind + libc)
+            e.linkLibCpp();
+        }
         e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
         e.want_lto = m.enable_lto;
@@ -103,18 +124,21 @@ pub fn build(b: *std.build.Builder) !void {
 
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
+    const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
     const llama = make.obj("llama", "llama.cpp");
-    const common = make.obj("common", "examples/common.cpp");
-    const console = make.obj("common", "examples/console.cpp");
-    const grammar_parser = make.obj("grammar-parser", "examples/grammar-parser.cpp");
+    const common = make.obj("common", "common/common.cpp");
+    const console = make.obj("console", "common/console.cpp");
+    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
+    const train = make.obj("train", "common/train.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, llama, common });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, llama, common });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, grammar_parser });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

codecov.yml

@@ -0,0 +1,14 @@
+comment: off
+
+coverage:
+  status:
+    project:
+      default:
+        target: auto
+        threshold: 0
+        base: auto
+    patch:
+      default:
+        target: auto
+        threshold: 0
+        base: auto

common/CMakeLists.txt

@@ -9,6 +9,8 @@ add_library(${TARGET} OBJECT
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    train.h
+    train.cpp
     )
 
 if (BUILD_SHARED_LIBS)

common/common.cpp

@@ -24,7 +24,9 @@
 
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
-#define NOMINMAX
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
 #include <codecvt>
 #include <locale>
 #include <windows.h>
@@ -55,7 +57,7 @@ int32_t get_num_physical_cores() {
             siblings.insert(line);
         }
     }
-    if (siblings.size() > 0) {
+    if (!siblings.empty()) {
         return static_cast<int32_t>(siblings.size());
     }
 #elif defined(__APPLE__) && defined(__MACH__)
@@ -127,6 +129,15 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.n_threads <= 0) {
                 params.n_threads = std::thread::hardware_concurrency();
             }
+        } else if (arg == "-tb" || arg == "--threads-batch") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch = std::stoi(argv[i]);
+            if (params.n_threads_batch <= 0) {
+                params.n_threads_batch = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -156,8 +167,10 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+            // store the external file name in params
+            params.prompt_file = argv[i];
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
-            if (params.prompt.back() == '\n') {
+            if (!params.prompt.empty() && params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
         } else if (arg == "-n" || arg == "--n-predict") {
@@ -282,7 +295,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.cfg_negative_prompt));
-            if (params.cfg_negative_prompt.back() == '\n') {
+            if (!params.cfg_negative_prompt.empty() && params.cfg_negative_prompt.back() == '\n') {
                 params.cfg_negative_prompt.pop_back();
             }
         } else if (arg == "--cfg-scale") {
@@ -303,18 +316,42 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_keep = std::stoi(argv[i]);
+        } else if (arg == "--draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_draft = std::stoi(argv[i]);
         } else if (arg == "--chunks") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.n_chunks = std::stoi(argv[i]);
+        } else if (arg == "-np" || arg == "--parallel") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_parallel = std::stoi(argv[i]);
+        } else if (arg == "-ns" || arg == "--sequences") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_sequences = std::stoi(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "-md" || arg == "--model-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_draft = argv[i];
         } else if (arg == "-a" || arg == "--alias") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -326,7 +363,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter = argv[i];
+            params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f));
+            params.use_mmap = false;
+        } else if (arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i])));
             params.use_mmap = false;
         } else if (arg == "--lora-base") {
             if (++i >= argc) {
@@ -342,10 +391,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
+        } else if (arg == "--infill") {
+            params.infill = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
             params.simple_io = true;
+        } else if (arg == "-cb" || arg == "--cont-batching") {
+            params.cont_batching = true;
         } else if (arg == "--color") {
             params.use_color = true;
         } else if (arg == "--mlock") {
@@ -360,6 +413,17 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #else
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+        } else if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers_draft = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
         } else if (arg == "--main-gpu" || arg == "-mg") {
             if (++i >= argc) {
@@ -400,21 +464,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.mul_mat_q = false;
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
-#endif // GGML_USE_CUBLAS
-        } else if (arg == "--low-vram" || arg == "-lv") {
-#ifdef GGML_USE_CUBLAS
-            params.low_vram = true;
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
-        } else if (arg == "--mtest") {
-            params.mem_test = true;
         } else if (arg == "--numa") {
             params.numa = true;
-        } else if (arg == "--export") {
-            params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -433,8 +487,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.logdir.back() != DIRECTORY_SEPARATOR) {
                 params.logdir += DIRECTORY_SEPARATOR;
             }
-        } else if (arg == "--perplexity") {
-            params.perplexity = true;
+        } else if (arg == "--perplexity" || arg == "--all-logits") {
+            params.logits_all = true;
         } else if (arg == "--ppl-stride") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -480,6 +534,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, default_params);
+#ifndef LOG_DISABLE_LOGS
+            log_print_usage();
+#endif // LOG_DISABLE_LOGS
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -519,6 +576,25 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 std::istreambuf_iterator<char>(),
                 std::back_inserter(params.grammar)
             );
+#ifndef LOG_DISABLE_LOGS
+        // Parse args for logging parameters
+        } else if ( log_param_single_parse( argv[i] ) ) {
+            // Do nothing, log_param_single_parse automatically does it's thing
+            //  and returns if a match was found and parsed.
+        } else if ( log_param_pair_parse( /*check_but_dont_parse*/ true, argv[i] ) ) {
+            // We have a matching known parameter requiring an argument,
+            //  now we need to check if there is anything after this argv
+            //  and flag invalid_param or parse it.
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            if( !log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i-1], argv[i]) ) {
+                invalid_param = true;
+                break;
+            }
+        // End of Parse args for logging parameters
+#endif // LOG_DISABLE_LOGS
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);
@@ -542,112 +618,135 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
         process_escapes(params.input_suffix);
+        for (auto & antiprompt : params.antiprompt) {
+            process_escapes(antiprompt);
+        }
     }
 
     return true;
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stdout, "  -i, --interactive     run in interactive mode\n");
-    fprintf(stdout, "  --interactive-first   run in interactive mode and wait for input right away\n");
-    fprintf(stdout, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    fprintf(stdout, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
-    fprintf(stdout, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stdout, "                        halt generation at PROMPT, return control in interactive mode\n");
-    fprintf(stdout, "                        (can be specified more than once for multiple prompts).\n");
-    fprintf(stdout, "  --color               colorise output to distinguish prompt and user input from generations\n");
-    fprintf(stdout, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stdout, "  -p PROMPT, --prompt PROMPT\n");
-    fprintf(stdout, "                        prompt to start generation with (default: empty)\n");
-    fprintf(stdout, "  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stdout, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
-    fprintf(stdout, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
-    fprintf(stdout, "                        not supported with --interactive or other interactive options\n");
-    fprintf(stdout, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
-    fprintf(stdout, "  --random-prompt       start with a randomized prompt.\n");
-    fprintf(stdout, "  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
-    fprintf(stdout, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
-    fprintf(stdout, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
-    fprintf(stdout, "  -f FNAME, --file FNAME\n");
-    fprintf(stdout, "                        prompt file to start generation.\n");
-    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
-    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stdout, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
-    fprintf(stdout, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
-    fprintf(stdout, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
-    fprintf(stdout, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
-    fprintf(stdout, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
-    fprintf(stdout, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    fprintf(stdout, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
-    fprintf(stdout, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
-    fprintf(stdout, "  --mirostat N          use Mirostat sampling.\n");
-    fprintf(stdout, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    fprintf(stdout, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
-    fprintf(stdout, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
-    fprintf(stdout, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
-    fprintf(stdout, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
-    fprintf(stdout, "                        modifies the likelihood of token appearing in the completion,\n");
-    fprintf(stdout, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
-    fprintf(stdout, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
-    fprintf(stdout, "  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
-    fprintf(stdout, "  --grammar-file FNAME  file to read grammar from\n");
-    fprintf(stdout, "  --cfg-negative-prompt PROMPT\n");
-    fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
-    fprintf(stdout, "  --cfg-negative-prompt-file FNAME\n");
-    fprintf(stdout, "                        negative prompt file to use for guidance. (default: empty)\n");
-    fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
-    fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
-    fprintf(stdout, "  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
-    fprintf(stdout, "  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
-    fprintf(stdout, "  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
-    fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --interactive-first   run in interactive mode and wait for input right away\n");
+    printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
+    printf("  -r PROMPT, --reverse-prompt PROMPT\n");
+    printf("                        halt generation at PROMPT, return control in interactive mode\n");
+    printf("                        (can be specified more than once for multiple prompts).\n");
+    printf("  --color               colorise output to distinguish prompt and user input from generations\n");
+    printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+    printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
+    printf("  -tb N, --threads-batch N\n");
+    printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
+    printf("  -p PROMPT, --prompt PROMPT\n");
+    printf("                        prompt to start generation with (default: empty)\n");
+    printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    printf("  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
+    printf("  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
+    printf("                        not supported with --interactive or other interactive options\n");
+    printf("  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
+    printf("  --random-prompt       start with a randomized prompt.\n");
+    printf("  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
+    printf("  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    printf("  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
+    printf("  -f FNAME, --file FNAME\n");
+    printf("                        prompt file to start generation.\n");
+    printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    printf("  --mirostat N          use Mirostat sampling.\n");
+    printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
+    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
+    printf("                        modifies the likelihood of token appearing in the completion,\n");
+    printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
+    printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
+    printf("  --grammar-file FNAME  file to read grammar from\n");
+    printf("  --cfg-negative-prompt PROMPT\n");
+    printf("                        negative prompt to use for guidance. (default: empty)\n");
+    printf("  --cfg-negative-prompt-file FNAME\n");
+    printf("                        negative prompt file to use for guidance. (default: empty)\n");
+    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
+    printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
+    printf("  --no-penalize-nl      do not penalize newline token\n");
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
+    printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
+    printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
+    printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
+    printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
+    printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     if (llama_mlock_supported()) {
-        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported()) {
-        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
-    fprintf(stdout, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
-    fprintf(stdout, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ngld N, --n-gpu-layers-draft N\n");
+    printf("                        number of layers to store in VRAM for the draft model\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
 #ifdef GGML_USE_CUBLAS
-    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
-    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif // GGML_USE_CUBLAS
 #endif
-    fprintf(stdout, "  --mtest               compute maximum memory usage\n");
-    fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
-    fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
+    printf("  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
-    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "  -ld LOGDIR, --logdir LOGDIR\n");
-    fprintf(stdout, "                        path under which to save YAML logs (no logging if unset)\n");
-    fprintf(stdout, "\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -md FNAME, --model-draft FNAME\n");
+    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
+    printf("  -ld LOGDIR, --logdir LOGDIR\n");
+    printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("\n");
+}
+
+std::string get_system_info(const gpt_params & params) {
+    std::ostringstream os;
+
+    os << "system_info: n_threads = " << params.n_threads;
+    if (params.n_threads_batch != -1) {
+        os << " (n_threads_batch = " << params.n_threads_batch << ")";
+    }
+    os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+
+    return os.str();
 }
 
 std::string gpt_random_prompt(std::mt19937 & rng) {
@@ -663,58 +762,74 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
         case 7: return "He";
         case 8: return "She";
         case 9: return "They";
-        default: return "To";
     }
 
-    return "The";
+    GGML_UNREACHABLE();
 }
 
 //
 // Model utils
 //
 
+struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
+    auto mparams = llama_model_default_params();
+
+    if (params.n_gpu_layers != -1) {
+        mparams.n_gpu_layers = params.n_gpu_layers;
+    }
+    mparams.main_gpu        = params.main_gpu;
+    mparams.tensor_split    = params.tensor_split;
+    mparams.use_mmap        = params.use_mmap;
+    mparams.use_mlock       = params.use_mlock;
+
+    return mparams;
+}
+
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
-    auto lparams = llama_context_default_params();
+    auto cparams = llama_context_default_params();
 
-    lparams.n_ctx           = params.n_ctx;
-    lparams.n_batch         = params.n_batch;
-    lparams.n_gpu_layers    = params.n_gpu_layers;
-    lparams.main_gpu        = params.main_gpu;
-    lparams.tensor_split    = params.tensor_split;
-    lparams.low_vram        = params.low_vram;
-    lparams.mul_mat_q       = params.mul_mat_q;
-    lparams.seed            = params.seed;
-    lparams.f16_kv          = params.memory_f16;
-    lparams.use_mmap        = params.use_mmap;
-    lparams.use_mlock       = params.use_mlock;
-    lparams.logits_all      = params.perplexity;
-    lparams.embedding       = params.embedding;
-    lparams.rope_freq_base  = params.rope_freq_base;
-    lparams.rope_freq_scale = params.rope_freq_scale;
+    cparams.n_ctx           = params.n_ctx;
+    cparams.n_batch         = params.n_batch;
+    cparams.n_threads       = params.n_threads;
+    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q       = params.mul_mat_q;
+    cparams.seed            = params.seed;
+    cparams.f16_kv          = params.memory_f16;
+    cparams.logits_all      = params.logits_all;
+    cparams.embedding       = params.embedding;
+    cparams.rope_freq_base  = params.rope_freq_base;
+    cparams.rope_freq_scale = params.rope_freq_scale;
 
-    return lparams;
+    return cparams;
 }
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
-    auto lparams = llama_context_params_from_gpt_params(params);
+    auto mparams = llama_model_params_from_gpt_params(params);
 
-    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model  = llama_load_model_from_file(params.model.c_str(), mparams);
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
         return std::make_tuple(nullptr, nullptr);
     }
 
-    llama_context * lctx = llama_new_context_with_model(model, lparams);
+    auto cparams = llama_context_params_from_gpt_params(params);
+
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
         return std::make_tuple(nullptr, nullptr);
     }
 
-    if (!params.lora_adapter.empty()) {
+    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
+        const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]);
+        float lora_scale = std::get<1>(params.lora_adapter[i]);
         int err = llama_model_apply_lora_from_file(model,
-                                             params.lora_adapter.c_str(),
-                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             lora_adapter.c_str(),
+                                             lora_scale,
+                                             ((i > 0) || params.lora_base.empty())
+                                                ? NULL
+                                                : params.lora_base.c_str(),
                                              params.n_threads);
         if (err != 0) {
             fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
@@ -728,6 +843,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
     }
 
+    {
+        LOG("warming up the model with an empty run\n");
+
+        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_reset_timings(lctx);
+    }
+
     return std::make_tuple(model, lctx);
 }
 
@@ -736,16 +860,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 //
 
 std::vector<llama_token> llama_tokenize(
-        struct llama_context * ctx,
+  const struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_bos) {
+    return llama_tokenize(llama_get_model(ctx), text, add_bos);
+}
+
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_bos) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + add_bos;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(ctx, text.c_str(), result.data(), result.size(), add_bos);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(ctx, text.c_str(), result.data(), result.size(), add_bos);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -755,10 +886,10 @@ std::vector<llama_token> llama_tokenize(
 
 std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -797,9 +928,135 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
         result += piece;
     }
 
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
     return result;
 }
 
+//
+// Sampling utils
+//
+
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx) {
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+    const float   temp            = params.temp;
+    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float   top_p           = params.top_p;
+    const float   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
+    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    const float   repeat_penalty  = params.repeat_penalty;
+    const float   alpha_presence  = params.presence_penalty;
+    const float   alpha_frequency = params.frequency_penalty;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+    const bool    penalize_nl     = params.penalize_nl;
+
+    llama_token id = 0;
+
+    float * logits = llama_get_logits_ith(ctx, idx);
+
+    // Apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    candidates.clear();
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    }
+
+    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+
+    if (ctx_guidance) {
+        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+    }
+
+    // apply penalties
+    if (!last_tokens.empty()) {
+        const float nl_logit = logits[llama_token_nl(ctx)];
+        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
+
+        llama_sample_repetition_penalty(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, repeat_penalty);
+        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, alpha_frequency, alpha_presence);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
+                }
+            }
+        }
+    }
+
+    if (grammar != NULL) {
+        llama_sample_grammar(ctx, &cur_p, grammar);
+    }
+
+    if (temp <= 0) {
+        // Greedy sampling
+        id = llama_sample_token_greedy(ctx, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            const int mirostat_m = 100;
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+        } else if (mirostat == 2) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+        } else {
+            // Temperature sampling
+            size_t min_keep = std::max(1, params.n_probs);
+            llama_sample_top_k      (ctx, &cur_p, top_k, min_keep);
+            llama_sample_tail_free  (ctx, &cur_p, tfs_z, min_keep);
+            llama_sample_typical    (ctx, &cur_p, typical_p, min_keep);
+            llama_sample_top_p      (ctx, &cur_p, top_p, min_keep);
+            llama_sample_temp(ctx, &cur_p, temp);
+
+            {
+                const int n_top = 10;
+                LOG("top %d candidates:\n", n_top);
+
+                for (int i = 0; i < n_top; i++) {
+                    const llama_token id = cur_p.data[i].id;
+                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
+                }
+            }
+
+            id = llama_sample_token(ctx, &cur_p);
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+        }
+    }
+    // printf("`%d`", candidates_p.size);
+
+    if (grammar != NULL) {
+        llama_grammar_accept_token(ctx, grammar, id);
+    }
+
+    return id;
+}
+
+//
+// YAML utils
+//
+
 // returns true if successful, false otherwise
 bool create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -975,7 +1232,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 #endif // NDEBUG
 
     fprintf(stream, "model_desc: %s\n", model_desc);
-    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(lctx));
+    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
 
 #ifdef __OPTIMIZE__
     fprintf(stream, "optimize: true\n");
@@ -999,13 +1256,12 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
-    fprintf(stream, "export: %s # default: false\n", params.export_cgraph ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
     fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
     dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
     fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
-    fprintf(stream, "hellaswag_tasks: %ld # default: 400\n", params.hellaswag_tasks);
+    fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
     const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx));
     const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY;
@@ -1028,9 +1284,21 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
         fprintf(stream, "  %d: %f", lb.first, lb.second);
     }
 
-    fprintf(stream, "lora: %s\n", params.lora_adapter.c_str());
+    fprintf(stream, "lora:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) != 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
+    }
+    fprintf(stream, "lora_scaled:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) == 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
+    }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
-    fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);
@@ -1038,9 +1306,9 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta);
     fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
     fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
-    fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false");
+    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
     fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
-    fprintf(stream, "n_gpu_layers: %d # default: 0\n", params.n_gpu_layers);
+    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
     fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
@@ -1073,6 +1341,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
     fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
+    fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", params.temp);
 
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);

common/common.h

@@ -4,6 +4,9 @@
 
 #include "llama.h"
 
+#define LOG_NO_FILE_LINE_FUNCTION
+#include "log.h"
+
 #include <string>
 #include <vector>
 #include <random>
@@ -17,6 +20,14 @@
 #define DIRECTORY_SEPARATOR '/'
 #endif // _WIN32
 
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define print_build_info() do {                                                             \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+} while(0)
+
 //
 // CLI argument parsing
 //
@@ -25,18 +36,23 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                           = -1;   // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_predict                       = -1;   // new tokens to predict
     int32_t n_ctx                           = 512;  // context size
     int32_t n_batch                         = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
+    int32_t n_draft                         = 16;   // number of tokens to draft during speculative decoding
     int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
-    int32_t n_gpu_layers                    = 0;    // number of layers to store in VRAM
+    int32_t n_parallel                      = 1;    // number of parallel sequences to decode
+    int32_t n_sequences                     = 1;    // number of sequences to decode
+    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
     int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
     int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
-    float   rope_freq_base                  = 10000.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
+    float   rope_freq_base                  = 0.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
 
     // sampling parameters
     int32_t top_k             = 40;    // <= 0 to use vocab size
@@ -60,8 +76,10 @@ struct gpt_params {
     float       cfg_scale         = 1.f;   // How strong is guidance
 
     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
@@ -69,8 +87,8 @@ struct gpt_params {
     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 lora_adapter = "";  // lora adapter path
-    std::string lora_base    = "";  // base model path for the lora adapter
+    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
 
     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
@@ -79,7 +97,6 @@ struct gpt_params {
     bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
@@ -93,31 +110,36 @@ 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 input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool penalize_nl       = true;  // consider newlines as a repeatable token
-    bool perplexity        = false; // compute perplexity over the prompt
+    bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool mem_test          = false; // compute maximum memory usage
     bool numa              = false; // attempt optimizations that help on some NUMA systems
-    bool export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool infill            = false; // use infill mode
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
 
+std::string get_system_info(const gpt_params & params);
+
 std::string gpt_random_prompt(std::mt19937 & rng);
 
+void process_escapes(std::string& input);
+
 //
 // Model utils
 //
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_model_params   llama_model_params_from_gpt_params(const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
 //
@@ -127,7 +149,12 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 // tokenizes a string into a vector of tokens
 // should work similar to Python's `tokenizer.encode`
 std::vector<llama_token> llama_tokenize(
-        struct llama_context * ctx,
+  const struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_bos);
+
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_bos);
 
@@ -153,6 +180,40 @@ std::string llama_detokenize_bpe(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens);
 
+//
+// Sampling utils
+//
+
+// this is a common sampling function used across the examples for convenience
+// it can serve as a starting point for implementing your own sampling function
+//
+// required:
+//  - ctx:    context to use for sampling
+//  - params: sampling parameters
+//
+// optional:
+//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
+//  - grammar:       grammar to use for sampling, ignore if NULL
+//  - last_tokens:   needed for repetition penalty, ignore if empty
+//  - idx:           sample from llama_get_logits_ith(ctx, idx)
+//
+// returns:
+//  - token:      sampled token
+//  - candidates: vector of candidate tokens
+//
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx = 0);
+
+//
+// YAML utils
+//
+
 bool create_directory_with_parents(const std::string & path);
 void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
 void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);

common/console.cpp

@@ -158,7 +158,7 @@ namespace console {
         }
     }
 
-    char32_t getchar32() {
+    static char32_t getchar32() {
 #if defined(_WIN32)
         HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
         wchar_t high_surrogate = 0;
@@ -212,7 +212,7 @@ namespace console {
 #endif
     }
 
-    void pop_cursor() {
+    static void pop_cursor() {
 #if defined(_WIN32)
         if (hConsole != NULL) {
             CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
@@ -233,15 +233,16 @@ namespace console {
         putc('\b', out);
     }
 
-    int estimateWidth(char32_t codepoint) {
+    static int estimateWidth(char32_t codepoint) {
 #if defined(_WIN32)
+        (void)codepoint;
         return 1;
 #else
         return wcwidth(codepoint);
 #endif
     }
 
-    int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+    static int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
 #if defined(_WIN32)
         CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
         if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
@@ -302,7 +303,7 @@ namespace console {
 #endif
     }
 
-    void replace_last(char ch) {
+    static void replace_last(char ch) {
 #if defined(_WIN32)
         pop_cursor();
         put_codepoint(&ch, 1, 1);
@@ -311,7 +312,7 @@ namespace console {
 #endif
     }
 
-    void append_utf8(char32_t ch, std::string & out) {
+    static void append_utf8(char32_t ch, std::string & out) {
         if (ch <= 0x7F) {
             out.push_back(static_cast<unsigned char>(ch));
         } else if (ch <= 0x7FF) {
@@ -332,7 +333,7 @@ namespace console {
     }
 
     // Helper function to remove the last UTF-8 character from a string
-    void pop_back_utf8_char(std::string & line) {
+    static void pop_back_utf8_char(std::string & line) {
         if (line.empty()) {
             return;
         }
@@ -348,7 +349,7 @@ namespace console {
         line.erase(pos);
     }
 
-    bool readline_advanced(std::string & line, bool multiline_input) {
+    static bool readline_advanced(std::string & line, bool multiline_input) {
         if (out != stdout) {
             fflush(stdout);
         }
@@ -451,7 +452,7 @@ namespace console {
         return has_more;
     }
 
-    bool readline_simple(std::string & line, bool multiline_input) {
+    static bool readline_simple(std::string & line, bool multiline_input) {
 #if defined(_WIN32)
         std::wstring wline;
         if (!std::getline(std::wcin, wline)) {

common/grammar-parser.cpp

@@ -9,7 +9,7 @@
 namespace grammar_parser {
     // NOTE: assumes valid utf8 (but checks for overrun)
     // copied from llama.cpp
-    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
         static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
         uint8_t  first_byte = static_cast<uint8_t>(*src);
         uint8_t  highbits   = first_byte >> 4;
@@ -24,19 +24,19 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
         return result.first->second;
     }
 
-    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
         return next_id;
     }
 
-    void add_rule(
+    static void add_rule(
             parse_state & state,
             uint32_t      rule_id,
             const std::vector<llama_grammar_element> & rule) {
@@ -46,11 +46,11 @@ namespace grammar_parser {
         state.rules[rule_id] = rule;
     }
 
-    bool is_word_char(char c) {
+    static bool is_word_char(char c) {
         return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
     }
 
-    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
         const char * pos   = src;
         const char * end   = src + size;
         uint32_t     value = 0;
@@ -73,7 +73,7 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    const char * parse_space(const char * src, bool newline_ok) {
+    static const char * parse_space(const char * src, bool newline_ok) {
         const char * pos = src;
         while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
                 (newline_ok && (*pos == '\r' || *pos == '\n'))) {
@@ -88,7 +88,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_name(const char * src) {
+    static const char * parse_name(const char * src) {
         const char * pos = src;
         while (is_word_char(*pos)) {
             pos++;
@@ -99,7 +99,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    std::pair<uint32_t, const char *> parse_char(const char * src) {
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
         if (*src == '\\') {
             switch (src[1]) {
                 case 'x': return parse_hex(src + 2, 2);
@@ -129,7 +129,7 @@ namespace grammar_parser {
             uint32_t            rule_id,
             bool                is_nested);
 
-    const char * parse_sequence(
+    static const char * parse_sequence(
             parse_state                        & state,
             const char                         * src,
             const std::string                  & rule_name,
@@ -247,7 +247,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_rule(parse_state & state, const char * src) {
+    static const char * parse_rule(parse_state & state, const char * src) {
         const char * name_end = parse_name(src);
         const char * pos      = parse_space(name_end, false);
         size_t       name_len = name_end - src;
@@ -285,7 +285,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_grammar_char(FILE * file, uint32_t c) {
+    static void print_grammar_char(FILE * file, uint32_t c) {
         if (0x20 <= c && c <= 0x7f) {
             fprintf(file, "%c", static_cast<char>(c));
         } else {
@@ -294,7 +294,7 @@ namespace grammar_parser {
         }
     }
 
-    bool is_char_element(llama_grammar_element elem) {
+    static bool is_char_element(llama_grammar_element elem) {
         switch (elem.type) {
             case LLAMA_GRETYPE_CHAR:           return true;
             case LLAMA_GRETYPE_CHAR_NOT:       return true;
@@ -304,7 +304,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+    static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
         for (auto elem : rule) {
             switch (elem.type) {
                 case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
@@ -334,7 +334,7 @@ namespace grammar_parser {
         fprintf(file, "\n");
     }
 
-    void print_rule(
+    static void print_rule(
             FILE     * file,
             uint32_t   rule_id,
             const std::vector<llama_grammar_element> & rule,
@@ -415,6 +415,7 @@ namespace grammar_parser {
 
     std::vector<const llama_grammar_element *> parse_state::c_rules() {
         std::vector<const llama_grammar_element *> ret;
+        ret.reserve(rules.size());
         for (const auto & rule : rules) {
             ret.push_back(rule.data());
         }

common/log.h

@@ -0,0 +1,643 @@
+#pragma once
+
+#include <chrono>
+#include <cstring>
+#include <sstream>
+#include <iostream>
+#include <thread>
+#include <vector>
+#include <algorithm>
+#include <cinttypes>
+
+// --------------------------------
+//
+// Basic usage:
+//
+// --------
+//
+//  The LOG() and LOG_TEE() macros are ready to go by default
+//   they do not require any initialization.
+//
+//  LOGLN() and LOG_TEELN() are variants which automatically
+//   include \n character at the end of the log string.
+//
+//  LOG() behaves exactly like printf, by default writing to a logfile.
+//  LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
+//
+//  Default logfile is named
+//   "llama.<threadID>.log"
+//  Default LOG_TEE() secondary output target is
+//   stderr
+//
+//  Logs can be dynamically disabled or enabled using functions:
+//   log_disable()
+//  and
+//   log_enable()
+//
+//  A log target can be changed with:
+//   log_set_target( string )
+//    creating and opening, or re-opening a file by string filename
+//  or
+//   log_set_target( FILE* )
+//    allowing to point at stderr, stdout, or any valid FILE* file handler.
+//
+// --------
+//
+// End of Basic usage.
+//
+// --------------------------------
+
+// Specifies a log target.
+//  default uses log_handler() with "llama.log" log file
+//  this can be changed, by defining LOG_TARGET
+//  like so:
+//
+//  #define LOG_TARGET (a valid FILE*)
+//  #include "log.h"
+//
+//  or it can be simply redirected to stdout or stderr
+//  like so:
+//
+//  #define LOG_TARGET stderr
+//  #include "log.h"
+//
+//  The log target can also be redirected to a diffrent function
+//  like so:
+//
+//  #define LOG_TARGET log_handler_diffrent()
+//  #include "log.h"
+//
+//  FILE* log_handler_diffrent()
+//  {
+//      return stderr;
+//  }
+//
+//  or:
+//
+//  #define LOG_TARGET log_handler_another_one("somelog.log")
+//  #include "log.h"
+//
+//  FILE* log_handler_another_one(char*filename)
+//  {
+//      static FILE* logfile = nullptr;
+//      (...)
+//      if( !logfile )
+//      {
+//          fopen(...)
+//      }
+//      (...)
+//      return logfile
+//  }
+//
+#ifndef LOG_TARGET
+    #define LOG_TARGET log_handler()
+#endif
+
+#ifndef LOG_TEE_TARGET
+    #define LOG_TEE_TARGET stderr
+#endif
+
+// Utility to obtain "pid" like unique process id and use it when creating log files.
+inline std::string log_get_pid()
+{
+    static std::string pid;
+    if (pid.empty())
+    {
+        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+        //  it's not the same as "pid" but is unique enough to solve multiple instances
+        //  trying to write to the same log.
+        std::stringstream ss;
+        ss << std::this_thread::get_id();
+        pid = ss.str();
+    }
+
+    return pid;
+}
+
+// Utility function for generating log file names with unique id based on thread id.
+//  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
+//  where the number is a runtime id of the current thread.
+
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+
+// INTERNAL, DO NOT USE
+inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+{
+    std::stringstream buf;
+
+    buf << log_file_basename;
+    buf << ".";
+    buf << log_get_pid();
+    buf << ".";
+    buf << log_file_extension;
+
+    return buf.str();
+}
+
+#ifndef LOG_DEFAULT_FILE_NAME
+    #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
+#endif
+
+// Utility for turning #define values into string literals
+//  so we can have a define for stderr and
+//  we can print "stderr" instead of literal stderr, etc.
+#define LOG_STRINGIZE1(s) #s
+#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
+
+#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
+
+// Allows disabling timestamps.
+//  in order to disable, define LOG_NO_TIMESTAMPS
+//  like so:
+//
+//  #define LOG_NO_TIMESTAMPS
+//  #include "log.h"
+//
+#ifndef LOG_NO_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TIMESTAMP_FMT "%s"
+    #define LOG_TIMESTAMP_VAL ,""
+#endif
+
+#ifdef LOG_TEE_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TEE_TIMESTAMP_FMT "%s"
+    #define LOG_TEE_TIMESTAMP_VAL ,""
+#endif
+
+// Allows disabling file/line/function prefix
+//  in order to disable, define LOG_NO_FILE_LINE_FUNCTION
+//  like so:
+//
+//  #define LOG_NO_FILE_LINE_FUNCTION
+//  #include "log.h"
+//
+#ifndef LOG_NO_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_FLF_FMT "%s"
+    #define LOG_FLF_VAL ,""
+#endif
+
+#ifdef LOG_TEE_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_TEE_FLF_FMT "%s"
+    #define LOG_TEE_FLF_VAL ,""
+#endif
+
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
+// INTERNAL, DO NOT USE
+//  USE LOG() INSTEAD
+//
+#ifndef _MSC_VER
+    #define LOG_IMPL(str, ...)                                                                                      \
+    do {                                                                                                            \
+        if (LOG_TARGET != nullptr)                                                                                  \
+        {                                                                                                           \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                     \
+        }                                                                                                           \
+    } while (0)
+#else
+    #define LOG_IMPL(str, ...)                                                                                           \
+    do {                                                                                                                 \
+        if (LOG_TARGET != nullptr)                                                                                       \
+        {                                                                                                                \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                          \
+        }                                                                                                                \
+    } while (0)
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG_TEE() INSTEAD
+//
+#ifndef _MSC_VER
+    #define LOG_TEE_IMPL(str, ...)                                                                                                      \
+    do {                                                                                                                                \
+        if (LOG_TARGET != nullptr)                                                                                                      \
+        {                                                                                                                               \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                         \
+        }                                                                                                                               \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                         \
+        {                                                                                                                               \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                     \
+        }                                                                                                                               \
+    } while (0)
+#else
+    #define LOG_TEE_IMPL(str, ...)                                                                                                           \
+    do {                                                                                                                                     \
+        if (LOG_TARGET != nullptr)                                                                                                           \
+        {                                                                                                                                    \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                              \
+        }                                                                                                                                    \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                              \
+        {                                                                                                                                    \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                          \
+        }                                                                                                                                    \
+    } while (0)
+#endif
+
+// The '\0' as a last argument, is a trick to bypass the silly
+//  "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
+//  so we can have a single macro which can be called just like printf.
+
+// Main LOG macro.
+//  behaves like printf, and supports arguments the exact same way.
+//
+#ifndef _MSC_VER
+    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// Main TEE macro.
+//  does the same as LOG
+//  and
+//  simultaneously writes stderr.
+//
+// Secondary target can be changed just like LOG_TARGET
+//  by defining LOG_TEE_TARGET
+//
+#ifndef _MSC_VER
+    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// LOG macro variants with auto endline.
+#ifndef _MSC_VER
+    #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")
+#endif
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+{
+    static bool _initialized{false};
+    static bool _disabled{(filename.empty() && target == nullptr)};
+    static std::string log_current_filename{filename};
+    static FILE *log_current_target{target};
+    static FILE *logfile = nullptr;
+
+    if (change)
+    {
+        if (disable == LogTriStateTrue)
+        {
+            // Disable primary target
+            _disabled = true;
+        }
+        // If previously disabled, only enable, and keep previous target
+        else if (disable == LogTriStateFalse)
+        {
+            _disabled = false;
+        }
+        // Otherwise, process the arguments
+        else if (log_current_filename != filename || log_current_target != target)
+        {
+            _initialized = false;
+        }
+    }
+
+    if (_disabled)
+    {
+        // Log is disabled
+        return nullptr;
+    }
+
+    if (_initialized)
+    {
+        // with fallback in case something went wrong
+        return logfile ? logfile : stderr;
+    }
+
+    // do the (re)initialization
+    if (target != nullptr)
+    {
+        if (logfile != nullptr && logfile != stdout && logfile != stderr)
+        {
+            fclose(logfile);
+        }
+
+        log_current_filename = LOG_DEFAULT_FILE_NAME;
+        log_current_target = target;
+
+        logfile = target;
+    }
+    else
+    {
+        if (log_current_filename != filename)
+        {
+            if (logfile != nullptr && logfile != stdout && logfile != stderr)
+            {
+                fclose(logfile);
+            }
+        }
+
+        logfile = fopen(filename.c_str(), "w");
+    }
+
+    if (!logfile)
+    {
+        //  Verify whether the file was opened, otherwise fallback to stderr
+        logfile = stderr;
+
+        fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
+        fflush(stderr);
+
+        // At this point we let the init flag be to true below, and let the target fallback to stderr
+        //  otherwise we would repeatedly fopen() which was already unsuccessful
+    }
+
+    _initialized = true;
+
+    return logfile ? logfile : stderr;
+}
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+{
+    return log_handler1_impl(change, disable, filename, target);
+}
+
+// Disables logs entirely at runtime.
+//  Makes LOG() and LOG_TEE() produce no output,
+//  untill enabled back.
+#define log_disable() log_disable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_disable_impl()
+{
+    return log_handler1_impl(true, LogTriStateTrue);
+}
+
+// Enables logs at runtime.
+#define log_enable() log_enable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_enable_impl()
+{
+    return log_handler1_impl(true, LogTriStateFalse);
+}
+
+// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
+#define log_set_target(target) log_set_target_impl(target)
+
+// INTERNAL, DO NOT USE
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler() { return log_handler1_impl(); }
+
+inline void log_test()
+{
+    log_disable();
+    LOG("01 Hello World to nobody, because logs are disabled!\n");
+    log_enable();
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
+    log_set_target(stderr);
+    LOG("04 Hello World to stderr!\n");
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("06 Hello World to default log file!\n");
+    log_set_target(stdout);
+    LOG("07 Hello World to stdout!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("08 Hello World to default log file again!\n");
+    log_disable();
+    LOG("09 Hello World _1_ into the void!\n");
+    log_enable();
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
+    log_disable();
+    log_set_target("llama.anotherlog.log");
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
+    log_enable();
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
+    log_set_target("llama.yetanotherlog.log");
+    LOG("13 Hello World this time in yet new file?\n");
+    log_set_target(log_filename_generator("llama_autonamed", "log"));
+    LOG("14 Hello World in log with generated filename!\n");
+#ifdef _MSC_VER
+    LOG_TEE("15 Hello msvc TEE without arguments\n");
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n");
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
+    LOG("19 Hello msvc LOG without arguments\n");
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
+    LOGLN("21 Hello msvc LOGLN without arguments\n");
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
+#endif
+}
+
+inline bool log_param_single_parse(const std::string & param)
+{
+    if ( param == "--log-test")
+    {
+        log_test();
+        return true;
+    }
+
+    if ( param == "--log-disable")
+    {
+        log_disable();
+        return true;
+    }
+
+    if ( param == "--log-enable")
+    {
+        log_enable();
+        return true;
+    }
+
+    return false;
+}
+
+inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
+{
+    if ( param == "--log-file")
+    {
+        if (!check_but_dont_parse)
+        {
+            log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
+inline void log_print_usage()
+{
+    printf("log options:\n");
+    /* format
+    printf("  -h, --help            show this help message and exit\n");*/
+    /* spacing
+    printf("__-param----------------Description\n");*/
+    printf("  --log-test            Run simple logging test\n");
+    printf("  --log-disable         Disable trace logs\n");
+    printf("  --log-enable          Enable trace logs\n");
+    printf("  --log-file            Specify a log filename (without extension)\n");
+    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+}
+
+#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
+
+// INTERNAL, DO NOT USE
+inline void log_dump_cmdline_impl(int argc, char **argv)
+{
+    std::stringstream buf;
+    for (int i = 0; i < argc; ++i)
+    {
+        if (std::string(argv[i]).find(' ') != std::string::npos)
+        {
+            buf << " \"" << argv[i] <<"\"";
+        }
+        else
+        {
+            buf << " " << argv[i];
+        }
+    }
+    LOGLN("Cmd:%s", buf.str().c_str());
+}
+
+#define log_tostr(var) log_var_to_string_impl(var).c_str()
+
+inline std::string log_var_to_string_impl(bool var)
+{
+    return var ? "true" : "false";
+}
+
+inline std::string log_var_to_string_impl(std::string var)
+{
+    return var;
+}
+
+inline std::string log_var_to_string_impl(const std::vector<int> & var)
+{
+    std::stringstream buf;
+    buf << "[ ";
+    bool first = true;
+    for (auto e : var)
+    {
+        if (first)
+        {
+            first = false;
+        }
+        else
+        {
+            buf << ", ";
+        }
+        buf << std::to_string(e);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
+    [&tokens, &ctx]()                                                        \
+    {                                                                        \
+        std::stringstream buf;                                               \
+        buf << "[ ";                                                         \
+                                                                             \
+        bool first = true;                                                   \
+        for (const auto &token : tokens)                                     \
+        {                                                                    \
+            if (!first)                                                      \
+                buf << ", ";                                                 \
+            else                                                             \
+                first = false;                                               \
+                                                                             \
+            auto detokenized = llama_token_to_piece(ctx, token);             \
+                                                                             \
+            detokenized.erase(                                               \
+                std::remove_if(                                              \
+                    detokenized.begin(),                                     \
+                    detokenized.end(),                                       \
+                    [](const unsigned char c) { return !std::isprint(c); }), \
+                detokenized.end());                                          \
+                                                                             \
+            buf                                                              \
+                << "'" << detokenized << "'"                                 \
+                << ":" << std::to_string(token);                             \
+        }                                                                    \
+        buf << " ]";                                                         \
+                                                                             \
+        return buf.str();                                                    \
+    }()                                                                      \
+        .c_str()
+
+#ifdef LOG_DISABLE_LOGS
+
+#undef LOG
+#define LOG(...) // dummy stub
+#undef LOGLN
+#define LOGLN(...) // dummy stub
+
+#undef LOG_TEE
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_TEELN
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_DISABLE
+#define LOG_DISABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_SET_TARGET
+#define LOG_SET_TARGET(...) // dummy stub
+
+#undef LOG_DUMP_CMDLINE
+#define LOG_DUMP_CMDLINE(...) // dummy stub
+
+#endif // LOG_DISABLE_LOGS

common/train.h

@@ -0,0 +1,230 @@
+// Various helper functions and utilities for training
+
+#pragma once
+
+#include <string>
+#include <random>
+#include <vector>
+
+#include "ggml.h"
+#include "llama.h"
+
+typedef std::string mt19937_state;
+
+struct train_state {
+    struct ggml_opt_context * opt;
+
+    uint64_t train_its;
+    uint64_t train_samples;
+    uint64_t train_tokens;
+    uint64_t train_epochs;
+
+    size_t        shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
+    mt19937_state shuffle_rng_state_current;
+    mt19937_state shuffle_rng_state_next;
+    size_t        shuffle_sample_count;
+    size_t        shuffle_next_sample;
+};
+
+struct train_params_common {
+    const char * fn_train_data;
+    const char * fn_checkpoint_in;
+    const char * fn_checkpoint_out;
+    const char * pattern_fn_it;
+    const char * fn_latest;
+
+    bool print_usage;
+
+    int save_every;
+
+    uint32_t seed;
+
+    int n_ctx;
+    int n_threads;
+    int n_batch;
+    int n_gradient_accumulation;
+    int n_epochs;
+
+    bool custom_n_ctx;
+
+    bool use_flash;
+    bool use_checkpointing;
+
+    std::string sample_start;
+    bool include_sample_start;
+    bool escape;
+    bool overlapping_samples;
+    bool fill_with_next_samples;
+    bool separate_with_eos;
+    bool separate_with_bos;
+    bool sample_random_offsets;
+
+    bool force_reshuffle;
+
+    int   warmup;
+    int   cos_decay_steps;
+    float cos_decay_restart;
+    float cos_decay_min;
+    bool  enable_restart;
+
+    int   opt_past;
+    float opt_delta;
+    int   opt_max_no_improvement;
+
+    int   adam_n_iter;
+    float adam_alpha;
+    float adam_min_alpha;
+    float adam_decay;
+    int   adam_decay_min_ndim;
+    float adam_beta1;
+    float adam_beta2;
+    float adam_gclip;
+    float adam_eps_f;
+};
+
+typedef void (*save_train_files_callback)(void * data, struct train_state * train);
+
+struct train_opt_callback_data {
+    struct train_params_common * params;
+    struct train_state         * train;
+    save_train_files_callback    save_cb;
+    void                       * save_data;
+    struct llama_context       * lctx;
+    int                          last_save_iter;
+    llama_token                * tokens_data;
+    size_t                       tokens_size;
+    size_t                     * samples_begin;
+    size_t                     * samples_size;
+    size_t                     * shuffled_samples_offs;
+    size_t                     * shuffled_samples_begin;
+    size_t                     * shuffled_samples_size;
+    size_t                       samples_count;
+    struct ggml_tensor         * tokens_input;
+    struct ggml_tensor         * target_probs;
+    int                          first_iter;
+    int                          first_epoch;
+    int                          iter_at_last_epoch;
+    int64_t                      last_time;
+    double                       millis_per_iter;
+};
+
+struct train_state * init_train_state();
+void free_train_state(struct train_state  * state);
+
+struct train_params_common get_default_train_params_common();
+void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
+
+bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
+void finish_processing_train_args(struct train_params_common * params);
+
+struct random_normal_distribution;
+struct random_uniform_distribution;
+
+struct random_normal_distribution  * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
+struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
+
+void free_random_normal_distribution (struct random_normal_distribution  * rnd);
+void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
+
+struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
+struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
+
+// generate random float in interval [0,1)
+float frand();
+float frand_normal (struct random_normal_distribution * rnd);
+float frand_uniform(struct random_uniform_distribution * rnd);
+
+int   clamp (const int v, const int min, const int max);
+float fclamp(const float v, const float min, const float max);
+
+void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
+void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
+void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
+void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
+
+size_t tokenize_file(
+        struct llama_context     * lctx,
+        const char               * filename,
+        const std::string        & sample_start,
+        bool                       include_sample_start,
+        bool                       overlapping_samples,
+        unsigned                   context_length,
+        std::vector<llama_token> & out_tokens,
+        std::vector<size_t>      & out_samples_begin,
+        std::vector<size_t>      & out_samples_size);
+
+int64_t get_example_targets_batch(
+        struct llama_context * lctx,
+        struct ggml_tensor   * tokens_input,
+        struct ggml_tensor   * target_probs,
+        int64_t                example_id,
+        const size_t         * samples_offs,
+        const size_t         * samples_begin,
+        const size_t         * samples_size,
+              size_t           samples_count,
+        const llama_token    * train_data,
+        size_t                 n_train_data,
+        bool                   separate_with_eos,
+        bool                   separate_with_bos,
+        bool                   fill_with_next_samples,
+        bool                   sample_random_offsets);
+
+
+void          mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
+mt19937_state mt19937_get_state(const std::mt19937& rng);
+mt19937_state mt19937_seed_to_state(unsigned seed);
+
+mt19937_state shuffle_samples(
+        const mt19937_state & rng_state,
+        size_t              * shuffled_offs,
+        size_t              * shuffled_begins,
+        size_t              * shuffled_sizes,
+        const size_t        * begins,
+        const size_t        * sizes,
+        size_t                count);
+
+size_t hash_combine(size_t h1, size_t h2);
+
+size_t compute_samples_hash(
+    const char* fn,
+    const size_t* samples_begin,
+    const size_t* samples_size,
+    size_t sample_count);
+
+
+std::string replace_str(const char * s, const char * needle, const char * replacement);
+
+void print_duration(double milliseconds);
+
+float cosine_decay(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum);
+
+float cosine_decay_restart(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum,
+    float   restart_step_mult);
+
+float learning_schedule(
+    int64_t step,
+    int64_t warmup_steps,
+    int64_t decay_steps,
+    float   learning_rate,
+    float   overall_minimum,
+    float   cos_decay_minimum,
+    float   cos_decay_restart_step_mult,
+    bool    enable_restart);
+
+void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
+
+void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
+void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
+
+bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
+void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
+
+std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
+
+void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);

convert-baichuan-hf-to-gguf.py

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

convert-falcon-hf-to-gguf.py

@@ -1,45 +1,30 @@
 #!/usr/bin/env python3
 # HF falcon--> gguf conversion
 
-import gguf
-import os
-import sys
-import struct
+from __future__ import annotations
+
+import argparse
+import contextlib
 import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
 import numpy as np
 import torch
+from transformers import AutoTokenizer  # type: ignore[import]
 
-from typing import Any, List
-from pathlib import Path
-from transformers import AutoTokenizer
-
-def bytes_to_unicode():
-    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path, prefix: str) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
+        if filename.startswith(prefix):
             num_parts += 1
 
     if num_parts > 0:
@@ -47,17 +32,34 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -65,35 +67,37 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
-if hparams["architectures"][0] != "RWForCausalLM":
+if hparams["architectures"][0] != "FalconForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
-    sys.exit()
+    sys.exit(1)
 
 # get number of model parts
-num_parts = count_model_parts(dir_model)
+num_parts = count_model_parts(dir_model, "model-00")
+if num_parts:
+    is_safetensors = True
+    from safetensors import safe_open
+else:
+    is_safetensors = False
+    num_parts = count_model_parts(dir_model, "pytorch_model-")
 
 ARCH=gguf.MODEL_ARCH.FALCON
 gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
-block_count = hparams["n_layer"]
+block_count = hparams["num_hidden_layers"]
 
 gguf_writer.add_name("Falcon")
 gguf_writer.add_context_length(2048) # not in config.json
@@ -101,9 +105,9 @@ gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
-if "n_head_kv" in hparams:
-    gguf_writer.add_head_count_kv(hparams["n_head_kv"])
+gguf_writer.add_head_count(hparams["num_attention_heads"])
+if "num_kv_heads" in hparams:
+    gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
 else:
     gguf_writer.add_head_count_kv(1)
 gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
@@ -113,85 +117,44 @@ gguf_writer.add_file_type(ftype)
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[str] = []
-scores: List[float] = []
-toktypes: List[int] = []
-merges: List[str] = []
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
 
-if Path(dir_model + "/tokenizer.json").is_file():
-    # gpt2 tokenizer
-    gguf_writer.add_tokenizer_model("gpt2")
+print("gguf: get gpt2 tokenizer vocab")
 
-    print("gguf: get gpt2 tokenizer merges")
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
 
-    gguf_writer.add_token_merges(merges)
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
-    print("gguf: get gpt2 tokenizer vocab")
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i])
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
 
-    vocab_size = len(tokenizer_json["model"]["vocab"])
-
-    # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-    tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-    reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-    byte_encoder = bytes_to_unicode()
-    byte_decoder = {v: k for k, v in byte_encoder.items()}
-
-    for i in range(vocab_size):
-        if i in reverse_vocab:
-            try:
-                text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-            except KeyError:
-                text = bytearray()
-                for c in reverse_vocab[i]:
-                    if ord(c) < 256:  # single byte character
-                        text.append(byte_decoder[ord(c)])
-                    else:  # multibyte special token character
-                        text.extend(c.encode('utf-8'))
-        else:
-            print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-            pad_token = f"[PAD{i}]".encode("utf8")
-            text = bytearray(pad_token)
-
-        tokens.append(text)
-        scores.append(0.0)                      # dymmy
-        toktypes.append(gguf.TokenType.NORMAL)  # dummy
-
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-print("gguf: get special token ids")
-# Look for special tokens in config.json
-
-if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-    gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-    gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-    gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-    gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-    gguf_writer.add_pad_token_id(hparams["pad_token_id"])
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
 tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 
 # params for qkv transform
-n_head    = hparams["n_head"]
-n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+n_head    = hparams["num_attention_heads"]
+n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
 
 head_dim = hparams["hidden_size"] // n_head
 
@@ -199,81 +162,89 @@ head_dim = hparams["hidden_size"] // n_head
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
+elif is_safetensors:
+    part_names = (
+        f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1)
+    )
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+    if is_safetensors:
+        ctx = safe_open(dir_model / part_name, framework="pt", device="cpu")
+    else:
+        ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu"))
 
-    for name in model_part.keys():
-        data = model_part[name]
+    with ctx as model_part:
+        for name in model_part.keys():
+            data = model_part.get_tensor(name) if is_safetensors else model_part[name]
 
-        old_dtype = data.dtype
+            old_dtype = data.dtype
 
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
+            # convert any unsupported data types to float32
+            if data.dtype != torch.float16 and data.dtype != torch.float32:
+                data = data.to(torch.float32)
 
-        # QKV tensor transform
-        # The original query_key_value tensor contains n_head_kv "kv groups",
-        # each consisting of n_head/n_head_kv query weights followed by one key
-        # and one value weight (shared by all query heads in the kv group).
-        # This layout makes it a big pain to work with in GGML.
-        # So we rearrange them here,, so that we have n_head query weights
-        # followed by n_head_kv key weights followed by n_head_kv value weights,
-        # in contiguous fashion.
-        # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
+            # QKV tensor transform
+            # The original query_key_value tensor contains n_head_kv "kv groups",
+            # each consisting of n_head/n_head_kv query weights followed by one key
+            # and one value weight (shared by all query heads in the kv group).
+            # This layout makes it a big pain to work with in GGML.
+            # So we rearrange them here,, so that we have n_head query weights
+            # followed by n_head_kv key weights followed by n_head_kv value weights,
+            # in contiguous fashion.
+            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
 
-        if "query_key_value" in name:
-            qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
-            q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
-            k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
-            v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
-            data = torch.cat((q,k,v)).reshape_as(data)
+            if "query_key_value" in name:
+                qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+                q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
+                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                data = torch.cat((q,k,v)).reshape_as(data)
 
-        data = data.squeeze().numpy()
+            data = data.squeeze().numpy()
 
-        # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+            if new_name is None:
+                print("Can not map tensor '" + name + "'")
+                sys.exit()
 
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
 
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
+            # if f32 desired, convert any float16 to float32
+            if ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
 
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
 
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+            print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+            gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")

convert-gptneox-hf-to-gguf.py

@@ -1,44 +1,26 @@
 #!/usr/bin/env python3
 # HF gptneox--> gguf conversion
 
-import gguf
-import os
-import sys
-import struct
+from __future__ import annotations
+
+import argparse
 import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
 import numpy as np
 import torch
+from transformers import AutoTokenizer  # type: ignore[import]
 
-from typing import Any, List
-from pathlib import Path
-from transformers import AutoTokenizer
-
-# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
         if filename.startswith("pytorch_model-"):
@@ -49,17 +31,34 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -67,19 +66,15 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
 if hparams["architectures"][0] != "GPTNeoXForCausalLM":
@@ -97,7 +92,7 @@ print("gguf: get model metadata")
 
 block_count = hparams["num_hidden_layers"]
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name(dir_model.name)
 gguf_writer.add_context_length(hparams["max_position_embeddings"])
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
@@ -111,86 +106,36 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[str] = []
-merges: List[str] = []
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
 
-if Path(dir_model + "/tokenizer.json").is_file():
-    # gpt2 tokenizer
-    gguf_writer.add_tokenizer_model("gpt2")
+print("gguf: get gpt2 tokenizer vocab")
 
-    print("gguf: get gpt2 tokenizer merges")
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
 
-    gguf_writer.add_token_merges(merges)
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
-    print("gguf: get gpt2 tokenizer vocab")
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
 
-    vocab_size = len(tokenizer_json["model"]["vocab"])
-
-    # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-    tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-    reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-    byte_encoder = bytes_to_unicode()
-    byte_decoder = {v: k for k, v in byte_encoder.items()}
-
-    for i in range(vocab_size):
-        if i in reverse_vocab:
-            try:
-                text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-            except KeyError:
-                text = bytearray()
-                for c in reverse_vocab[i]:
-                    if ord(c) < 256:  # single byte character
-                        text.append(byte_decoder[ord(c)])
-                    else:  # multibyte special token character
-                        text.extend(c.encode('utf-8'))
-        else:
-            print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-            pad_token = f"[PAD{i}]".encode("utf8")
-            text = bytearray(pad_token)
-
-        tokens.append(text)
-
-    gguf_writer.add_token_list(tokens)
-
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
-        print("gguf: get special token ids")
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        # find special token ids
-
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
@@ -200,13 +145,15 @@ tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
     model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
 
@@ -226,11 +173,8 @@ for part_name in part_names:
         data = data.squeeze().numpy()
 
         # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
             print("Can not map tensor '" + name + "'")
             sys.exit()
 
@@ -249,19 +193,20 @@ for part_name in part_names:
         if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
             data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+        gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")

convert-llama-7b-pth-to-gguf.py

@@ -1,308 +0,0 @@
-#!/usr/bin/env python3
-# 7b pth llama --> gguf conversion
-# Only models with a single datafile are supported, like 7B
-# HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model
-
-import gguf
-import os
-import sys
-import struct
-import json
-import numpy as np
-import torch
-
-from typing import Any, List
-from pathlib import Path
-from sentencepiece import SentencePieceProcessor
-
-#NDArray = np.ndarray[Any, Any]
-# compatible with python < 3.9
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
-
-
-def count_model_parts(dir_model: str) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("consolidated."):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-
-    sys.exit(1)
-
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-
-        sys.exit(1)
-
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "LlamaForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-if num_parts > 1:
-    print("gguf: Only models with a single datafile are supported.")
-
-    sys.exit()
-
-ARCH=gguf.MODEL_ARCH.LLAMA
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-head_count = hparams["num_attention_heads"]
-
-if "num_key_value_heads" in hparams:
-    head_count_kv = hparams["num_key_value_heads"]
-else:
-    head_count_kv = head_count
-
-if "_name_or_path" in hparams:
-    hf_repo = hparams["_name_or_path"]
-else:
-    hf_repo = ""
-
-if "max_sequence_length" in hparams:
-    ctx_length = hparams["max_sequence_length"]
-elif "max_position_embeddings" in hparams:
-    ctx_length = hparams["max_position_embeddings"]
-else:
-    print("gguf: can not find ctx length parameter.")
-
-    sys.exit()
-
-
-gguf_writer.add_name(last_dir)
-gguf_writer.add_source_hf_repo(hf_repo)
-gguf_writer.add_tensor_data_layout("Meta AI original pth")
-gguf_writer.add_context_length(ctx_length)
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
-gguf_writer.add_head_count(head_count)
-gguf_writer.add_head_count_kv(head_count_kv)
-gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
-
-if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
-    if "type" in hparams["rope_scaling"]:
-        if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
-
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: List[bytes] = []
-scores: List[float] = []
-toktypes: List[int] = []
-
-if Path(dir_model + "/tokenizer.model").is_file():
-    # vocab type sentencepiece
-    print("gguf: get sentencepiece tokenizer vocab and scores")
-
-    tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1  # defualt to normal token type
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-
-        # toktype = 4 is user-defined = tokens from added_tokens.json
-
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-
-    if Path(dir_model + "/added_tokens.json").is_file():
-        with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
-            addtokens_json = json.load(f)
-
-            print("gguf: get added tokens")
-
-            for key in addtokens_json:
-                tokens.append( key.encode("utf-8") )
-                scores.append(-1000.0)
-                toktypes.append(4) # user-defined token type
-
-    gguf_writer.add_tokenizer_model("llama")
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-
-print("gguf: get special token ids")
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts))
-
-for part_name in part_names:
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name == "rope.freqs":
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-
-print("gguf: model successfully exported to '" + fname_out + "'")
-print("")

convert-llama-ggml-to-gguf.py

@@ -1,9 +1,18 @@
 #!/usr/bin/env python3
-import sys, struct, math, argparse
+from __future__ import annotations
+
+import argparse
+import math
+import struct
+import sys
+from enum import IntEnum
 from pathlib import Path
 
 import numpy as np
 
+import os
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
 import gguf
 
 # Note: Does not support GGML_QKK_64
@@ -26,10 +35,35 @@ GGML_QUANT_SIZES = {
     gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8),
 }
 
+class GGMLFormat(IntEnum):
+    GGML = 0
+    GGMF = 1
+    GGJT = 2
+
+class GGMLFType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1
+    MOSTLY_Q4_0          = 2
+    MOSTLY_Q4_1          = 3
+    MOSTLY_Q4_1_SOME_F16 = 4
+    MOSTLY_Q8_0          = 7
+    MOSTLY_Q5_0          = 8
+    MOSTLY_Q5_1          = 9
+    MOSTLY_Q2_K          = 10
+    MOSTLY_Q3_K_S        = 11
+    MOSTLY_Q3_K_M        = 12
+    MOSTLY_Q3_K_L        = 13
+    MOSTLY_Q4_K_S        = 14
+    MOSTLY_Q4_K_M        = 15
+    MOSTLY_Q5_K_S        = 16
+    MOSTLY_Q5_K_M        = 17
+    MOSTLY_Q6_K          = 18
+
 class Hyperparameters:
     def __init__(self):
-        self.n_vocab = self.n_embd = self.n_mult = self.n_head = self.n_layer = self.n_rot = self.ftype = 0
-        self.n_ff = 0
+        self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
+        self.n_layer = self.n_rot = self.n_ff = 0
+        self.ftype = GGMLFType.ALL_F32
 
     def set_n_ff(self, model):
         ff_tensor_idx = model.tensor_map.get(b'layers.0.feed_forward.w1.weight')
@@ -45,16 +79,21 @@ class Hyperparameters:
             self.n_head,
             self.n_layer,
             self.n_rot,
-            self.ftype,
+            ftype,
         ) = struct.unpack('<7I', data[offset:offset + (4 * 7)])
+        try:
+            self.ftype = GGMLFType(ftype)
+        except ValueError:
+            raise ValueError(f'Invalid ftype {ftype}')
         return 4 * 7
 
     def __str__(self):
-        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype}>'
+        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'
 
 class Vocab:
-    def __init__(self):
+    def __init__(self, load_scores = True):
         self.items = []
+        self.load_scores = load_scores
 
     def load(self, data, offset, n_vocab):
         orig_offset = offset
@@ -62,20 +101,24 @@ class Vocab:
             itemlen = struct.unpack('<I', data[offset:offset + 4])[0]
             assert itemlen < 4096, 'Absurd vocab item length'
             offset += 4
-            vocab = bytes(data[offset:offset + itemlen])
+            item_text = bytes(data[offset:offset + itemlen])
             offset += itemlen
-            score = struct.unpack('<f', data[offset:offset + 4])[0]
-            offset += 4
-            self.items.append((vocab, score))
+            if self.load_scores:
+                item_score = struct.unpack('<f', data[offset:offset + 4])[0]
+                offset += 4
+            else:
+                item_score = 0.0
+            self.items.append((item_text, item_score))
         return offset - orig_offset
 
 class Tensor:
-    def __init__(self):
+    def __init__(self, use_padding = True):
         self.name = None
-        self.dims = ()
+        self.dims: tuple[int, ...] = ()
         self.dtype = None
         self.start_offset = 0
-        self.len_bytes = 0
+        self.len_bytes = np.int64(0)
+        self.use_padding = use_padding
 
     def load(self, data, offset):
         orig_offset = offset
@@ -91,7 +134,7 @@ class Tensor:
         offset += 4 * n_dims
         self.name = bytes(data[offset:offset + name_len])
         offset += name_len
-        pad = ((offset + 31) & ~31) - offset
+        pad = ((offset + 31) & ~31) - offset if self.use_padding else 0
         offset += pad
         n_elems = np.prod(self.dims)
         n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
@@ -101,7 +144,7 @@ class Tensor:
         # print(n_dims, name_len, dtype, self.dims, self.name, pad)
         return offset - orig_offset
 
-class GGMLV3Model:
+class GGMLModel:
     def __init__(self):
         self.hyperparameters = None
         self.vocab = None
@@ -109,20 +152,52 @@ class GGMLV3Model:
         self.tensors = []
 
     def validate_header(self, data, offset):
-        if bytes(data[offset:offset + 4]) != b'tjgg' or struct.unpack('<I', data[offset + 4:offset + 8])[0] != 3:
-            raise ValueError('Only GGJTv3 supported')
-        return 8
+        magic = bytes(data[offset:offset + 4])
+        if magic == b'GGUF':
+            raise ValueError('File is already in GGUF format.')
+        if magic == b'lmgg':
+            self.file_format = GGMLFormat.GGML
+            self.format_version = 1
+            return 4
+        version = struct.unpack('<I', data[offset + 4:offset + 8])[0]
+        if magic == b'fmgg':
+            if version != 1:
+                raise ValueError(f'Cannot handle unexpected GGMF file version {version}')
+            self.file_format = GGMLFormat.GGMF
+            self.format_version = version
+            return 8
+        if magic == b'tjgg':
+            if version < 1 or version > 3:
+                raise ValueError(f'Cannot handle unexpected GGJT file version {version}')
+            self.file_format = GGMLFormat.GGJT
+            self.format_version = version
+            return 8
+        raise ValueError(f"Unexpected file magic {magic!r}! This doesn't look like a GGML format file.")
+
+    def validate_conversion(self, ftype):
+        err = ''
+        if (self.file_format < GGMLFormat.GGJT or self.format_version < 2):
+            if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
+                err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
+        elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
+            if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                          GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+                err = 'Q4 and Q8 quantizations changed in GGJTv3.'
+        if len(err) > 0:
+            raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
 
     def load(self, data, offset):
         offset += self.validate_header(data, offset)
         hp = Hyperparameters()
         offset += hp.load(data, offset)
-        vocab = Vocab()
+        print(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
+        self.validate_conversion(hp.ftype)
+        vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
         offset += vocab.load(data, offset, hp.n_vocab)
-        tensors = []
+        tensors: list[Tensor] = []
         tensor_map = {}
         while offset < len(data):
-            tensor = Tensor()
+            tensor = Tensor(use_padding = self.file_format > GGMLFormat.GGMF)
             offset += tensor.load(data, offset)
             tensor_map[tensor.name] = len(tensors)
             tensors.append(tensor)
@@ -134,13 +209,14 @@ class GGMLV3Model:
         return offset
 
 class GGMLToGGUF:
-    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None):
+    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
         hp = ggml_model.hyperparameters
         self.model = ggml_model
         self.data = data
         self.cfg = cfg
         self.params_override = params_override
         self.vocab_override = vocab_override
+        self.special_vocab = special_vocab
         if params_override is not None:
             n_kv_head = params_override.n_head_kv
         else:
@@ -159,9 +235,14 @@ class GGMLToGGUF:
 
     def save(self):
         print('* Preparing to save GGUF file')
-        gguf_writer = gguf.GGUFWriter(self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+        gguf_writer = gguf.GGUFWriter(
+            self.cfg.output,
+            gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
+            use_temp_file = False )
         self.add_params(gguf_writer)
         self.add_vocab(gguf_writer)
+        if self.special_vocab is not None:
+            self.special_vocab.add_to_gguf(gguf_writer)
         self.add_tensors(gguf_writer)
         print("    gguf: write header")
         gguf_writer.write_header_to_file()
@@ -174,7 +255,10 @@ class GGMLToGGUF:
     def add_params(self, gguf_writer):
         hp = self.model.hyperparameters
         cfg = self.cfg
-        desc = cfg.desc if cfg.desc is not None else 'converted from legacy GGJTv3 format'
+        if cfg.desc is not None:
+            desc = cfg.desc
+        else:
+            desc = f'converted from legacy {self.model.file_format.name}v{self.model.format_version} {hp.ftype.name} format'
         try:
             # Filenames aren't necessarily valid UTF8.
             name = cfg.name if cfg.name is not None else cfg.input.name
@@ -184,6 +268,7 @@ class GGMLToGGUF:
         if name is not None:
             gguf_writer.add_name(name)
         gguf_writer.add_description(desc)
+        gguf_writer.add_file_type(int(hp.ftype))
         if self.params_override is not None:
             po = self.params_override
             assert po.n_embd == hp.n_embd, 'Model hyperparams mismatch'
@@ -220,7 +305,8 @@ class GGMLToGGUF:
                 tokens.append(vbytes)
                 scores.append(score)
                 toktypes.append(ttype)
-            assert len(tokens) == hp.n_vocab, f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
+            assert len(tokens) == hp.n_vocab, \
+                f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
             gguf_writer.add_token_list(tokens)
             gguf_writer.add_token_scores(scores)
             if len(toktypes) > 0:
@@ -259,27 +345,24 @@ class GGMLToGGUF:
         gguf_writer.add_eos_token_id(2)
 
     def add_tensors(self, gguf_writer):
-        nm = self.name_map
+        tensor_map = self.name_map
         data = self.data
         print(f'* Adding {len(self.model.tensors)} tensor(s)')
         for tensor in self.model.tensors:
             name = str(tensor.name, 'UTF-8')
-            if name.endswith('.weight'):
-                name = name[:-7]
-                suffix = '.weight'
-            elif name.endswith('.bias'):
-                name = name[:-5]
-                suffix = '.bias'
-            mapped_name = nm.get(name)
+            mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
             assert mapped_name is not None, f'Bad name {name}'
-            mapped_name += suffix
             tempdims = list(tensor.dims[:])
             if len(tempdims) > 1:
                 temp = tempdims[1]
                 tempdims[1] = tempdims[0]
                 tempdims[0] = temp
             # print(f'+ {tensor.name} | {mapped_name} {tensor.dims} :: {tempdims}')
-            gguf_writer.add_tensor(mapped_name, data[tensor.start_offset:tensor.start_offset + tensor.len_bytes], raw_shape = tempdims, raw_dtype = tensor.dtype)
+            gguf_writer.add_tensor(
+                mapped_name,
+                data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
+                raw_shape = tempdims,
+                raw_dtype = tensor.dtype )
 
 def handle_metadata(cfg, hp):
     import convert
@@ -301,43 +384,66 @@ def handle_metadata(cfg, hp):
         params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
     else:
         raise ValueError('Unable to load metadata')
-    vocab = convert.load_vocab(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, cfg.vocabtype)
+    vocab = convert.load_vocab(
+        cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
+        cfg.vocabtype )
+    # FIXME: Respect cfg.vocab_dir?
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
-    return (params, vocab)
+    return (params, vocab, svocab)
 
 def handle_args():
-    parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF')
-    parser.add_argument('--input', '-i', type = Path, help = 'Input GGMLv3 filename')
-    parser.add_argument('--output', '-o', type = Path, help ='Output GGUF filename')
-    parser.add_argument('--name', help = 'Set model name')
-    parser.add_argument('--desc', help = 'Set model description')
-    parser.add_argument('--gqa', type = int, default = 1, help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
-    parser.add_argument('--eps', default = '5.0e-06', help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
-    parser.add_argument('--context-length', '-c', type=int, default = 2048, help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
-    parser.add_argument('--model-metadata-dir', '-m', type = Path, help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
-    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"], help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)", default="spm")
+    parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
+    parser.add_argument('--input', '-i', type = Path, required = True,
+        help = 'Input GGMLv3 filename')
+    parser.add_argument('--output', '-o', type = Path, required = True,
+        help ='Output GGUF filename')
+    parser.add_argument('--name',
+        help = 'Set model name')
+    parser.add_argument('--desc',
+        help = 'Set model description')
+    parser.add_argument('--gqa', type = int, default = 1,
+        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+    parser.add_argument('--eps', default = '5.0e-06',
+        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+    parser.add_argument('--context-length', '-c', type=int, default = 2048,
+        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+    parser.add_argument('--model-metadata-dir', '-m', type = Path,
+        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+    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)")
     return parser.parse_args()
 
 def main():
     cfg = handle_args()
     print(f'* Using config: {cfg}')
     print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n')
+    if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
+        print('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
     data = np.memmap(cfg.input, mode = 'r')
-    model = GGMLV3Model()
+    model = GGMLModel()
     print('* Scanning GGML input file')
     offset = model.load(data, 0)
     print(f'* GGML model hyperparameters: {model.hyperparameters}')
     vocab_override = None
     params_override = None
+    special_vocab = None
     if cfg.model_metadata_dir is not None:
-        (params_override, vocab_override) = handle_metadata(cfg, model.hyperparameters)
+        (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
         print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
         print(f'* Overriding params: {params_override}')
         print(f'* Overriding vocab: {vocab_override}')
+        print(f'* Special vocab: {special_vocab}')
     else:
         print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
-    converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override)
+        if model.file_format == GGMLFormat.GGML:
+            print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
+    converter = GGMLToGGUF(model, data, cfg,
+        params_override = params_override,
+        vocab_override = vocab_override,
+        special_vocab = special_vocab )
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 

convert-llama-hf-to-gguf.py

@@ -1,328 +0,0 @@
-#!/usr/bin/env python3
-# HF llama --> gguf conversion
-
-import gguf
-import os
-import sys
-import struct
-import json
-import numpy as np
-import torch
-
-from typing import Any, List, Optional
-from pathlib import Path
-from sentencepiece import SentencePieceProcessor
-
-#NDArray = np.ndarray[Any, Any]
-# compatible with python < 3.9
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
-
-# reverse HF permute back to original pth layout
-# https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py
-
-
-def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray:
-    if n_kv_head is not None and n_head != n_kv_head:
-        n_head //= n_kv_head
-
-    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-            .swapaxes(1, 2)
-            .reshape(weights.shape))
-
-
-def count_model_parts(dir_model: str) -> int:
-    num_parts = 0
-
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-
-    return num_parts
-
-
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-
-    sys.exit(1)
-
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-
-        sys.exit(1)
-
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "LlamaForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.LLAMA
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-head_count = hparams["num_attention_heads"]
-
-if "num_key_value_heads" in hparams:
-    head_count_kv = hparams["num_key_value_heads"]
-else:
-    head_count_kv = head_count
-
-if "_name_or_path" in hparams:
-    hf_repo = hparams["_name_or_path"]
-else:
-    hf_repo = ""
-
-if "max_sequence_length" in hparams:
-    ctx_length = hparams["max_sequence_length"]
-elif "max_position_embeddings" in hparams:
-    ctx_length = hparams["max_position_embeddings"]
-else:
-    print("gguf: can not find ctx length parameter.")
-
-    sys.exit()
-
-
-gguf_writer.add_name(last_dir)
-gguf_writer.add_source_hf_repo(hf_repo)
-gguf_writer.add_tensor_data_layout("Meta AI original pth")
-gguf_writer.add_context_length(ctx_length)
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
-gguf_writer.add_head_count(head_count)
-gguf_writer.add_head_count_kv(head_count_kv)
-gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
-
-if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
-    if "type" in hparams["rope_scaling"]:
-        if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
-
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: List[bytes] = []
-scores: List[float] = []
-toktypes: List[int] = []
-
-if Path(dir_model + "/tokenizer.model").is_file():
-    # vocab type sentencepiece
-    print("gguf: get sentencepiece tokenizer vocab, scores and token types")
-
-    tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1  # defualt to normal token type
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-
-        # toktype = 4 is user-defined = tokens from added_tokens.json
-
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-
-    if Path(dir_model + "/added_tokens.json").is_file():
-        with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
-            addtokens_json = json.load(f)
-
-            print("gguf: get added tokens")
-
-            for key in addtokens_json:
-                tokens.append( key.encode("utf-8") )
-                scores.append(-1000.0)
-                toktypes.append(4) # user-defined token type
-
-
-    gguf_writer.add_tokenizer_model("llama")
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-
-print("gguf: get special token ids")
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name.endswith(".rotary_emb.inv_freq"):
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # reverse permute these
-        if name.endswith(".q_proj.weight"):
-            data = reverse_hf_permute(data, head_count)
-        if name.endswith(".k_proj.weight"):
-            data = reverse_hf_permute(data, head_count, head_count_kv)
-
-        # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-
-print("gguf: model successfully exported to '" + fname_out + "'")
-print("")

convert-lora-to-ggml.py

@@ -1,15 +1,17 @@
 #!/usr/bin/env python3
+from __future__ import annotations
+
 import json
 import os
 import re
 import struct
 import sys
-from typing import Any, Dict, Sequence, TextIO
+from typing import Any, BinaryIO, Sequence
 
 import numpy as np
 import torch
 
-NUMPY_TYPE_TO_FTYPE: Dict[str, int] = {"float32": 0, "float16": 1}
+NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
 
 
 HF_SUBLAYER_TO_GGML = {
@@ -46,7 +48,7 @@ def translate_tensor_name(t: str) -> str:
         sys.exit(1)
 
 
-def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
+def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
     fout.write(struct.pack("i", params["r"]))
@@ -60,7 +62,7 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
 
 
 def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: np.dtype
+    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
 ) -> None:
     sname = name.encode("utf-8")
     fout.write(

convert-persimmon-to-gguf.py

@@ -0,0 +1,130 @@
+import torch
+import os
+from pprint import pprint
+import sys
+import argparse
+from pathlib import Path
+from sentencepiece import SentencePieceProcessor
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+def _flatten_dict(dct, tensors, prefix=None):
+    assert isinstance(dct, dict)
+    for key in dct.keys():
+        new_prefix = prefix + '.' + key if prefix is not None else key
+        if isinstance(dct[key], torch.Tensor):
+            tensors[new_prefix] = dct[key]
+        elif isinstance(dct[key], dict):
+            _flatten_dict(dct[key], tensors, new_prefix)
+        else:
+            raise ValueError(type(dct[key]))
+    return None
+
+def _get_sentencepiece_tokenizer_info(dir_model: Path):
+    tokenizer_path = dir_model / 'adept_vocab.model'
+    print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
+    tokenizer = SentencePieceProcessor(str(tokenizer_path))
+    print('gguf: adding tokens')
+    tokens: list[bytes] = []
+    scores: list[float] = []
+    toktypes: list[int] = []
+
+    for i in range(tokenizer.vocab_size()):
+        text: bytes
+        score: float
+
+        piece = tokenizer.id_to_piece(i)
+        text = piece.encode("utf-8")
+        score = tokenizer.get_score(i)
+
+        toktype = 1
+        if tokenizer.is_unknown(i):
+            toktype = 2
+        if tokenizer.is_control(i):
+            toktype = 3
+        if tokenizer.is_unused(i):
+            toktype = 5
+        if tokenizer.is_byte(i):
+            toktype = 6
+
+        tokens.append(text)
+        scores.append(score)
+        toktypes.append(toktype)
+        pass
+    return tokens, scores, toktypes
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
+    parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
+    parser.add_argument("--ckpt-path",           type=Path, help="path to persimmon checkpoint .pt file")
+    parser.add_argument("--model-dir",           type=Path, help="directory containing model e.g. 8b_chat_model_release")
+    parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory")
+    args = parser.parse_args()
+    sys.path.append(str(args.adept_inference_dir))
+    persimmon_model = torch.load(args.ckpt_path)
+    hparams = persimmon_model['args']
+    pprint(hparams)
+    tensors = {}
+    _flatten_dict(persimmon_model['model'], tensors, None)
+
+    arch = gguf.MODEL_ARCH.PERSIMMON
+    gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
+
+    block_count = hparams.num_layers
+    head_count = hparams.num_attention_heads
+    head_count_kv = head_count
+    ctx_length = hparams.seq_length
+    hidden_size = hparams.hidden_size
+
+    gguf_writer.add_name('persimmon-8b-chat')
+    gguf_writer.add_context_length(ctx_length)
+    gguf_writer.add_embedding_length(hidden_size)
+    gguf_writer.add_block_count(block_count)
+    gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
+    gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+    gguf_writer.add_head_count(head_count)
+    gguf_writer.add_head_count_kv(head_count_kv)
+    gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
+    gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
+
+    tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
+    gguf_writer.add_tokenizer_model('llama')
+    gguf_writer.add_token_list(tokens)
+    gguf_writer.add_token_scores(scores)
+    gguf_writer.add_token_types(toktypes)
+    gguf_writer.add_bos_token_id(71013)
+    gguf_writer.add_eos_token_id(71013)
+
+    tensor_map = gguf.get_tensor_name_map(arch, block_count)
+    print(tensor_map)
+    for name in tensors.keys():
+        data = tensors[name]
+        if name.endswith(".self_attention.rotary_emb.inv_freq"):
+            continue
+        old_dtype = data.dtype
+        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
+        data = data.to(torch.float32).squeeze().numpy()
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+        n_dims = len(data.shape)
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        gguf_writer.add_tensor(new_name, data)
+    print("gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+    gguf_writer.close()
+
+    print(f"gguf: model successfully exported to '{args.outfile}'")
+    print("")
+
+
+
+if __name__ == '__main__':
+    main()

convert-refact-hf-to-gguf.py

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

convert-starcoder-hf-to-gguf.py

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

convert.py

@@ -1,9 +1,8 @@
 #!/usr/bin/env python3
+from __future__ import annotations
 
-import gguf
 import argparse
 import concurrent.futures
-from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
 import copy
 import enum
 import faulthandler
@@ -20,24 +19,29 @@ import struct
 import sys
 import time
 import zipfile
-import numpy as np
-
 from abc import ABCMeta, abstractmethod
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, TypeVar, Union)
-from sentencepiece import SentencePieceProcessor  # type: ignore
+from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar
+
+import numpy as np
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+import os
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 if TYPE_CHECKING:
-    from typing_extensions import TypeAlias
+    from typing import TypeAlias
 
 if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
     faulthandler.register(signal.SIGUSR1)
 
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 
-ARCH=gguf.MODEL_ARCH.LLAMA
-NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
+ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
 #
@@ -47,8 +51,8 @@ DEFAULT_CONCURRENCY = 8
 @dataclass(frozen=True)
 class DataType:
     name: str
-    dtype: 'np.dtype[Any]'
-    valid_conversions: List[str]
+    dtype: np.dtype[Any]
+    valid_conversions: list[str]
 
     def elements_to_bytes(self, n_elements: int) -> int:
         return n_elements * self.dtype.itemsize
@@ -65,7 +69,7 @@ DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_convers
 @dataclass(frozen=True)
 class QuantizedDataType(DataType):
     block_size: int
-    quantized_dtype: 'np.dtype[Any]'
+    quantized_dtype: np.dtype[Any]
     ggml_type: gguf.GGMLQuantizationType
 
     def quantize(self, arr: NDArray) -> NDArray:
@@ -84,7 +88,7 @@ class Q8_0QuantizedDataType(QuantizedDataType):
         n_blocks = arr.size // self.block_size
         blocks = arr.reshape((n_blocks, self.block_size))
         # Much faster implementation of block quantization contributed by @Cebtenzzre
-        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[Tuple[Any, Any]]:
+        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
             d = abs(blocks).max(axis = 1) / np.float32(127)
             with np.errstate(divide = 'ignore'):
                 qs = (blocks / d[:, None]).round()
@@ -98,13 +102,13 @@ DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
     quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
 
 # Quantized types skipped here because they may also map to np.float32
-NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = {}
+NUMPY_TYPE_TO_DATA_TYPE: dict[np.dtype[Any], DataType] = {}
 for dt in (DT_BF16, DT_F16, DT_F32, DT_I32):
     if dt.dtype in NUMPY_TYPE_TO_DATA_TYPE:
         raise ValueError(f'Invalid duplicate data type {dt}')
     NUMPY_TYPE_TO_DATA_TYPE[dt.dtype] = dt
 
-SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
+SAFETENSORS_DATA_TYPES: dict[str, DataType] = {
     'BF16': DT_BF16,
     'F16': DT_F16,
     'F32': DT_F32,
@@ -119,14 +123,14 @@ class GGMLFileType(enum.IntEnum):
     MostlyF16  = 1  # except 1d tensors
     MostlyQ8_0 = 7  # except 1d tensors
 
-    def type_for_tensor(self, name: str, tensor: 'LazyTensor') -> DataType:
+    def type_for_tensor(self, name: str, tensor: LazyTensor) -> DataType:
         dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
         if dt is None:
             raise ValueError(self)
         # 1D tensors are always F32.
         return dt if len(tensor.shape) > 1 else DT_F32
 
-GGML_FILE_TYPE_TO_DATA_TYPE: Dict[GGMLFileType, DataType] = {
+GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
     GGMLFileType.AllF32    : DT_F32,
     GGMLFileType.MostlyF16 : DT_F16,
     GGMLFileType.MostlyQ8_0: DT_Q8_0,
@@ -140,7 +144,6 @@ GGML_FILE_TYPE_TO_DATA_TYPE: Dict[GGMLFileType, DataType] = {
 class Params:
     n_vocab:    int
     n_embd:     int
-    n_mult:     int
     n_layer:    int
     n_ctx:      int
     n_ff:       int
@@ -148,25 +151,16 @@ class Params:
     n_head_kv:  int
     f_norm_eps: float
 
-    f_rope_freq_base: Optional[float] = None
-    f_rope_scale: Optional[float] = None
+    f_rope_freq_base: float | None = None
+    f_rope_scale: float | None = None
 
-    ftype: Optional[GGMLFileType] = None
+    ftype: GGMLFileType | None = None
 
     # path to the directory containing the model files
-    path_model: Optional['Path'] = None
+    path_model: Path | None = None
 
     @staticmethod
-    def find_n_mult(n_ff: int, n_embd: int) -> int:
-        # hardcoded magic range
-        for n_mult in range(8192, 1, -1):
-            calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
-            if calc_ff == n_ff:
-                return n_mult
-        raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).")
-
-    @staticmethod
-    def guessed(model: 'LazyModel') -> 'Params':
+    def guessed(model: LazyModel) -> Params:
         # try transformer naming first
         n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
 
@@ -192,7 +186,6 @@ class Params:
         return Params(
             n_vocab    = n_vocab,
             n_embd     = n_embd,
-            n_mult     = n_mult,
             n_layer    = n_layer,
             n_ctx      = -1,
             n_ff       = n_ff,
@@ -202,7 +195,7 @@ class Params:
         )
 
     @staticmethod
-    def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+    def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
         n_vocab          = config["vocab_size"]
@@ -220,8 +213,6 @@ class Params:
         else:
             f_rope_scale = None
 
-        n_mult = Params.find_n_mult(n_ff, n_embd)
-
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
         elif "max_position_embeddings" in config:
@@ -233,7 +224,6 @@ class Params:
         return Params(
             n_vocab          = n_vocab,
             n_embd           = n_embd,
-            n_mult           = n_mult,
             n_layer          = n_layer,
             n_ctx            = n_ctx,
             n_ff             = n_ff,
@@ -245,15 +235,14 @@ class Params:
         )
 
     # LLaMA v2 70B params.json
-    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1
+    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
     @staticmethod
-    def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+    def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
         n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
         n_embd           = config["dim"]
         n_layer          = config["n_layers"]
-        n_mult           = config["multiple_of"]
         n_ff             = -1
         n_head           = config["n_heads"]
         n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
@@ -261,7 +250,7 @@ class Params:
         f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
 
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if f_rope_freq_base and f_rope_freq_base == 1000000:
+        if f_rope_freq_base == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -280,7 +269,6 @@ class Params:
         return Params(
             n_vocab          = n_vocab,
             n_embd           = n_embd,
-            n_mult           = n_mult,
             n_layer          = n_layer,
             n_ctx            = n_ctx,
             n_ff             = n_ff,
@@ -291,7 +279,7 @@ class Params:
         )
 
     @staticmethod
-    def load(model_plus: 'ModelPlus') -> 'Params':
+    def load(model_plus: ModelPlus) -> Params:
         hf_config_path   = model_plus.paths[0].parent / "config.json"
         orig_config_path = model_plus.paths[0].parent / "params.json"
 
@@ -299,8 +287,10 @@ class Params:
             params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
         elif orig_config_path.exists():
             params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
-        else:
+        elif model_plus.format != 'none':
             params = Params.guessed(model_plus.model)
+        else:
+            raise ValueError('Cannot guess params when model format is none')
 
         params.path_model = model_plus.paths[0].parent
 
@@ -312,19 +302,31 @@ class Params:
 #
 
 class BpeVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        added_tokens: Dict[str, int]
+        added_tokens: dict[str, int]
         if fname_added_tokens is not None:
+            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
             added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
-            added_tokens = {}
+            # Fall back to trying to find the added tokens in tokenizer.json
+            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
+            if not tokenizer_json_file.is_file():
+                added_tokens = {}
+            else:
+                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
+                added_tokens = dict(
+                    (item['content'], item['id'])
+                    for item in tokenizer_json.get('added_tokens', [])
+                    # Added tokens here can be duplicates of the main vocabulary.
+                    if item['content'] not in self.bpe_tokenizer )
 
         vocab_size: int = len(self.bpe_tokenizer)
         expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids      = sorted(added_tokens.values())
         if expected_ids != actual_ids:
-            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
+            expected_end_id = vocab_size + len(actual_ids) - 1
+            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
         self.added_tokens_list    = [text for (text, idx) in items]
@@ -333,33 +335,31 @@ class BpeVocab:
         self.fname_tokenizer      = fname_tokenizer
         self.fname_added_tokens   = fname_added_tokens
 
-    def bpe_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.bpe_tokenizer
-        from transformers.models.gpt2 import tokenization_gpt2
-        byte_encoder = tokenization_gpt2.bytes_to_unicode()
-        byte_decoder = {v: k for k, v in byte_encoder.items()}
-        for i, item in enumerate(tokenizer):
-            text: bytes = item.encode("utf-8")
-            score: float = -i
-            yield text, score, gguf.TokenType.USER_DEFINED
+        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
 
-    def added_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+        for i, _ in enumerate(tokenizer):
+            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         for text in self.added_tokens_list:
             score = -1000.0
-            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
 
-    def all_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         yield from self.bpe_tokens()
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
 class SentencePieceVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        added_tokens: Dict[str, int]
+        added_tokens: dict[str, int]
         if fname_added_tokens is not None:
             added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
@@ -378,7 +378,7 @@ class SentencePieceVocab:
         self.fname_tokenizer = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
-    def sentencepiece_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.sentencepiece_tokenizer
         for i in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(i)
@@ -402,20 +402,19 @@ class SentencePieceVocab:
 
             yield text, score, toktype
 
-    def added_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         for text in self.added_tokens_list:
             score = -1000.0
             yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
 
-    def all_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         yield from self.sentencepiece_tokens()
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
-Vocab = Union[BpeVocab, SentencePieceVocab]
-
+Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 
 #
 # data loading
@@ -425,7 +424,7 @@ Vocab = Union[BpeVocab, SentencePieceVocab]
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
     #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
     if n_head_kv is not None and n_head != n_head_kv:
-        n_head //= n_head_kv
+        n_head = n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
                 .swapaxes(1, 2)
                 .reshape(weights.shape))
@@ -435,18 +434,18 @@ class Tensor(metaclass=ABCMeta):
     data_type: DataType
 
     @abstractmethod
-    def astype(self, data_type: DataType) -> 'Tensor': ...
+    def astype(self, data_type: DataType) -> Tensor: ...
     @abstractmethod
-    def permute(self, n_head: int, n_head_kv: int) -> 'Tensor': ...
+    def permute(self, n_head: int, n_head_kv: int) -> Tensor: ...
     @abstractmethod
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor: ...
     @abstractmethod
-    def part(self, n_part: int) -> 'UnquantizedTensor': ...
+    def part(self, n_part: int) -> UnquantizedTensor: ...
     @abstractmethod
-    def to_ggml(self) -> 'GGMLCompatibleTensor': ...
+    def to_ggml(self) -> GGMLCompatibleTensor: ...
 
 
-def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray:
+def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
     assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
     fp32_arr = bf16_arr.astype(np.uint32) << 16
     return fp32_arr.view(np.float32)
@@ -464,22 +463,22 @@ class UnquantizedTensor(Tensor):
             self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
 
-    def to_ggml(self) -> 'UnquantizedTensor':
+    def to_ggml(self) -> UnquantizedTensor:
         return self
 
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor:
         r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
 
-    def part(self, n_part: int) -> 'UnquantizedTensor':
+    def part(self, n_part: int) -> UnquantizedTensor:
         r = self.ndarray.shape[0] // 3
         return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
-    def permute(self, n_head: int, n_head_kv: int) -> 'UnquantizedTensor':
+    def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor:
         return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv))
 
 
-def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray:
+def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray:
     tensor = lazy_tensor.load()
     assert isinstance(tensor, UnquantizedTensor)
 
@@ -495,13 +494,13 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv
     return tensor.ndarray
 
 
-GGMLCompatibleTensor = Union[UnquantizedTensor]
+GGMLCompatibleTensor = UnquantizedTensor
 
 
 @dataclass
 class LazyTensor:
     _load: Callable[[], Tensor]
-    shape: List[int]
+    shape: list[int]
     data_type: DataType
     description: str
 
@@ -512,7 +511,7 @@ class LazyTensor:
                 (self.data_type, ret.data_type, self.description)
         return ret
 
-    def astype(self, data_type: DataType) -> 'LazyTensor':
+    def astype(self, data_type: DataType) -> LazyTensor:
         self.validate_conversion_to(data_type)
 
         def load() -> Tensor:
@@ -524,24 +523,24 @@ class LazyTensor:
             raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
 
 
-LazyModel = Dict[str, LazyTensor]
+LazyModel: TypeAlias = 'dict[str, LazyTensor]'
 
 
 @dataclass
 class ModelPlus:
     model: LazyModel
-    paths: List[Path]  # Where this was read from.
-    format: Literal['ggml', 'torch', 'safetensors']
-    vocab: Optional[Vocab]  # For GGML models (which have vocab built in), the vocab.
+    paths: list[Path]  # Where this was read from.
+    format: Literal['ggml', 'torch', 'safetensors', 'none']
+    vocab: Vocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
-def merge_sharded(models: List[LazyModel]) -> LazyModel:
+def merge_sharded(models: list[LazyModel]) -> LazyModel:
     # Original LLaMA models have each file contain one part of each tensor.
     # Use a dict instead of a set to preserve order.
     names = {name: None for model in models for name in model}
 
     def convert(name: str) -> LazyTensor:
-        lazy_tensors: List[LazyTensor] = [model[name] for model in models]
+        lazy_tensors: list[LazyTensor] = [model[name] for model in models]
         if len(lazy_tensors) == 1:
             # only one file; don't go through this procedure since there might
             # be quantized tensors
@@ -569,7 +568,7 @@ def merge_sharded(models: List[LazyModel]) -> LazyModel:
     return {name: convert(name) for name in names}
 
 
-def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
+def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
     formats = set(mp.format for mp in models_plus)
     assert len(formats) == 1, "different formats?"
     format = formats.pop()
@@ -597,12 +596,12 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe
         return lazy_tensor.load().permute(n_head, n_head_kv)
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
-def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
+def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute_part(n_part, n_head)
+        return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
     s = lazy_tensor.shape.copy()
     s[0] = s[0] // 3
-    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
 def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     def load() -> Tensor:
@@ -643,7 +642,7 @@ class LazyUnpickler(pickle.Unpickler):
         assert isinstance(pid[1], LazyStorageKind)
         data_type = pid[1].data_type
         filename_stem = pid[2]
-        filename = self.data_base_path + '/' + filename_stem
+        filename = f'{self.data_base_path}/{filename_stem}'
         info = self.zip_file.getinfo(filename)
 
         def load(offset: int, elm_count: int) -> NDArray:
@@ -657,9 +656,8 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         return LazyStorage(load=load, kind=pid[1], description=description)
 
-    # @staticmethod
+    @staticmethod
     def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
-                               # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
         assert isinstance(storage, LazyStorage)
 
@@ -669,13 +667,15 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'pickled storage_offset={storage_offset} in {storage.description}'
         return LazyTensor(load, list(size), storage.kind.data_type, description)
 
-    # @staticmethod
+    @staticmethod
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: Dict[Any, Any] = {
-        ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2,
-        ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2,
+    CLASSES: dict[tuple[str, str], Any] = {
+        # getattr used here as a workaround for mypy not being smart enough to detrmine
+        # the staticmethods have a __func__ attribute.
+        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
+        ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
         ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
         ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
         ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
@@ -704,15 +704,15 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
 
 def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
     header_size, = struct.unpack('<Q', fp.read(8))
-    header: Dict[str, Dict[str, Any]] = json.loads(fp.read(header_size))
+    header: dict[str, dict[str, Any]] = json.loads(fp.read(header_size))
     # Use mmap for the actual data to avoid race conditions with the file offset.
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
     byte_buf = mapped[8 + header_size:]
 
-    def convert(info: Dict[str, Any]) -> LazyTensor:
+    def convert(info: dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
         numpy_dtype = data_type.dtype
-        shape: List[int] = info['shape']
+        shape: list[int] = info['shape']
         begin, end = info['data_offsets']
         assert 0 <= begin <= end <= len(byte_buf)
         assert end - begin == math.prod(shape) * numpy_dtype.itemsize
@@ -751,7 +751,7 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, factory: Callable = ThreadPoolExecutor) -> Iterable[Out]:
+def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
     letting results pile up in memory.  Specifically, there is a max of one
@@ -760,8 +760,13 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
         yield from map(func, iterable)
         # Not reached.
     iterable = iter(iterable)
-    with factory(max_workers = max_workers) as executor:
-        futures: List[concurrent.futures.Future[Out]] = []
+    executor_class: type[ThreadPoolExecutor] | type[ProcessPoolExecutor]
+    if use_processpool_executor:
+        executor_class = ProcessPoolExecutor
+    else:
+        executor_class = ThreadPoolExecutor
+    with executor_class(max_workers = max_workers) as executor:
+        futures: list[concurrent.futures.Future[Out]] = []
         done = False
         for _ in range(concurrency):
             try:
@@ -803,10 +808,12 @@ class OutputFile:
 
     def add_meta_arch(self, params: Params) -> None:
         name = "LLaMA"
-        if (params.n_ctx == 4096):
+
+        # TODO: better logic to determine model name
+        if params.n_ctx == 4096:
             name = "LLaMA v2"
-            if params.path_model:
-                name = str(params.path_model.parent).split('/')[-1]
+        elif params.path_model is not None:
+            name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
         self.gguf.add_context_length      (params.n_ctx)
@@ -818,31 +825,38 @@ class OutputFile:
         self.gguf.add_head_count_kv       (params.n_head_kv)
         self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
 
-        if params.f_rope_freq_base:
+        if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
-        if params.f_rope_scale:
+        if params.f_rope_scale is not None:
             self.gguf.add_rope_scale_linear(params.f_rope_scale)
 
-        if params.ftype:
+        if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
         tokens = []
         scores = []
         toktypes = []
-        # NOTE: `all_tokens` returns the the base vocabulary and added tokens
-        # TODO: add special tokens?
+        # NOTE: `all_tokens` returns the base vocabulary and added tokens
         for text, score, toktype in vocab.all_tokens():
             tokens.append(text)
             scores.append(score)
             toktypes.append(toktype)
 
-        self.gguf.add_tokenizer_model("llama")
+        if isinstance(vocab, SentencePieceVocab):
+            self.gguf.add_tokenizer_model("llama")
+        elif isinstance(vocab, BpeVocab):
+            self.gguf.add_tokenizer_model("gpt2")
+        else:
+            raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab')
         self.gguf.add_token_list(tokens)
         self.gguf.add_token_scores(scores)
         self.gguf.add_token_types(toktypes)
 
+    def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None:
+        svocab.add_to_gguf(self.gguf)
+
     def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
         n_elements = int(np.prod(tensor.shape))
         raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
@@ -861,7 +875,7 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab) -> None:
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -869,25 +883,27 @@ class OutputFile:
         # meta data
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
+
         of.write_meta()
 
         of.close()
 
     @staticmethod
-    def do_item(item: Tuple[str, LazyTensor]) -> Tuple[DataType, NDArray]:
+    def do_item(item: tuple[str, LazyTensor]) -> tuple[DataType, NDArray]:
         name, lazy_tensor = item
         tensor = lazy_tensor.load().to_ggml()
         return (lazy_tensor.data_type, tensor.ndarray)
 
     @staticmethod
-    def maybe_do_quantize(item: Tuple[DataType, NDArray]) -> NDArray:
+    def maybe_do_quantize(item: tuple[DataType, NDArray]) -> NDArray:
         dt, arr = item
         if not isinstance(dt, QuantizedDataType):
             return arr
         return dt.quantize(arr)
 
     @staticmethod
-    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -895,6 +911,7 @@ class OutputFile:
         # meta data
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -906,7 +923,7 @@ class OutputFile:
         # 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, factory = ProcessPoolExecutor)
+            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)
 
@@ -920,8 +937,8 @@ class OutputFile:
 
         of.close()
 
-def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType:
-    wq_type = model[NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
+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):
         return GGMLFileType.AllF32
@@ -939,7 +956,8 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
-    tmap = gguf.get_tensor_name_map(ARCH, params.n_layer)
+    tmap = gguf.TensorNameMap(ARCH, params.n_layer)
+    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
@@ -955,37 +973,31 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
             tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
             tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
             tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy        (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
         else:
             break
 
     out: LazyModel = {}
     for name, lazy_tensor in model.items():
-        name_new = name
-
-        if name in tmap:
-            name_new = tmap[name]
-        elif name.endswith(".weight") and name[:-7] in tmap:
-            name_new = tmap[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tmap:
-            name_new = tmap[name[:-5]] + ".bias"
-        else:
+        tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
+        if name_new is None:
             raise Exception(f"Unexpected tensor name: {name}")
 
-        if gguf.should_skip_tensor_TMP(ARCH, params.n_layer, name_new):
+        if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
             continue
-        else:
-            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
-            out[name_new] = lazy_tensor
+
+        print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
+        out[name_new] = lazy_tensor
 
     return out
 
-def nth_multifile_path(path: Path, n: int) -> Optional[Path]:
+def nth_multifile_path(path: Path, n: int) -> Path | None:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the nth path in the model.
     '''
     # Support the following patterns:
-    patterns: List[Tuple[str, str]] = [
+    patterns: list[tuple[str, str]] = [
         # - x.00.pth, x.01.pth, etc.
         (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
         # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
@@ -1001,11 +1013,11 @@ def nth_multifile_path(path: Path, n: int) -> Optional[Path]:
     return None
 
 
-def find_multifile_paths(path: Path) -> List[Path]:
+def find_multifile_paths(path: Path) -> list[Path]:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the whole list of paths in the model.
     '''
-    ret: List[Path] = []
+    ret: list[Path] = []
     for i in itertools.count():
         nth_path = nth_multifile_path(path, i)
         if nth_path is None:
@@ -1036,7 +1048,7 @@ def load_some_model(path: Path) -> ModelPlus:
         path = files[0]
 
     paths = find_multifile_paths(path)
-    models_plus: List[ModelPlus] = []
+    models_plus: list[ModelPlus] = []
     for path in paths:
         print(f"Loading model file {path}")
         models_plus.append(lazy_load_file(path))
@@ -1045,7 +1057,7 @@ def load_some_model(path: Path) -> ModelPlus:
     return model_plus
 
 
-def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, SentencePieceVocab]:
+def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
     # Be extra-friendly and accept either a file or a directory.  Also, if it's
     # a directory, it might be the model directory, and tokenizer.model might
     # be in the parent of that.
@@ -1076,7 +1088,7 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, Sentence
         raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
 
-def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
     namestr = {
         GGMLFileType.AllF32:    "f32",
         GGMLFileType.MostlyF16: "f16",
@@ -1099,7 +1111,7 @@ def do_dump_model(model_plus: ModelPlus) -> None:
         print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}")
 
 
-def main(args_in: Optional[List[str]] = None) -> None:
+def main(args_in: list[str] | None = None) -> 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")
@@ -1116,8 +1128,16 @@ def main(args_in: Optional[List[str]] = None) -> None:
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
+        return
 
-    model_plus = load_some_model(args.model)
+    if not args.vocab_only:
+        model_plus = load_some_model(args.model)
+    else:
+        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
+
+    if args.dump:
+        do_dump_model(model_plus)
+        return
 
     params = Params.load(model_plus)
     if params.n_ctx == -1:
@@ -1139,33 +1159,34 @@ def main(args_in: Optional[List[str]] = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
         assert args.outfile, "need --outfile if using --vocab-only"
+        # FIXME: Try to respect vocab_dir somehow?
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
+        return
+
+    if model_plus.vocab is not None and args.vocab_dir is None:
+        vocab = model_plus.vocab
     else:
-        if args.dump:
-            do_dump_model(model_plus)
-            return
+        vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
+        vocab = load_vocab(vocab_dir, args.vocabtype)
+    # FIXME: Try to respect vocab_dir somehow?
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
 
-        if model_plus.vocab is not None and args.vocab_dir is None:
-            vocab = model_plus.vocab
-        else:
-            vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-            vocab = load_vocab(vocab_dir, args.vocabtype)
+    model   = model_plus.model
+    model   = convert_model_names(model, params)
+    ftype   = pick_output_type(model, args.outtype)
+    model   = convert_to_output_type(model, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
 
-        model   = model_plus.model
-        model   = convert_model_names(model, params)
-        ftype   = pick_output_type(model, args.outtype)
-        model   = convert_to_output_type(model, ftype)
-        outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+    params.ftype = ftype
+    print(f"Writing {outfile}, format {ftype}")
 
-        params.ftype = ftype
-        print(f"Writing {outfile}, format {ftype}")
-
-        OutputFile.write_all(outfile, ftype, params, model, vocab, concurrency = args.concurrency)
-        print(f"Wrote {outfile}")
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
+    print(f"Wrote {outfile}")
 
 
 if __name__ == '__main__':

docs/BLIS.md

@@ -48,7 +48,7 @@ make -j
 According to the BLIS documentation, we could set the following
 environment variables to modify the behavior of openmp:
 
-```
+```bash
 export GOMP_GPU_AFFINITY="0-19"
 export BLIS_NUM_THREADS=14
 ```

examples/CMakeLists.txt

@@ -21,15 +21,20 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(finetune)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(simple)
+    add_subdirectory(batched)
+    add_subdirectory(speculative)
+    add_subdirectory(parallel)
     add_subdirectory(embd-input)
     add_subdirectory(llama-bench)
-    add_subdirectory(beam_search)
+    add_subdirectory(beam-search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()
+    add_subdirectory(export-lora)
 endif()

examples/baby-llama/baby-llama.cpp

@@ -1,43 +1,24 @@
 #include "ggml.h"
+#include "train.h"
+
 #include <vector>
 #include <cassert>
-#include <random>
+#include <cstdlib>
 #include <cstring>
+#include <random>
+#include <vector>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #ifdef LLAMA_DEFAULT_RMS_EPS
-static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
+constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
 #else
-static const float rms_norm_eps = 5e-6f;
+constexpr float rms_norm_eps = 5e-6f;
 #endif
 
-float frand() {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-struct random_normal_distribution {
-    std::mt19937 gen;
-    std::normal_distribution<float> nd;
-    float min;
-    float max;
-};
-
-void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
-    rnd->gen = std::mt19937(seed);
-    rnd->nd = std::normal_distribution<float>{mean, std};
-    rnd->min = min;
-    rnd->max = max;
-}
-
-float frand_normal(struct random_normal_distribution * rnd) {
-    const float r = rnd->nd(rnd->gen);
-    return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
-}
-
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
@@ -48,13 +29,9 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
     ggml_graph_compute(graph, &plan);
 }
 
-struct ggml_tensor * randomize_tensor(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        float fmin,
-        float fmax) {
-
+static struct ggml_tensor * randomize_tensor(
+    struct ggml_tensor * tensor, int ndims, const int64_t ne[], float fmin, float fmax
+) {
     switch (ndims) {
         case 1:
             for (int i0 = 0; i0 < ne[0]; i0++) {
@@ -90,57 +67,7 @@ struct ggml_tensor * randomize_tensor(
             break;
         default:
             assert(false);
-    };
-
-    return tensor;
-}
-
-struct ggml_tensor * randomize_tensor_normal(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        struct random_normal_distribution * rnd) {
-    float scale = 1.0; // xavier
-    switch (ndims) {
-        case 1:
-            scale /= sqrtf(ne[0]);
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
-            }
-            break;
-        case 2:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
-                }
-            }
-            break;
-        case 3:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                    }
-                }
-            }
-            break;
-        case 4:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    };
+    }
 
     return tensor;
 }
@@ -159,7 +86,7 @@ struct llama_hparams {
     }
 };
 
-uint32_t get_n_ff(const struct llama_hparams* hparams) {
+static uint32_t get_n_ff(const struct llama_hparams* hparams) {
     const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
     return n_ff;
 }
@@ -260,7 +187,7 @@ struct llama_model_lora {
     std::vector<llama_layer_lora> layers;
 };
 
-void init_model(struct llama_model * model) {
+static void init_model(struct llama_model * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -297,7 +224,7 @@ void init_model(struct llama_model * model) {
 }
 
 
-void init_model_lora(struct llama_model_lora * model) {
+static void init_model_lora(struct llama_model_lora * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -340,7 +267,7 @@ void init_model_lora(struct llama_model_lora * model) {
     }
 }
 
-void set_param_model(struct llama_model * model) {
+static void set_param_model(struct llama_model * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -366,7 +293,7 @@ void set_param_model(struct llama_model * model) {
     }
 }
 
-void set_param_model_lora(struct llama_model_lora * model) {
+static void set_param_model_lora(struct llama_model_lora * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -397,69 +324,109 @@ void set_param_model_lora(struct llama_model_lora * model) {
     }
 }
 
-void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->output,         model->output->n_dims,         model->output->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings , rnd);
+    randomize_tensor_normal(model->norm           , rnd);
+    randomize_tensor_normal(model->output         , rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
 
-        randomize_tensor_normal(layer.wq, layer.wq->n_dims, layer.wq->ne, &rnd);
-        randomize_tensor_normal(layer.wk, layer.wk->n_dims, layer.wk->ne, &rnd);
-        randomize_tensor_normal(layer.wv, layer.wv->n_dims, layer.wv->ne, &rnd);
-        randomize_tensor_normal(layer.wo, layer.wo->n_dims, layer.wo->ne, &rnd);
+        randomize_tensor_normal(layer.wq, rnd);
+        randomize_tensor_normal(layer.wk, rnd);
+        randomize_tensor_normal(layer.wv, rnd);
+        randomize_tensor_normal(layer.wo, rnd);
 
-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
+
+    free_random_normal_distribution(rnd);
 }
 
 
-void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model_lora(
+    struct llama_model_lora * model, int seed, float mean, float std, float min, float max
+) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->outputa,        model->outputa->n_dims,        model->outputa->ne,         &rnd);
-    randomize_tensor_normal(model->outputb,        model->outputb->n_dims,        model->outputb->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings, rnd);
+    randomize_tensor_normal(model->norm          , rnd);
+    randomize_tensor_normal(model->outputa       , rnd);
+    randomize_tensor_normal(model->outputb       , rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
 
-        randomize_tensor_normal(layer.wqa, layer.wqa->n_dims, layer.wqa->ne, &rnd);
-        randomize_tensor_normal(layer.wqb, layer.wqb->n_dims, layer.wqb->ne, &rnd);
-        randomize_tensor_normal(layer.wka, layer.wka->n_dims, layer.wka->ne, &rnd);
-        randomize_tensor_normal(layer.wkb, layer.wkb->n_dims, layer.wkb->ne, &rnd);
-        randomize_tensor_normal(layer.wva, layer.wva->n_dims, layer.wva->ne, &rnd);
-        randomize_tensor_normal(layer.wvb, layer.wvb->n_dims, layer.wvb->ne, &rnd);
-        randomize_tensor_normal(layer.woa, layer.woa->n_dims, layer.woa->ne, &rnd);
-        randomize_tensor_normal(layer.wob, layer.wob->n_dims, layer.wob->ne, &rnd);
+        randomize_tensor_normal(layer.wqa, rnd);
+        randomize_tensor_normal(layer.wqb, rnd);
+        randomize_tensor_normal(layer.wka, rnd);
+        randomize_tensor_normal(layer.wkb, rnd);
+        randomize_tensor_normal(layer.wva, rnd);
+        randomize_tensor_normal(layer.wvb, rnd);
+        randomize_tensor_normal(layer.woa, rnd);
+        randomize_tensor_normal(layer.wob, rnd);
 
-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
+
+    free_random_normal_distribution(rnd);
 }
 
-bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+static void init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+    const auto & hparams = model->hparams;
+
+    const uint32_t n_ctx   = hparams.n_ctx;
+    const uint32_t n_embd  = hparams.n_embd;
+    const uint32_t n_layer = hparams.n_layer;
+
+    const int64_t n_mem      = n_layer*n_ctx*n_batch;
+    const int64_t n_elements = n_embd*n_mem;
+
+    // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+
+    // struct ggml_init_params params;
+    // params.mem_size   = cache.buf.size;
+    // params.mem_buffer = cache.buf.addr;
+    // params.no_alloc   = false;
+    if (!cache->ctx) {
+        struct ggml_init_params params;
+        params.mem_size   = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024;
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        cache->ctx = ggml_init(params);
+
+        if (!cache->ctx) {
+            fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
+            exit(1);
+        }
+    }
+
+    cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
+    cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
+}
+
+static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_ctx   = hparams.n_ctx;
@@ -495,51 +462,15 @@ bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int
     return true;
 }
 
-bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
-    const auto & hparams = model->hparams;
-
-    const uint32_t n_ctx   = hparams.n_ctx;
-    const uint32_t n_embd  = hparams.n_embd;
-    const uint32_t n_layer = hparams.n_layer;
-
-    const int64_t n_mem      = n_layer*n_ctx*n_batch;
-    const int64_t n_elements = n_embd*n_mem;
-
-    // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
-
-    // struct ggml_init_params params;
-    // params.mem_size   = cache.buf.size;
-    // params.mem_buffer = cache.buf.addr;
-    // params.no_alloc   = false;
-    if (!cache->ctx) {
-        struct ggml_init_params params;
-        params.mem_size   = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024;
-        params.mem_buffer = NULL;
-        params.no_alloc   = false;
-
-        cache->ctx = ggml_init(params);
-
-        if (!cache->ctx) {
-            fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
-            return false;
-        }
-    }
-
-    cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-    cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-
-    return true;
-}
-
-struct ggml_tensor * forward(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past) {
-
+static struct ggml_tensor * forward(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -556,6 +487,14 @@ struct ggml_tensor * forward(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N,1,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     for (int il = 0; il < n_layer; ++il) {
@@ -583,8 +522,8 @@ struct ggml_tensor * forward(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, 1]
             // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {
@@ -756,42 +695,16 @@ struct ggml_tensor * forward(
     return inpL;
 }
 
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-}
-
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-}
-
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-}
-
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == ne3);
-}
-
-struct ggml_tensor * forward_batch(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past,
-        const  int              n_batch) {
-
+static struct ggml_tensor * forward_batch(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past,
+    const  int              n_batch
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -810,9 +723,18 @@ struct ggml_tensor * forward_batch(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N*n_batch,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     assert_shape_2d(inpL, n_embd, N*n_batch);
+
     for (int il = 0; il < n_layer; ++il) {
         struct ggml_tensor * inpSA = inpL;
 
@@ -840,8 +762,8 @@ struct ggml_tensor * forward_batch(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, n_batch]
             // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
             assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
             assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
 
@@ -1073,16 +995,15 @@ struct ggml_tensor * forward_batch(
     return inpL;
 }
 
-
-struct ggml_tensor * forward_lora(
-        struct llama_model_lora * model,
-        struct llama_kv_cache   * cache,
-        struct ggml_context     * ctx0,
-        struct ggml_cgraph      * gf,
-        struct ggml_tensor      * tokens_input,
-        const  int                n_tokens,
-        const  int                n_past) {
-
+static struct ggml_tensor * forward_lora(
+    struct llama_model_lora * model,
+    struct llama_kv_cache   * cache,
+    struct ggml_context     * ctx0,
+    struct ggml_cgraph      * gf,
+    struct ggml_tensor      * tokens_input,
+    const  int                n_tokens,
+    const  int                n_past
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -1100,6 +1021,14 @@ struct ggml_tensor * forward_lora(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N,1,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     for (int il = 0; il < n_layer; ++il) {
@@ -1133,7 +1062,7 @@ struct ggml_tensor * forward_lora(
                                                         model->layers[il].wqb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
             struct ggml_tensor * Kcur = ggml_rope(ctx0,
                                             ggml_reshape_3d(ctx0,
                                                 ggml_mul_mat(ctx0,
@@ -1142,7 +1071,7 @@ struct ggml_tensor * forward_lora(
                                                         model->layers[il].wkb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {
@@ -1328,7 +1257,7 @@ struct ggml_tensor * forward_lora(
     return inpL;
 }
 
-void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
     assert(logits->n_dims == 2);
     assert(probs->n_dims == 2);
     assert(best_samples->n_dims == 1);
@@ -1359,7 +1288,10 @@ void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, str
     }
 }
 
-void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax_batch(
+    struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
+    struct ggml_tensor * best_samples
+) {
     GGML_ASSERT(best_samples->n_dims == 2);
     GGML_ASSERT(logits->n_dims == 3);
     GGML_ASSERT(probs->n_dims == 3);
@@ -1393,7 +1325,7 @@ void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits
     }
 }
 
-void print_row(struct ggml_tensor * probs, int i) {
+static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
         printf(" %.2f", p);
@@ -1401,7 +1333,7 @@ void print_row(struct ggml_tensor * probs, int i) {
     printf("\n");
 }
 
-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
     assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
@@ -1412,7 +1344,7 @@ void print_matrix(struct ggml_tensor * probs) {
     }
 }
 
-void print_token(int token, int n_vocab) {
+static void print_token(int token, int n_vocab) {
     for (int k = 0; k < token; ++k) {
         printf(" ");
     }
@@ -1423,14 +1355,14 @@ void print_token(int token, int n_vocab) {
     printf("\n");
 }
 
-void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
+static void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
     for (int i=0; i<tokens->ne[0]; ++i) {
         int token = ggml_get_i32_1d(tokens, i);
         print_token(token, n_vocab);
     }
 }
 
-void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     float randomness = 0.0f;
@@ -1451,7 +1383,9 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru
     }
 }
 
-void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets_batch(
+    struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
+) {
     GGML_ASSERT(tokens_input->n_dims == 2);
     GGML_ASSERT(     targets->n_dims == 3);
     int n_tokens = tokens_input->ne[0];
@@ -1474,7 +1408,7 @@ void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct
     }
 }
 
-void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
+static void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     for (int i=0; i<n_tokens-n_shift; ++i) {
@@ -1485,12 +1419,16 @@ void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * tar
     }
 }
 
-struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * square_error_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
     // todo: instead of a-b: a[1:]-b[:-1]
     return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, a, b)));
 }
 
-struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * cross_entropy_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
     const float eps = 1e-3f;
     return
         ggml_sum(ctx,
@@ -1617,15 +1555,10 @@ int main(int argc, char ** argv) {
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM);
         struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
-        opt_params_adam.print_forward_graph = false;
-        opt_params_adam.print_backward_graph = false;
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
-        opt_params_adam.adam.n_iter = 16;
         opt_params_lbfgs.lbfgs.n_iter = 16;
-        // ggml_opt(ctx0, opt_params_adam, e);
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
         ggml_build_forward_expand(&gf, e);

examples/batched/CMakeLists.txt

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

@@ -0,0 +1,44 @@
+# llama.cpp/example/batched
+
+The example demonstrates batched generation from a given prompt
+
+```bash
+./batched ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 4
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 4, n_kv_req = 113
+
+ Hello my name is
+
+main: generating 4 sequences ...
+
+main: stream 0 finished
+main: stream 1 finished
+main: stream 2 finished
+main: stream 3 finished
+
+sequence 0:
+
+Hello my name is Shirley. I am a 25-year-old female who has been working for over 5 years as a b
+
+sequence 1:
+
+Hello my name is Renee and I'm a 32 year old female from the United States. I'm looking for a man between
+
+sequence 2:
+
+Hello my name is Diana. I am looking for a housekeeping job. I have experience with children and have my own transportation. I am
+
+sequence 3:
+
+Hello my name is Cody. I am a 3 year old neutered male. I am a very friendly cat. I am very playful and
+
+main: decoded 108 tokens in 3.57 s, speed: 30.26 t/s
+
+llama_print_timings:        load time =   587.00 ms
+llama_print_timings:      sample time =     2.56 ms /   112 runs   (    0.02 ms per token, 43664.72 tokens per second)
+llama_print_timings: prompt eval time =  4089.11 ms /   118 tokens (   34.65 ms per token,    28.86 tokens per second)
+llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_print_timings:       total time =  4156.04 ms
+```

examples/batched/batched.cpp

@@ -0,0 +1,255 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
+        return 1 ;
+    }
+
+    int n_parallel = 1;
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        params.prompt = argv[2];
+    }
+
+    if (argc >= 4) {
+        n_parallel = std::atoi(argv[3]);
+    }
+
+    if (params.prompt.empty()) {
+        params.prompt = "Hello my name is";
+    }
+
+    // total length of the sequences including the prompt
+    const int n_len = 32;
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(model, params.prompt, true);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
+
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = n_kv_req;
+    ctx_params.n_batch = std::max(n_len, 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;
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    const int n_ctx    = llama_n_ctx(ctx);
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        return 1;
+    }
+
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
+
+    for (auto id : tokens_list) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
+
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0);
+
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
+    // 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);
+    }
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+    }
+
+    // main loop
+
+    // we will store the parallel decoded sequences in this vector
+    std::vector<std::string> streams(n_parallel);
+
+    // remember the batch index of the last token for each parallel sequence
+    // we need this to determine which logits to sample from
+    std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
+
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
+        // prepare the next batch
+        batch.n_tokens = 0;
+
+        // sample the next token for each parallel sequence / stream
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            if (i_batch[i] < 0) {
+                // the stream has already finished
+                continue;
+            }
+
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
+
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            const int   top_k = 40;
+            const float top_p = 0.9f;
+            const float temp  = 0.4f;
+
+            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+            llama_sample_temp (ctx, &candidates_p, temp);
+
+            const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);
+
+            //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream? -> mark the stream as finished
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                i_batch[i] = -1;
+                LOG_TEE("\n");
+                if (n_parallel > 1) {
+                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                }
+
+                continue;
+            }
+
+            // if there is only one stream, we print immediately to stdout
+            if (n_parallel == 1) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                fflush(stdout);
+            }
+
+            streams[i] += llama_token_to_piece(ctx, new_token_id);
+
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = i;
+            batch.logits[batch.n_tokens] = true;
+
+            i_batch[i] = batch.n_tokens;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
+        }
+
+        // all streams are finished
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n");
+
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/beam-search/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET beam-search)
+add_executable(${TARGET} beam-search.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/beam-search/beam-search.cpp

@@ -1,10 +1,5 @@
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -22,7 +17,9 @@
 #include <unistd.h>
 #elif defined (_WIN32)
 #define WIN32_LEAN_AND_MEAN
-#define NOMINMAX
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
 #include <windows.h>
 #include <signal.h>
 #endif
@@ -32,7 +29,8 @@ struct ostream_beam_view {
     llama_context * ctx;
     llama_beam_view beam_view;
 };
-std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) {
+
+static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) {
     os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
     for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
         os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
@@ -48,7 +46,7 @@ struct beam_search_callback_data {
 
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
-bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const size_t n_tokens) {
+static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
     return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
 }
 
@@ -58,7 +56,7 @@ bool is_at_eob(const beam_search_callback_data & callback_data, const llama_toke
 //  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
 //    This is also called when the stop condition is met.
 //    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
-void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
+static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
     auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
     // Mark beams as EOS as needed.
     for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
@@ -73,7 +71,7 @@ void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_stat
         assert(0u < beams_state.n_beams);
         const llama_token * tokens = beams_state.beam_views[0].tokens;
         std::copy(tokens, tokens + n, callback_data.response.end() - n);
-        printf("%lu", n);
+        printf("%zu", n);
     }
     fflush(stdout);
 #if 1 // DEBUG: print current beams for this iteration
@@ -145,7 +143,7 @@ int main(int argc, char ** argv)
 
     if (tokens_list.size() > max_tokens_list_size)
     {
-        fprintf( stderr , "%s: error: prompt too long (%lu tokens, max %lu)\n" ,
+        fprintf( stderr , "%s: error: prompt too long (%zu tokens, max %zu)\n" ,
              __func__ , tokens_list.size() , max_tokens_list_size );
         return 1;
     }
@@ -160,8 +158,9 @@ int main(int argc, char ** argv)
     }
     std::cout << std::flush;
 
-    int n_past = llama_get_kv_cache_token_count(ctx);
-    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
+    int n_past = 0;
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens_list.data(), tokens_list.size(), n_past, 0)))
     {
         fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
         return 1;
@@ -171,7 +170,7 @@ int main(int argc, char ** argv)
     beam_search_callback_data callback_data{ctx, {}};
     size_t const beam_width = static_cast<size_t>(params.n_beams);
     int const n_predict = 256;
-    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads);
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict);
 
     std::cout << "\n\n";
     for (llama_token const token_id : callback_data.response) {

examples/benchmark/CMakeLists.txt

@@ -1,7 +1,8 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

examples/benchmark/benchmark-matmult.cpp

@@ -1,5 +1,6 @@
-#include "ggml.h"
 #include "build-info.h"
+#include "common.h"
+#include "ggml.h"
 
 #include <locale.h>
 #include <assert.h>
@@ -20,7 +21,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
@@ -31,19 +32,19 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
     ggml_graph_compute(graph, &plan);
 }
 
-float tensor_sum_elements(const ggml_tensor * tensor) {
-    float sum = 0;
-    if (tensor->type==GGML_TYPE_F32) {
+static float tensor_sum_elements(const ggml_tensor * tensor) {
+    double sum = 0;
+    if (tensor->type == GGML_TYPE_F32) {
         for (int j = 0; j < tensor->ne[1]; j++) {
             for (int k = 0; k < tensor->ne[0]; k++) {
-                sum +=  ((float *) tensor->data)[j*tensor->ne[0]+k];
+                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
             }
         }
     }
     return sum;
 }
 
-void tensor_dump(const ggml_tensor * tensor, const char * name) {
+static void tensor_dump(const ggml_tensor * tensor, const char * name) {
     printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
         tensor->type, ggml_type_name(tensor->type),
         tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
@@ -58,7 +59,7 @@ struct benchmark_params_struct {
     int32_t n_iterations  = 10;
 };
 
-void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
+static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
@@ -99,7 +100,7 @@ int main(int argc, char ** argv)  {
         exit(1);
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
     printf("Starting Test\n");
 
     // create the ggml context
@@ -125,12 +126,15 @@ int main(int argc, char ** argv)  {
 
     //printf("Memsize required = %i\n", sizex*sizex);
 
+    // TODO: perform the bench for all types or for a user specified type
+    const ggml_type qtype = GGML_TYPE_Q4_1;
+
     size_t ctx_size = 0;
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
     ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
@@ -163,7 +167,7 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
     ggml_set_f32(m2, 2.0f);
 
-    printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
     // printf("Creating new tensor m11xm2\n");
     struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
 
@@ -181,17 +185,16 @@ int main(int argc, char ** argv)  {
 
     TENSOR_DUMP(gf.nodes[0]);
 
-    printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
     int32_t nelements = sizex*sizey;
-    int32_t ne[2] = { 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, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -202,8 +205,8 @@ 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, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
@@ -220,7 +223,7 @@ int main(int argc, char ** argv)  {
     printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
 
 
-    // Let's use the F32 result from above as a reference for the q4_0 multiplication
+    // Let's use the F32 result from above as a reference for the quantized multiplication
     float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
@@ -250,7 +253,7 @@ int main(int argc, char ** argv)  {
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
         float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
-        float delta = abs(sum_of_Q4_result - sum_of_F32_reference);
+        float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 
         if (delta > allowed_delta)  {

examples/chat-persistent.sh

@@ -9,7 +9,7 @@ if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then
     exit 1
 fi
 
-MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}"
+MODEL="${MODEL:-./models/llama-13b/ggml-model-q4_0.gguf}"
 PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}"
 USER_NAME="${USER_NAME:-User}"
 AI_NAME="${AI_NAME:-ChatLLaMa}"
@@ -61,9 +61,9 @@ fi
 
 if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
     echo 'Prompt cache does not exist, building...'
-    # Default batch_size to 8 here for better user feedback during initial prompt processing
+    # Default batch_size to 64 here for better user feedback during initial prompt processing
     ./main 2>>"$LOG" \
-        --batch_size 8 \
+        --batch_size 64 \
         "${OPTS[@]}" \
         --prompt-cache "$PROMPT_CACHE_FILE" \
         --file "$CUR_PROMPT_FILE" \
@@ -132,7 +132,7 @@ while read -e line; do
     # HACK get num tokens from debug message
     # TODO get both messages in one go
     if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
-        ! sample_time_msg="$( tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
+        ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
         echo >&2 "Couldn't get number of tokens from ./main output!"
         exit 1
     fi

examples/chat.sh

@@ -11,6 +11,6 @@ cd ..
 #
 #   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
 #
-./main -m ./models/7B/ggml-model-q4_0.bin -c 512 -b 1024 -n 256 --keep 48 \
+./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
     --repeat_penalty 1.0 --color -i \
     -r "User:" -f prompts/chat-with-bob.txt

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

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "llama.h"
+#include "common.h"
 
 #include <unordered_map>
 #include <vector>
@@ -75,7 +76,7 @@ typedef struct {
     int seq_len; // max sequence length
 } Config;
 
-typedef struct {
+struct TransformerWeights {
     // token embedding table
     float* token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
@@ -97,9 +98,24 @@ typedef struct {
     // float* freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
     float* wcls;
-} TransformerWeights;
 
-void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
+    ~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;
+    }
+};
+
+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);
@@ -142,7 +158,7 @@ void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
     }
 }
 
-int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_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;
@@ -173,22 +189,7 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shar
     return 0;
 }
 
-void free_weights(TransformerWeights* w) {
-    delete w->token_embedding_table;
-    delete w->rms_att_weight;
-    delete w->rms_ffn_weight;
-    delete w->wq;
-    delete w->wk;
-    delete w->wv;
-    delete w->wo;
-    delete w->w1;
-    delete w->w2;
-    delete w->w3;
-    delete w->rms_final_weight;
-    if (w->wcls) delete w->wcls;
-}
-
-void print_sample_weights(TransformerWeights *w){
+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]);
@@ -323,7 +324,7 @@ struct train_params {
     int mem_compute1_gb;
 };
 
-void print_params(struct my_llama_hparams * params) {
+static void print_params(struct my_llama_hparams * params) {
     printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
     printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
     printf("%s: n_embd:  %d\n", __func__, params->n_embd);
@@ -334,7 +335,7 @@ void print_params(struct my_llama_hparams * params) {
     printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }
 
-void init_model(struct my_llama_model * model) {
+static void init_model(struct my_llama_model * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -407,17 +408,17 @@ void init_model(struct my_llama_model * model) {
     }
 }
 
-float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
     float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
     return *ptr;
 }
 
-int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
     int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
     return *ptr;
 }
 
-void print_row(struct ggml_tensor * probs, int i) {
+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);
@@ -425,7 +426,7 @@ void print_row(struct ggml_tensor * probs, int i) {
     printf("\n");
 }
 
-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
     assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
@@ -499,10 +500,10 @@ struct llama_file {
         errno = 0;
         std::size_t ret = std::fread(ptr, size, 1, fp);
         if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
+            die_fmt("fread failed: %s", strerror(errno));
         }
         if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+            die("unexpectedly reached end of file");
         }
     }
 
@@ -530,7 +531,7 @@ struct llama_file {
     }
 };
 
-bool is_ggml_file(const char *filename) {
+static bool is_ggml_file(const char * filename) {
     llama_file file(filename, "rb");
     if (file.size < 4) {
         return false;
@@ -539,7 +540,7 @@ bool is_ggml_file(const char *filename) {
     return magic == GGUF_MAGIC;
 }
 
-static std::string llama_escape_whitespaces(const std::string& text) {
+static std::string llama_escape_whitespaces(const std::string & text) {
     std::ostringstream out;
     for (char c : text) {
         if (c == ' ') out << "\xe2\x96\x81";
@@ -548,7 +549,7 @@ static std::string llama_escape_whitespaces(const std::string& text) {
     return out.str();
 }
 
-void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
     if (is_ggml_file(filename)) {
         struct ggml_context * ctx_data = NULL;
 
@@ -596,6 +597,9 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
         // assume llama2.c vocabulary
         printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
         llama_file file(filename, "rb");
+        if (!file.fp) {
+            die_fmt("%s: %s", strerror(errno), filename);
+        }
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
@@ -633,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
     }
 }
 
-void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * karpathy_weights){
+static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
     int ct;
     switch (gg_weights->n_dims){
         case 1:
@@ -669,14 +673,16 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * kar
     }
 }
 
-void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
-    // stuff AK weights into GG weights one by one.
+static void save_as_llama_model(
+    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+) {
+    // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
-    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
-    stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+    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);
 
-    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
     //print_row(model->norm, 0);
 
     // for rms-att-weight
@@ -686,18 +692,18 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
         // 1d
-        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
-        stuff_karpathy_weights_into_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
+        convert_weights_ak_to_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
+        convert_weights_ak_to_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
 
         // from 3d matrix layer x dim x dim to 2d matrix dim x dim
-        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
-        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+        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]);
 
-        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
-        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
-        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
+        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]);
+        convert_weights_ak_to_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
     }
 
     struct gguf_context * ctx = gguf_init_empty();
@@ -781,7 +787,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
     gguf_free(ctx);
 }
 
-struct train_params get_default_train_params() {
+static struct train_params get_default_train_params() {
     struct train_params params;
     params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
@@ -831,7 +837,7 @@ struct train_params get_default_train_params() {
     return params;
 }
 
-void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
+static void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
@@ -842,7 +848,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
     fprintf(stderr, "\n");
 }
 
-bool params_parse(int argc, char ** argv, struct train_params * params) {
+static bool params_parse(int argc, char ** argv, struct train_params * params) {
     bool invalid_param = false;
     bool reqd_param_found = false;
     std::string arg;
@@ -897,8 +903,8 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
     return true;
 }
 
-std::string basename(const std::string &path) {
-    size_t pos = path.find_last_of("/");
+static std::string basename(const std::string &path) {
+    size_t pos = path.find_last_of("/\\");
     if (pos == std::string::npos) {
         return path;
     }
@@ -911,7 +917,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
     Config config;
-    TransformerWeights weights;
+    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; }
@@ -953,6 +959,5 @@ int main(int argc, char ** argv) {
     printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
-    free_weights(&weights);
     return 0;
 }

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

@@ -1,8 +1,5 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
+#include "build-info.h"
+#include "common.h"
 #include "embd-input.h"
 
 #include <cassert>
@@ -23,11 +20,11 @@ extern "C" {
 struct MyModel* create_mymodel(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return nullptr;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = uint32_t(time(NULL));
@@ -51,8 +48,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
     struct MyModel * ret = new MyModel();
     ret->ctx = ctx;
@@ -74,7 +70,7 @@ bool eval_float(void * model, float * input, int N){
     MyModel * mymodel = (MyModel*)model;
     llama_context * ctx = mymodel->ctx;
     gpt_params params = mymodel->params;
-    int n_emb = llama_n_embd(ctx);
+    int n_emb = llama_n_embd(llama_get_model(ctx));
     int n_past = mymodel->n_past;
     int n_batch = N; // params.n_batch;
 
@@ -83,7 +79,8 @@ bool eval_float(void * model, float * input, int N){
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_eval_embd(ctx, (input+i*n_emb), n_eval, n_past, params.n_threads)) {
+        llama_batch batch = {  int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
+        if (llama_decode(ctx, batch)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
@@ -104,7 +101,7 @@ bool eval_tokens(void * model, std::vector<llama_token> tokens) {
         if (n_eval > params.n_batch) {
             n_eval = params.n_batch;
         }
-        if (llama_eval(ctx, &tokens[i], n_eval, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&tokens[i], n_eval, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
@@ -135,7 +132,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
 
     // out of user input, sample next token
     const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
+    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : params.top_k;
     const float   top_p           = params.top_p;
     const float   tfs_z           = params.tfs_z;
     const float   typical_p       = params.typical_p;
@@ -151,7 +148,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
     llama_token id = 0;
     {
         auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
+        auto n_vocab = llama_n_vocab(llama_get_model(ctx));
 
         // Apply params.logit_bias map
         for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
@@ -186,11 +183,11 @@ llama_token sampling_id(struct MyModel* mymodel) {
             if (mirostat == 1) {
                 static float mirostat_mu = 2.0f * mirostat_tau;
                 const int mirostat_m = 100;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
             } else if (mirostat == 2) {
                 static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
             } else {
                 // Temperature sampling
@@ -198,7 +195,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
                 llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
                 llama_sample_typical(ctx, &candidates_p, typical_p, 1);
                 llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token(ctx, &candidates_p);
             }
         }

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

@@ -8,7 +8,7 @@ int main(int argc, char** argv) {
     auto mymodel = create_mymodel(argc, argv);
     int N = 10;
     int max_tgt_len = 500;
-    int n_embd = llama_n_embd(mymodel->ctx);
+    int n_embd = llama_n_embd(llama_get_model(mymodel->ctx));
 
     // add random float embd to test evaluation
     float * data = new float[N*n_embd];

examples/embd-input/embd-input.h

@@ -3,7 +3,6 @@
 
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 extern "C" {
 

examples/embedding/README.md

@@ -1,3 +1,21 @@
-# embedding
+# llama.cpp/example/embedding
 
-TODO
+This example demonstrates generate high-dimensional embedding vector of a given text with llama.cpp.
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+```
+
+### Windows:
+
+```powershell
+embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+```
+
+The above command will output space-separated float values.

examples/embedding/embedding.cpp

@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <ctime>
 
@@ -11,18 +11,13 @@
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
     params.embedding = true;
 
-    if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
-                "expect poor results\n", __func__, params.n_ctx);
-    }
-
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -47,11 +42,18 @@ int main(int argc, char ** argv) {
         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, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
     int n_past = 0;
@@ -69,15 +71,15 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)params.n_ctx) {
+    if (embd_inp.size() > (size_t)n_ctx) {
         fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), params.n_ctx);
+                __func__, embd_inp.size(), n_ctx);
         return 1;
     }
 
     while (!embd_inp.empty()) {
         int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return 1;
         }
@@ -85,8 +87,8 @@ int main(int argc, char ** argv) {
         embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
     }
 
-    const int n_embd = llama_n_embd(ctx);
-    const auto embeddings = llama_get_embeddings(ctx);
+    const int n_embd = llama_n_embd(model);
+    const auto * embeddings = llama_get_embeddings(ctx);
 
     for (int i = 0; i < n_embd; i++) {
         printf("%f ", embeddings[i]);

examples/export-lora/CMakeLists.txt

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

@@ -0,0 +1,26 @@
+# export-lora
+
+Apply LORA adapters to base model and export the resulting model.
+
+```
+usage: export-lora [options]
+
+options:
+  -h, --help                         show this help message and exit
+  -m FNAME, --model-base FNAME       model path from which to load base model (default '')
+  -o FNAME, --model-out FNAME        path to save exported model (default '')
+  -l FNAME, --lora FNAME             apply LoRA adapter
+  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S
+  -t N, --threads N                  number of threads to use during computation (default: 4)
+```
+
+For example:
+
+```bash
+./bin/export-lora \
+    -m open-llama-3b-v2-q8_0.gguf \
+    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
+    -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
+```
+
+Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters.

examples/export-lora/export-lora.cpp

@@ -0,0 +1,474 @@
+
+#include "common.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#include <vector>
+#include <string>
+#include <thread>
+
+static const size_t tensor_alignment = 32;
+
+struct lora_info {
+    std::string filename;
+    float scale;
+};
+
+struct export_lora_params {
+    std::string fn_model_base;
+    std::string fn_model_out;
+    std::vector<struct lora_info> lora;
+    int n_threads;
+};
+
+struct lora_data {
+    struct lora_info     info;
+    std::vector<uint8_t> data;
+    struct ggml_context * ctx;
+
+    uint32_t lora_r;
+    uint32_t lora_alpha;
+};
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    bool eof() {
+        return tell() >= size;
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static struct export_lora_params get_default_export_lora_params() {
+    struct export_lora_params result;
+    result.fn_model_base = "";
+    result.fn_model_out  = "";
+    result.n_threads = GGML_DEFAULT_N_THREADS;
+    return result;
+}
+
+static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                         show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model-base FNAME       model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
+    fprintf(stderr, "  -o FNAME, --model-out FNAME        path to save exported model (default '%s')\n", params->fn_model_out.c_str());
+    fprintf(stderr, "  -l FNAME, --lora FNAME             apply LoRA adapter\n");
+    fprintf(stderr, "  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S\n");
+    fprintf(stderr, "  -t N, --threads N                  number of threads to use during computation (default: %d)\n", params->n_threads);
+}
+
+static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct export_lora_params default_params = get_default_export_lora_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-m" || arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "-o" || arg == "--model-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_out = argv[i];
+        } else if (arg == "-l" || arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            lora.scale = 1.0f;
+            params->lora.push_back(lora);
+        } else if (arg == "-s" || arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            lora.scale = std::stof(argv[i]);
+            params->lora.push_back(lora);
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_threads = std::stoi(argv[i]);
+            if (params->n_threads <= 0) {
+                params->n_threads = std::thread::hardware_concurrency();
+            }
+        } else {
+            fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
+            export_lora_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+
+    if (params->fn_model_base == default_params.fn_model_base) {
+        fprintf(stderr, "error: please specify a filename for model-base.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (params->fn_model_out == default_params.fn_model_out) {
+        fprintf(stderr, "error: please specify a filename for model-out.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    return true;
+}
+
+static void free_lora(struct lora_data * lora) {
+    if (lora->ctx != NULL) {
+        ggml_free(lora->ctx);
+    }
+    delete lora;
+}
+
+static struct lora_data * load_lora(struct lora_info * info) {
+    struct lora_data * result = new struct lora_data;
+    result->info = *info;
+    result->ctx = NULL;
+    result->lora_r     = 1;
+    result->lora_alpha = 1;
+
+    struct llama_file file(info->filename.c_str(), "rb");
+    if (file.fp == NULL) {
+        fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
+            info->filename.c_str());
+        free_lora(result);
+        return NULL;
+    }
+
+    struct ggml_init_params params_ggml;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
+    params_ggml.mem_buffer = NULL;
+    params_ggml.no_alloc   = true;
+    result->ctx = ggml_init(params_ggml);
+
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
+    uint32_t magic   = file.read_u32();
+    if (magic != LLAMA_FILE_MAGIC_LORA) {
+        die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
+    }
+    uint32_t version = file.read_u32();
+    if (version != 1) {
+        die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
+    }
+    result->lora_r     = file.read_u32();
+    result->lora_alpha = file.read_u32();
+    // read tensor infos from file
+    std::vector<char> name_buf;
+    std::vector<struct ggml_tensor *> tensors;
+    std::vector<size_t> tensors_offset;
+    size_t total_nbytes_pad = 0;
+    while(!file.eof()) {
+        int64_t ne[4]   = {1,1,1,1};
+        uint32_t n_dims  = file.read_u32();
+        uint32_t namelen = file.read_u32();
+        uint32_t type    = file.read_u32();
+        for (uint32_t k = 0; k < n_dims; ++k) {
+            ne[k] = (int64_t)file.read_u32();
+        }
+        name_buf.clear();
+        name_buf.resize(namelen + 1, '\0');
+        file.read_raw(name_buf.data(), namelen);
+        file.seek((0-file.tell()) & 31, SEEK_CUR);
+        size_t offset = file.tell();
+        struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
+        ggml_set_name(tensor, name_buf.data());
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        total_nbytes_pad += nbytes_pad;
+        tensors.push_back(tensor);
+        tensors_offset.push_back(offset);
+        file.seek(nbytes, SEEK_CUR);
+    }
+    // read tensor data
+    result->data.resize(total_nbytes_pad);
+    size_t data_offset = 0;
+    for (size_t i = 0; i < tensors.size(); ++i) {
+        struct ggml_tensor * tensor = tensors[i];
+        size_t offset     = tensors_offset[i];
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        file.seek(offset, SEEK_SET);
+        tensor->data = result->data.data() + data_offset;
+        file.read_raw(tensor->data, nbytes);
+        data_offset += nbytes_pad;
+    }
+    return result;
+}
+
+
+static struct ggml_cgraph * build_graph_lora(
+    struct ggml_context * ctx,
+    struct ggml_tensor * tensor,
+    struct ggml_tensor * lora_a,
+    struct ggml_tensor * lora_b,
+    float scaling
+) {
+    struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
+    if (scaling != 1.0f) {
+        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+    }
+    struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
+
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand (gf, res);
+    return gf;
+}
+
+static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
+    if (lora->ctx == NULL) {
+        return false;
+    }
+    std::string name = ggml_get_name(tensor);
+    std::string name_a = name + std::string(".loraA");
+    std::string name_b = name + std::string(".loraB");
+    struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
+    struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
+    if (lora_a == NULL || lora_b == NULL) {
+        return false;
+    }
+
+    float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
+
+    struct ggml_init_params params;
+    params.mem_size   = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_buffer = NULL;
+    params.no_alloc   = true;
+    struct ggml_context * ctx = NULL;
+    struct ggml_allocr * alloc = NULL;
+    struct ggml_cgraph * gf = NULL;
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+    ggml_free(ctx);
+
+    static std::vector<uint8_t> data_compute;
+    data_compute.resize(alloc_size + tensor_alignment);
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
+    static std::vector<uint8_t> data_work;
+    data_work.resize(cplan.work_size);
+    cplan.work_data = data_work.data();
+
+    ggml_graph_compute(gf, &cplan);
+
+    ggml_free(ctx);
+    return true;
+}
+
+static void export_lora(struct export_lora_params * params) {
+    // load all loras
+    std::vector<struct lora_data *> loras;
+    for (size_t i = 0; i < params->lora.size(); ++i) {
+        struct lora_data * lora = load_lora(&params->lora[i]);
+        if (lora != NULL) {
+            loras.push_back(lora);
+        }
+    }
+    if (loras.size() == 0) {
+        fprintf(stderr, "warning: no lora adapters will be applied.\n");
+    }
+
+    // open input file
+    struct llama_file fin(params->fn_model_base.c_str(), "rb");
+    if (!fin.fp) {
+        die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
+    }
+
+    // open base model gguf, read tensors without their data
+    struct ggml_context * ctx_in;
+    struct gguf_init_params params_gguf;
+    params_gguf.no_alloc = true;
+    params_gguf.ctx      = &ctx_in;
+    struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
+
+    // create new gguf
+    struct gguf_context * gguf_out = gguf_init_empty();
+
+    // copy meta data from base model: kv and tensors
+    gguf_set_kv(gguf_out, gguf_in);
+    int n_tensors = gguf_get_n_tensors(gguf_in);
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+        gguf_add_tensor(gguf_out, tensor);
+    }
+
+    // create output file
+    struct llama_file fout(params->fn_model_out.c_str(), "wb");
+    if (!fout.fp) {
+        die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
+    }
+
+    // write gguf meta data
+    std::vector<uint8_t> meta;
+    meta.resize(gguf_get_meta_size(gguf_out));
+    gguf_get_meta_data(gguf_out, meta.data());
+    fout.write_raw(meta.data(), meta.size());
+
+    std::vector<uint8_t> data;
+    std::vector<uint8_t> padding;
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+
+        // read tensor data
+        data.resize(ggml_nbytes(tensor));
+        tensor->data = data.data();
+        size_t offset = gguf_get_tensor_offset(gguf_in, i);
+        fin.seek(offset + meta.size(), SEEK_SET);
+        fin.read_raw(data.data(), data.size());
+
+        // apply all loras
+        for (size_t k = 0; k < loras.size(); ++k) {
+            apply_lora(tensor, loras[k], params->n_threads);
+        }
+
+        // write tensor data + padding
+        padding.clear();
+        padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
+
+        GGML_ASSERT(fout.tell() == offset + meta.size());
+        // fout.seek(offset + meta.size(), SEEK_SET);
+        fout.write_raw(data.data(), data.size());
+        fout.write_raw(padding.data(), padding.size());
+
+        if (i % 2 == 0) {
+            printf(".");
+        }
+    }
+    printf("\n");
+
+    // close gguf
+    gguf_free(gguf_out);
+    gguf_free(gguf_in);
+
+    // free loras
+    for (size_t i = 0; i < loras.size(); ++i) {
+        free_lora(loras[i]);
+    }
+}
+
+int main(int argc, char ** argv) {
+    struct export_lora_params params = get_default_export_lora_params();
+
+    if (!export_lora_params_parse(argc, argv, &params)) {
+        return 1;
+    }
+
+    export_lora(&params);
+
+    return 0;
+}

examples/finetune/CMakeLists.txt

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

@@ -0,0 +1,90 @@
+# finetune
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# finetune LORA adapter
+./bin/finetune \
+        --model-base open-llama-3b-v2-q8_0.gguf \
+        --checkpoint-in  chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \
+        --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \
+        --train-data "shakespeare.txt" \
+        --save-every 10 \
+        --threads 6 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing
+
+# predict
+./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+```
+
+Finetune output files will be saved every N iterations (config with `--save-every N`).
+The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
+So in above example after 10 iterations these files will be written:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+After 10 more iterations:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter.
+
+llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`.
+These LORA adapters can then be used by `main` together with the base model, like in the 'predict' example command above.
+
+In `main` you can also load multiple LORA adapters, which will then be mixed together.
+
+For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \
+  --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin
+```
+
+You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`.
+
+For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \
+  --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
+```
+
+The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values to big will sometimes result in worse output. Play around to find good values.
+
+Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
+If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.
+
+The default LORA rank can be specified with `--lora-r N`.
+The LORA rank can be configured for each model tensor type separately with these command line options:
+
+```bash
+  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4)
+  --rank-att-norm N          LORA rank for attention norm tensor (default 1)
+  --rank-ffn-norm N          LORA rank for feed-forward norm tensor (default 1)
+  --rank-out-norm N          LORA rank for output norm tensor (default 1)
+  --rank-tok-embd N          LORA rank for token embeddings tensor (default 4)
+  --rank-out N               LORA rank for output tensor (default 4)
+  --rank-wq N                LORA rank for wq tensor (default 4)
+  --rank-wk N                LORA rank for wk tensor (default 4)
+  --rank-wv N                LORA rank for wv tensor (default 4)
+  --rank-wo N                LORA rank for wo tensor (default 4)
+  --rank-w1 N                LORA rank for w1 tensor (default 4)
+  --rank-w2 N                LORA rank for w2 tensor (default 4)
+  --rank-w3 N                LORA rank for w3 tensor (default 4)
+```
+
+The LORA rank of 'norm' tensors should always be 1.
+
+To see all available options use `finetune --help`.

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

@@ -0,0 +1,489 @@
+#!/usr/bin/env python3
+# finetune checkpoint --> gguf conversion
+
+import argparse
+import gguf
+import os
+import struct
+import sys
+import numpy as np
+from pathlib import Path
+
+# gguf constants
+LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
+LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"
+LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
+LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version"
+LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count"
+LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count"
+LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count"
+LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized"
+LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss"
+LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss"
+LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count"
+LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
+LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end"
+LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
+
+LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
+
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction"
+LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"
+
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"
+
+LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output"
+LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm"
+LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q"
+LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k"
+LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v"
+LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output"
+LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm"
+LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate"
+LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down"
+LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up"
+
+class Tensor:
+    def __init__(self, dtype='f', ne=None):
+        if ne is None:
+            ne = []
+        self.dtype = dtype
+        self.ne = ne
+        self.nbytes = 0
+        if self.dtype == 'f':
+            if len(self.ne) == 0:
+                self.nbytes = 0
+            else:
+                self.nbytes = int(np.product(self.ne)) * 4
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+    def load(self, data, offset):
+        nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        assert(nd == len(self.ne))
+        ne = []
+        for d in range(nd):
+            n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+            ne.append(n)
+
+        if tuple(ne) != tuple(self.ne):
+            raise ValueError(f"Tensor.load: Expected number of elements {str(self.ne)} does not match what is read from file {str(ne)}")
+
+        if self.dtype == 'f':
+            assert(dtype == 0)
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+        self.name = bytes(data[offset:offset+namelen]); offset += namelen
+        # 32-byte alignment
+        offset += (0 - offset) & 31
+        self.data = data[offset:offset+self.nbytes]
+        offset += self.nbytes
+        return offset
+
+    def max_storage_size(self):
+        result = 0
+        result += 4 # nd
+        result += 4 # namelen
+        result += 4 # dtype
+        result += len(self.ne)*8 # ne
+        result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
+        result += 31 # 32-byte alignment
+        result += self.nbytes
+        return result
+
+    def save_gguf(self, gguf_writer, name):
+        gguf_writer.add_tensor(
+            name=name,
+            tensor=self.data,
+            raw_shape=np.array(list(reversed(self.ne))),
+            raw_dtype=gguf.GGMLQuantizationType.F32)
+
+class OptimizationContext:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
+        offset += 4
+
+        if self.version != 1:
+            raise ValueError('Invalid version of optimization context in checkpoint file')
+
+        self.past    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.nx      = struct.unpack('N',  bytes(data[offset:offset + 8]))[0];  offset += 8
+        self.iter    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+
+        self.adam_m  = Tensor('f', [self.nx])
+        self.adam_v  = Tensor('f', [self.nx])
+        self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+        self.lbfgs_x    = Tensor('f', [self.nx])
+        self.lbfgs_xp   = Tensor('f', [self.nx])
+        self.lbfgs_g    = Tensor('f', [self.nx])
+        self.lbfgs_gp   = Tensor('f', [self.nx])
+        self.lbfgs_d    = Tensor('f', [self.nx])
+        self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+        self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+        self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+        # forgot to save type in version 1:
+        # guess self.type from number of remaining bytes
+        size_type_0 = 12 + sum([t.max_storage_size() for t in
+                                [self.adam_m, self.adam_v]
+                                +([self.adam_pf] if (self.past > 0) else [])])
+        size_type_1 = 24 + sum([t.max_storage_size() for t in
+                                [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
+                                 self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
+                                 self.lbfgs_lmal, self.lbfgs_lmys,
+                                 self.lbfgs_lms, self.lbfgs_lmy]
+                                 +([self.lbfgs_pf] if (self.past > 0) else [])])
+        # due to alignment padding the size might not by exact
+        # but the difference in size for both types is significant,
+        # so we can just use whichever is closest
+        remaining = len(data) - offset
+        if abs(remaining - size_type_0) < abs(remaining - size_type_1):
+            self.type = 0
+        else:
+            self.type = 1
+
+        if self.type == 0:
+            offset = self.adam_m.load(data, offset)
+            offset = self.adam_v.load(data, offset)
+            offset = self.adam_pf.load(data,offset)
+
+            self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        elif self.type == 1:
+            offset = self.lbfgs_x.load(data, offset)
+            offset = self.lbfgs_xp.load(data, offset)
+            offset = self.lbfgs_g.load(data, offset)
+            offset = self.lbfgs_gp.load(data, offset)
+            offset = self.lbfgs_d.load(data, offset)
+            offset = self.lbfgs_pf.load(data, offset)
+            offset = self.lbfgs_lmal.load(data, offset)
+            offset = self.lbfgs_lmys.load(data, offset)
+            offset = self.lbfgs_lms.load(data, offset)
+            offset = self.lbfgs_lmy.load(data, offset)
+
+            self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        else:
+            raise ValueError(f"Invalid optimizer type '{self.type}'")
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
+        gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
+        gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
+
+        if self.type == 0:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
+
+            self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
+            self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
+            if self.past > 0:
+                self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
+
+        elif self.type == 1:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
+
+            self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
+            self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
+            self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
+            self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
+            self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
+            if self.past > 0:
+                self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
+            self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
+            self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
+            self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
+            self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
+        else:
+            raise ValueError('Unknown optimizer type')
+
+class LoraParams:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.n_rank_attention_norm  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wq              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wk              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wv              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wo              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_ffn_norm        = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w1              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w2              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w3              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_tok_embeddings  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_norm            = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_output          = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,  self.n_rank_tok_embeddings)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, self.n_rank_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT,      self.n_rank_output)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,   self.n_rank_attention_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_Q,      self.n_rank_wq)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_K,      self.n_rank_wk)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_V,      self.n_rank_wv)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,    self.n_rank_wo)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_NORM,    self.n_rank_ffn_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_GATE,    self.n_rank_w1)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,    self.n_rank_w2)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_UP,      self.n_rank_w3)
+
+class ModelParams:
+    def __init__(self, n_ff = None):
+        self.n_ff = n_ff
+
+    def load(self, data, offset):
+        self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_embd  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_mult  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_head  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rot   = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def get_n_ff(self):
+        if self.n_ff is None:
+            # struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
+            return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
+        else:
+            return self.n_ff
+
+    def save_gguf(self, gguf_writer):
+        # self.n_vocab not saved
+        gguf_writer.add_embedding_length(self.n_embd)
+        gguf_writer.add_head_count(self.n_head)
+        gguf_writer.add_block_count(self.n_layer)
+        gguf_writer.add_rope_dimension_count(self.n_rot)
+        gguf_writer.add_feed_forward_length(self.get_n_ff())
+
+def tensor_name(key, bid=None, suffix=".weight"):
+    return gguf.TENSOR_NAMES[key].format(bid=bid) + suffix
+
+class Layer:
+    def __init__(self, params, lora_params, bid):
+        self.bid = bid
+        self.att_norm_a = Tensor('f', [lora_params.n_rank_attention_norm, params.n_embd])
+        self.att_norm_b = Tensor('f', [lora_params.n_rank_attention_norm, 1])
+        self.wq_a       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wq_b       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wk_a       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wk_b       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wv_a       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wv_b       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wo_a       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.wo_b       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.ffn_norm_a = Tensor('f', [lora_params.n_rank_ffn_norm, params.n_embd])
+        self.ffn_norm_b = Tensor('f', [lora_params.n_rank_ffn_norm, 1])
+        self.w1_a       = Tensor('f', [lora_params.n_rank_w1, params.n_embd])
+        self.w1_b       = Tensor('f', [lora_params.n_rank_w1, params.get_n_ff()])
+        self.w2_a       = Tensor('f', [lora_params.n_rank_w2, params.get_n_ff()])
+        self.w2_b       = Tensor('f', [lora_params.n_rank_w2, params.n_embd])
+        self.w3_a       = Tensor('f', [lora_params.n_rank_w3, params.n_embd])
+        self.w3_b       = Tensor('f', [lora_params.n_rank_w3, params.get_n_ff()])
+
+    def load(self, data, offset):
+        offset = self.att_norm_a.load(data, offset)
+        offset = self.att_norm_b.load(data, offset)
+        offset = self.wq_a.load(data, offset)
+        offset = self.wq_b.load(data, offset)
+        offset = self.wk_a.load(data, offset)
+        offset = self.wk_b.load(data, offset)
+        offset = self.wv_a.load(data, offset)
+        offset = self.wv_b.load(data, offset)
+        offset = self.wo_a.load(data, offset)
+        offset = self.wo_b.load(data, offset)
+        offset = self.ffn_norm_a.load(data, offset)
+        offset = self.ffn_norm_b.load(data, offset)
+        offset = self.w1_a.load(data, offset)
+        offset = self.w1_b.load(data, offset)
+        offset = self.w2_a.load(data, offset)
+        offset = self.w2_b.load(data, offset)
+        offset = self.w3_a.load(data, offset)
+        offset = self.w3_b.load(data, offset)
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.att_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_a"))
+        self.att_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_b"))
+        self.wq_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_a"))
+        self.wq_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_b"))
+        self.wk_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_a"))
+        self.wk_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_b"))
+        self.wv_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_a"))
+        self.wv_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_b"))
+        self.wo_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_a"))
+        self.wo_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_b"))
+        self.ffn_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_a"))
+        self.ffn_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_b"))
+        self.w1_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_a"))
+        self.w1_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_b"))
+        self.w2_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_a"))
+        self.w2_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_b"))
+        self.w3_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_a"))
+        self.w3_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_b"))
+
+class LoraModel:
+    def __init__(self, n_ff = None):
+        self.params = ModelParams(n_ff = n_ff)
+        self.lora_params = LoraParams()
+        self.layers = []
+
+    def load(self, data, offset):
+        offset = self.params.load(data, offset)
+        offset = self.lora_params.load(data, offset)
+
+        self.tok_embd_a = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_embd])
+        self.tok_embd_b = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_vocab])
+        self.norm_a     = Tensor('f', [self.lora_params.n_rank_norm, self.params.n_embd])
+        self.norm_b     = Tensor('f', [self.lora_params.n_rank_norm, 1])
+        self.output_a   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_embd])
+        self.output_b   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_vocab])
+
+        offset = self.tok_embd_a.load(data, offset)
+        offset = self.tok_embd_b.load(data, offset)
+        offset = self.norm_a.load(data, offset)
+        offset = self.norm_b.load(data, offset)
+        offset = self.output_a.load(data, offset)
+        offset = self.output_b.load(data, offset)
+
+        self.layers.clear()
+        for bid in range(self.params.n_layer):
+            layer = Layer(self.params, self.lora_params, bid)
+            offset = layer.load(data, offset)
+            self.layers.append(layer)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.params.save_gguf(gguf_writer)
+        self.lora_params.save_gguf(gguf_writer)
+
+        self.tok_embd_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_a"))
+        self.tok_embd_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_b"))
+        self.norm_a.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_a"))
+        self.norm_b.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_b"))
+        self.output_a.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_a"))
+        self.output_b.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_b"))
+
+        for layer in self.layers:
+            layer.save_gguf(gguf_writer)
+
+class LoraCheckpoint:
+    def __init__(self, n_ff = None):
+        self.model = LoraModel(n_ff = n_ff)
+        self.opt_ctx = OptimizationContext()
+
+    def load(self, data, offset):
+        magic   = bytes(reversed(data[offset:offset + 4])); offset += 4
+        if magic != b'ggcl':
+            raise ValueError(f"File header magic indicates, that this is no finetune-lora checkpoint file. Expected 'ggcl', Got '{str(magic)}'")
+
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        if self.version != 0:
+            raise ValueError('Invalid version of checkpoint file')
+
+        self.train_its     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_tokens  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        offset = self.model.load(data, offset)
+        offset = self.opt_ctx.load(data, offset)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
+        gguf_writer.add_layer_norm_rms_eps(1e-5)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_FINETUNE_LORA)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
+        self.model.save_gguf(gguf_writer)
+        self.opt_ctx.save_gguf(gguf_writer)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert finetune checkpoints to GGUF')
+    parser.add_argument('--input',  '-i', type = Path, help = 'Input finetune checkpoint filename', required=True)
+    parser.add_argument('--output', '-o', type = Path, help = 'Output GGUF filename', required=True)
+    parser.add_argument('--ff', type = int, help = "Feedforward size, if not provided compute from n_mult. Provide this if you get 'ValueError: Tensor.load: Expected number of elements does not match what is read from file'", required=False)
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    print(cfg)
+    data = np.memmap(cfg.input, mode = 'r')
+    chk = LoraCheckpoint(n_ff = cfg.ff)
+    offset = 0
+    offset = chk.load(data, offset)
+    # we should have read all available data
+    assert(offset == len(data))
+
+    gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+    chk.save_gguf(gguf_writer)
+    print("    gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("    gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("    gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+    gguf_writer.close()
+
+if __name__ == '__main__':
+    main()

examples/gguf/gguf.cpp

@@ -13,14 +13,14 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-template<typename T>
+template <typename T>
 static std::string to_string(const T & val) {
     std::stringstream ss;
     ss << val;
     return ss.str();
 }
 
-bool gguf_ex_write(const std::string & fname) {
+static bool gguf_ex_write(const std::string & fname) {
     struct gguf_context * ctx = gguf_init_empty();
 
     gguf_set_val_u8  (ctx, "some.parameter.uint8",    0x12);
@@ -76,7 +76,7 @@ bool gguf_ex_write(const std::string & fname) {
 
     gguf_write_to_file(ctx, fname.c_str(), false);
 
-    fprintf(stdout, "%s: wrote file '%s;\n", __func__, fname.c_str());
+    printf("%s: wrote file '%s;\n", __func__, fname.c_str());
 
     ggml_free(ctx_data);
     gguf_free(ctx);
@@ -85,7 +85,7 @@ bool gguf_ex_write(const std::string & fname) {
 }
 
 // just read tensor info
-bool gguf_ex_read_0(const std::string & fname) {
+static bool gguf_ex_read_0(const std::string & fname) {
     struct gguf_init_params params = {
         /*.no_alloc = */ false,
         /*.ctx      = */ NULL,
@@ -93,20 +93,20 @@ bool gguf_ex_read_0(const std::string & fname) {
 
     struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
 
-    fprintf(stdout, "%s: version:      %d\n", __func__, gguf_get_version(ctx));
-    fprintf(stdout, "%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
-    fprintf(stdout, "%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
+    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
+    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
+    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
 
     // kv
     {
         const int n_kv = gguf_get_n_kv(ctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
 
@@ -116,10 +116,10 @@ bool gguf_ex_read_0(const std::string & fname) {
 
         const int keyidx = gguf_find_key(ctx, findkey);
         if (keyidx == -1) {
-            fprintf(stdout, "%s: find key: %s not found.\n", __func__, findkey);
+            printf("%s: find key: %s not found.\n", __func__, findkey);
         } else {
             const char * key_value = gguf_get_val_str(ctx, keyidx);
-            fprintf(stdout, "%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);
+            printf("%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);
         }
     }
 
@@ -127,13 +127,13 @@ bool gguf_ex_read_0(const std::string & fname) {
     {
         const int n_tensors = gguf_get_n_tensors(ctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
 
@@ -143,7 +143,7 @@ bool gguf_ex_read_0(const std::string & fname) {
 }
 
 // read and create ggml_context containing the tensors and their data
-bool gguf_ex_read_1(const std::string & fname) {
+static bool gguf_ex_read_1(const std::string & fname) {
     struct ggml_context * ctx_data = NULL;
 
     struct gguf_init_params params = {
@@ -153,20 +153,20 @@ bool gguf_ex_read_1(const std::string & fname) {
 
     struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
 
-    fprintf(stdout, "%s: version:      %d\n", __func__, gguf_get_version(ctx));
-    fprintf(stdout, "%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
-    fprintf(stdout, "%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
+    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
+    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
+    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
 
     // kv
     {
         const int n_kv = gguf_get_n_kv(ctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
 
@@ -174,13 +174,13 @@ bool gguf_ex_read_1(const std::string & fname) {
     {
         const int n_tensors = gguf_get_n_tensors(ctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
 
@@ -189,13 +189,13 @@ bool gguf_ex_read_1(const std::string & fname) {
         const int n_tensors = gguf_get_n_tensors(ctx);
 
         for (int i = 0; i < n_tensors; ++i) {
-            fprintf(stdout, "%s: reading tensor %d data\n", __func__, i);
+            printf("%s: reading tensor %d data\n", __func__, i);
 
             const char * name = gguf_get_tensor_name(ctx, i);
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            fprintf(stdout, "%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;
@@ -219,7 +219,7 @@ bool gguf_ex_read_1(const std::string & fname) {
         }
     }
 
-    fprintf(stdout, "%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));
+    printf("%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));
 
     ggml_free(ctx_data);
     gguf_free(ctx);
@@ -229,7 +229,7 @@ bool gguf_ex_read_1(const std::string & fname) {
 
 int main(int argc, char ** argv) {
     if (argc < 3) {
-        fprintf(stdout, "usage: %s data.gguf r|w\n", argv[0]);
+        printf("usage: %s data.gguf r|w\n", argv[0]);
         return -1;
     }
 

examples/gptneox-wip/falcon-main.cpp

@@ -305,9 +305,9 @@ struct ggml_tensor * get_tensor_ex( struct ggml_context * ctx, std::string name)
 
     struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
     if( cur == NULL ) {
-        fprintf(stdout, "%s: tensor '%s' not found!\n", __func__, name.c_str());
+        printf("%s: tensor '%s' not found!\n", __func__, name.c_str());
     } else {
-//        fprintf(stdout, "%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
+//        printf("%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
     }
 
     return cur;
@@ -333,21 +333,21 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         return false;
     }
 
-    fprintf(stdout, "%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    fprintf(stdout, "%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    fprintf(stdout, "%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
+    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
+    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
+    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
 
     // print all kv
     #if 0
     {
         const int n_kv = gguf_get_n_kv(ggufctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ggufctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
     #endif
@@ -357,21 +357,21 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         int keyidx;
 
         keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { fprintf(stdout, "%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { fprintf(stdout, "%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { fprintf(stdout, "%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { fprintf(stdout, "%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { fprintf(stdout, "%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { fprintf(stdout, "%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { fprintf(stdout, "%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
-        if (keyidx != -1) { fprintf(stdout, "%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
+        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
     }
 
     // check required metadata
@@ -382,11 +382,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "general.architecture");
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "falcon") != 0) {
-                fprintf(stdout, "%s: model architecture not supported!\n", __func__);
+                printf("%s: model architecture not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: gguf model architecture not found!\n", __func__);
+            printf("%s: gguf model architecture not found!\n", __func__);
             return false;
         }
 
@@ -394,11 +394,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "falcon.tensor_data_layout");
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "jploski") != 0) {
-                fprintf(stdout, "%s: model tensor data layout not supported!\n", __func__);
+                printf("%s: model tensor data layout not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: gguf model tensor data layout not found!\n", __func__);
+            printf("%s: gguf model tensor data layout not found!\n", __func__);
             return false;
         }
 
@@ -455,11 +455,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
 
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
-                fprintf(stdout, "%s: tokenizer model not supported!\n", __func__);
+                printf("%s: tokenizer model not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: tokenizer model not found!\n", __func__);
+            printf("%s: tokenizer model not found!\n", __func__);
             return false;
         }
 
@@ -467,22 +467,22 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
 
         if (tokens_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            printf("%s: gpt2 tokenizer vocab not found!\n", __func__);
             return false;
         }
 
         int merges_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.merges");
 
         if (merges_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer merges not found!\n", __func__);
+            printf("%s: gpt2 tokenizer merges not found!\n", __func__);
             return false;
         }
 
         hparams.n_vocab = gguf_get_arr_n(ggufctx,tokens_keyidx);
         hparams.n_merges = gguf_get_arr_n(ggufctx,merges_keyidx);
 
-        fprintf(stdout, "%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
-        fprintf(stdout, "%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
+        printf("%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
+        printf("%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
 
         for (size_t i = 0; i < hparams.n_vocab; i++) {
             std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
@@ -523,12 +523,12 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) {   vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
 
-        if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
-        if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
-        if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
-        if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
-        if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
-        if( vocab.linefeed_id    != -1 ) { fprintf(stdout, "%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
+        if( vocab.special_bos_id != -1 ) { printf("%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
+        if( vocab.special_eos_id != -1 ) { printf("%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
+        if( vocab.special_unk_id != -1 ) { printf("%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
+        if( vocab.special_sep_id != -1 ) { printf("%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
+        if( vocab.special_pad_id != -1 ) { printf("%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
+        if( vocab.linefeed_id    != -1 ) { printf("%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
 
     }
 
@@ -543,13 +543,13 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
     {
         const int n_tensors = gguf_get_n_tensors(ggufctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ggufctx, i);
             const size_t offset = gguf_get_tensor_offset(ggufctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
     #endif
@@ -953,7 +953,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 

examples/gptneox-wip/gptneox-main.cpp

@@ -318,9 +318,9 @@ struct ggml_tensor * get_tensor_ex( struct ggml_context * ctx, std::string name)
 
     struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
     if( cur == NULL ) {
-        fprintf(stdout, "%s: tensor '%s' not found!\n", __func__, name.c_str());
+        printf("%s: tensor '%s' not found!\n", __func__, name.c_str());
     } else {
-//        fprintf(stdout, "%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
+//        printf("%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
     }
 
     return cur;
@@ -346,21 +346,21 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         return false;
     }
 
-    fprintf(stdout, "%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    fprintf(stdout, "%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    fprintf(stdout, "%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
+    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
+    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
+    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
 
     // print all kv
     #if 0
     {
         const int n_kv = gguf_get_n_kv(ggufctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ggufctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
     #endif
@@ -370,21 +370,21 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         int keyidx;
 
         keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { fprintf(stdout, "%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { fprintf(stdout, "%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { fprintf(stdout, "%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { fprintf(stdout, "%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { fprintf(stdout, "%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { fprintf(stdout, "%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { fprintf(stdout, "%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
-        if (keyidx != -1) { fprintf(stdout, "%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
+        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
     }
 
     // check required metadata
@@ -395,11 +395,11 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         keyidx = gguf_find_key(ggufctx, "general.architecture");
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gptneox") != 0) {
-                fprintf(stdout, "%s: model architecture not supported!\n", __func__);
+                printf("%s: model architecture not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: gguf model architecture not found!\n", __func__);
+            printf("%s: gguf model architecture not found!\n", __func__);
             return false;
         }
 
@@ -456,11 +456,11 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
 
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
-                fprintf(stdout, "%s: tokenizer model not supported!\n", __func__);
+                printf("%s: tokenizer model not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: tokenizer model not found!\n", __func__);
+            printf("%s: tokenizer model not found!\n", __func__);
             return false;
         }
 
@@ -468,22 +468,22 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
 
         if (tokens_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            printf("%s: gpt2 tokenizer vocab not found!\n", __func__);
             return false;
         }
 
         int merges_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.merges");
 
         if (merges_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer merges not found!\n", __func__);
+            printf("%s: gpt2 tokenizer merges not found!\n", __func__);
             return false;
         }
 
         hparams.n_vocab = gguf_get_arr_n(ggufctx,tokens_keyidx);
         hparams.n_merges = gguf_get_arr_n(ggufctx,merges_keyidx);
 
-        fprintf(stdout, "%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
-        fprintf(stdout, "%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
+        printf("%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
+        printf("%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
 
         for (size_t i = 0; i < hparams.n_vocab; i++) {
             std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
@@ -524,12 +524,12 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) {   vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
 
-        if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
-        if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
-        if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
-        if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
-        if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
-        if( vocab.linefeed_id    != -1 ) { fprintf(stdout, "%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
+        if( vocab.special_bos_id != -1 ) { printf("%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
+        if( vocab.special_eos_id != -1 ) { printf("%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
+        if( vocab.special_unk_id != -1 ) { printf("%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
+        if( vocab.special_sep_id != -1 ) { printf("%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
+        if( vocab.special_pad_id != -1 ) { printf("%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
+        if( vocab.linefeed_id    != -1 ) { printf("%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
     }
 
 
@@ -543,13 +543,13 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
     {
         const int n_tensors = gguf_get_n_tensors(ggufctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ggufctx, i);
             const size_t offset = gguf_get_tensor_offset(ggufctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
     #endif
@@ -660,9 +660,10 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
 ggml_tensor * gpt_neox_ff(
         const gpt_neox_block &block,
         ggml_context * ctx0,
-        ggml_tensor * inp) {
+        ggml_tensor * inp,
+        const gpt_neox_hparams &hparams) {
 
-    ggml_tensor * cur = ggml_norm(ctx0, inp);
+    ggml_tensor * cur = ggml_norm(ctx0, inp, hparams.norm_eps);
 
     cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, block.ln_2_g, cur), cur), ggml_repeat(ctx0, block.ln_2_b, cur));
     cur = ggml_mul_mat(ctx0, block.c_mlp_fc_w, cur);
@@ -753,7 +754,7 @@ bool gpt_neox_eval(
         // self-attention
         {
             {
-                cur = ggml_norm(ctx0, inpL);
+                cur = ggml_norm(ctx0, inpL, hparams.norm_eps);
 
                 cur = ggml_add(ctx0,
                         ggml_mul(ctx0, ggml_repeat(ctx0, model.blocks[il].ln_1_g, cur), cur),
@@ -844,7 +845,7 @@ bool gpt_neox_eval(
         if (hparams.par_res == 0) {
             struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
 
-            cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF);
+            cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF, hparams);
 
             // input for next layer
             inpL = ggml_add(ctx0, cur, inpFF);
@@ -853,7 +854,7 @@ bool gpt_neox_eval(
 
             // this is independent of the self-attention result, so it could be done in parallel to the self-attention
             // note here we pass inpL instead of cur
-            cur = gpt_neox_ff(model.blocks[il], ctx0, inpL);
+            cur = gpt_neox_ff(model.blocks[il], ctx0, inpL, hparams);
 
             // layer input + FF
             cur  = ggml_add(ctx0, cur, inpFF);
@@ -867,7 +868,7 @@ bool gpt_neox_eval(
 
     // norm
     {
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, hparams.norm_eps);
 
         // inpL = ln_f_g*inpL + ln_f_b
         inpL = ggml_add(ctx0,
@@ -924,7 +925,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 

examples/infill/CMakeLists.txt

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

examples/infill/README.md

@@ -0,0 +1,41 @@
+# llama.cpp/example/infill
+
+This example shows how to use the infill mode with Code Llama models supporting infill mode.
+Currently the 7B and 13B models support infill mode.
+
+Infill supports most of the options available in the main example.
+
+For further information have a look at the main README.md in llama.cpp/example/main/README.md
+
+## Common Options
+
+In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models:
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+
+## Input Prompts
+
+The `infill` program provides several ways to interact with the LLaMA models using input prompts:
+
+-   `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option.
+-   `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option.
+-   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
+
+## Interaction
+
+The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first`
+
+### Interaction Options
+
+-   `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model.
+-   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
+-   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
+
+### Example
+
+```bash
+./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
+```

examples/infill/infill.cpp

@@ -0,0 +1,769 @@
+#include "common.h"
+
+#include "console.h"
+#include "llama.h"
+#include "build-info.h"
+#include "grammar-parser.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static gpt_params               * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
+static bool is_interacting = false;
+
+
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = get_sortable_timestamp();
+
+    const bool success = create_directory_with_parents(params.logdir);
+    if (!success) {
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: infill\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+
+    llama_dump_timing_info_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (!is_interacting) {
+            is_interacting = true;
+        } else {
+            console::cleanup();
+            printf("\n");
+            llama_print_timings(*g_ctx);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+            _exit(130);
+        }
+    }
+}
+#endif
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    g_params = &params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("infill", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    if (params.logits_all) {
+        printf("\n************\n");
+        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.embedding) {
+        printf("\n************\n");
+        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+    if (params.instruct) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for instruct mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.antiprompt.empty()) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
+        printf("\n************\n");
+        printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (params.random_prompt) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for random prompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.path_prompt_cache.empty()) {
+        printf("\n************\n");
+        printf("%s: infill does not support prompt caching\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    LOG("%s: llama backend init\n", __func__);
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+    llama_context * ctx_guidance = NULL;
+    g_model = &model;
+    g_ctx = &ctx;
+
+    // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (params.cfg_scale > 1.f) {
+        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
+        ctx_guidance = llama_new_context_with_model(model, lparams);
+    }
+
+    if (model == NULL) {
+        LOG_TEE("%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);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
+    // print system information
+    {
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", get_system_info(params).c_str());
+    }
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
+
+    std::vector<llama_token> embd_inp;
+    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    embd_inp = inp_pfx;
+    embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+    embd_inp.push_back(llama_token_middle(ctx));
+
+    LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
+    LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    }
+
+    // Tokenize negative prompt
+    std::vector<llama_token> guidance_inp;
+    int guidance_offset = 0;
+    int original_prompt_len = 0;
+    if (ctx_guidance) {
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+
+        original_prompt_len = original_inp.size();
+        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
+    }
+
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
+        params.n_keep = (int)embd_inp.size();
+    }
+
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+
+
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
+        params.interactive = true;
+    }
+
+    if (params.verbose_prompt) {
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+
+        if (ctx_guidance) {
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            for (int i = 0; i < (int) guidance_inp.size(); i++) {
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
+            }
+        }
+
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            }
+            LOG_TEE("'\n");
+        }
+        LOG_TEE("\n");
+    }
+
+    if (params.interactive) {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+
+        LOG_TEE("%s: interactive mode on.\n", __func__);
+
+        if (params.input_prefix_bos) {
+            LOG_TEE("Input prefix with BOS\n");
+        }
+
+        if (!params.input_prefix.empty()) {
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+        }
+
+        if (!params.input_suffix.empty()) {
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
+        }
+    }
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
+
+    struct llama_grammar * grammar = NULL;
+    grammar_parser::parse_state parsed_grammar;
+
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        LOG_TEE("%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        LOG_TEE("\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos(ctx));
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+    LOG_TEE("\n#####  Infill mode  #####\n\n");
+    if (params.infill) {
+        printf("\n************\n");
+        printf("no need to specify '--infill', always running infill\n");
+        printf("************\n\n");
+    }
+    if (params.interactive) {
+        const char *control_message;
+        if (params.multiline_input) {
+            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"
+                              " - 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";
+        }
+        LOG_TEE("== Running in interactive mode. ==\n");
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
+#endif
+        LOG_TEE(       "%s\n", control_message);
+
+        is_interacting = params.interactive_first;
+    }
+
+    bool input_echo           = true;
+
+    int n_past             = 0;
+    int n_remain           = params.n_predict;
+    int n_consumed         = 0;
+    int n_past_guidance    = 0;
+
+    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
+    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
+    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+
+    // the first thing we will do is to output the prompt, so set color accordingly
+    console::set_display(console::prompt);
+
+    std::vector<llama_token> embd;
+    std::vector<llama_token> embd_guidance;
+
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    while (n_remain != 0 || params.interactive) {
+        // predict
+        if (!embd.empty()) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            int max_embd_size = n_ctx - 4;
+
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
+                console::set_display(console::error);
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console::set_display(console::reset);
+                fflush(stdout);
+            }
+
+            // infinite text generation via context swapping
+            // if we run out of context:
+            // - take the n_keep first tokens from the original prompt (via n_past)
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                if (params.n_predict == -2) {
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    break;
+                }
+
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
+
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+            }
+
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+
+            if (ctx_guidance) {
+                int input_size = 0;
+                llama_token * input_buf = NULL;
+
+                if (n_past_guidance < (int) guidance_inp.size()) {
+                    // Guidance context should have the same data with these modifications:
+                    //
+                    // * Replace the initial prompt
+                    // * Shift everything by guidance_offset
+                    embd_guidance = guidance_inp;
+                    if (embd.begin() + original_prompt_len < embd.end()) {
+                        embd_guidance.insert(
+                            embd_guidance.end(),
+                            embd.begin() + original_prompt_len,
+                            embd.end()
+                        );
+                    }
+
+                    input_buf  = embd_guidance.data();
+                    input_size = embd_guidance.size();
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                } else {
+                    input_buf  = embd.data();
+                    input_size = embd.size();
+                }
+
+                for (int i = 0; i < input_size; i += params.n_batch) {
+                    int n_eval = std::min(input_size - i, params.n_batch);
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
+                        LOG_TEE("%s : failed to eval\n", __func__);
+                        return 1;
+                    }
+
+                    n_past_guidance += n_eval;
+                }
+            }
+
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                    LOG_TEE("%s : failed to eval\n", __func__);
+                    return 1;
+                }
+
+                n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
+            }
+
+        }
+
+        embd.clear();
+        embd_guidance.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
+
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
+
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+
+            embd.push_back(id);
+
+            // echo this to console
+            input_echo = true;
+
+            // decrement remaining sampling budget
+            --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
+        } else {
+            // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        if (input_echo) {
+            for (auto id : embd) {
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                printf("%s", token_str.c_str());
+
+                if (embd.size() > 1) {
+                    input_tokens.push_back(id);
+                } else {
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
+            }
+            fflush(stdout);
+        }
+        // reset color to default if we there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
+            console::set_display(console::reset);
+        }
+
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
+
+            // deal with eot token in infill mode
+            if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
+                if(is_interacting && !params.interactive_first) {
+                    // print an eot token
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                }
+                fflush(stdout);
+                printf("\n");
+                console::set_display(console::user_input);
+                std::string buffer;
+                std::string line;
+                bool another_line=true;
+                // set a new prefix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line, if so we use the old input
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_prefix = buffer;
+                }
+                buffer.clear();
+                // set a new suffix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_suffix = buffer;
+                }
+                buffer.clear();
+                // done taking input, reset color
+                console::set_display(console::reset);
+                // tokenize new prefix and suffix
+                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                embd_inp = inp_pfx;
+                embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+                embd_inp.push_back(llama_token_middle(ctx));
+                embd.clear();
+                embd_guidance.clear();
+                n_remain = params.n_predict;
+                n_past = 0;
+                n_consumed = 0;
+                // LOG_TEE("took new input\n");
+                is_interacting = false;
+            }
+            // deal with end of text token in interactive mode
+            else if (last_tokens.back() == llama_token_eos(ctx)) {
+                LOG("found EOS token\n");
+
+                if (params.interactive) {
+
+                    is_interacting = true;
+                    printf("\n");
+                    console::set_display(console::user_input);
+                    fflush(stdout);
+               }
+            }
+
+            if (n_past > 0 && is_interacting && !params.interactive) {
+                LOG("waiting for user input\n");
+
+                if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
+                    embd_inp.push_back(llama_token_bos(ctx));
+                }
+
+                std::string buffer;
+                if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    buffer += params.input_prefix;
+                    printf("%s", buffer.c_str());
+                }
+
+                std::string line;
+                bool another_line = true;
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
+                    // append input suffix if any
+                    if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        buffer += params.input_suffix;
+                        printf("%s", params.input_suffix.c_str());
+                    }
+
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        output_tokens.push_back(token);
+                        output_ss << llama_token_to_piece(ctx, token);
+                    }
+
+                    n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
+                }
+
+                input_echo = false; // do not echo this again
+            }
+
+            if (n_past > 0) {
+                if (is_interacting) {
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
+                }
+                is_interacting = false;
+            }
+        }
+
+        // end of text token
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+            break;
+        }
+
+        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
+            n_remain = params.n_predict;
+            is_interacting = true;
+        }
+    }
+    if (!params.interactive && n_remain <= 0) {
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        fflush(stdout);
+    }
+
+    llama_print_timings(ctx);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
+
+    if (ctx_guidance) { llama_free(ctx_guidance); }
+    llama_free(ctx);
+    llama_free_model(model);
+
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
+    llama_backend_free();
+
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
+    return 0;
+}
+

examples/jeopardy/README.md

@@ -2,7 +2,7 @@
 
 This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer.
 
-The jeopardy test can be used to compare the fact knowledge of different models and compare them to eachother. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
+The jeopardy test can be used to compare the fact knowledge of different models and compare them to each other. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
 
 
 Step 1: Open jeopardy.sh and modify the following:

examples/llama-bench/README.md

@@ -0,0 +1,271 @@
+# llama.cpp/example/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [SQL](#sql)
+
+## Syntax
+
+```
+usage: ./llama-bench [options]
+
+options:
+  -h, --help
+  -m, --model <filename>            (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                (default: 512)
+  -n, --n-gen <n>                   (default: 128)
+  -b, --batch-size <n>              (default: 512)
+  --memory-f32 <0|1>                (default: 0)
+  -t, --threads <n>                 (default: 16)
+  -ngl N, --n-gpu-layers <n>        (default: 99)
+  -mg i, --main-gpu <i>             (default: 0)
+  -mmq, --mul-mat-q <0|1>           (default: 1)
+  -ts, --tensor_split <ts0/ts1/..>
+  -r, --repetitions <n>             (default: 5)
+  -o, --output <csv|json|md|sql>    (default: md)
+  -v, --verbose                     (default: 0)
+
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+```
+
+llama-bench can perform two types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+For a description of the other options, see the [main example](../main/README.md).
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cuda,opencl,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 512,
+    "n_gen": 0,
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+  },
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 0,
+    "n_gen": 128,
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+  }
+]
+```
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cuda INTEGER,
+  opencl INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_threads INTEGER,
+  f16_kv INTEGER,
+  n_gpu_layers INTEGER,
+  main_gpu INTEGER,
+  mul_mat_q INTEGER,
+  tensor_split TEXT,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+```

examples/llama-bench/llama-bench.cpp

@@ -74,14 +74,6 @@ static T stdev(const std::vector<T> & v) {
     return stdev;
 }
 
-static bool ggml_cpu_has_metal() {
-#if defined(GGML_USE_METAL)
-    return true;
-#else
-    return false;
-#endif
-}
-
 static std::string get_cpu_info() {
     std::string id;
 #ifdef __linux__
@@ -140,7 +132,6 @@ struct cmd_params {
     std::vector<int> n_gpu_layers;
     std::vector<int> main_gpu;
     std::vector<bool> mul_mat_q;
-    std::vector<bool> low_vram;
     std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
     int reps;
     bool verbose;
@@ -157,7 +148,6 @@ static const cmd_params cmd_params_defaults = {
     /* n_gpu_layers  */ {99},
     /* main_gpu      */ {0},
     /* mul_mat_q     */ {true},
-    /* low_vram      */ {false},
     /* tensor_split  */ {{}},
     /* reps          */ 5,
     /* verbose       */ false,
@@ -165,26 +155,25 @@ static const cmd_params cmd_params_defaults = {
 };
 
 static void print_usage(int /* argc */, char ** argv) {
-    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help\n");
-    fprintf(stdout, "  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    fprintf(stdout, "  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    fprintf(stdout, "  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    fprintf(stdout, "  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    fprintf(stdout, "  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
-    fprintf(stdout, "  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    fprintf(stdout, "  -ngl N, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    fprintf(stdout, "  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    fprintf(stdout, "  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
-    fprintf(stdout, "  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    fprintf(stdout, "  -ts, --tensor_split <ts0/ts1/..>               \n");
-    fprintf(stdout, "  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
-    fprintf(stdout, "  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
-    fprintf(stdout, "  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
-    fprintf(stdout, "\n");
-    fprintf(stdout, "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help\n");
+    printf("  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    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/..>               \n");
+    printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
+    printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("\n");
+    printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 
 }
 
@@ -263,13 +252,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.main_gpu = split<int>(argv[i], split_delim);
-        } else if (arg == "-lv" || arg == "--low-vram") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.low_vram.insert(params.low_vram.end(), p.begin(), p.end());
         } else if (arg == "-mmq" || arg == "--mul-mat-q") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -344,7 +326,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
-    if (params.low_vram.empty())     { params.low_vram = cmd_params_defaults.low_vram; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
 
@@ -361,21 +342,34 @@ struct cmd_params_instance {
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
-    bool low_vram;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
 
-    llama_context_params to_llama_params() const {
-        llama_context_params lparams = llama_context_default_params();
-        lparams.n_ctx = n_prompt + n_gen;
-        lparams.n_batch = n_batch;
-        lparams.f16_kv = !f32_kv;
-        lparams.n_gpu_layers = n_gpu_layers;
-        lparams.main_gpu = main_gpu;
-        lparams.mul_mat_q = mul_mat_q;
-        lparams.low_vram = low_vram;
-        lparams.tensor_split = tensor_split.data();
+    llama_model_params to_llama_mparams() const {
+        llama_model_params mparams = llama_model_default_params();
 
-        return lparams;
+        mparams.n_gpu_layers = n_gpu_layers;
+        mparams.main_gpu = main_gpu;
+        mparams.tensor_split = tensor_split.data();
+
+        return mparams;
+    }
+
+    bool equal_mparams(const cmd_params_instance & other) const {
+        return model == other.model &&
+               n_gpu_layers == other.n_gpu_layers &&
+               main_gpu == other.main_gpu &&
+               tensor_split == other.tensor_split;
+    }
+
+    llama_context_params to_llama_cparams() const {
+        llama_context_params cparams = llama_context_default_params();
+
+        cparams.n_ctx = n_prompt + n_gen;
+        cparams.n_batch = n_batch;
+        cparams.f16_kv = !f32_kv;
+        cparams.mul_mat_q = mul_mat_q;
+
+        return cparams;
     }
 };
 
@@ -383,13 +377,12 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
     std::vector<cmd_params_instance> instances;
 
     for (const auto & m : params.model)
-    for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
     for (const auto & nl : params.n_gpu_layers)
     for (const auto & mg : params.main_gpu)
-    for (const auto & mmq : params.mul_mat_q)
-    for (const auto & lv : params.low_vram)
     for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         cmd_params_instance instance = {
             /* .model        = */ m,
@@ -401,7 +394,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
             /* .n_gpu_layers = */ nl,
             /* .main_gpu     = */ mg,
             /* .mul_mat_q    = */ mmq,
-            /* .low_vram     = */ lv,
             /* .tensor_split = */ ts,
         };
         instances.push_back(instance);
@@ -412,6 +404,56 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
 static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
     std::vector<cmd_params_instance> instances;
 
+#if 1
+    // this ordering minimizes the number of times that each model needs to be reloaded
+    for (const auto & m : params.model)
+    for (const auto & nl : params.n_gpu_layers)
+    for (const auto & mg : params.main_gpu)
+    for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
+    for (const auto & nt : params.n_threads) {
+        for (const auto & n_prompt : params.n_prompt) {
+            if (n_prompt == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_prompt,
+                /* .n_gen        = */ 0,
+                /* .n_batch      = */ nb,
+                /* .f32_kv       = */ fk,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .main_gpu     = */ mg,
+                /* .mul_mat_q    = */ mmq,
+                /* .tensor_split = */ ts,
+            };
+            instances.push_back(instance);
+        }
+
+        for (const auto & n_gen : params.n_gen) {
+            if (n_gen == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ 0,
+                /* .n_gen        = */ n_gen,
+                /* .n_batch      = */ nb,
+                /* .f32_kv       = */ fk,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .main_gpu     = */ mg,
+                /* .mul_mat_q    = */ mmq,
+                /* .tensor_split = */ ts,
+            };
+            instances.push_back(instance);
+        }
+    }
+#else
+    // this ordering separates the prompt and generation tests
     for (const auto & n_prompt : params.n_prompt) {
         if (n_prompt == 0) {
             continue;
@@ -427,6 +469,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
         auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
         instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
     }
+#endif
 
     return instances;
 }
@@ -451,7 +494,6 @@ struct test {
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
-    bool low_vram;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
     int n_prompt;
     int n_gen;
@@ -471,7 +513,6 @@ struct test {
         n_gpu_layers = inst.n_gpu_layers;
         main_gpu = inst.main_gpu;
         mul_mat_q = inst.mul_mat_q;
-        low_vram = inst.low_vram;
         tensor_split = inst.tensor_split;
         n_prompt = inst.n_prompt;
         n_gen = inst.n_gen;
@@ -532,7 +573,7 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "f16_kv",
-            "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
+            "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -551,7 +592,7 @@ struct test {
             return INT;
         }
         if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "mul_mat_q" || field == "low_vram") {
+            field == "f16_kv" || field == "mul_mat_q") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -582,7 +623,7 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
-            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
+            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -614,9 +655,9 @@ struct printer {
     virtual ~printer() {}
 
     FILE * fout;
-    virtual void print_header(const cmd_params & params) { (void) params; };
+    virtual void print_header(const cmd_params & params) { (void) params; }
     virtual void print_test(const test & t) = 0;
-    virtual void print_footer() { };
+    virtual void print_footer() { }
 };
 
 struct csv_printer : public printer {
@@ -774,9 +815,6 @@ struct markdown_printer : public printer {
         if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
             fields.push_back("mul_mat_q");
         }
-        if (params.low_vram.size() > 1 || params.low_vram != cmd_params_defaults.low_vram) {
-            fields.push_back("low_vram");
-        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.push_back("tensor_split");
         }
@@ -897,21 +935,27 @@ struct sql_printer : public printer {
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
     std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
     int n_processed = 0;
+
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        llama_eval(ctx, tokens.data(), n_tokens, n_past + n_processed, n_threads);
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
         n_processed += n_tokens;
     }
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
     llama_token token = llama_token_bos(ctx);
+
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     for (int i = 0; i < n_gen; i++) {
-        llama_eval(ctx, &token, 1, n_past + i, n_threads);
+        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
     }
 }
 
-static void llama_null_log_callback(enum llama_log_level level, const char * text, void * user_data) {
+static void llama_null_log_callback(enum ggml_log_level level, const char * text, void * user_data) {
     (void) level;
     (void) text;
     (void) user_data;
@@ -966,17 +1010,25 @@ int main(int argc, char ** argv) {
 
     std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
 
-    for (const auto & inst : params_instances) {
-        // TODO: keep the model between tests when possible
-        llama_context_params lparams = inst.to_llama_params();
+    llama_model * lmodel = nullptr;
+    const cmd_params_instance * prev_inst = nullptr;
 
-        llama_model * lmodel  = llama_load_model_from_file(inst.model.c_str(), lparams);
-        if (lmodel == NULL) {
-            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
-            return 1;
+    for (const auto & inst : params_instances) {
+        // keep the same model between tests when possible
+        if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
+            if (lmodel) {
+                llama_free_model(lmodel);
+            }
+
+            lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
+            if (lmodel == NULL) {
+                fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
+                return 1;
+            }
+            prev_inst = &inst;
         }
 
-        llama_context * ctx = llama_new_context_with_model(lmodel, lparams);
+        llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
             llama_free_model(lmodel);
@@ -985,10 +1037,19 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         // warmup run
-        test_gen(ctx, 1, 0, t.n_threads);
+        if (t.n_prompt > 0) {
+            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
+        }
+        if (t.n_gen > 0) {
+            test_gen(ctx, 1, 0, t.n_threads);
+        }
 
         for (int i = 0; i < params.reps; i++) {
+            llama_kv_cache_tokens_rm(ctx, -1, -1);
+
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {
                 test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
@@ -1005,9 +1066,10 @@ int main(int argc, char ** argv) {
         llama_print_timings(ctx);
 
         llama_free(ctx);
-        llama_free_model(lmodel);
     }
 
+    llama_free_model(lmodel);
+
     p->print_footer();
 
     llama_backend_free();

examples/llm.vim

@@ -8,7 +8,7 @@ function! Llm()
   let buffer_content = join(getline(1, '$'), "\n")
 
   " Create the JSON payload
-  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":10,"stream": v:false}
+  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":256,"stop": ["\n\n\n"],"stream": v:false}
   let json_payload.prompt = buffer_content
 
   " Define the curl command
@@ -25,3 +25,4 @@ function! Llm()
 endfunction
 
 command! Llm call Llm()
+noremap <F2> :Llm<CR>

examples/main-cmake-pkg/.gitignore

@@ -0,0 +1,51 @@
+# Prerequisites
+*.d
+
+# Compiled Object files
+*.slo
+*.lo
+*.o
+*.obj
+
+# Precompiled Headers
+*.gch
+*.pch
+
+# Compiled Dynamic libraries
+*.so
+*.dylib
+*.dll
+
+# Fortran module files
+*.mod
+*.smod
+
+# Compiled Static libraries
+*.lai
+*.la
+*.a
+*.lib
+
+# Executables
+*.exe
+*.out
+*.app
+
+*.gguf
+
+*.log
+.DS_Store
+.build/
+.cache/
+.direnv/
+.envrc
+.swiftpm
+.venv
+.clang-tidy
+.vs/
+.vscode/
+
+build*/
+out/
+tmp/
+

examples/main-cmake-pkg/CMakeLists.txt

@@ -0,0 +1,46 @@
+cmake_minimum_required(VERSION 3.12)
+project("main-cmake-pkg" C CXX)
+set(TARGET main-cmake-pkg)
+
+find_package(Llama 0.0.1 REQUIRED)
+
+# Bake common functionality in with target. Because applications
+# using the relocatable Llama package should be outside of the
+# source tree, main-cmake-pkg pretends the dependencies are built-in.
+
+set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
+add_library(common OBJECT
+    ${_common_path}/common.h
+    ${_common_path}/common.cpp
+    ${_common_path}/console.h
+    ${_common_path}/console.cpp
+    ${_common_path}/grammar-parser.h
+    ${_common_path}/grammar-parser.cpp
+    )
+
+# WARNING: because build-info.h is auto-generated, it will only
+# be available after the user has built the llama.cpp sources.
+#
+configure_file(${_common_path}/../build-info.h
+    ${CMAKE_CURRENT_BINARY_DIR}/build-info.h
+    COPYONLY)
+
+target_include_directories(common PUBLIC ${LLAMA_INCLUDE_DIR}
+    ${CMAKE_CURRENT_BINARY_DIR})
+
+# If the common project was part of "main-cmake-pkg" the transient
+# defines would automatically be attached. Because the common func-
+# tionality is separate, but dependent upon the defines, it must be
+# explicitly extracted from the "llama" target.
+#
+get_target_property(_llama_transient_defines llama
+    INTERFACE_COMPILE_DEFINITIONS)
+
+target_compile_definitions(common PRIVATE "${_llama_transient_defines}")
+
+add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
+target_include_directories(${TARGET} PRIVATE ${_common_path})
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+

examples/main-cmake-pkg/README.md

@@ -0,0 +1,37 @@
+# llama.cpp/example/main-cmake-pkg
+
+This program builds the [main](../main) application using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+
+## Building
+
+Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions.
+
+### 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_.
+
+### Build llama.cpp and install to C:\LlamaCPP directory
+
+In this case, CLBlast was already installed so the CMake package is referenced in `CMAKE_PREFIX_PATH`.
+
+```cmd
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+mkdir build
+cd build
+cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64
+cmake --build . --config Release
+cmake --install . --prefix C:/LlamaCPP
+```
+
+### Build main-cmake-pkg
+
+
+```cmd
+cd ..\examples\main-cmake-pkg
+mkdir build
+cd build
+cmake .. -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
+cmake --build . --config Release
+cmake --install . --prefix C:/MyLlamaApp
+```

examples/main/README.md

@@ -34,7 +34,7 @@ For an interactive experience, try this command:
 #### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " \
+./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -p \
 'User: Hi
 AI: Hello. I am an AI chatbot. Would you like to talk?
 User: Sure!
@@ -45,7 +45,7 @@ User:'
 #### Windows:
 
 ```powershell
-main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -e --prompt "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
+main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -e -p "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
 ```
 
 The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -144,7 +144,7 @@ The `--ctx-size` option allows you to set the size of the prompt context used by
 
 Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
 
-- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+-   `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
 
 ### Keep Prompt
 
@@ -262,7 +262,8 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Number of Threads
 
--   `-t N, --threads N`: Set the number of threads to use during computation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. In some systems, it is beneficial to use a higher number of threads during batch processing than during generation. If not specified, the number of threads used for batch processing will be the same as the number of threads used for generation.
 
 ### Mlock
 
@@ -274,7 +275,7 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### NUMA support
 
--   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop\_caches' as root.
+-   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
 ### Memory Float 32
 
@@ -302,10 +303,8 @@ 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.
--   `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
--   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -1,9 +1,5 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
 #include "common.h"
+
 #include "console.h"
 #include "llama.h"
 #include "build-info.h"
@@ -45,10 +41,12 @@ static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;
 
-void write_logfile(
-    const llama_context * ctx, const gpt_params & params, const llama_model * model,
-    const std::vector<llama_token> input_tokens, const std::string output, const std::vector<llama_token> output_tokens) {
 
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
     if (params.logdir.empty()) {
         return;
     }
@@ -89,7 +87,7 @@ void write_logfile(
 }
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-void sigint_handler(int signo) {
+static void sigint_handler(int signo) {
     if (signo == SIGINT) {
         if (!is_interacting) {
             is_interacting = true;
@@ -108,16 +106,25 @@ int main(int argc, char ** argv) {
     gpt_params params;
     g_params = &params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("main", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // TODO: Dump params ?
+    //LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity));
+
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
 
-    if (params.perplexity) {
+    if (params.logits_all) {
         printf("\n************\n");
         printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
         printf("************\n\n");
@@ -133,35 +140,34 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    if (params.rope_freq_base != 10000.0) {
-        fprintf(stderr, "%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
-    }
-
-    if (params.rope_freq_scale != 1.0) {
-        fprintf(stderr, "%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
-    }
-
-    if (params.n_ctx > 2048) {
-        // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
-        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
-    } else if (params.n_ctx < 8) {
-        fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    LOG("%s: llama backend init\n", __func__);
     llama_backend_init(params.numa);
 
     llama_model * model;
@@ -171,6 +177,7 @@ int main(int argc, char ** argv) {
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (params.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
@@ -178,84 +185,71 @@ int main(int argc, char ** argv) {
     }
 
     if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        LOG_TEE("%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);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_TEE("%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, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
-    }
-
-    // determine the maximum memory usage needed to do inference for the given n_batch and n_ctx parameters
-    // uncomment the "used_mem" line in llama.cpp to see the results
-    if (params.mem_test) {
-        {
-            fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
-
-            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos(ctx));
-            llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
-        }
-
-        llama_print_timings(ctx);
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
-    }
-
-    // export the cgraph and exit
-    if (params.export_cgraph) {
-        llama_eval_export(ctx, "llama.ggml");
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", get_system_info(params).c_str());
     }
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
 
     if (!path_session.empty()) {
-        fprintf(stderr, "%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
+        LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
 
         // fopen to check for existing session
         FILE * fp = std::fopen(path_session.c_str(), "rb");
         if (fp != NULL) {
             std::fclose(fp);
 
-            session_tokens.resize(params.n_ctx);
+            session_tokens.resize(n_ctx);
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
-                fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
+                LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
                 return 1;
             }
             session_tokens.resize(n_token_count_out);
             llama_set_rng_seed(ctx, params.seed);
 
-            fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
+            LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
         } else {
-            fprintf(stderr, "%s: session file does not exist, will create\n", __func__);
+            LOG_TEE("%s: session file does not exist, will create\n", __func__);
         }
     }
 
-    // Add BOS if SPM tokenizer
-    const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
 
-    // tokenize the prompt
     std::vector<llama_token> embd_inp;
 
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
+        LOG("tokenize the prompt\n");
         embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
     } else {
+        LOG("use session tokens\n");
         embd_inp = session_tokens;
     }
 
+    LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+
     // Should not run without any tokens
     if (embd_inp.empty()) {
         embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
     }
 
     // Tokenize negative prompt
@@ -263,23 +257,28 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+
         guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
 
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
     }
 
-    const int n_ctx = llama_n_ctx(ctx);
-
     if ((int) embd_inp.size() > n_ctx - 4) {
-        fprintf(stderr, "%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
     }
 
     // debug message about similarity of saved session, if applicable
     size_t n_matching_session_tokens = 0;
-    if (session_tokens.size()) {
+    if (!session_tokens.empty()) {
         for (llama_token id : session_tokens) {
             if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
                 break;
@@ -287,22 +286,27 @@ int main(int argc, char ** argv) {
             n_matching_session_tokens++;
         }
         if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
-            fprintf(stderr, "%s: using full prompt from session file\n", __func__);
+            LOG_TEE("%s: using full prompt from session file\n", __func__);
         } else if (n_matching_session_tokens >= embd_inp.size()) {
-            fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__);
+            LOG_TEE("%s: session file has exact match for prompt!\n", __func__);
         } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
-            fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+            LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         } else {
-            fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n",
+            LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         }
     }
 
+    LOGLN(
+            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu",
+            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size());
+
     // if we will use the cache for the full prompt without reaching the end of the cache, force
     // reevaluation of the last token token to recalculate the cached logits
-    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
-            session_tokens.size() > embd_inp.size()) {
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
+        LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1);
+
         session_tokens.resize(embd_inp.size() - 1);
     }
 
@@ -315,6 +319,9 @@ int main(int argc, char ** argv) {
     const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
     const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
 
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
         params.interactive_first = true;
@@ -327,30 +334,30 @@ int main(int argc, char ** argv) {
     }
 
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (ctx_guidance) {
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
-            fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
             }
         }
 
         if (params.n_keep > 0) {
-        fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            fprintf(stderr, "'\n");
+            LOG_TEE("'\n");
         }
-        fprintf(stderr, "\n");
+        LOG_TEE("\n");
     }
 
     if (params.interactive) {
@@ -367,47 +374,48 @@ int main(int argc, char ** argv) {
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-        fprintf(stderr, "%s: interactive mode on.\n", __func__);
+        LOG_TEE("%s: interactive mode on.\n", __func__);
 
-        if (params.antiprompt.size()) {
-            for (auto antiprompt : params.antiprompt) {
-                fprintf(stderr, "Reverse prompt: '%s'\n", antiprompt.c_str());
+        if (!params.antiprompt.empty()) {
+            for (const auto & antiprompt : params.antiprompt) {
+                LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
             }
         }
 
         if (params.input_prefix_bos) {
-            fprintf(stderr, "Input prefix with BOS\n");
+            LOG_TEE("Input prefix with BOS\n");
         }
 
         if (!params.input_prefix.empty()) {
-            fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
         }
 
         if (!params.input_suffix.empty()) {
-            fprintf(stderr, "Input suffix: '%s'\n", params.input_suffix.c_str());
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    fprintf(stderr, "sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
             params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
-    fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-    fprintf(stderr, "\n\n");
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
 
+    struct llama_grammar * grammar = NULL;
     grammar_parser::parse_state parsed_grammar;
-    llama_grammar *             grammar = NULL;
+
     if (!params.grammar.empty()) {
         parsed_grammar = grammar_parser::parse(params.grammar.c_str());
         // will be empty (default) if there are parse errors
         if (parsed_grammar.rules.empty()) {
             return 1;
         }
-        fprintf(stderr, "%s: grammar:\n", __func__);
+        LOG_TEE("%s: grammar:\n", __func__);
         grammar_parser::print_grammar(stderr, parsed_grammar);
-        fprintf(stderr, "\n");
+        LOG_TEE("\n");
 
         {
             auto it = params.logit_bias.find(llama_token_eos(ctx));
             if (it != params.logit_bias.end() && it->second == -INFINITY) {
-                fprintf(stderr, "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
             }
         }
 
@@ -417,8 +425,8 @@ int main(int argc, char ** argv) {
     }
 
     // TODO: replace with ring-buffer
-    std::vector<llama_token> last_n_tokens(n_ctx);
-    std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
 
     if (params.interactive) {
         const char *control_message;
@@ -430,11 +438,11 @@ int main(int argc, char ** argv) {
                               " - 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";
         }
-        fprintf(stderr, "== Running in interactive mode. ==\n"
+        LOG_TEE("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-               " - Press Ctrl+C to interject at any time.\n"
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-               "%s\n", control_message);
+        LOG_TEE(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -459,27 +467,27 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    // do one empty run to warm up the model
-    {
-        const std::vector<llama_token> tmp = { llama_token_bos(ctx), };
-        llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
-        llama_reset_timings(ctx);
-    }
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
-        if (embd.size() > 0) {
+        if (!embd.empty()) {
             // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
-            auto max_embd_size = n_ctx - 4;
+            int max_embd_size = n_ctx - 4;
+
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
-            if ((int)embd.size() > max_embd_size) {
-                auto skipped_tokens = embd.size() - max_embd_size;
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
                 console::set_display(console::error);
-                printf("<<input too long: skipped %zu token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
                 console::set_display(console::reset);
                 fflush(stdout);
-                embd.resize(max_embd_size);
             }
 
             // infinite text generation via context swapping
@@ -488,28 +496,31 @@ int main(int argc, char ** argv) {
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
                 if (params.n_predict == -2) {
-                    fprintf(stderr, "\n\n%s: context full, stopping generation\n", __func__);
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
                 }
 
-                const int n_left = n_past - params.n_keep;
-                // always keep the first token - BOS
-                n_past = std::max(1, params.n_keep);
-                n_past_guidance = std::max(1, params.n_keep + guidance_offset);
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
 
-                // insert n_left/2 tokens at the start of embd from last_n_tokens
-                embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                // stop saving session if we run out of context
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                LOG("clear session path\n");
                 path_session.clear();
-
-                //printf("\n---\n");
-                //printf("resetting: '");
-                //for (int i = 0; i < (int) embd.size(); i++) {
-                //    printf("%s", llama_token_to_piece(ctx, embd[i]));
-                //}
-                //printf("'\n");
-                //printf("\n---\n");
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
@@ -532,6 +543,9 @@ int main(int argc, char ** argv) {
                 if (i > 0) {
                     embd.erase(embd.begin(), embd.begin() + i);
                 }
+
+                // remove any "future" tokens that we might have inherited from the session from the KV cache
+                llama_kv_cache_tokens_rm(ctx, n_past, -1);
             }
 
             // evaluate tokens in batches
@@ -539,7 +553,7 @@ int main(int argc, char ** argv) {
 
             if (ctx_guidance) {
                 int input_size = 0;
-                llama_token* input_buf = NULL;
+                llama_token * input_buf = NULL;
 
                 if (n_past_guidance < (int) guidance_inp.size()) {
                     // Guidance context should have the same data with these modifications:
@@ -555,22 +569,19 @@ int main(int argc, char ** argv) {
                         );
                     }
 
-                    input_buf = embd_guidance.data();
+                    input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
-                    //fprintf(stderr, "\n---------------------\n");
-                    //for (int i = 0; i < (int) embd_guidance.size(); i++) {
-                        //fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_guidance[i]));
-                    //}
-                    //fprintf(stderr, "\n---------------------\n");
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
                 } else {
-                    input_buf = embd.data();
+                    input_buf  = embd.data();
                     input_size = embd.size();
                 }
 
                 for (int i = 0; i < input_size; i += params.n_batch) {
                     int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
-                        fprintf(stderr, "%s : failed to eval\n", __func__);
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
+                        LOG_TEE("%s : failed to eval\n", __func__);
                         return 1;
                     }
 
@@ -583,14 +594,20 @@ int main(int argc, char ** argv) {
                 if (n_eval > params.n_batch) {
                     n_eval = params.n_batch;
                 }
-                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
-                    fprintf(stderr, "%s : failed to eval\n", __func__);
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                    LOG_TEE("%s : failed to eval\n", __func__);
                     return 1;
                 }
+
                 n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
             }
 
-            if (embd.size() > 0 && !path_session.empty()) {
+            if (!embd.empty() && !path_session.empty()) {
                 session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
                 n_session_consumed = session_tokens.size();
             }
@@ -600,106 +617,21 @@ int main(int argc, char ** argv) {
         embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            // out of user input, sample next token
-            const float   temp            = params.temp;
-            const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
-            const float   top_p           = params.top_p;
-            const float   tfs_z           = params.tfs_z;
-            const float   typical_p       = params.typical_p;
-            const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
-            const float   repeat_penalty  = params.repeat_penalty;
-            const float   alpha_presence  = params.presence_penalty;
-            const float   alpha_frequency = params.frequency_penalty;
-            const int     mirostat        = params.mirostat;
-            const float   mirostat_tau    = params.mirostat_tau;
-            const float   mirostat_eta    = params.mirostat_eta;
-            const bool    penalize_nl     = params.penalize_nl;
-
             // optionally save the session on first sample (for faster prompt loading next time)
             if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
                 need_to_save_session = false;
                 llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+
+                LOG("saved session to %s\n", path_session.c_str());
             }
 
-            llama_token id = 0;
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
 
-            {
-                auto logits  = llama_get_logits(ctx);
-                auto n_vocab = llama_n_vocab(ctx);
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
 
-                // Apply params.logit_bias map
-                for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-                    logits[it->first] += it->second;
-                }
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
 
-                std::vector<llama_token_data> candidates;
-                candidates.reserve(n_vocab);
-                for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                    candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-                }
-
-                llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-                if (ctx_guidance) {
-                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale);
-                }
-
-                // Apply penalties
-                float nl_logit = logits[llama_token_nl(ctx)];
-                auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-                llama_sample_repetition_penalty(ctx, &candidates_p,
-                    last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                    last_n_repeat, repeat_penalty);
-                llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-                    last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                    last_n_repeat, alpha_frequency, alpha_presence);
-                if (!penalize_nl) {
-                    for (size_t idx = 0; idx < candidates_p.size; idx++) {
-                        if (candidates_p.data[idx].id == llama_token_nl(ctx)) {
-                            candidates_p.data[idx].logit = nl_logit;
-                            break;
-                        }
-                    }
-                }
-
-                if (grammar != NULL) {
-                    llama_sample_grammar(ctx, &candidates_p, grammar);
-                }
-
-                if (temp <= 0) {
-                    // Greedy sampling
-                    id = llama_sample_token_greedy(ctx, &candidates_p);
-                } else {
-                    if (mirostat == 1) {
-                        static float mirostat_mu = 2.0f * mirostat_tau;
-                        const int mirostat_m = 100;
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-                    } else if (mirostat == 2) {
-                        static float mirostat_mu = 2.0f * mirostat_tau;
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-                    } else {
-                        // Temperature sampling
-                        llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-                        llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
-                        llama_sample_typical(ctx, &candidates_p, typical_p, 1);
-                        llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token(ctx, &candidates_p);
-                    }
-                }
-                // printf("`%d`", candidates_p.size);
-
-                if (grammar != NULL) {
-                    llama_grammar_accept_token(ctx, grammar, id);
-                }
-
-                last_n_tokens.erase(last_n_tokens.begin());
-                last_n_tokens.push_back(id);
-            }
-
-            // add it to the context
             embd.push_back(id);
 
             // echo this to console
@@ -707,12 +639,15 @@ int main(int argc, char ** argv) {
 
             // decrement remaining sampling budget
             --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
         } else {
             // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-                last_n_tokens.erase(last_n_tokens.begin());
-                last_n_tokens.push_back(embd_inp[n_consumed]);
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -735,18 +670,17 @@ int main(int argc, char ** argv) {
             }
             fflush(stdout);
         }
-        // reset color to default if we there is no pending user input
-        if (input_echo && (int)embd_inp.size() == n_consumed) {
+        // reset color to default if there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
         }
 
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
-
             // check for reverse prompt
-            if (params.antiprompt.size()) {
+            if (!params.antiprompt.empty()) {
                 std::string last_output;
-                for (auto id : last_n_tokens) {
+                for (auto id : last_tokens) {
                     last_output += llama_token_to_piece(ctx, id);
                 }
 
@@ -760,22 +694,26 @@ int main(int argc, char ** argv) {
                         ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
                         : 0;
 
-                    if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
+                    if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
                         if (params.interactive) {
                             is_interacting = true;
-                            console::set_display(console::user_input);
                         }
                         is_antiprompt = true;
-                        fflush(stdout);
                         break;
                     }
                 }
+
+                if (is_antiprompt) {
+                    LOG("found antiprompt: %s\n", last_output.c_str());
+                }
             }
 
             // deal with end of text token in interactive mode
-            if (last_n_tokens.back() == llama_token_eos(ctx)) {
+            if (last_tokens.back() == llama_token_eos(ctx)) {
+                LOG("found EOS token\n");
+
                 if (params.interactive) {
-                    if (params.antiprompt.size() != 0) {
+                    if (!params.antiprompt.empty()) {
                         // tokenize and inject first reverse prompt
                         const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
                         embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
@@ -784,28 +722,33 @@ int main(int argc, char ** argv) {
 
                     is_interacting = true;
                     printf("\n");
-                    console::set_display(console::user_input);
-                    fflush(stdout);
                 } else if (params.instruct) {
                     is_interacting = true;
                 }
             }
 
             if (n_past > 0 && is_interacting) {
+                LOG("waiting for user input\n");
+
                 if (params.instruct) {
                     printf("\n> ");
                 }
 
                 if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
                     embd_inp.push_back(llama_token_bos(ctx));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                     buffer += params.input_prefix;
                     printf("%s", buffer.c_str());
                 }
 
+                // color user input only
+                console::set_display(console::user_input);
+
                 std::string line;
                 bool another_line = true;
                 do {
@@ -821,23 +764,30 @@ int main(int argc, char ** argv) {
                 if (buffer.length() > 1) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                         buffer += params.input_suffix;
                         printf("%s", params.input_suffix.c_str());
                     }
 
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
                     const size_t original_size = embd_inp.size();
 
                     // instruct mode: insert instruction prefix
                     if (params.instruct && !is_antiprompt) {
+                        LOG("inserting instruction prefix\n");
                         n_consumed = embd_inp.size();
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
 
-                    auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
                     // instruct mode: insert response suffix
                     if (params.instruct) {
+                        LOG("inserting instruction suffix\n");
                         embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
                     }
 
@@ -848,6 +798,9 @@ int main(int argc, char ** argv) {
                     }
 
                     n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again
@@ -859,7 +812,7 @@ int main(int argc, char ** argv) {
                     if (grammar != NULL) {
                         llama_grammar_free(grammar);
 
-                        std::vector<const llama_grammar_element *> grammar_rules( parsed_grammar.c_rules());
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
                         grammar = llama_grammar_init(
                             grammar_rules.data(), grammar_rules.size(),
                             parsed_grammar.symbol_ids.at("root"));
@@ -871,7 +824,7 @@ int main(int argc, char ** argv) {
 
         // end of text token
         if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
-            fprintf(stderr, " [end of text]\n");
+            LOG_TEE(" [end of text]\n");
             break;
         }
 
@@ -884,7 +837,7 @@ int main(int argc, char ** argv) {
     }
 
     if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
-        fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
+        LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
         llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
@@ -900,5 +853,9 @@ int main(int argc, char ** argv) {
     }
     llama_backend_free();
 
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
     return 0;
 }

examples/make-ggml.py

@@ -1,22 +1,25 @@
 #!/usr/bin/env python3
 """
-This script converts Hugging Face llama models to GGML and quantizes them.
+This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them.
 
 Usage:
-python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
 
 Arguments:
-- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.
 - --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
 - --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
 - --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
 - --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
 
-Quant types:
+Old quant types (some base model types require these):
 - Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
 - Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
 - Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
 - Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+
+New quant types (recommended):
 - Q2_K: smallest, extreme quality loss - not recommended
 - Q3_K: alias for Q3_K_M
 - Q3_K_S: very small, very high quality loss
@@ -40,9 +43,7 @@ import argparse
 import os
 from huggingface_hub import snapshot_download
 
-def main(model, outname, outdir, quants, keep_fp16):
-    ggml_version = "v3"
-
+def main(model, model_type, outname, outdir, quants, keep_fp16):
     if not os.path.isdir(model):
         print(f"Model not found at {model}. Downloading...")
         try:
@@ -63,17 +64,20 @@ def main(model, outname, outdir, quants, keep_fp16):
     print("Building llama.cpp")
     subprocess.run(f"cd .. && make quantize", shell=True, check=True)
 
-    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+    fp16 = f"{outdir}/{outname}.gguf.fp16.bin"
 
-    print(f"Making unquantised GGML at {fp16}")
+    print(f"Making unquantised GGUF at {fp16}")
     if not os.path.isfile(fp16):
-        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+        if model_type != "llama":
+            subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True)
+        else:
+            subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
     else:
         print(f"Unquantised GGML already exists at: {fp16}")
 
     print("Making quants")
     for type in quants:
-        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        outfile = f"{outdir}/{outname}.gguf.{type}.bin"
         print(f"Making {type} : {outfile}")
         subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
 
@@ -81,8 +85,9 @@ def main(model, outname, outdir, quants, keep_fp16):
         os.remove(fp16)
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
-    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.')
     parser.add_argument('--outname', default=None, help='Output model(s) name')
     parser.add_argument('--outdir', default=None, help='Output directory')
     parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
@@ -90,4 +95,4 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)
+    main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16)

examples/parallel/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET parallel)
+add_executable(${TARGET} parallel.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/parallel/README.md

@@ -0,0 +1,3 @@
+# llama.cpp/example/parallel
+
+Simplified simluation for serving incoming requests in parallel

examples/parallel/parallel.cpp

@@ -0,0 +1,426 @@
+// A basic application simulating a server with multiple clients.
+// The clients submite requests to the server and they are processed in parallel.
+
+#include "build-info.h"
+
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+#include <ctime>
+
+// trim whitespace from the beginning and end of a string
+static std::string trim(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+
+    while (start < end && isspace(str[start])) {
+        start += 1;
+    }
+
+    while (end > start && isspace(str[end - 1])) {
+        end -= 1;
+    }
+
+    return str.substr(start, end - start);
+}
+
+static std::string k_system =
+R"(Transcript of a never ending dialog, where the User interacts with an Assistant.
+The Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.
+
+User: Recommend a nice restaurant in the area.
+Assistant: I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
+User: Who is Richard Feynman?
+Assistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
+User:)";
+
+static std::vector<std::string> k_prompts = {
+    "What is the meaning of life?",
+    "Tell me an interesting fact about llamas.",
+    "What is the best way to cook a steak?",
+    "Are you familiar with the Special Theory of Relativity and can you explain it to me?",
+    "Recommend some interesting books to read.",
+    "What is the best way to learn a new language?",
+    "How to get a job at Google?",
+    "If you could have any superpower, what would it be?",
+    "I want to learn how to play the piano.",
+};
+
+struct client {
+    int32_t id = 0;
+
+    llama_seq_id seq_id = -1;
+
+    llama_token sampled;
+
+    int64_t t_start_prompt;
+    int64_t t_start_gen;
+
+    int32_t n_prompt  = 0;
+    int32_t n_decoded = 0;
+    int32_t i_batch   = -1;
+
+    std::string input;
+    std::string prompt;
+    std::string response;
+
+    std::vector<llama_token> tokens_prev;
+};
+
+static void print_date_time() {
+    std::time_t current_time = std::time(nullptr);
+    std::tm* local_time = std::localtime(&current_time);
+    char buffer[80];
+    strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", local_time);
+
+    printf("\n\033[35mrun parameters as at %s\033[0m\n", buffer);
+}
+
+// Define a split string function to ...
+static std::vector<std::string> split_string(const std::string& input, char delimiter) {
+    std::vector<std::string> tokens;
+    std::istringstream stream(input);
+    std::string token;
+    while (std::getline(stream, token, delimiter)) {
+        tokens.push_back(token);
+    }
+    return tokens;
+}
+
+int main(int argc, char ** argv) {
+    srand(1234);
+
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    // number of simultaneous "clients" to simulate
+    const int32_t n_clients = params.n_parallel;
+
+    // requests to simulate
+    const int32_t n_seq = params.n_sequences;
+
+    // insert new requests as soon as the previous one is done
+    const bool cont_batching = params.cont_batching;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("parallel", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    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
+    if (params.prompt.empty()) {
+        printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
+    } else {
+        // Output each line of the input params.prompts vector and copy to k_prompts
+        int index = 0;
+        printf("\n\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
+
+        std::vector<std::string> prompts = split_string(params.prompt, '\n');
+        for (const auto& prompt : prompts) {
+            k_prompts.resize(index + 1);
+            k_prompts[index] = prompt;
+            index++;
+            printf("%3d prompt: %s\n", index, prompt.c_str());
+        }
+    }
+
+    fprintf(stderr, "\n\n");
+    fflush(stderr);
+
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<client> clients(n_clients);
+    for (size_t i = 0; i < clients.size(); ++i) {
+        auto & client = clients[i];
+        client.id = i;
+        client.tokens_prev.resize(std::max(256, params.n_predict));
+        std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+    }
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    std::vector<llama_token> tokens_system;
+    tokens_system = ::llama_tokenize(ctx, k_system, true);
+    const int32_t n_tokens_system = tokens_system.size();
+
+    llama_seq_id g_seq_id = 0;
+
+    // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
+    // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
+    llama_batch batch = llama_batch_init(n_ctx, 0);
+
+    int32_t n_total_prompt = 0;
+    int32_t n_total_gen    = 0;
+    int32_t n_cache_miss   = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
+    LOG_TEE("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    LOG_TEE("\n");
+
+    {
+        LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
+
+        batch.n_tokens = n_tokens_system;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            batch.token[i]  = tokens_system[i];
+            batch.pos[i]    = i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        // 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);
+        }
+
+        LOG_TEE("\n");
+    }
+
+    LOG_TEE("Processing requests ...\n\n");
+
+    while (true) {
+        batch.n_tokens = 0;
+
+        // decode any currently ongoing sequences
+        for (auto & client : clients) {
+            if (client.seq_id == -1) {
+                continue;
+            }
+
+            batch.token [batch.n_tokens] = client.sampled;
+            batch.pos   [batch.n_tokens] = n_tokens_system + client.n_prompt + client.n_decoded;
+            batch.seq_id[batch.n_tokens] = client.id;
+            batch.logits[batch.n_tokens] = true;
+
+            client.n_decoded += 1;
+            client.i_batch = batch.n_tokens;
+
+            batch.n_tokens += 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);
+            }
+
+            LOG_TEE("%s: clearing the KV cache\n", __func__);
+        }
+
+        // insert new sequences for decoding
+        if (cont_batching || batch.n_tokens == 0) {
+            for (auto & client : clients) {
+                if (client.seq_id == -1 && g_seq_id < n_seq) {
+                    client.seq_id = g_seq_id;
+
+                    client.t_start_prompt = ggml_time_us();
+                    client.t_start_gen    = 0;
+
+                    client.input    = k_prompts[rand() % k_prompts.size()];
+                    client.prompt   = client.input + "\nAssistant:";
+                    client.response = "";
+
+                    std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+
+                    // do not prepend BOS because we have a system prompt!
+                    std::vector<llama_token> tokens_prompt;
+                    tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
+
+                    for (size_t i = 0; i < tokens_prompt.size(); ++i) {
+                        batch.token [batch.n_tokens] = tokens_prompt[i];
+                        batch.pos   [batch.n_tokens] = i + n_tokens_system;
+                        batch.seq_id[batch.n_tokens] = client.id;
+                        batch.logits[batch.n_tokens] = false;
+                        batch.n_tokens += 1;
+                    }
+
+                    // extract the logits only for the last token
+                    if (batch.n_tokens > 0) {
+                        batch.logits[batch.n_tokens - 1] = true;
+                    }
+
+                    client.n_prompt  = tokens_prompt.size();
+                    client.n_decoded = 0;
+                    client.i_batch   = batch.n_tokens - 1;
+
+                    LOG_TEE("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+
+                    g_seq_id += 1;
+
+                    // insert new requests one-by-one
+                    //if (cont_batching) {
+                    //    break;
+                    //}
+                }
+            }
+        }
+
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        // process in chunks of params.n_batch
+        int32_t n_batch = params.n_batch;
+
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+            // experiment: process in powers of 2
+            //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
+            //    n_batch /= 2;
+            //    i -= n_batch;
+            //    continue;
+            //}
+
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            llama_batch batch_view = {
+                n_tokens,
+                batch.token  + i,
+                nullptr,
+                batch.pos    + i,
+                batch.seq_id + i,
+                batch.logits + i,
+                0, 0, 0, // unused
+            };
+
+            const int ret = llama_decode(ctx, batch_view);
+            if (ret != 0) {
+                if (n_batch == 1 || ret < 0) {
+                    // if you get here, it means the KV cache is full - try increasing it via the context size
+                    LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
+                    return 1;
+                }
+
+                LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+
+                n_cache_miss += 1;
+
+                // retry with half the batch size to try to find a free slot in the KV cache
+                n_batch /= 2;
+                i -= n_batch;
+
+                continue;
+            }
+
+            LOG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
+
+            for (auto & client : clients) {
+                if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
+                    continue;
+                }
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
+                //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
+
+                const llama_token id = llama_sample_token(ctx, NULL, NULL, params, client.tokens_prev, candidates, client.i_batch - i);
+
+                if (client.n_decoded == 1) {
+                    // start measuring generation time after the first token to make sure all concurrent clients
+                    // have their prompt already processed
+                    client.t_start_gen = ggml_time_us();
+                }
+
+                // remember which tokens were sampled - used for repetition penalties during sampling
+                client.tokens_prev.erase(client.tokens_prev.begin());
+                client.tokens_prev.push_back(id);
+
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                client.response += token_str;
+                client.sampled = id;
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d: %s\n",
+                //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
+
+                if (client.n_decoded > 2 &&
+                        (id == llama_token_eos(ctx) ||
+                         (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
+                         client.response.find("User:") != std::string::npos ||
+                         client.response.find('\n') != std::string::npos)) {
+                    // basic reverse prompt
+                    const size_t pos = client.response.find("User:");
+                    if (pos != std::string::npos) {
+                        client.response = client.response.substr(0, pos);
+                    }
+
+                    // 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);
+
+                    const auto t_main_end = ggml_time_us();
+
+                    LOG_TEE("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \nInput:    %s\n\033[35mResponse: %s\033[0m\n\n",
+                            client.id, client.seq_id, n_seq, client.n_prompt, client.n_decoded,
+                            (t_main_end - client.t_start_prompt) / 1e6,
+                            (double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
+                            n_cache_miss,
+                            ::trim(client.input).c_str(),
+                            ::trim(client.response).c_str());
+
+                    n_total_prompt += client.n_prompt;
+                    n_total_gen    += client.n_decoded;
+
+                    client.seq_id = -1;
+                }
+
+                client.i_batch = -1;
+            }
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    print_date_time();
+
+    LOG_TEE("\n%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    if (params.prompt_file.empty()) {
+        params.prompt_file = "used built-in defaults";
+    }
+    LOG_TEE("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
+    LOG_TEE("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
+
+    LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Total speed (AVG):   %6s  speed: %5.2f t/s\n", "",             (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Cache misses:        %6d\n", n_cache_miss);
+
+    LOG_TEE("\n");
+
+    llama_print_timings(ctx);
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/perplexity/README.md

@@ -1,3 +1,21 @@
 # perplexity
 
 TODO
+
+## Llama 2 70B Scorechart
+Quantization | Model size (GiB) | Perplexity | Delta to fp16
+-- | -- | -- | --
+Q4_0 | 36.20 | 3.5550 | 3.61%
+Q4_1 | 40.20 | 3.5125 | 2.37%
+Q5_0 | 44.20 | 3.4744 | 1.26%
+Q2_K | 27.27 | 3.7339 | 8.82%
+Q3_K_S | 27.86 | 3.7019 | 7.89%
+Q3_K_M | 30.83 | 3.5932 | 4.72%
+Q3_K_L | 33.67 | 3.5617 | 3.80%
+Q4_K_S | 36.39 | 3.4852 | 1.57%
+Q4_K_M | 38.54 | 3.4725 | 1.20%
+Q5_K_S | 44.20 | 3.4483 | 0.50%
+Q5_K_M | 45.41 | 3.4451 | 0.40%
+Q6_K | 52.70 | 3.4367 | 0.16%
+fp16 | 128.5 | 3.4313 | -
+

examples/perplexity/perplexity.cpp

@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <cmath>
 #include <cstdio>
@@ -28,9 +28,10 @@ struct results_log_softmax {
     float  prob;
 };
 
-void write_logfile(const llama_context * ctx, const gpt_params & params,
-                   const llama_model * model, const struct results_perplexity & results) {
-
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const struct results_perplexity & results
+) {
     if (params.logdir.empty()) {
         return;
     }
@@ -76,10 +77,12 @@ void write_logfile(const llama_context * ctx, const gpt_params & params,
     fclose(logfile);
 }
 
-std::vector<float> softmax(const std::vector<float>& logits) {
+static std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];
-    for (float v : logits) max_logit = std::max(max_logit, v);
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
     double sum_exp = 0.0;
     for (size_t i = 0; i < logits.size(); i++) {
         // Subtract the maximum logit value from the current logit value for numerical stability
@@ -88,25 +91,33 @@ std::vector<float> softmax(const std::vector<float>& logits) {
         sum_exp += exp_logit;
         probs[i] = exp_logit;
     }
-    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
     return probs;
 }
 
-results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
     float max_logit = logits[0];
-    for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
     double sum_exp = 0.0;
-    for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit);
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
     return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
 }
 
-void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
-        double & nll, double & nll2, float * logit_history, float * prob_history) {
-
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history
+) {
     std::mutex mutex;
     int counter = 0;
     auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
-        double local_nll = 0, local_nll2 = 0;
+        double local_nll  = 0;
+        double local_nll2 = 0;
         while (true) {
             std::unique_lock<std::mutex> lock(mutex);
             int i = counter++;
@@ -124,26 +135,39 @@ void process_logits(int n_vocab, const float * logits, const int * tokens, int n
             prob_history[i]  = results.prob;
         }
     };
-    for (auto & w : workers) w = std::thread(compute);
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
     compute();
-    for (auto & w : workers) w.join();
-
+    for (auto & w : workers) {
+        w.join();
+    }
 }
 
-results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
+static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     const bool add_bos = is_spm;
 
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
-    std::vector<float>       logit_history;
-    std::vector<float>       prob_history;
+
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
+    std::vector<float> logit_history;
+    std::vector<float> prob_history;
 
     logit_history.resize(tokens.size());
     prob_history.resize(tokens.size());
@@ -153,20 +177,20 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
         return {tokens, -1, logit_history, prob_history};
     }
 
-    const int calc_chunk = params.n_ctx;
+    const int calc_chunk = n_ctx;
 
     fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
 
     if (int(tokens.size()) <= calc_chunk) {
         fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
-                tokens.size(), params.n_ctx, params.ppl_stride);
+                tokens.size(), n_ctx, params.ppl_stride);
         return {tokens, -1, logit_history, prob_history};
     }
 
     const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -185,12 +209,15 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             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);
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            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};
             }
@@ -225,7 +252,7 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
         }
 
         //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
-        for (int j = params.n_ctx - params.ppl_stride - 1; j < params.n_ctx - 1; ++j) {
+        for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {
 
             // Calculate probability of next token, given the previous ones.
             const std::vector<float> tok_logits(
@@ -252,8 +279,7 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
     return {tokens, std::exp(nll / count), logit_history, prob_history};
 }
 
-results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
-
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     if (params.ppl_stride > 0) {
         return perplexity_v2(ctx, params);
     }
@@ -263,8 +289,9 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     const bool add_bos = is_spm;
+    const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
@@ -274,16 +301,23 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     auto tim2 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
     std::vector<float> logit_history;
     logit_history.resize(tokens.size());
 
     std::vector<float> prob_history;
     prob_history.resize(tokens.size());
 
-    const int n_chunk_max = tokens.size() / params.n_ctx;
+    const int n_chunk_max = tokens.size() / n_ctx;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -295,15 +329,18 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
     for (int i = 0; i < n_chunk; ++i) {
-        const int start =     i * params.n_ctx;
-        const int end   = start + params.n_ctx;
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
 
-        const int num_batches = (params.n_ctx + n_batch - 1) / n_batch;
+        const int num_batches = (n_ctx + n_batch - 1) / n_batch;
 
         std::vector<float> logits;
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -316,7 +353,7 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
                 tokens[batch_start] = llama_token_bos(ctx);
             }
 
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            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};
             }
@@ -324,7 +361,7 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
             // restore the original token in case it was set to BOS
             tokens[batch_start] = token_org;
 
-            const auto batch_logits = llama_get_logits(ctx);
+            const auto * batch_logits = llama_get_logits(ctx);
             logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
         }
 
@@ -353,10 +390,10 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
         // 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 = std::min(512, params.n_ctx/2);
-        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first,
+        const int first = n_ctx/2;
+        process_logits(n_vocab, logits.data() + 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 += params.n_ctx - first - 1;
+        count += n_ctx - first - 1;
 
         // perplexity is e^(average negative log-likelihood)
         if (params.ppl_output_type == 0) {
@@ -365,7 +402,7 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
             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*params.n_ctx, std::exp(nll / count), av, av2);
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
         }
         fflush(stdout);
     }
@@ -385,15 +422,16 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     return {tokens, ppl, logit_history, prob_history};
 }
 
-std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
-        int n_vocab, int n_thread) {
+static std::vector<float> hellaswag_evaluate_tokens(
+    llama_context * ctx, std::vector<int> & tokens, int n_past, int n_batch, int n_vocab
+) {
     std::vector<float> result;
     result.reserve(tokens.size() * n_vocab);
     size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch;
     for (size_t i_chunk = 0; i_chunk < n_chunk; ++i_chunk) {
         size_t n_tokens = tokens.size() - i_chunk * n_batch;
         n_tokens = std::min(n_tokens, size_t(n_batch));
-        if (llama_eval(ctx, tokens.data() + i_chunk * n_batch, n_tokens, n_past, n_thread)) {
+        if (llama_decode(ctx, llama_batch_get_one(tokens.data() + i_chunk * n_batch, n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return {};
         }
@@ -406,7 +444,7 @@ std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vec
     return result;
 }
 
-void hellaswag_score(llama_context * ctx, const gpt_params & params) {
+static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     // Calculates hellaswag score (acc_norm) from prompt
     //
     // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
@@ -440,7 +478,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     size_t hs_task_count = prompt_lines.size()/6;
     fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     fprintf(stderr, "================================= is_spm = %d\n", is_spm);
 
     // This is needed as usual for LLaMA models
@@ -495,7 +533,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     printf("\ntask\tacc_norm\n");
 
     double acc = 0.0f;
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
 
     std::vector<std::vector<int>> ending_tokens(4);
 
@@ -523,7 +562,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         auto query_size = query_embd.size();
 
         // Stop if query wont fit the ctx window
-        if (query_size > (size_t)params.n_ctx) {
+        if (query_size > (size_t)n_ctx) {
             fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
             return;
         }
@@ -533,7 +572,10 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             query_embd.resize(32);
         }
 
-        auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab, params.n_threads);
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
+        auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab);
         if (logits.empty()) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return;
@@ -570,7 +612,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             query_size = query_embd.size();
 
             // Stop if query wont fit the ctx window
-            if (context_size + query_size > (size_t)params.n_ctx) {
+            if (context_size + query_size > (size_t)n_ctx) {
                 fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
                 return;
             }
@@ -582,7 +624,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             //}
 
             // Evaluate the query
-            logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab, params.n_threads);
+            logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab);
             if (logits.empty()) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;
@@ -640,11 +682,11 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     params.n_batch = 512;
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
-    params.perplexity = true;
+    params.logits_all = true;
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     if (params.ppl_stride > 0) {
@@ -653,12 +695,7 @@ int main(int argc, char ** argv) {
         params.n_ctx += params.ppl_stride/2;
     }
 
-    if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
-                "expect poor results\n", __func__, params.n_ctx);
-    }
-
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -683,11 +720,16 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(model);
+    if (params.n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
     struct results_perplexity results;

examples/quantize-stats/CMakeLists.txt

@@ -2,4 +2,5 @@ set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize-stats/quantize-stats.cpp

@@ -1,7 +1,7 @@
-#include "ggml.h"
-#include "build-info.h"
-
 #define LLAMA_API_INTERNAL
+#include "build-info.h"
+#include "common.h"
+#include "ggml.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -34,8 +34,8 @@ struct quantize_stats_params {
     std::vector<enum ggml_type> include_types;
 };
 
-const size_t HISTOGRAM_BUCKETS = 150;
-const double HISTOGRAM_RANGE = 0.03;
+constexpr size_t HISTOGRAM_BUCKETS = 150;
+constexpr double HISTOGRAM_RANGE = 0.03;
 
 struct error_stats {
     size_t num_samples;
@@ -44,8 +44,7 @@ struct error_stats {
     uint64_t error_histogram[HISTOGRAM_BUCKETS];
 };
 
-
-void quantize_stats_print_usage(int /*argc*/, char ** argv) {
+static void quantize_stats_print_usage(int /*argc*/, char ** argv) {
     quantize_stats_params params;
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
@@ -71,7 +70,7 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) {
 }
 
 // Check if a layer is included/excluded by command line
-bool layer_included(const quantize_stats_params params, const std::string & layer) {
+static bool layer_included(const quantize_stats_params & params, const std::string & layer) {
     for (const auto& excluded : params.exclude_layers) {
         if (std::regex_search(layer, std::regex(excluded))) {
             return false;
@@ -86,7 +85,7 @@ bool layer_included(const quantize_stats_params params, const std::string & laye
 }
 
 // Update error statistics given vectors with the before/after result of quantization
-void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
+static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
     for (int64_t i = 0; i < nelements; i++) {
         double diff = input[i] - output[i];
         stats.total_error += diff * diff;
@@ -96,14 +95,14 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
     stats.num_samples += nelements;
 }
 
-void combine_error_stats(error_stats & into, const error_stats & from) {
+static void combine_error_stats(error_stats & into, const error_stats & from) {
     into.num_samples += from.num_samples;
     into.total_error += from.total_error;
     if (from.max_error > into.max_error) into.max_error = from.max_error;
     for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
 }
 
-double find_quantile(const error_stats & stats, double quantile) {
+static double find_quantile(const error_stats & stats, double quantile) {
     double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
 
     double accum = 0;
@@ -116,7 +115,7 @@ double find_quantile(const error_stats & stats, double quantile) {
     return INFINITY;
 }
 
-void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
+static void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
     double rmse = sqrt(stats.total_error / (double) stats.num_samples);
     double median = find_quantile(stats, .5);
     double pct95 = find_quantile(stats, .95);
@@ -143,17 +142,10 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
-void test_roundtrip_on_chunk(
-        const ggml_tensor * layer,
-        int64_t offset,
-        int64_t chunk_size,
-        const ggml_type_traits_t & qfns,
-        bool use_reference,
-        float * input_scratch,
-        char * quantized_scratch,
-        float * output_scratch,
-        error_stats & stats) {
-
+static void test_roundtrip_on_chunk(
+    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference,
+    float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
+) {
     if (layer->type == GGML_TYPE_F16) {
         for (int i = 0; i < chunk_size; i++) {
             input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
@@ -174,18 +166,11 @@ void test_roundtrip_on_chunk(
 
 
 // Run quantization function for a single layer and update error stats
-void test_roundtrip_on_layer(
-        std::string & name,
-        bool print_layer_stats,
-        const ggml_type_traits_t & qfns,
-        bool use_reference,
-        const ggml_tensor * layer,
-        std::vector<float> & input_scratch,
-        std::vector<char> & quantized_scratch,
-        std::vector<float> & output_scratch,
-        error_stats & total_error,
-        int max_thread = 0) {
-
+static void test_roundtrip_on_layer(
+    std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
+    const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
+    std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
+) {
     assert(tensor_is_contiguous(layer));
     error_stats layer_error {};
     uint64_t nelements = ggml_nelements(layer);
@@ -314,7 +299,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     // load the model
     fprintf(stderr, "Loading model\n");
@@ -324,21 +309,22 @@ int main(int argc, char ** argv) {
     llama_context * ctx;
 
     {
-        auto lparams = llama_context_default_params();
+        auto mparams = llama_model_default_params();
+        mparams.use_mlock  = false;
 
-        lparams.n_ctx      = 256;
-        lparams.seed       = 1;
-        lparams.f16_kv     = false;
-        lparams.use_mlock  = false;
-
-        model = llama_load_model_from_file(params.model.c_str(), lparams);
+        model = llama_load_model_from_file(params.model.c_str(), mparams);
 
         if (model == NULL) {
             fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
             return 1;
         }
 
-        ctx = llama_new_context_with_model(model, lparams);
+        auto cparams = llama_context_default_params();
+        cparams.n_ctx      = 256;
+        cparams.seed       = 1;
+        cparams.f16_kv     = false;
+
+        ctx = llama_new_context_with_model(model, cparams);
 
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());

examples/quantize/CMakeLists.txt

@@ -2,6 +2,7 @@ set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

examples/quantize/README.md

@@ -1,3 +1,44 @@
 # quantize
 
 TODO
+
+## Llama 2 7B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.35
+Q3_K_S | 3.50
+Q3_K_M | 3.91
+Q3_K_L | 4.27
+Q4_K_S | 4.58
+Q4_K_M | 4.84
+Q5_K_S | 5.52
+Q5_K_M | 5.68
+Q6_K | 6.56
+
+## Llama 2 13B
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.34
+Q3_K_S | 3.48
+Q3_K_M | 3.89
+Q3_K_L | 4.26
+Q4_K_S | 4.56
+Q4_K_M | 4.83
+Q5_K_S | 5.51
+Q5_K_M | 5.67
+Q6_K | 6.56
+
+# Llama 2 70B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.40
+Q3_K_S | 3.47
+Q3_K_M | 3.85
+Q3_K_L | 4.19
+Q4_K_S | 4.53
+Q4_K_M | 4.80
+Q5_K_S | 5.50
+Q5_K_M | 5.65
+Q6_K | 6.56

examples/quantize/quantize.cpp

@@ -1,5 +1,5 @@
 #include "build-info.h"
-
+#include "common.h"
 #include "llama.h"
 
 #include <cstdio>
@@ -35,10 +35,12 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
+    // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
+    { "COPY",   LLAMA_FTYPE_ALL_F32,       "only copy tensors, no quantizing", },
 };
 
 
-bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
     std::string ftype_str;
 
     for (auto ch : ftype_str_in) {
@@ -70,13 +72,19 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std:
 // usage:
 //  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
-void usage(const char * executable) {
-    fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
-    fprintf(stderr, "  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
-    fprintf(stderr, "  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
-    fprintf(stderr, "\nAllowed quantization types:\n");
+[[noreturn]]
+static void usage(const char * executable) {
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] 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("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
-        printf("  %2d  or  %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
+        if (it.name != "COPY") {
+            printf("  %2d  or  ", it.ftype);
+        } else {
+            printf("          ");
+        }
+        printf("%-6s : %s\n", it.name.c_str(), it.desc.c_str());
     }
     exit(1);
 }
@@ -121,6 +129,9 @@ int main(int argc, char ** argv) {
         // export as [inp path]/ggml-model-[ftype].gguf
         fname_out = fpath + "ggml-model-" + ftype_str + ".gguf";
         arg_idx++;
+        if (ftype_str == "COPY") {
+            params.only_copy = true;
+        }
     }
     else {
         fname_out = argv[arg_idx];
@@ -134,6 +145,9 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
             return 1;
         }
+        if (ftype_str == "COPY") {
+           params.only_copy = true;
+        }
         arg_idx++;
     }
 
@@ -148,7 +162,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
     if (params.nthread > 0) {

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

@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <vector>
 #include <cstdio>
@@ -13,38 +13,32 @@ int main(int argc, char ** argv) {
     params.repeat_last_n = 64;
     params.prompt = "The quick brown fox";
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.n_predict < 0) {
         params.n_predict = 16;
     }
 
-    auto lparams = llama_context_default_params();
-
-    lparams.n_ctx     = params.n_ctx;
-    lparams.seed      = params.seed;
-    lparams.f16_kv    = params.memory_f16;
-    lparams.use_mmap  = params.use_mmap;
-    lparams.use_mlock = params.use_mlock;
-
     auto n_past = 0;
     auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
 
     // init
-    auto model = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
     if (model == nullptr) {
         return 1;
     }
-    auto ctx = llama_new_context_with_model(model, lparams);
     if (ctx == nullptr) {
         llama_free_model(model);
         return 1;
     }
-    auto tokens = llama_tokenize(ctx, params.prompt.c_str(), true);
+    auto tokens = llama_tokenize(ctx, params.prompt, true);
     auto n_prompt_tokens = tokens.size();
     if (n_prompt_tokens < 1) {
         fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
@@ -54,7 +48,7 @@ int main(int argc, char ** argv) {
     }
 
     // evaluate prompt
-    llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads);
+    llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0));
 
     last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
     n_past += n_prompt_tokens;
@@ -78,8 +72,8 @@ int main(int argc, char ** argv) {
     printf("\n%s", params.prompt.c_str());
 
     for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
+        auto * logits = llama_get_logits(ctx);
+        auto n_vocab = llama_n_vocab(model);
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
         for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -91,7 +85,7 @@ int main(int argc, char ** argv) {
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx);
             llama_free_model(model);
@@ -106,7 +100,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
 
     // make new context
-    auto ctx2 = llama_new_context_with_model(model, lparams);
+    auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
     // Load state (rng, logits, embedding and kv_cache) from file
     {
@@ -138,8 +132,8 @@ int main(int argc, char ** argv) {
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx2);
-        auto n_vocab = llama_n_vocab(ctx2);
+        auto * logits = llama_get_logits(ctx2);
+        auto n_vocab = llama_n_vocab(model);
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
         for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -151,7 +145,7 @@ int main(int argc, char ** argv) {
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);

examples/server/README.md

@@ -4,14 +4,14 @@ This example demonstrates a simple HTTP API server and a simple web front end to
 
 Command line options:
 
--   `--threads N`, `-t N`: Set the number of threads to use during computation.
+-   `--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.
 -   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
 -   `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
 -   `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. 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.
--   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
 -   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 -   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
@@ -114,9 +114,9 @@ node index.js
 
     `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
 
-    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
 
-    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: 128, -1 = infinity).
+    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
 
     `n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
     By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
@@ -156,6 +156,8 @@ node index.js
 
     `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
 
+    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
+
 -   **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
@@ -176,6 +178,16 @@ node index.js
 
     `content`: Set the text to process.
 
+    **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+
+    *Options:*
+
+    `input_prefix`: Set the prefix of the code to infill.
+
+    `input_suffix`: Set the suffix of the code to infill.
+
+    It also accepts all the options of `/completion` except `stream` and `prompt`.
+
 ## More examples
 
 ### Interactive mode

examples/server/api_like_OAI.py

@@ -27,10 +27,10 @@ def is_present(json, key):
         buf = json[key]
     except KeyError:
         return False
+    if json[key] == None:
+        return False
     return True
 
-
-
 #convert chat to prompt
 def convert_chat(messages):
     prompt = "" + args.chat_prompt.replace("\\n", "\n")

examples/server/public/index.html

@@ -145,7 +145,29 @@
       color: #888;
     }
 
+
+    @keyframes loading-bg-wipe {
+      0% {
+        background-position: 0%;
+      }
+      100% {
+        background-position: 100%;
+      }
+    }
+
+    .loading {
+      --loading-color-1: #eeeeee00;
+      --loading-color-2: #eeeeeeff;
+      background-size: 50% 100%;
+      background-image: linear-gradient(90deg, var(--loading-color-1), var(--loading-color-2), var(--loading-color-1));
+      animation: loading-bg-wipe 2s linear infinite;
+    }
+
     @media (prefers-color-scheme: dark) {
+      .loading {
+        --loading-color-1: #22222200;
+        --loading-color-2: #222222ff;
+      }
       .popover-content {
         background-color: black;
       }
@@ -321,7 +343,10 @@
     const llamaStats = signal(null)
     const controller = signal(null)
 
-    const generating = computed(() => controller.value == null )
+    // currently generating a completion?
+    const generating = computed(() => controller.value != null)
+
+    // has the user started a chat?
     const chatStarted = computed(() => session.value.transcript.length > 0)
 
     const transcriptUpdate = (transcript) => {
@@ -430,11 +455,19 @@
       return html`
         <form onsubmit=${submit}>
           <div>
-            <textarea type="text" rows=2 onkeypress=${enterSubmits} value="${message}" oninput=${(e) => message.value = e.target.value} placeholder="Say something..."/>
+            <textarea
+               className=${generating.value ? "loading" : null}
+               oninput=${(e) => message.value = e.target.value}
+               onkeypress=${enterSubmits}
+               placeholder="Say something..."
+               rows=2
+               type="text"
+               value="${message}"
+            />
           </div>
           <div class="right">
-            <button type="submit" disabled=${!generating.value} >Send</button>
-            <button onclick=${stop} disabled=${generating}>Stop</button>
+            <button type="submit" disabled=${generating.value}>Send</button>
+            <button onclick=${stop} disabled=${!generating.value}>Stop</button>
             <button onclick=${reset}>Reset</button>
           </div>
         </form>

examples/server/server.cpp

@@ -17,6 +17,8 @@
 #include "completion.js.hpp"
 #include "json-schema-to-grammar.mjs.hpp"
 
+#include <cstddef>
+
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
 #endif
@@ -116,7 +118,7 @@ static void server_log(const char *level, const char *function, int line,
     }
 
     const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
-    fprintf(stdout, "%.*s\n", (int)str.size(), str.data());
+    printf("%.*s\n", (int)str.size(), str.data());
     fflush(stdout);
 }
 
@@ -137,7 +139,7 @@ static std::string tokens_to_output_formatted_string(const llama_context *ctx, c
 }
 
 // convert a vector of completion_token_output to json
-static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> probs)
+static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> & probs)
 {
     json out = json::array();
     for (const auto &prob : probs)
@@ -198,6 +200,7 @@ struct llama_server_context
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
     gpt_params params;
+    int n_ctx;
 
     grammar_parser::parse_state parsed_grammar;
     llama_grammar *grammar = nullptr;
@@ -237,7 +240,7 @@ struct llama_server_context
         num_prompt_tokens = 0;
         num_tokens_predicted = 0;
         generated_text = "";
-        generated_text.reserve(params.n_ctx);
+        generated_text.reserve(n_ctx);
         generated_token_probs.clear();
         truncated = false;
         stopped_eos = false;
@@ -263,13 +266,13 @@ struct llama_server_context
             LOG_ERROR("unable to load model", {{"model", params_.model}});
             return false;
         }
-
-        last_n_tokens.resize(params.n_ctx);
+        n_ctx = llama_n_ctx(ctx);
+        last_n_tokens.resize(n_ctx);
         std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
         return true;
     }
 
-    std::vector<llama_token> tokenize(json json_prompt, bool add_bos)
+    std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
     {
         // If `add_bos` is true, we only add BOS, when json_prompt is a string,
         // or the first element of the json_prompt array is a string.
@@ -339,9 +342,15 @@ struct llama_server_context
         return true;
     }
 
-    void loadPrompt()
+    void loadInfill()
     {
-        auto prompt_tokens = tokenize(prompt, true);  // always add BOS
+        auto prefix_tokens = tokenize(params.input_prefix, true);  // always add BOS
+        auto suffix_tokens = tokenize(params.input_suffix, true);  // always add BOS
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+        prefix_tokens.push_back(llama_token_middle(ctx));
+        auto prompt_tokens = prefix_tokens;
 
         num_prompt_tokens = prompt_tokens.size();
 
@@ -354,6 +363,8 @@ struct llama_server_context
         // if input prompt is too big, truncate like normal
         if (num_prompt_tokens >= (size_t)params.n_ctx)
         {
+            printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
+            // todo we probably want to cut from both sides
             const int n_left = (params.n_ctx - params.n_keep) / 2;
             std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
             const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
@@ -379,6 +390,66 @@ struct llama_server_context
 
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
+        embd = prompt_tokens;
+        if (n_past == num_prompt_tokens)
+        {
+            // we have to evaluate at least 1 token to generate logits.
+            printf("we have to evaluate at least 1 token to generate logits\n");
+            n_past--;
+        }
+
+        LOG_VERBOSE("prompt ingested", {
+                                           {"n_past", n_past},
+                                           {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
+                                           {"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
+                                       });
+
+        has_next_token = true;
+    }
+    void loadPrompt()
+    {
+        auto prompt_tokens = tokenize(prompt, true);  // always add BOS
+
+        num_prompt_tokens = prompt_tokens.size();
+
+        if (params.n_keep < 0)
+        {
+            params.n_keep = (int)num_prompt_tokens;
+        }
+        params.n_keep = std::min(n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (num_prompt_tokens >= (size_t)n_ctx)
+        {
+            const int n_left = (n_ctx - params.n_keep) / 2;
+            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
+            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
+            std::copy(prompt_tokens.end() - n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+
+            LOG_VERBOSE("input truncated", {
+                                               {"n_ctx", n_ctx},
+                                               {"n_keep", params.n_keep},
+                                               {"n_left", n_left},
+                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                                           });
+
+            truncated = true;
+            prompt_tokens = new_tokens;
+        }
+        else
+        {
+            const size_t ps = num_prompt_tokens;
+            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
+            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
         embd = prompt_tokens;
         if (n_past == num_prompt_tokens)
         {
@@ -407,37 +478,47 @@ struct llama_server_context
         completion_token_output result;
         result.tok = -1;
 
-        if (embd.size() >= (size_t)params.n_ctx)
+        if (embd.size() >= (size_t)n_ctx)
         {
-            // Reset context
-            const int n_left = (params.n_ctx - params.n_keep) / 2;
+            // Shift context
+
+            const int n_left    = n_past - params.n_keep - 1;
+            const int n_discard = n_left/2;
+
+            llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+            llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+            for (size_t i = params.n_keep + 1 + n_discard; i < embd.size(); i++)
+            {
+                embd[i - n_discard] = embd[i];
+            }
+            embd.resize(embd.size() - n_discard);
+
+            n_past -= n_discard;
 
-            std::vector<llama_token> new_tokens(embd.begin(), embd.begin() + params.n_keep);
-            new_tokens.insert(new_tokens.end(), embd.end() - n_left, embd.end());
-            embd = new_tokens;
-            n_past = params.n_keep;
             truncated = true;
             LOG_VERBOSE("input truncated", {
-                                               {"n_ctx", params.n_ctx},
+                                               {"n_ctx", n_ctx},
                                                {"n_keep", params.n_keep},
                                                {"n_left", n_left},
-                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
                                            });
         }
 
+        bool tg = true;
         while (n_past < embd.size())
         {
             int n_eval = (int)embd.size() - n_past;
+            tg = n_eval == 1;
             if (n_eval > params.n_batch)
             {
                 n_eval = params.n_batch;
             }
-            if (llama_eval(ctx, &embd[n_past], n_eval, n_past, params.n_threads))
+
+            if (llama_decode(ctx, llama_batch_get_one(&embd[n_past], n_eval, n_past, 0)))
             {
                 LOG_ERROR("failed to eval", {
                                                 {"n_eval", n_eval},
                                                 {"n_past", n_past},
-                                                {"n_threads", params.n_threads},
                                                 {"embd", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
                                             });
                 has_next_token = false;
@@ -453,98 +534,20 @@ struct llama_server_context
             return result;
         }
 
-        // out of user input, sample next token
-        const float temp = params.temp;
-        const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
-        const float top_p = params.top_p;
-        const float tfs_z = params.tfs_z;
-        const float typical_p = params.typical_p;
-        const int32_t repeat_last_n = params.repeat_last_n < 0 ? params.n_ctx : params.repeat_last_n;
-        const float repeat_penalty = params.repeat_penalty;
-        const float alpha_presence = params.presence_penalty;
-        const float alpha_frequency = params.frequency_penalty;
-        const int mirostat = params.mirostat;
-        const float mirostat_tau = params.mirostat_tau;
-        const float mirostat_eta = params.mirostat_eta;
-        const bool penalize_nl = params.penalize_nl;
-        const int32_t n_probs = params.n_probs;
-
         {
-            auto *logits = llama_get_logits(ctx);
-            auto n_vocab = llama_n_vocab(ctx);
-
-            // Apply params.logit_bias map
-            for (const auto &it : params.logit_bias)
-            {
-                logits[it.first] += it.second;
-            }
-
+            // out of user input, sample next token
             std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++)
+            candidates.reserve(llama_n_vocab(model));
+
+            result.tok = llama_sample_token(ctx, NULL, grammar, params, last_n_tokens, candidates);
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            const int32_t n_probs = params.n_probs;
+            if (params.temp <= 0 && n_probs > 0)
             {
-                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-            }
-
-            llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
-
-            // Apply penalties
-            float nl_logit = logits[llama_token_nl(ctx)];
-            auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), params.n_ctx);
-            llama_sample_repetition_penalty(ctx, &candidates_p,
-                                            last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                                            last_n_repeat, repeat_penalty);
-            llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-                                                          last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                                                          last_n_repeat, alpha_frequency, alpha_presence);
-            if (!penalize_nl)
-            {
-                logits[llama_token_nl(ctx)] = nl_logit;
-            }
-
-            if (grammar != nullptr) {
-                llama_sample_grammar(ctx, &candidates_p, grammar);
-            }
-
-            if (temp <= 0)
-            {
-                // Greedy sampling
-                result.tok = llama_sample_token_greedy(ctx, &candidates_p);
-                if (n_probs > 0)
-                {
-                    llama_sample_softmax(ctx, &candidates_p);
-                }
-            }
-            else
-            {
-                if (mirostat == 1)
-                {
-                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    const int mirostat_m = 100;
-                    llama_sample_temperature(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-                }
-                else if (mirostat == 2)
-                {
-                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    llama_sample_temperature(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-                }
-                else
-                {
-                    // Temperature sampling
-                    size_t min_keep = std::max(1, n_probs);
-                    llama_sample_top_k(ctx, &candidates_p, top_k, min_keep);
-                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, min_keep);
-                    llama_sample_typical(ctx, &candidates_p, typical_p, min_keep);
-                    llama_sample_top_p(ctx, &candidates_p, top_p, min_keep);
-                    llama_sample_temperature(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token(ctx, &candidates_p);
-                }
-            }
-
-            if (grammar != nullptr) {
-                llama_grammar_accept_token(ctx, grammar, result.tok);
+                // For llama_sample_token_greedy we need to sort candidates
+                llama_sample_softmax(ctx, &candidates_p);
             }
 
             for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
@@ -554,7 +557,9 @@ struct llama_server_context
 
             last_n_tokens.erase(last_n_tokens.begin());
             last_n_tokens.push_back(result.tok);
-            num_tokens_predicted++;
+            if (tg) {
+                num_tokens_predicted++;
+            }
         }
 
         // add it to the context
@@ -609,7 +614,7 @@ struct llama_server_context
 
     completion_token_output doCompletion()
     {
-        const completion_token_output token_with_probs = nextToken();
+        auto token_with_probs = nextToken();
 
         const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
         generated_text += token_text;
@@ -675,7 +680,7 @@ struct llama_server_context
 
     std::vector<float> getEmbedding()
     {
-        static const int n_embd = llama_n_embd(ctx);
+        static const int n_embd = llama_n_embd(model);
         if (!params.embedding)
         {
             LOG_WARNING("embedding disabled", {
@@ -692,50 +697,49 @@ struct llama_server_context
 static void server_print_usage(const char *argv0, const gpt_params &params,
                                const server_params &sparams)
 {
-    fprintf(stdout, "usage: %s [options]\n", argv0);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stdout, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("usage: %s [options]\n", argv0);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  --rope-freq-base N    RoPE base frequency (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: loaded from model)\n");
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported())
     {
-        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported())
     {
-        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
-    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
-    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "  -a ALIAS, --alias ALIAS\n");
-    fprintf(stdout, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stdout, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
-    fprintf(stdout, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
-    fprintf(stdout, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
-    fprintf(stdout, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
-    fprintf(stdout, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
-    fprintf(stdout, "\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -a ALIAS, --alias ALIAS\n");
+    printf("                        set an alias for the model, will be added as `model` field in completion response\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+    printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
+    printf("\n");
 }
 
 static void server_params_parse(int argc, char **argv, server_params &sparams,
@@ -903,14 +907,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
 #else
             LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
-#endif // GGML_USE_CUBLAS
-        }
-        else if (arg == "--low-vram" || arg == "-lv")
-        {
-#ifdef GGML_USE_CUBLAS
-            params.low_vram = true;
-#else
-            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--no-mul-mat-q" || arg == "-nommq")
@@ -941,7 +937,23 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter = argv[i];
+            params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f));
+            params.use_mmap = false;
+        }
+        else if (arg == "--lora-scaled")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i])));
             params.use_mmap = false;
         }
         else if (arg == "--lora-base")
@@ -1000,7 +1012,7 @@ static json format_generation_settings(llama_server_context &llama)
                             eos_bias->second < 0.0f && std::isinf(eos_bias->second);
 
     return json{
-        {"n_ctx", llama.params.n_ctx},
+        {"n_ctx", llama.n_ctx},
         {"model", llama.params.model_alias},
         {"seed", llama.params.seed},
         {"temp", llama.params.temp},
@@ -1038,8 +1050,6 @@ static json format_timings(llama_server_context &llama)
 {
     const auto timings = llama_get_timings(llama.ctx);
 
-    assert(timings.n_eval == llama.num_tokens_predicted);
-
     return json{
         {"prompt_n", timings.n_p_eval},
         {"prompt_ms", timings.t_p_eval_ms},
@@ -1081,8 +1091,9 @@ static json format_final_response(llama_server_context &llama, const std::string
     return res;
 }
 
-static json format_partial_response(llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs)
-{
+static json format_partial_response(
+    llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs
+) {
     json res = json{
         {"content", content},
         {"stop", false},
@@ -1159,7 +1170,7 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     const auto &logit_bias = body.find("logit_bias");
     if (logit_bias != body.end() && logit_bias->is_array())
     {
-        const int n_vocab = llama_n_vocab(llama.ctx);
+        const int n_vocab = llama_n_vocab(llama.model);
         for (const auto &el : *logit_bias)
         {
             if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
@@ -1196,6 +1207,27 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
 }
 
+static void parse_options_infill(const json &body, llama_server_context &llama)
+{
+    if (body.count("input_prefix") != 0)
+    {
+        llama.params.input_prefix = body["input_prefix"];
+    }
+    else
+    {
+        llama.params.input_prefix = "";
+    }
+    if (body.count("input_suffix") != 0)
+    {
+        llama.params.input_suffix = body["input_suffix"];
+    }
+    else
+    {
+        llama.params.input_suffix = "";
+    }
+    parse_options_completion(body, llama);
+}
+
 static void log_server_request(const Request &req, const Response &res)
 {
     LOG_INFO("request", {
@@ -1213,7 +1245,7 @@ static void log_server_request(const Request &req, const Response &res)
                            });
 }
 
-bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) {
+static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
     return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
 }
 
@@ -1223,7 +1255,7 @@ bool is_at_eob(llama_server_context & server_context, const llama_token * tokens
 //  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
 //    This is also called when the stop condition is met.
 //    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
-void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
+static void beam_search_callback(void *callback_data, llama_beams_state beams_state) {
     auto & llama = *static_cast<llama_server_context*>(callback_data);
     // Mark beams as EOS as needed.
     for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
@@ -1239,7 +1271,7 @@ void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
         const llama_token * tokens = beams_state.beam_views[0].tokens;
         const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
         std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
-        printf("%lu", n);
+        printf("%zu", n);
     }
     fflush(stdout);
 #if 0 // DEBUG: print current beams for this iteration
@@ -1253,10 +1285,11 @@ void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
 struct token_translator {
     llama_context * ctx;
     std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); }
-    std::string operator()(completion_token_output cto) const { return (*this)(cto.tok); }
+    std::string operator()(const completion_token_output & cto) const { return (*this)(cto.tok); }
 };
 
-void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) {
+static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama)
+{
     auto & gtps = llama.generated_token_probs;
     auto translator = token_translator{llama.ctx};
     auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
@@ -1291,6 +1324,7 @@ int main(int argc, char **argv)
                             {"commit", BUILD_COMMIT}});
     LOG_INFO("system info", {
                                 {"n_threads", params.n_threads},
+                                {"n_threads_batch", params.n_threads_batch},
                                 {"total_threads", std::thread::hardware_concurrency()},
                                 {"system_info", llama_print_system_info()},
                             });
@@ -1354,7 +1388,7 @@ int main(int argc, char **argv)
             if (llama.params.n_beams) {
                 // Fill llama.generated_token_probs vector with final beam.
                 llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams,
-                                  llama.n_past, llama.n_remain, llama.params.n_threads);
+                                  llama.n_past, llama.n_remain);
                 // Translate llama.generated_token_probs to llama.generated_text.
                 append_to_generated_text_from_generated_token_probs(llama);
             } else {
@@ -1377,7 +1411,13 @@ int main(int argc, char **argv)
                 }
             }
 
-            const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
+            auto probs = llama.generated_token_probs;
+            if (llama.params.n_probs > 0 && llama.stopped_word) {
+                const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
+                probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
+            }
+
+            const json data = format_final_response(llama, llama.generated_text, probs);
 
             llama_print_timings(llama.ctx);
 
@@ -1454,7 +1494,11 @@ int main(int argc, char **argv)
 
                     if (!llama.has_next_token) {
                         // Generation is done, send extra information.
-                        const json data = format_final_response(llama, "", llama.generated_token_probs);
+                        const json data = format_final_response(
+                            llama,
+                            "",
+                            std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
+                        );
 
                         const std::string str =
                             "data: " +
@@ -1484,6 +1528,127 @@ int main(int argc, char **argv)
             res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         } });
 
+    svr.Post("/infill", [&llama](const Request &req, Response &res)
+             {
+        auto lock = llama.lock();
+
+        llama.rewind();
+
+        llama_reset_timings(llama.ctx);
+
+        parse_options_infill(json::parse(req.body), llama);
+
+        if (!llama.loadGrammar())
+        {
+            res.status = 400;
+            return;
+        }
+        llama.loadInfill();
+        llama.beginCompletion();
+        const auto chunked_content_provider = [&](size_t, DataSink & sink) {
+            size_t sent_count = 0;
+            size_t sent_token_probs_index = 0;
+
+            while (llama.has_next_token) {
+                const completion_token_output token_with_probs = llama.doCompletion();
+                if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
+                    continue;
+                }
+                const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
+
+                size_t pos = std::min(sent_count, llama.generated_text.size());
+
+                const std::string str_test = llama.generated_text.substr(pos);
+                bool is_stop_full = false;
+                size_t stop_pos =
+                    llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
+                if (stop_pos != std::string::npos) {
+                    is_stop_full = true;
+                    llama.generated_text.erase(
+                        llama.generated_text.begin() + pos + stop_pos,
+                        llama.generated_text.end());
+                    pos = std::min(sent_count, llama.generated_text.size());
+                } else {
+                    is_stop_full = false;
+                    stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
+                        STOP_PARTIAL);
+                }
+
+                if (
+                    stop_pos == std::string::npos ||
+                    // Send rest of the text if we are at the end of the generation
+                    (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+                ) {
+                    const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
+
+                    sent_count += to_send.size();
+
+                    std::vector<completion_token_output> probs_output = {};
+
+                    if (llama.params.n_probs > 0) {
+                        const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                        size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                        size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                        if (probs_pos < probs_stop_pos) {
+                            probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                        }
+                        sent_token_probs_index = probs_stop_pos;
+                    }
+
+                    const json data = format_partial_response(llama, to_send, probs_output);
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+
+                if (!llama.has_next_token) {
+                    // Generation is done, send extra information.
+                    const json data = format_final_response(
+                        llama,
+                        "",
+                        std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
+                    );
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+            }
+
+            llama_print_timings(llama.ctx);
+            sink.done();
+            return true;
+        };
+        const auto on_complete = [&](bool) {
+            llama.mutex.unlock();
+        };
+        lock.release();
+        res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
+        });
+
     svr.Get("/model.json", [&llama](const Request &, Response &res)
             {
         const json data = format_generation_settings(llama);
@@ -1548,7 +1713,7 @@ int main(int argc, char **argv)
 
     svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep)
                               {
-        const auto * fmt = "500 Internal Server Error\n%s";
+        const char fmt[] = "500 Internal Server Error\n%s";
         char buf[BUFSIZ];
         try {
             std::rethrow_exception(std::move(ep));
@@ -1583,7 +1748,7 @@ int main(int argc, char **argv)
     svr.set_base_dir(sparams.public_path);
 
     // to make it ctrl+clickable:
-    fprintf(stdout, "\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+    printf("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
 
     LOG_INFO("HTTP server listening", {
                                           {"hostname", sparams.hostname},

examples/simple/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/simple/README.md

@@ -0,0 +1,21 @@
+# llama.cpp/example/simple
+
+The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
+
+```bash
+./simple ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is"
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 1, n_kv_req = 32
+
+ Hello my name is Shawn and I'm a 20 year old male from the United States. I'm a 20 year old
+
+main: decoded 27 tokens in 2.31 s, speed: 11.68 t/s
+
+llama_print_timings:        load time =   579.15 ms
+llama_print_timings:      sample time =     0.72 ms /    28 runs   (    0.03 ms per token, 38888.89 tokens per second)
+llama_print_timings: prompt eval time =   655.63 ms /    10 tokens (   65.56 ms per token,    15.25 tokens per second)
+llama_print_timings:        eval time =  2180.97 ms /    27 runs   (   80.78 ms per token,    12.38 tokens per second)
+llama_print_timings:       total time =  2891.13 ms
+```

examples/simple/simple.cpp

@@ -1,9 +1,3 @@
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
 
@@ -32,35 +26,62 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    // total length of the sequence including the prompt
+    const int n_len = 32;
+
     // init LLM
 
     llama_backend_init(params.numa);
 
-    llama_context_params ctx_params = llama_context_default_params();
+    // initialize the model
 
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
+    llama_model_params model_params = llama_model_default_params();
+
+    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = 2048;
+    ctx_params.n_threads = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(ctx, params.prompt, true);
 
-    const int max_context_size     = llama_n_ctx(ctx);
-    const int max_tokens_list_size = max_context_size - 4;
+    const int n_ctx    = llama_n_ctx(ctx);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
 
-    if ((int) tokens_list.size() > max_tokens_list_size) {
-        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
         return 1;
     }
 
-    fprintf(stderr, "\n\n");
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
 
     for (auto id : tokens_list) {
         fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
@@ -68,63 +89,104 @@ int main(int argc, char ** argv) {
 
     fflush(stderr);
 
+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
+
+    llama_batch batch = llama_batch_init(512, 0);
+
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
     // main loop
 
-    // The LLM keeps a contextual cache memory of previous token evaluation.
-    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
-    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
-    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
 
-    const int n_gen = std::min(32, max_context_size);
+    const auto t_main_start = ggml_time_us();
 
-    while (llama_get_kv_cache_token_count(ctx) < n_gen) {
-        // evaluate the transformer
+    while (n_cur <= n_len) {
+        // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
 
-        if (llama_eval(ctx, tokens_list.data(), int(tokens_list.size()), llama_get_kv_cache_token_count(ctx), params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            // sample the most likely token
+            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
+
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
+
+            // prepare the next batch
+            batch.n_tokens = 0;
+
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = 0;
+            batch.logits[batch.n_tokens] = true;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
-
-        tokens_list.clear();
-
-        // sample the next token
-
-        llama_token new_token_id = 0;
-
-        auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
-
-        // is it an end of stream ?
-        if (new_token_id == llama_token_eos(ctx)) {
-            fprintf(stderr, " [end of text]\n");
-            break;
-        }
-
-        // print the new token :
-        printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
-        fflush(stdout);
-
-        // push this new token for next evaluation
-        tokens_list.push_back(new_token_id);
     }
 
+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
     llama_free(ctx);
     llama_free_model(model);
 
     llama_backend_free();
 
-    fprintf(stderr, "\n\n");
-
     return 0;
 }

examples/speculative/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET speculative)
+add_executable(${TARGET} speculative.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/speculative/speculative.cpp

@@ -0,0 +1,314 @@
+#include "build-info.h"
+
+#include "common.h"
+#include "llama.h"
+#include "grammar-parser.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.model_draft.empty()) {
+        fprintf(stderr, "%s: error: --model-draft is required\n", __func__);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("speculative", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model_tgt = NULL;
+    llama_model * model_dft = NULL;
+
+    llama_context * ctx_tgt = NULL;
+    llama_context * ctx_dft = NULL;
+
+    // load the target model
+    params.logits_all = true;
+    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+
+    // load the draft model
+    params.model = params.model_draft;
+    params.n_gpu_layers = params.n_gpu_layers_draft;
+    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+
+    // tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
+
+    const int max_context_size     = llama_n_ctx(ctx_tgt);
+    const int max_tokens_list_size = max_context_size - 4;
+
+    if ((int) inp.size() > max_tokens_list_size) {
+        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
+        return 1;
+    }
+
+    fprintf(stderr, "\n\n");
+
+    for (auto id : inp) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx_tgt, id).c_str());
+    }
+
+    fflush(stderr);
+
+    const int n_input = inp.size();
+
+    const auto t_enc_start = ggml_time_us();
+
+    // eval the prompt with both models
+    llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input,     0,           0));
+
+    const auto t_enc_end = ggml_time_us();
+
+    // the 2 models should have the same vocab
+    const int n_ctx   = llama_n_ctx(ctx_tgt);
+    const int n_vocab = llama_n_vocab(model_tgt);
+    //GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
+
+    // how many tokens to draft each time
+    int n_draft = params.n_draft;
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    int n_past_tgt = inp.size();
+    int n_past_dft = inp.size();
+
+    std::vector<llama_token> drafted;
+
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+
+    for (auto & id : inp) {
+        last_tokens.erase(last_tokens.begin());
+        last_tokens.push_back(id);
+    }
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    // used to determine end of generation
+    bool has_eos = false;
+
+    // grammar stuff
+    struct llama_grammar * grammar_dft = NULL;
+    struct llama_grammar * grammar_tgt = NULL;
+
+    grammar_parser::parse_state parsed_grammar;
+
+    // if requested - load the grammar, error checking is omitted for brevity
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));
+
+        int i_dft = 0;
+
+        while (true) {
+            // sample from the target model
+            llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
+
+            // remember which tokens were sampled - used for repetition penalties during sampling
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
+
+            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens));
+
+            const std::string token_str = llama_token_to_piece(ctx_tgt, id);
+            printf("%s", token_str.c_str());
+            fflush(stdout);
+
+            if (id == llama_token_eos(ctx_tgt)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
+
+            // check if the draft matches the target
+            if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past_tgt;
+                ++n_past_dft;
+                ++i_dft;
+
+                continue;
+            }
+
+            // the drafted token was rejected or we are out of drafted tokens
+
+            if (i_dft < (int) drafted.size()) {
+                LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
+                        i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
+            } else {
+                LOG("out of drafted tokens\n");
+            }
+
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
+            ++n_past_dft;
+
+            // heuristic for n_draft
+            {
+                const int  n_draft_cur  = (int) drafted.size();
+                const bool all_accepted = i_dft == n_draft_cur;
+
+                LOG("n_draft      = %d\n", n_draft);
+                LOG("n_draft_cur  = %d\n", n_draft_cur);
+                LOG("i_dft        = %d\n", i_dft);
+                LOG("all_accepted = %d\n", all_accepted);
+
+                if (all_accepted && n_draft == n_draft_cur) {
+                    LOG(" - max drafted tokens accepted - n_draft += 8\n");
+                    n_draft = std::min(30, n_draft + 8);
+                } else if (all_accepted) {
+                    LOG(" - partially drafted tokens accepted - no change\n");
+                } else {
+                    LOG(" - drafted token rejected - n_draft -= 1\n");
+                    n_draft = std::max(2, n_draft - 1);
+                }
+            }
+
+            drafted.clear();
+            drafted.push_back(id);
+
+            break;
+        }
+
+        if (n_predict > params.n_predict || has_eos) {
+            break;
+        }
+
+        if (grammar_tgt) {
+            if (grammar_dft) {
+                llama_grammar_free(grammar_dft);
+            }
+            grammar_dft = llama_grammar_copy(grammar_tgt);
+
+            LOG("copied target grammar to draft grammar\n");
+        }
+
+        // sample n_draft tokens from the draft model using greedy decoding
+        int n_past_cur = n_past_dft;
+        for (int i = 0; i < n_draft; ++i) {
+            float * logits = llama_get_logits(ctx_dft);
+
+            candidates.clear();
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+            }
+
+            llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+
+            if (grammar_dft != NULL) {
+                llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
+            }
+
+            // computes softmax and sorts the candidates
+            llama_sample_softmax(ctx_dft, &cur_p);
+
+            for (int i = 0; i < 3; ++i) {
+                LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
+            }
+
+            // TODO: better logic?
+            if (cur_p.data[0].p < 2*cur_p.data[1].p) {
+                LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
+                break;
+            }
+
+            // drafted token
+            const llama_token id = cur_p.data[0].id;
+
+            drafted.push_back(id);
+            ++n_drafted;
+
+            // no need to evaluate the last drafted token, since we won't use the result
+            if (i == n_draft - 1) {
+                break;
+            }
+
+            // evaluate the drafted token on the draft model
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
+            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
+            ++n_past_cur;
+
+            if (grammar_dft != NULL) {
+                llama_grammar_accept_token(ctx_dft, grammar_dft, id);
+            }
+        }
+
+        // evaluate the target model on the drafted tokens
+        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
+        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
+        ++n_past_tgt;
+
+        // the first token is always proposed by the traget model before the speculation loop
+        drafted.erase(drafted.begin());
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    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));
+
+    // TODO: make sure these numbers are computed correctly
+    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("n_accept  = %d\n", n_accept);
+    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_TEE("\ndraft:\n");
+    llama_print_timings(ctx_dft);
+
+    LOG_TEE("\ntarget:\n");
+    llama_print_timings(ctx_tgt);
+
+    llama_free(ctx_tgt);
+    llama_free_model(model_tgt);
+
+    llama_free(ctx_dft);
+    llama_free_model(model_dft);
+
+    if (grammar_dft != NULL) {
+        llama_grammar_free(grammar_dft);
+        llama_grammar_free(grammar_tgt);
+    }
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}