simple : fixes

ggml-ci

.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: "

.editorconfig

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

.github/workflows/build.yml

@@ -18,7 +18,6 @@ on:
 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
@@ -53,7 +52,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -88,7 +87,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -122,7 +121,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -150,7 +149,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -175,7 +174,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -198,6 +197,62 @@ jobs:
           cd build
           ctest --verbose --timeout 900
 
+  macOS-latest-cmake-ios:
+    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
+          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
+
+  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
+          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
+
   windows-latest-cmake:
     runs-on: windows-latest
 
@@ -210,22 +265,24 @@ jobs:
       matrix:
         include:
           - build: 'noavx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
+            defines: '-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_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
+            defines: '-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'
           - build: 'clblast'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$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_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
@@ -335,27 +392,29 @@ 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 .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
           cmake --build . --config Release
 
       - name: Determine tag name
@@ -385,27 +444,11 @@ 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: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -414,6 +457,22 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ 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'
+        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
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -430,7 +489,9 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
       - name: Determine tag name
         id: tag
@@ -488,7 +549,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -512,7 +573,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -536,7 +597,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -566,7 +627,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -605,7 +666,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -651,7 +712,7 @@ 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 +1,43 @@
+# 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: poetry build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        password: ${{ secrets.PYPI_API_TOKEN }}

.gitignore

@@ -6,6 +6,10 @@
 *.exe
 *.dll
 *.log
+*.gcov
+*.gcno
+*.gcda
+*.dot
 .DS_Store
 .build/
 .cache/
@@ -17,6 +21,9 @@
 .vs/
 .vscode/
 
+lcov-report/
+gcovr-report/
+
 build*/
 out/
 tmp/
@@ -24,24 +31,30 @@ tmp/
 models/*
 models-mnt
 
-/main
-/quantize
-/quantize-stats
-/result
-/perplexity
-/embedding
-/train-text-from-scratch
-/convert-llama2c-to-ggml
-/simple
-/benchmark-matmult
-/vdot
-/server
 /Pipfile
+/baby-llama
+/beam-search
+/benchmark-matmult
+/convert-llama2c-to-ggml
 /embd-input-test
+/embedding
 /gguf
 /gguf-llama-simple
 /libllama.so
 /llama-bench
+/main
+/metal
+/perplexity
+/quantize
+/quantize-stats
+/result
+/save-load-state
+/server
+/simple
+/speculative
+/parallel
+/train-text-from-scratch
+/vdot
 build-info.h
 arm_neon.h
 compile_commands.json

CMakeLists.txt

@@ -36,6 +36,12 @@ 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)
@@ -74,9 +80,12 @@ 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)
@@ -128,6 +137,7 @@ 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)
@@ -158,6 +168,32 @@ if (APPLE AND LLAMA_ACCELERATE)
     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 +270,8 @@ 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 +287,8 @@ 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 +306,7 @@ 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)
@@ -293,34 +332,12 @@ 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})
@@ -343,7 +360,8 @@ 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 +389,15 @@ 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)
 
@@ -403,15 +423,21 @@ if (LLAMA_ALL_WARNINGS)
             -Wpointer-arith
             -Wmissing-prototypes
             -Werror=implicit-int
+            -Wno-unused-function
         )
         set(cxx_flags
             -Wall
             -Wextra
             -Wpedantic
             -Wcast-qual
+            -Wmissing-declarations
             -Wno-unused-function
             -Wno-multichar
         )
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+            # g++ only
+            set(cxx_flags ${cxx_flags} -Wno-format-truncation -Wno-array-bounds)
+        endif()
     else()
         # todo : msvc
     endif()
@@ -423,7 +449,7 @@ if (LLAMA_ALL_WARNINGS)
 
 endif()
 
-if (MSVC)
+if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 
     if (BUILD_SHARED_LIBS)
@@ -445,6 +471,13 @@ 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)
@@ -460,25 +493,33 @@ if (NOT MSVC)
     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)
@@ -535,27 +576,84 @@ 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_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 $<TARGET_OBJECTS:ggml>)
 if (BUILD_SHARED_LIBS)
@@ -585,14 +683,53 @@ if (BUILD_SHARED_LIBS)
     if (LLAMA_METAL)
         set_target_properties(llama PROPERTIES RESOURCE "${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})
+
+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

@@ -1,27 +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 save-load-state server embd-input-test gguf llama-bench baby-llama beam_search tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative parallel 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
 
-default: $(BUILD_TARGETS)
-
-test:
-	@echo "Running tests..."
-	@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 [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
-			continue; \
-		elif [ "$$test_target" = "tests/test-tokenizer-1" ]; then \
-			continue; \
-		else \
-			./$$test_target; \
-		fi; \
-	done
-	@echo "All tests have been run."
-
-all: $(BUILD_TARGETS) $(TEST_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)
@@ -35,12 +19,13 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
-CXXV := $(shell $(CXX) --version | head -n 1)
-
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(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)
@@ -51,64 +36,164 @@ 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 [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
+			continue; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; 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
+
+CCV := $(shell $(CC) --version | head -n 1)
+CXXV := $(shell $(CXX) --version | head -n 1)
+
 #
 # 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 -Werror=implicit-int
-CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
+MK_CFLAGS    += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
+				-Wmissing-prototypes -Werror=implicit-int -Wno-unused-function
+MK_CXXFLAGS  += -Wall -Wextra -Wpedantic -Wcast-qual -Wmissing-declarations -Wno-unused-function -Wno-multichar
+
+# TODO(cebtenzzre): remove this once PR #2632 gets merged
+TTFS_CXXFLAGS = $(CXXFLAGS) -Wno-missing-declarations
+
+ifneq '' '$(findstring clang,$(shell $(CXX) --version))'
+	# clang++ only
+	MK_CXXFLAGS   += -Wmissing-prototypes
+	TTFS_CXXFLAGS += -Wno-missing-prototypes
+else
+	# g++ only
+	MK_CXXFLAGS += -Wno-format-truncation -Wno-array-bounds
+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
@@ -134,104 +219,117 @@ 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_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$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
@@ -270,6 +368,11 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(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
@@ -277,19 +380,20 @@ 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) -Wno-pedantic -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
 
@@ -304,15 +408,13 @@ 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
@@ -322,10 +424,12 @@ 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
@@ -338,29 +442,36 @@ 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
 
-ifdef LLAMA_DISABLE_LOGS
-	CFLAGS   += -DLOG_DISABLE_LOGS
-	CXXFLAGS += -DLOG_DISABLE_LOGS
-endif # LLAMA_DISABLE_LOGS
+# 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:        $(CCV))
+$(info I CXX:       $(CXXV))
 $(info )
 
 #
@@ -391,7 +502,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
@@ -403,22 +514,22 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
+simple: examples/simple/simple.cpp                            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)
+quantize: examples/quantize/quantize.cpp                      ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o common.o $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o common.o $(OBJS)
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
@@ -435,7 +546,7 @@ 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 ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(TTFS_CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -446,12 +557,14 @@ llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o co
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.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)
+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)
 
-ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
-BUILD_TARGETS += metal
-endif
+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)
@@ -459,7 +572,7 @@ metal: examples/metal/metal.cpp ggml.o $(OBJS)
 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 \
@@ -472,7 +585,7 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 
 tests: $(TEST_TARGETS)
 
-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	./$@
 
@@ -509,7 +622,7 @@ tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h gg
 tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+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

Package.swift

@@ -2,8 +2,30 @@
 
 import PackageDescription
 
+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v11),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_SWIFT"),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
 let package = Package(
     name: "llama",
+    platforms: platforms,
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -11,14 +33,23 @@ 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,
             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")
+            ] + additionalSettings,
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]
-        ),
+        )
     ],
     cxxLanguageStandard: .cxx11
 )

README.md

@@ -11,21 +11,9 @@ 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
+- 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
 
 ----
 
@@ -114,11 +102,13 @@ as the main playground for developing new features for the [ggml](https://github
 - 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)
 
 ---
 
@@ -278,29 +268,11 @@ 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
 
@@ -419,17 +391,18 @@ 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...
@@ -463,6 +436,8 @@ 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>
 
@@ -480,10 +455,27 @@ 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
@@ -497,21 +489,32 @@ 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 --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.
 
@@ -723,12 +726,12 @@ 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)
+  - [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
 
@@ -842,8 +845,17 @@ 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
 

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/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__)
@@ -76,7 +78,7 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-void process_escapes(std::string& input) {
+static void process_escapes(std::string& input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
 
@@ -303,18 +305,42 @@ 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;
@@ -346,6 +372,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             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 +388,17 @@ 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) {
@@ -409,12 +448,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #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 +468,8 @@ 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;
@@ -570,106 +605,112 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }
 
 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 computation (default: %d)\n", params.n_threads);
+    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)\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 (default: %g)\n", 1.0f/params.rope_freq_scale);
+    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
+    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
+    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");
+    printf("  -lv, --low-vram       don't allocate VRAM scratch buffer\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-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 gpt_random_prompt(std::mt19937 & rng) {
@@ -700,7 +741,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 
     lparams.n_ctx           = params.n_ctx;
     lparams.n_batch         = params.n_batch;
-    lparams.n_gpu_layers    = params.n_gpu_layers;
+    if (params.n_gpu_layers != -1) {
+        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;
@@ -709,7 +752,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     lparams.f16_kv          = params.memory_f16;
     lparams.use_mmap        = params.use_mmap;
     lparams.use_mlock       = params.use_mlock;
-    lparams.logits_all      = params.perplexity;
+    lparams.logits_all      = params.logits_all;
     lparams.embedding       = params.embedding;
     lparams.rope_freq_base  = params.rope_freq_base;
     lparams.rope_freq_scale = params.rope_freq_scale;
@@ -750,6 +793,15 @@ 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), params.n_threads);
+        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_reset_timings(lctx);
+    }
+
     return std::make_tuple(model, lctx);
 }
 
@@ -764,10 +816,10 @@ std::vector<llama_token> llama_tokenize(
     // 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(ctx, 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(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -822,6 +874,130 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
     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(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(ctx) + idx * n_vocab;
+
+    // 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
+            llama_sample_top_k      (ctx, &cur_p, top_k, 1);
+            llama_sample_tail_free  (ctx, &cur_p, tfs_z, 1);
+            llama_sample_typical    (ctx, &cur_p, typical_p, 1);
+            llama_sample_top_p      (ctx, &cur_p, top_p, 1);
+            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
@@ -1021,13 +1197,12 @@ 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;
@@ -1060,9 +1235,9 @@ 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");
@@ -1095,6 +1270,7 @@ 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

@@ -3,6 +3,7 @@
 #pragma once
 
 #include "llama.h"
+#include "build-info.h"
 
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
@@ -20,6 +21,14 @@
 #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
 //
@@ -32,8 +41,12 @@ struct gpt_params {
     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.
@@ -63,6 +76,7 @@ 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 path_prompt_cache = "";  // path to file for saving/loading prompt eval state
@@ -96,17 +110,16 @@ 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
 };
 
@@ -156,6 +169,40 @@ 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(ctx) + idx * n_vocab
+//
+// 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

@@ -154,7 +154,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
 //  #include "log.h"
 //
 #ifndef LOG_NO_TIMESTAMPS
-    #ifndef _WIN32
+    #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
@@ -167,7 +167,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
 #endif
 
 #ifdef LOG_TEE_TIMESTAMPS
-    #ifndef _WIN32
+    #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
@@ -187,7 +187,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
 //  #include "log.h"
 //
 #ifndef LOG_NO_FILE_LINE_FUNCTION
-    #ifndef _WIN32
+    #ifndef _MSC_VER
         #define LOG_FLF_FMT "[%24s:%5d][%24s] "
         #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
@@ -200,7 +200,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
 #endif
 
 #ifdef LOG_TEE_FILE_LINE_FUNCTION
-    #ifndef _WIN32
+    #ifndef _MSC_VER
         #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
         #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
@@ -224,7 +224,7 @@ enum LogTriState
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
-#ifndef _WIN32
+#ifndef _MSC_VER
     #define LOG_IMPL(str, ...)                                                                                          \
     {                                                                                                               \
         if (LOG_TARGET != nullptr)                                                                                  \
@@ -247,7 +247,7 @@ enum LogTriState
 // INTERNAL, DO NOT USE
 //  USE LOG_TEE() INSTEAD
 //
-#ifndef _WIN32
+#ifndef _MSC_VER
     #define LOG_TEE_IMPL(str, ...)                                                                                                          \
     {                                                                                                                                   \
         if (LOG_TARGET != nullptr)                                                                                                      \
@@ -284,7 +284,7 @@ enum LogTriState
 // Main LOG macro.
 //  behaves like printf, and supports arguments the exact same way.
 //
-#ifndef _WIN32
+#ifndef _MSC_VER
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
     #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
@@ -298,14 +298,14 @@ enum LogTriState
 // Secondary target can be changed just like LOG_TARGET
 //  by defining LOG_TEE_TARGET
 //
-#ifndef _WIN32
+#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 _WIN32
+#ifndef _MSC_VER
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else
@@ -341,14 +341,14 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
         }
     }
 
+    if (_disabled)
+    {
+        // Log is disabled
+        return nullptr;
+    }
+
     if (_initialized)
     {
-        if (_disabled)
-        {
-            // Log is disabled
-            return nullptr;
-        }
-
         // with fallback in case something went wrong
         return logfile ? logfile : stderr;
     }
@@ -461,7 +461,7 @@ inline void log_test()
     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 _WIN32
+#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")
@@ -513,16 +513,16 @@ inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string &
 
 inline void log_print_usage()
 {
-    fprintf(stdout, "log options:\n");
+    printf("log options:\n");
     /* format
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");*/
+    printf("  -h, --help            show this help message and exit\n");*/
     /* spacing
-    fprintf(stdout, "__-param----------------Description\n");*/
-    fprintf(stdout, "  --log-test            Run simple logging test\n");
-    fprintf(stdout, "  --log-disable         Disable trace logs\n");
-    fprintf(stdout, "  --log-enable          Enable trace logs\n");
-    fprintf(stdout, "  --log-file            Specify a log filename (without extension)\n");
-    fprintf(stdout, "                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+    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)

convert-baichuan-hf-to-gguf.py

@@ -1,28 +1,31 @@
 #!/usr/bin/env python3
-# HF llama --> gguf conversion
+# HF baichuan --> 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 TYPE_CHECKING, Any
+import itertools
+import gguf
 import numpy as np
 import torch
-import argparse
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
 
-from typing import Any, List, Optional, TypeAlias
-from pathlib import Path
-from sentencepiece import SentencePieceProcessor
 
-#NDArray = np.ndarray[Any, Any]
-# compatible with python < 3.9
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
+if TYPE_CHECKING:
+    from typing import TypeAlias
+
+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:
+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
 
@@ -30,6 +33,13 @@ def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] =
             .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
@@ -44,12 +54,25 @@ def count_model_parts(dir_model: str) -> int:
     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],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    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()
@@ -77,16 +100,16 @@ 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] != "LlamaForCausalLM":
+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)
-
-ARCH=gguf.MODEL_ARCH.LLAMA
+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")
@@ -108,6 +131,8 @@ 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.")
 
@@ -136,9 +161,9 @@ if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[bytes] = []
-scores: List[float] = []
-toktypes: List[int] = []
+tokens: list[bytes] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
 tokenizer_model_file = dir_model / 'tokenizer.model'
 if not tokenizer_model_file.is_file():
@@ -210,15 +235,24 @@ else:
         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
@@ -231,12 +265,6 @@ for part_name in part_names:
 
         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
         new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
         if new_name is None:
@@ -258,8 +286,7 @@ 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(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
+        print(name + " -> " +  new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
         gguf_writer.add_tensor(new_name, data)
 
 

convert-falcon-hf-to-gguf.py

@@ -1,18 +1,24 @@
 #!/usr/bin/env python3
 # HF falcon--> 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
-import argparse
+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
 
-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
@@ -49,10 +55,22 @@ def count_model_parts(dir_model: Path) -> int:
 
 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],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    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()
@@ -114,9 +132,9 @@ gguf_writer.add_file_type(ftype)
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[bytearray] = []
-scores: List[float] = []
-toktypes: List[int] = []
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -131,7 +149,9 @@ with open(tokenizer_json_file, "r", encoding="utf-8") as f:
 
 print("gguf: get gpt2 tokenizer vocab")
 
-vocab_size = len(tokenizer_json["model"]["vocab"])
+# 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["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])
 
 # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
 tokenizer = AutoTokenizer.from_pretrained(dir_model)

convert-gptneox-hf-to-gguf.py

@@ -1,18 +1,23 @@
 #!/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
-import argparse
+from transformers import AutoTokenizer  # type: ignore[import]
 
-from typing import Any, List
-from pathlib import Path
-from transformers import AutoTokenizer
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
 
@@ -51,10 +56,22 @@ def count_model_parts(dir_model: Path) -> int:
 
 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],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    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()
@@ -112,7 +129,7 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[bytearray] = []
+tokens: list[bytearray] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():

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

@@ -1,258 +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
-import argparse
-
-from typing import Any, List, TypeAlias
-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: Path) -> 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
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a PyTorch 7B 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],   help="output format - use 0 for float32, 1 for float16", 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] != "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(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 and scores")
-
-tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
-
-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)
-
-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")
-
-part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts))
-
-for part_name in part_names:
-    if args.vocab_only:
-        break
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name == "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
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(new_name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-if not args.vocab_only:
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-print(f"gguf: model successfully exported to '{fname_out}'")
-print("")

convert-llama-ggml-to-gguf.py

@@ -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 = 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)
@@ -160,7 +235,10 @@ 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:
@@ -177,7 +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
@@ -187,6 +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'
@@ -223,7 +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:
@@ -275,7 +358,11 @@ class GGMLToGGUF:
                 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
@@ -297,32 +384,46 @@ 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, 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}')
@@ -337,7 +438,12 @@ def main():
         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, special_vocab = special_vocab)
+        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-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, BinaryIO
+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: BinaryIO, 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"]))

convert-starcoder-hf-to-gguf.py

@@ -0,0 +1,248 @@
+#!/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 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
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
+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] = []
+
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# 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["vocab_size"] if "vocab_size" in hparams else 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)
+
+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
@@ -20,21 +19,27 @@ 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, Type, 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]
@@ -47,8 +52,8 @@ 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 +70,7 @@ 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 +89,7 @@ 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 +103,13 @@ 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 +124,14 @@ 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 +145,6 @@ 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 +152,16 @@ 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 +187,6 @@ 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 +196,7 @@ 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 +214,6 @@ 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 +225,6 @@ 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 +236,14 @@ 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 +251,7 @@ 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 +270,6 @@ 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 +280,7 @@ 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"
 
@@ -314,19 +303,31 @@ 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]
@@ -335,22 +336,34 @@ 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
+        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
         byte_encoder = tokenization_gpt2.bytes_to_unicode()
         byte_decoder = {v: k for k, v in byte_encoder.items()}
+        score = 0.0
         for i, item in enumerate(tokenizer):
             text: bytes = item.encode("utf-8")
-            score: float = -i
-            yield text, score, gguf.TokenType.USER_DEFINED
+            # FIXME: These shouldn't be hardcoded, but it's probably better than the current behavior?
+            if i <= 258 and text.startswith(b'<') and text.endswith(b'>'):
+                if i == 0 and text == b'<unk>':
+                    toktype = gguf.TokenType.UNKNOWN
+                elif i == 1 or i == 2:
+                    toktype = gguf.TokenType.CONTROL
+                elif i >= 3 and text.startswith(b'<0x'):
+                    toktype = gguf.TokenType.BYTE
+                else:
+                    toktype = gguf.TokenType.NORMAL
+            else:
+                toktype = gguf.TokenType.NORMAL
+            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.bpe_tokens()
         yield from self.added_tokens()
 
@@ -359,9 +372,9 @@ class BpeVocab:
 
 
 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:
@@ -380,7 +393,7 @@ 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)
@@ -404,19 +417,19 @@ 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
@@ -436,15 +449,15 @@ 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, n_head_kv: 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[Any, np.dtype[np.uint16]]) -> NDArray:
@@ -465,22 +478,22 @@ 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, n_head_kv: 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, 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)
 
@@ -496,13 +509,13 @@ 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
 
@@ -513,7 +526,7 @@ 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:
@@ -525,24 +538,24 @@ 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.
+    paths: list[Path]  # Where this was read from.
     format: Literal['ggml', 'torch', 'safetensors', 'none']
-    vocab: Optional[Vocab]  # For GGML models (which have vocab built in), the vocab.
+    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
@@ -570,7 +583,7 @@ 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()
@@ -644,7 +657,7 @@ 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:
@@ -660,7 +673,6 @@ class LazyUnpickler(pickle.Unpickler):
 
     @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)
 
@@ -674,7 +686,7 @@ class LazyUnpickler(pickle.Unpickler):
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: Dict[Tuple[str, str], Any] = {
+    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__'),
@@ -707,15 +719,15 @@ 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
@@ -754,7 +766,7 @@ 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, use_processpool_executor: bool = False) -> 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
@@ -763,13 +775,13 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
         yield from map(func, iterable)
         # Not reached.
     iterable = iter(iterable)
-    executor_class: Union[Type[ThreadPoolExecutor], Type[ProcessPoolExecutor]]
+    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]] = []
+        futures: list[concurrent.futures.Future[Out]] = []
         done = False
         for _ in range(concurrency):
             try:
@@ -811,10 +823,12 @@ 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)
@@ -826,21 +840,20 @@ 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)
@@ -892,13 +905,13 @@ class OutputFile:
         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
@@ -939,7 +952,7 @@ class OutputFile:
 
         of.close()
 
-def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType:
+def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
     wq_type = model[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):
@@ -959,7 +972,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
 
 def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
-    should_skip: Set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
+    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
@@ -994,12 +1007,12 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
 
     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.
@@ -1015,11 +1028,11 @@ 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:
@@ -1050,7 +1063,7 @@ 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))
@@ -1059,7 +1072,7 @@ 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.
@@ -1090,7 +1103,7 @@ 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",
@@ -1113,7 +1126,7 @@ 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")

examples/CMakeLists.txt

@@ -23,9 +23,11 @@ else()
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(simple)
+    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()

examples/baby-llama/baby-llama.cpp

@@ -9,12 +9,12 @@
 #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() {
+static float frand() {
     return (float)rand()/(float)RAND_MAX;
 }
 
@@ -25,19 +25,21 @@ struct random_normal_distribution {
     float max;
 };
 
-void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
+static 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) {
+static 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 +50,9 @@ 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++) {
@@ -95,11 +93,9 @@ struct ggml_tensor * randomize_tensor(
     return tensor;
 }
 
-struct ggml_tensor * randomize_tensor_normal(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        struct random_normal_distribution * rnd) {
+static 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:
@@ -159,7 +155,7 @@ 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 +256,7 @@ 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 +293,7 @@ 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 +336,7 @@ 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 +362,7 @@ 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;
@@ -397,7 +393,7 @@ 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;
@@ -426,7 +422,9 @@ void randomize_model(struct llama_model * model, int seed, float mean, float std
 }
 
 
-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;
@@ -459,7 +457,7 @@ void randomize_model_lora(struct llama_model_lora * model, int seed, float mean,
     }
 }
 
-bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+static bool 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;
@@ -495,7 +493,7 @@ 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) {
+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;
@@ -531,15 +529,15 @@ bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora *
     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 +554,14 @@ 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 +589,8 @@ 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
             {
@@ -756,25 +762,25 @@ struct ggml_tensor * forward(
     return inpL;
 }
 
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
+static 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) {
+static 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) {
+static 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) {
+static 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);
@@ -782,16 +788,16 @@ void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int6
     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 +816,18 @@ 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 +855,8 @@ 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 +1088,15 @@ 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 +1114,14 @@ 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 +1155,7 @@ 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 +1164,7 @@ 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 +1350,7 @@ 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 +1381,10 @@ 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 +1418,7 @@ 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 +1426,7 @@ 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 +1437,7 @@ 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 +1448,14 @@ 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 +1476,9 @@ 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 +1501,7 @@ 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 +1512,16 @@ 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 +1648,10 @@ 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/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), params.n_threads))
     {
         fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
         return 1;

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 +1,5 @@
+#include "common.h"
 #include "ggml.h"
-#include "build-info.h"
 
 #include <locale.h>
 #include <assert.h>
@@ -99,7 +99,7 @@ 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

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 +1,4 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
+#include "common.h"
 #include "embd-input.h"
 
 #include <cassert>
@@ -23,11 +19,11 @@ 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));
@@ -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, params.n_threads)) {
             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), params.n_threads)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
@@ -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.h

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

examples/embedding/embedding.cpp

@@ -1,6 +1,5 @@
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <ctime>
 
@@ -11,18 +10,13 @@
 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);
@@ -47,6 +41,12 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(ctx);
+    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");
@@ -77,7 +77,7 @@ int main(int argc, char ** argv) {
 
     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), params.n_threads)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return 1;
         }

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      = %s\n", __func__, gguf_get_val_str(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)); }
+        if (keyidx != -1) { printf("%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 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      = %s\n", __func__, gguf_get_val_str(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)); }
+        if (keyidx != -1) { printf("%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 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/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__
@@ -165,26 +157,26 @@ 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, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").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");
 
 }
 
@@ -899,7 +891,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
     int n_processed = 0;
     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_threads);
         n_processed += n_tokens;
     }
 }
@@ -907,7 +899,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
     llama_token token = llama_token_bos(ctx);
     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), n_threads);
     }
 }
 
@@ -985,10 +977,19 @@ 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);

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 +1,36 @@
+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}
+    ${CMAKE_CURRENT_BINARY_DIR})
+
+add_executable(${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
 
@@ -274,7 +274,7 @@ 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
 
@@ -302,7 +302,6 @@ 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.

examples/main/main.cpp

@@ -1,8 +1,3 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
 #include "common.h"
 
 #include "console.h"
@@ -46,10 +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;
     }
@@ -90,7 +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;
@@ -109,14 +106,14 @@ 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);
+    log_dump_cmdline(argc, argv);
 #endif // LOG_DISABLE_LOGS
 
     // TODO: Dump params ?
@@ -127,7 +124,7 @@ int main(int argc, char ** argv) {
     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");
@@ -151,15 +148,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%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
-        LOG_TEE("%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) {
-        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
-        params.n_ctx = 8;
-    }
-
     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);
@@ -194,6 +184,15 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(ctx);
+    if (params.n_ctx > n_ctx_train) {
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    } else if (params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+
     // print system information
     {
         LOG_TEE("\n");
@@ -201,32 +200,6 @@ int main(int argc, char ** argv) {
                 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) {
-        {
-            LOG_TEE("%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;
-    }
-
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
 
@@ -304,7 +277,7 @@ int main(int argc, char ** argv) {
 
     // debug message about similarity of saved session, if applicable
     size_t n_matching_session_tokens = 0;
-    if (session_tokens.size() > 0) {
+    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;
@@ -402,7 +375,7 @@ int main(int argc, char ** argv) {
 
         LOG_TEE("%s: interactive mode on.\n", __func__);
 
-        if (params.antiprompt.size()) {
+        if (!params.antiprompt.empty()) {
             for (const auto & antiprompt : params.antiprompt) {
                 LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
             }
@@ -425,8 +398,9 @@ int main(int argc, char ** argv) {
     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
@@ -450,8 +424,8 @@ 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;
@@ -492,17 +466,14 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    {
-        LOG("warming up the model with an empty run\n");
+    const int n_vocab = llama_n_vocab(ctx);
 
-        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);
-    }
+    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.
             int max_embd_size = n_ctx - 4;
@@ -528,18 +499,23 @@ int main(int argc, char ** argv) {
                     break;
                 }
 
-                const int n_left = n_past - params.n_keep;
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d\n", n_past, n_left, n_ctx, params.n_keep);
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
 
-                // 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);
+                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);
 
-                // 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("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
                 LOG("clear session path\n");
@@ -600,7 +576,7 @@ int main(int argc, char ** argv) {
 
                 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)) {
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0), params.n_threads)) {
                         LOG_TEE("%s : failed to eval\n", __func__);
                         return 1;
                     }
@@ -617,7 +593,7 @@ int main(int argc, char ** argv) {
 
                 LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
-                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0), params.n_threads)) {
                     LOG_TEE("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -627,7 +603,7 @@ int main(int argc, char ** argv) {
                 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();
             }
@@ -637,20 +613,6 @@ int main(int argc, char ** argv) {
         embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            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;
@@ -659,98 +621,12 @@ int main(int argc, char ** argv) {
                 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;
-                }
-
-                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 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
-                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, &cur_p,
-                    last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                    last_n_repeat, repeat_penalty);
-                llama_sample_frequency_and_presence_penalties(ctx, &cur_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 < 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_temperature(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_temperature(ctx, &cur_p, temp);
-                        id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-                    } else {
-                        // Temperature sampling
-                        llama_sample_top_k      (ctx, &cur_p, top_k, 1);
-                        llama_sample_tail_free  (ctx, &cur_p, tfs_z, 1);
-                        llama_sample_typical    (ctx, &cur_p, typical_p, 1);
-                        llama_sample_top_p      (ctx, &cur_p, top_p, 1);
-                        llama_sample_temperature(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);
-                }
-
-                last_n_tokens.erase(last_n_tokens.begin());
-                last_n_tokens.push_back(id);
-
-                LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_n_tokens));
-            }
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
 
             embd.push_back(id);
 
@@ -766,8 +642,8 @@ int main(int argc, char ** argv) {
             LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-                last_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;
@@ -798,9 +674,9 @@ int main(int argc, char ** argv) {
         // 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);
                 }
 
@@ -831,11 +707,11 @@ int main(int argc, char ** argv) {
             }
 
             // 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());
@@ -933,7 +809,7 @@ 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"));

examples/parallel/CMakeLists.txt

@@ -1,5 +1,5 @@
-set(TARGET beam_search)
-add_executable(${TARGET} beam_search.cpp)
+set(TARGET parallel)
+add_executable(${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)

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 +1,379 @@
+// 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>
+
+// 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;
+};
+
+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);
+
+    fprintf(stderr, "\n\n");
+    fflush(stderr);
+
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(ctx);
+
+    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(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(params.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, params.n_threads) != 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[1mClient %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, params.n_threads);
+            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) || 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, n_ctx);
+
+                    const auto t_main_end = ggml_time_us();
+
+                    LOG_TEE("\033[1mClient %3d, seq %4d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \n\nInput:    %s\nResponse: %s\n\n",
+                            client.id, client.seq_id, 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();
+
+    LOG_TEE("\n\n");
+    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\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/perplexity.cpp

@@ -1,6 +1,5 @@
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <cmath>
 #include <cstdio>
@@ -28,9 +27,10 @@ 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 +76,12 @@ 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 +90,33 @@ 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++;
@@ -124,13 +134,16 @@ 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]`
@@ -150,8 +163,8 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
         return {std::move(tokens), 0., {}, {}};
     }
 
-    std::vector<float>       logit_history;
-    std::vector<float>       prob_history;
+    std::vector<float> logit_history;
+    std::vector<float> prob_history;
 
     logit_history.resize(tokens.size());
     prob_history.resize(tokens.size());
@@ -193,12 +206,15 @@ 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), params.n_threads)) {
                 //fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
@@ -260,8 +276,7 @@ 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);
     }
@@ -319,6 +334,9 @@ results_perplexity perplexity(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);
@@ -331,7 +349,7 @@ 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), params.n_threads)) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
@@ -368,7 +386,7 @@ 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);
+        const int first = params.n_ctx/2;
         process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first,
                        workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
         count += params.n_ctx - first - 1;
@@ -400,15 +418,16 @@ 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, int n_thread
+) {
     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), n_thread)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return {};
         }
@@ -421,7 +440,7 @@ 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
@@ -548,6 +567,9 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             query_embd.resize(32);
         }
 
+        // 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, params.n_threads);
         if (logits.empty()) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
@@ -655,11 +677,11 @@ 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) {
@@ -668,12 +690,7 @@ 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);
@@ -698,6 +715,12 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(ctx);
+    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");

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 +1,6 @@
-#include "ggml.h"
-#include "build-info.h"
-
 #define LLAMA_API_INTERNAL
+#include "common.h"
+#include "ggml.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -34,8 +33,8 @@ 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 +43,7 @@ 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 +69,7 @@ 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 +84,7 @@ 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 +94,14 @@ 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 +114,7 @@ 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 +141,10 @@ 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 +165,11 @@ 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 +298,7 @@ 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");

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/quantize.cpp

@@ -1,5 +1,4 @@
-#include "build-info.h"
-
+#include "common.h"
 #include "llama.h"
 
 #include <cstdio>
@@ -35,10 +34,12 @@ 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 +71,18 @@ 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");
+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 +127,9 @@ 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 +143,9 @@ 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 +160,7 @@ 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 +1,5 @@
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <vector>
 #include <cstdio>
@@ -13,11 +12,11 @@ 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;
@@ -35,16 +34,16 @@ int main(int argc, char ** argv) {
     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);
+    auto * model = llama_load_model_from_file(params.model.c_str(), lparams);
     if (model == nullptr) {
         return 1;
     }
-    auto ctx = llama_new_context_with_model(model, lparams);
+    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 +53,7 @@ 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), params.n_threads);
 
     last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
     n_past += n_prompt_tokens;
@@ -78,7 +77,7 @@ 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 * logits = llama_get_logits(ctx);
         auto n_vocab = llama_n_vocab(ctx);
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
@@ -91,7 +90,7 @@ 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), params.n_threads)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx);
             llama_free_model(model);
@@ -106,7 +105,7 @@ 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, lparams);
 
     // Load state (rng, logits, embedding and kv_cache) from file
     {
@@ -138,7 +137,7 @@ int main(int argc, char ** argv) {
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx2);
+        auto * logits = llama_get_logits(ctx2);
         auto n_vocab = llama_n_vocab(ctx2);
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
@@ -151,7 +150,7 @@ 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), params.n_threads)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);

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)
@@ -269,7 +271,7 @@ struct llama_server_context
         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.
@@ -432,7 +434,7 @@ struct llama_server_context
             {
                 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), params.n_threads))
             {
                 LOG_ERROR("failed to eval", {
                                                 {"n_eval", n_eval},
@@ -521,13 +523,13 @@ struct llama_server_context
                 {
                     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);
                     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);
+                    llama_sample_temp(ctx, &candidates_p, temp);
                     result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                 }
                 else
@@ -538,7 +540,7 @@ struct llama_server_context
                     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);
+                    llama_sample_temp(ctx, &candidates_p, temp);
                     result.tok = llama_sample_token(ctx, &candidates_p);
                 }
             }
@@ -609,7 +611,7 @@ 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;
@@ -692,50 +694,50 @@ 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: %.1f)\n", params.rope_freq_base);
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    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("  -lv, --low-vram       don't allocate VRAM scratch buffer\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,
@@ -1038,7 +1040,7 @@ static json format_timings(llama_server_context &llama)
 {
     const auto timings = llama_get_timings(llama.ctx);
 
-    assert(timings.n_eval == llama.num_tokens_predicted);
+    assert(timings.n_eval == ptrdiff_t(llama.num_tokens_predicted));
 
     return json{
         {"prompt_n", timings.n_p_eval},
@@ -1081,8 +1083,9 @@ 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},
@@ -1213,7 +1216,7 @@ 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 +1226,7 @@ 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 +1242,7 @@ 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 +1256,11 @@ 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(); };
@@ -1377,7 +1381,13 @@ 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 +1464,11 @@ 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: " +
@@ -1548,7 +1562,7 @@ 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 +1597,7 @@ 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 +1,67 @@
+# 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.
+The example demonstrates single-batch as well as parallel generation.
+
+## Single-batch generation
+
+```bash
+./simple ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 1
+
+...
+
+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
+```
+
+## Parallel generation
+
+```bash
+./simple ./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/simple/simple.cpp

@@ -1,12 +1,7 @@
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
 
+#include <algorithm>
 #include <cmath>
 #include <cstdio>
 #include <string>
@@ -16,10 +11,12 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT]\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
         return 1 ;
     }
 
+    int n_parallel = 1;
+
     if (argc >= 2) {
         params.model = argv[1];
     }
@@ -28,16 +25,28 @@ int main(int argc, char ** argv) {
         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);
 
     llama_context_params ctx_params = llama_context_default_params();
 
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = n_len*n_parallel; // FIXME: use n_kv_req instead (tokenize with model after #3301)
+    ctx_params.n_batch = std::max(n_len, n_parallel);
+    // ctx_params.n_gpu_layers = 99; // offload all layers to the GPU
+
     llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
 
     if (model == NULL) {
@@ -47,20 +56,31 @@ int main(int argc, char ** argv) {
 
     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())*n_parallel;
 
-    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_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;
     }
 
-    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 +88,159 @@ 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(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, params.n_threads) != 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
 
-    // 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.
+    // we will store the parallel decoded sequences in this vector
+    std::vector<std::string> streams(n_parallel);
 
-    const int n_gen = std::min(32, max_context_size);
+    // 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);
 
-    while (llama_get_kv_cache_token_count(ctx) < n_gen) {
-        // evaluate the transformer
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
 
-        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__);
-            return 1;
+    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(ctx);
+            auto * logits  = llama_get_logits(ctx) + i_batch[i] * n_vocab;
+
+            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;
         }
 
-        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");
+        // all streams are finished
+        if (batch.n_tokens == 0) {
             break;
         }
 
-        // print the new token :
-        printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
-        fflush(stdout);
+        n_cur += 1;
 
-        // push this new token for next evaluation
-        tokens_list.push_back(new_token_id);
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch, params.n_threads)) {
+            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();
 
-    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), params.n_threads);
+    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0), params.n_threads);
+    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input,     0,           0), params.n_threads);
+
+    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(ctx_tgt);
+    //GGML_ASSERT(n_vocab == llama_n_vocab(ctx_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, n_ctx);
+            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0), params.n_threads);
+            ++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, n_ctx);
+            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0), params.n_threads);
+            ++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, n_ctx);
+        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0), params.n_threads);
+        ++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;
+}

examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py

@@ -2,13 +2,16 @@
 # train-text-from-scratch checkpoint --> gguf conversion
 
 import argparse
-import gguf
 import os
 import struct
 import sys
 import numpy as np
 from pathlib import Path
 
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / '..' / '..' / 'gguf-py' / 'gguf'))
+import gguf
+
 # gguf constants
 LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
 LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -169,10 +169,6 @@ struct my_llama_hparams {
 
     float rope_freq_base  = 10000.0f;
     float rope_freq_scale = 1.0f;
-
-    bool operator!=(const my_llama_hparams& other) const {
-        return memcmp(this, &other, sizeof(my_llama_hparams));
-    }
 };
 
 struct my_llama_layer {
@@ -683,15 +679,23 @@ struct ggml_tensor * llama_build_train_graphs(
         }
     };
 
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // rope has so much parameters that we make a custom function for it
-    auto rope = [ctx, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
+    auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
                 (struct ggml_tensor * t) -> struct ggml_tensor * {
         // not capturing these, to silcence warnings
-        const int n_past    = 0;
         const int rope_mode = 0;
 
         return ggml_rope_custom(ctx,
-            t, n_past, n_rot, rope_mode, n_ctx,
+            t, KQ_pos, n_rot, rope_mode, n_ctx,
             rope_freq_base, rope_freq_scale);
     };
 
@@ -791,6 +795,8 @@ struct ggml_tensor * llama_build_train_graphs(
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
         // input gradient
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+        // KQ_pos
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
         GGML_ASSERT(t36->grad->data == NULL && !ggml_is_view(t36->grad));
         ggml_allocr_alloc(alloc, t36->grad);
         // gradient tensors (will be set to zero by ggml_graph_reset)
@@ -929,28 +935,6 @@ void get_example_targets_batch(struct llama_context * lctx, const int * train_sa
     }
 }
 
-
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 int tokenize_file(struct llama_context * lctx, const char * filename, std::vector<llama_token>& out) {
     FILE * fp = std::fopen(filename, "rb");
     if (fp == NULL) {
@@ -983,18 +967,18 @@ int tokenize_file(struct llama_context * lctx, const char * filename, std::vecto
     out.resize(size+1);
 
     if (std::fread(buf.data(), size, 1, fp) != 1) {
-        throw std::runtime_error(std::string("unexpectedly reached end of file"));
+        die("unexpectedly reached end of file");
     }
     if (ferror(fp)) {
-        throw std::runtime_error(format("read error: %s", strerror(errno)));
+        die_fmt("fread failed: %s", strerror(errno));
     }
 
     buf[size] = '\0';
 
-    int n_tokens = llama_tokenize(lctx, buf.data(), out.data(), out.size(), false);
+    int n_tokens = llama_tokenize(lctx, buf.data(), buf.size(), out.data(), out.size(), false);
     if (n_tokens < 0) {
         out.resize(-n_tokens);
-        n_tokens = llama_tokenize(lctx, buf.data(), out.data(), out.size(), false);
+        n_tokens = llama_tokenize(lctx, buf.data(), buf.size(), out.data(), out.size(), false);
     }
     GGML_ASSERT(n_tokens >= 0);
     out.resize(n_tokens);
@@ -1047,11 +1031,11 @@ void shuffle_ints(int * begin, int * end) {
     if (kid >= 0) { \
         enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
         if (ktype != (type)) { \
-            throw std::runtime_error(format("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype))); \
+            die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
         } \
         (dst) = func(ctx, kid); \
     } else if (req) { \
-        throw std::runtime_error(format("key not found in model: %s", skey.c_str())); \
+        die_fmt("key not found in model: %s", skey.c_str()); \
     } \
 }
 
@@ -1136,7 +1120,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
         read_tensor_by_name(opt->lbfgs.lms,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
         read_tensor_by_name(opt->lbfgs.lmy,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
     } else {
-        throw std::runtime_error("unknown optimizer type\n");
+        die("unknown optimizer type");
     }
 }
 
@@ -1315,20 +1299,20 @@ void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_mod
 
         const int token_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_LIST));
         if (token_idx == -1) {
-            throw std::runtime_error("cannot find tokenizer vocab in model file\n");
+            die("cannot find tokenizer vocab in model file");
         }
         const uint32_t n_vocab = gguf_get_arr_n(vctx, token_idx);
 
         const int score_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_SCORES));
         if (score_idx == -1) {
-            throw std::runtime_error("cannot find tokenizer scores in model file\n");
+            die("cannot find tokenizer scores in model file");
         }
 
         const float * scores = (const float * ) gguf_get_arr_data(vctx, score_idx);
 
         const int toktype_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE));
         if (toktype_idx == -1) {
-            throw std::runtime_error("cannot find token type list in GGUF file\n");
+            die("cannot find token type list in GGUF file");
         }
 
         const int * toktypes = (const int * ) gguf_get_arr_data(vctx, toktype_idx);
@@ -1356,7 +1340,7 @@ void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_mod
             // read and copy bpe merges
             const int merges_keyidx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_MERGES));
             if (merges_keyidx == -1) {
-                throw std::runtime_error("cannot find tokenizer merges in model file\n");
+                die("cannot find tokenizer merges in model file");
             }
 
             const int n_merges = gguf_get_arr_n(vctx, merges_keyidx);
@@ -1988,7 +1972,7 @@ void opt_callback(void * vdata, float * sched) {
     float min_sched = params->adam_min_alpha / params->adam_alpha;
     *sched = min_sched + *sched * (1.0f - min_sched);
 
-    int impr_plot = std::isnan(opt->loss_after) ? 0 : -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
+    int impr_plot = std::isnan(opt->loss_after) ? 0 : -std::lround(1 + (opt->loss_before - opt->loss_after) * 10.0f);
     printf("%s: iter=%*d, sched=%f loss0=%f loss=%f | improvement: %*d>\n", __func__, 6, opt->iter, *sched, opt->loss_before, opt->loss_after, impr_plot, (int)0);
 
     if (data->shuffle_countdown < n_batch) {

flake.nix

@@ -34,7 +34,7 @@
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ];
+        nativeBuildInputs = with pkgs; [ cmake ninja pkg-config ];
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
         postPatch = ''
@@ -45,6 +45,8 @@
         postInstall = ''
           mv $out/bin/main $out/bin/llama
           mv $out/bin/server $out/bin/llama-server
+          mkdir -p $out/include
+          cp ${src}/llama.h $out/include/
         '';
         cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
       in
@@ -93,6 +95,10 @@
           type = "app";
           program = "${self.packages.${system}.default}/bin/quantize";
         };
+        apps.train-text-from-scratch = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/train-text-from-scratch";
+        };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
           buildInputs = [ llama-python ];

ggml-alloc.c

@@ -6,6 +6,26 @@
 #include <stdlib.h>
 #include <string.h>
 
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/types.h>
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <memoryapi.h>
+#endif
+
+
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
@@ -99,19 +119,28 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-
-static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     return ggml_nbytes(tensor);
 
     UNUSED(alloc);
 }
 
+// check if a tensor is allocated by this buffer
+static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+}
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
 #ifdef GGML_ALLOCATOR_DEBUG
-    GGML_ASSERT(ggml_is_view(tensor) == false); // views generally get data pointer from one of their sources
+    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
     GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
 #endif
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -135,14 +164,14 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
-            max_avail = MAX(max_avail, block->size);
         } else {
             fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
                     __func__, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
-        return;
+            return;
         }
     }
     struct free_block * block = &alloc->free_blocks[best_fit_block];
@@ -177,17 +206,17 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     void * ptr = tensor->data;
 
-    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+    if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
         return;
     }
 
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
 
@@ -281,17 +310,68 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-// address and size of the buffer when measuring
-// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
-static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
-static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+// OS specific functions to allocate and free uncommitted virtual memory
+static void * alloc_vmem(size_t size) {
+#if defined(_WIN32)
+    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
+#elif defined(_POSIX_MAPPED_FILES)
+    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+    return ptr;
+#else
+    // use a fixed address for other platforms
+    uintptr_t base_addr = (uintptr_t)-size - 0x100;
+    return (void *)base_addr;
+#endif
+}
+
+static void free_vmem(void * base_addr, size_t size) {
+#if defined(_WIN32)
+    VirtualFree(base_addr, 0, MEM_RELEASE);
+    UNUSED(size);
+#elif defined(_POSIX_MAPPED_FILES)
+    munmap(base_addr, size);
+#else
+    // nothing to do
+    UNUSED(base_addr);
+    UNUSED(size);
+#endif
+}
+
+// allocate uncommitted virtual memory to measure the size of the graph
+static void alloc_measure_vmem(void ** base_addr, size_t * size) {
+    // 128GB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
+    do {
+        *base_addr = alloc_vmem(*size);
+        if (*base_addr != NULL) {
+            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
+            return;
+        }
+        // try again with half the size
+        *size /= 2;
+    } while (*size > 0);
+
+    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+}
+
+static void free_measure_vmem(void * base_addr, size_t size) {
+    free_vmem(base_addr, size);
+}
 
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
     struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
 
+    void * base_addr;
+    size_t size;
+
+    alloc_measure_vmem(&base_addr, &size);
+
     *alloc = (struct ggml_allocr){
-        /*.data          = */ MEASURE_BASE_ADDR,
-        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.data          = */ base_addr,
+        /*.size          = */ size,
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -311,6 +391,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
+    if (alloc->measure) {
+        free_measure_vmem(alloc->data, alloc->size);
+    }
     free(alloc);
 }
 
@@ -320,10 +403,6 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 
 //////////// compute graph allocator
 
-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
     if (a->type != b->type) {
         return false;
@@ -380,8 +459,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     }
 
                     // if the node's data is external, then we cannot re-use it
-                    if ((char *) parent->data < (char *) alloc->data ||
-                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                    if (ggml_allocr_is_own(alloc, parent) == false) {
                         AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
                         continue;
                     }
@@ -415,7 +493,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-static size_t ggml_allocator_alloc_graph_tensors_n(
+static size_t ggml_allocr_alloc_graph_tensors_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -493,11 +571,10 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 AT_PRINTF("\n");
             }
 
-
             // update parents
             // update immediately if there is no parse_seq
             // update only at barriers if there is parse_seq
-            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) {
+            if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) {
                 int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
                 int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
                 for (int i = update_start; i < update_end; i++) {
@@ -521,12 +598,12 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                                 view_src_hn->n_views -= 1;
                                 AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
                                 if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, view_src);
+                                    ggml_allocr_free_tensor(alloc, view_src);
                                 }
                             }
                             else {
                                 if (parent->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, parent);
+                                    ggml_allocr_free_tensor(alloc, parent);
                                 }
                             }
                         }
@@ -543,7 +620,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             for (int i = 0; outputs[g][i] != NULL; i++) {
                 struct ggml_tensor * output = outputs[g][i];
                 AT_PRINTF("output: %s\n", output->name);
-                ggml_allocator_free_tensor(alloc, output);
+                ggml_allocr_free_tensor(alloc, output);
             }
         }
     }
@@ -552,5 +629,5 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
 }

ggml-cuda.h

@@ -31,6 +31,7 @@ GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tens
 
 GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
+GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
 
 GGML_API void   ggml_cuda_set_main_device(int main_device);
 GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);

ggml-metal.m

@@ -63,7 +63,10 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
     GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(soft_max_4);
     GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DECL_KERNEL(get_rows_f32);
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
@@ -75,7 +78,10 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(norm);
+    GGML_METAL_DECL_KERNEL(mul_mat_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_l4);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
@@ -84,6 +90,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
@@ -93,7 +100,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(rope_f32);
+    GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
@@ -116,22 +124,47 @@ static NSString * const msl_library_source = @"see metal.metal";
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
     metal_printf("%s: allocating\n", __func__);
 
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    id <MTLDevice> device;
+    NSString * s;
 
+#if TARGET_OS_OSX
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
+    for (device in devices) {
+        s = [device name];
+        metal_printf("%s: found device: %s\n", __func__, [s UTF8String]);
+    }
+#endif
+
+    // Pick and show default Metal device
+    device = MTLCreateSystemDefaultDevice();
+    s = [device name];
+    metal_printf("%s: picking default device: %s\n", __func__, [s UTF8String]);
+
+    // Configure context
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
-    ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
     ctx->concur_list_len = 0;
 
-    ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+    ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-#if 0
-    // compile from source string and show compile log
+#ifdef GGML_SWIFT
+    // load the default.metallib file
     {
         NSError * error = nil;
 
-        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+        NSString * llamaBundlePath = [bundle pathForResource:@"llama_llama" ofType:@"bundle"];
+        NSBundle * llamaBundle = [NSBundle bundleWithPath:llamaBundlePath];
+        NSString * libPath = [llamaBundle pathForResource:@"default" ofType:@"metallib"];
+        NSURL * libURL = [NSURL fileURLWithPath:libPath];
+
+        // Load the metallib file into a Metal library
+        ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+
         if (error) {
             metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
@@ -146,7 +179,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
         //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
         NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        NSString * path   = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
         metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
 
         NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
@@ -192,7 +225,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
         GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(soft_max_4);
         GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf_8);
+        GGML_METAL_ADD_KERNEL(get_rows_f32);
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
@@ -204,7 +240,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(norm);
+        GGML_METAL_ADD_KERNEL(mul_mat_f32_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_l4);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
@@ -213,6 +252,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
@@ -222,7 +262,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(rope_f32);
+        GGML_METAL_ADD_KERNEL(rope_f16);
         GGML_METAL_ADD_KERNEL(alibi_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
@@ -231,13 +272,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
     metal_printf("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+#if TARGET_OS_OSX
+    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
     if (ctx->device.maxTransferRate != 0) {
         metal_printf("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
         metal_printf("%s: maxTransferRate               = built-in GPU\n", __func__);
     }
+#endif
 
     return ctx;
 }
@@ -257,7 +300,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
     GGML_METAL_DEL_KERNEL(soft_max);
+    GGML_METAL_DEL_KERNEL(soft_max_4);
     GGML_METAL_DEL_KERNEL(diag_mask_inf);
+    GGML_METAL_DEL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DEL_KERNEL(get_rows_f32);
     GGML_METAL_DEL_KERNEL(get_rows_f16);
     GGML_METAL_DEL_KERNEL(get_rows_q4_0);
     GGML_METAL_DEL_KERNEL(get_rows_q4_1);
@@ -269,7 +315,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_q6_K);
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(norm);
+    GGML_METAL_DEL_KERNEL(mul_mat_f32_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_l4);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
@@ -278,6 +327,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
@@ -287,7 +337,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DEL_KERNEL(rope);
+    GGML_METAL_DEL_KERNEL(rope_f32);
+    GGML_METAL_DEL_KERNEL(rope_f16);
     GGML_METAL_DEL_KERNEL(alibi_f32);
     GGML_METAL_DEL_KERNEL(cpy_f32_f16);
     GGML_METAL_DEL_KERNEL(cpy_f32_f32);
@@ -310,7 +361,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
 
 void * ggml_metal_host_malloc(size_t n) {
     void * data = NULL;
-    const int result = posix_memalign((void **) &data, getpagesize(), n);
+    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
     if (result != 0) {
         metal_printf("%s: error: posix_memalign failed\n", __func__);
         return NULL;
@@ -348,6 +399,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
     for (int i = 0; i < ctx->n_buffers; ++i) {
         const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
 
+        //metal_printf("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
         if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
@@ -384,7 +436,7 @@ bool ggml_metal_add_buffer(
             }
         }
 
-        const size_t size_page = getpagesize();
+        const size_t size_page = sysconf(_SC_PAGESIZE);
 
         size_t size_aligned = size;
         if ((size_aligned % size_page) != 0) {
@@ -437,6 +489,7 @@ bool ggml_metal_add_buffer(
             }
         }
 
+#if TARGET_OS_OSX
         metal_printf(", (%8.2f / %8.2f)",
                 ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
                 ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
@@ -446,6 +499,9 @@ bool ggml_metal_add_buffer(
         } else {
             metal_printf("\n");
         }
+#else
+        metal_printf(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
     }
 
     return true;
@@ -680,9 +736,21 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_ADD:
                         {
-                            if (ggml_nelements(src1) == ne10) {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));
+
+                            bool bcast_row = false;
+
+                            int64_t nb = ne00;
+
+                            if (ggml_nelements(src1) == ne10 && ne00 % 4 == 0) {
                                 // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
+
+                                nb = ne00 / 4;
                                 [encoder setComputePipelineState:ctx->pipeline_add_row];
+
+                                bcast_row = true;
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_add];
                             }
@@ -690,15 +758,53 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
 
-                            const int64_t n = ggml_nelements(dst);
+                            if (bcast_row) {
+                                const int64_t n = ggml_nelements(dst)/4;
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } else {
+                                const int nth = MIN(1024, ne0);
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            }
                         } break;
                     case GGML_OP_MUL:
                         {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));
+
+                            // utilize float4
+                            GGML_ASSERT(ne00 % 4 == 0);
+                            const int64_t nb = ne00/4;
+
                             if (ggml_nelements(src1) == ne10) {
                                 // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
                                 [encoder setComputePipelineState:ctx->pipeline_mul_row];
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_mul];
@@ -706,14 +812,16 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&nb     length:sizeof(nb) atIndex:3];
 
-                            const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst)/4;
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_SCALE:
                         {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+
                             const float scale = *(const float *) src1->data;
 
                             [encoder setComputePipelineState:ctx->pipeline_scale];
@@ -721,7 +829,7 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst)/4;
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
@@ -733,7 +841,7 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
@@ -753,7 +861,7 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
@@ -765,15 +873,18 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_SOFT_MAX:
                         {
-                            const int nth = 32;
+                            const int nth = MIN(32, ne00);
 
-                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            if (ne00%4 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -781,14 +892,23 @@ void ggml_metal_graph_compute(
                         {
                             const int n_past = ((int32_t *)(dst->op_params))[0];
 
-                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            if (ne00%8 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf_8];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            if (ne00%8 == 0) {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
+                            else {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
                         } break;
                     case GGML_OP_MUL_MAT:
                         {
@@ -801,13 +921,14 @@ void ggml_metal_graph_compute(
 
                             // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                             // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                            if (ggml_is_contiguous(src0) &&
-                                ggml_is_contiguous(src1) &&
+                            if (!ggml_is_transposed(src0) &&
+                                !ggml_is_transposed(src1) &&
                                 src1t == GGML_TYPE_F32 &&
                                 [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
                                 ne00%32 == 0 &&
                                 ne11 > 1) {
                                 switch (src0->type) {
+                                    case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32];  break;
                                     case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
                                     case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
                                     case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
@@ -827,22 +948,39 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
                                 [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
                                 [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
-                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:8];
-                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:9];
-                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:10];
+                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
+                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
+                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:13];
                                 [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                             } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
+                                int nrows = 1;
 
                                 // use custom matrix x vector kernel
                                 switch (src0t) {
+                                    case GGML_TYPE_F32:
+                                        {
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f32_f32];
+                                            nrows = 4;
+                                        } break;
                                     case GGML_TYPE_F16:
                                         {
-                                            nth0 = 64;
+                                            nth0 = 32;
                                             nth1 = 1;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            if (ne11 * ne12 < 4) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row];
+                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_l4];
+                                                nrows = ne11;
+                                            } else {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                                nrows = 4;
+                                            }
                                         } break;
                                     case GGML_TYPE_Q4_0:
                                         {
@@ -894,8 +1032,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 2;
-                                            nth1 = 32;
+                                            nth0 = 4; //1;
+                                            nth1 = 8; //32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -943,9 +1081,12 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
-                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
+                                    src0t == GGML_TYPE_Q2_K) {// || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -959,14 +1100,15 @@ void ggml_metal_graph_compute(
                                 else if (src0t == GGML_TYPE_Q6_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
-                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    int64_t ny = (ne11 + nrows - 1)/nrows;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
                         } break;
                     case GGML_OP_GET_ROWS:
                         {
                             switch (src0->type) {
+                                case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f32];  break;
                                 case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                 case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
@@ -982,9 +1124,9 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
-                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:5];
 
                             const int64_t n = ggml_nelements(src1);
 
@@ -995,7 +1137,7 @@ void ggml_metal_graph_compute(
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 512;
+                            const int nth = MIN(512, ne00);
 
                             [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -1014,7 +1156,7 @@ void ggml_metal_graph_compute(
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 256;
+                            const int nth = MIN(256, ne00);
 
                             [encoder setComputePipelineState:ctx->pipeline_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
@@ -1032,6 +1174,8 @@ void ggml_metal_graph_compute(
                         {
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
+                            const int nth = MIN(1024, ne00);
+
                             const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
                             const int n_head = ((int32_t *) dst->op_params)[1];
                             float max_bias;
@@ -1065,12 +1209,14 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
                             [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
 
-                            const int nth = 32;
-
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ROPE:
                         {
+                            GGML_ASSERT(ne10 == ne02);
+
+                            const int nth = MIN(1024, ne00);
+
                             const int n_past = ((int32_t *) dst->op_params)[0];
                             const int n_dims = ((int32_t *) dst->op_params)[1];
                             const int mode   = ((int32_t *) dst->op_params)[2];
@@ -1080,38 +1226,44 @@ void ggml_metal_graph_compute(
                             memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
                             memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
-                            [encoder setComputePipelineState:ctx->pipeline_rope];
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:18];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
+                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_rope_f16]; break;
+                                default: GGML_ASSERT(false);
+                            };
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:2];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:22];
+                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_DUP:
                     case GGML_OP_CPY:
                     case GGML_OP_CONT:
                         {
-                            const int nth = 32;
+                            const int nth = MIN(1024, ne00);
 
                             switch (src0t) {
                                 case GGML_TYPE_F32:

ggml-opencl.cpp

@@ -1334,7 +1334,7 @@ void ggml_cl_free_data(const struct ggml_tensor* tensor) {
         return;
     }
 
-    cl_mem mem = (cl_mem)tensor->data;
+    cl_mem mem = (cl_mem)tensor->extra;
     clReleaseMemObject(mem);
 }
 
@@ -1393,7 +1393,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
     size_t d_size;
 
     cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
-    cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
     cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
 
 
@@ -1491,9 +1491,9 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
     } else {
-        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
+        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
@@ -1567,7 +1567,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
     }
@@ -1697,7 +1697,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 events.emplace_back();
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
             } else if (src0->backend == GGML_BACKEND_GPU) {
-                d_Q = (cl_mem) src0->data;
+                d_Q = (cl_mem) src0->extra;
             } else {
                 GGML_ASSERT(false);
             }
@@ -1860,6 +1860,6 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
 
     CL_CHECK(clFinish(queue));
 
-    tensor->data = dst;
+    tensor->extra = dst;
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }

ggml.h

@@ -195,6 +195,14 @@
 #    define GGML_DEPRECATED(func, hint) func
 #endif
 
+#ifndef __GNUC__
+#    define GGML_ATTRIBUTE_FORMAT(...)
+#elif defined(__MINGW32__)
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -685,6 +693,7 @@ extern "C" {
 
     GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_ATTRIBUTE_FORMAT(2, 3)
     GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
@@ -1040,7 +1049,6 @@ extern "C" {
             size_t                nb1,
             size_t                offset);
 
-
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
             struct ggml_context * ctx,
@@ -1063,6 +1071,33 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, with new shape
+    GGML_API struct ggml_tensor * ggml_cont_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
+    GGML_API struct ggml_tensor * ggml_cont_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1);
+
+    GGML_API struct ggml_tensor * ggml_cont_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2);
+
+    GGML_API struct ggml_tensor * ggml_cont_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1210,14 +1245,15 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements
+    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
     // if mode & 2 == 1, GPT-NeoX style
     // if mode & 4 == 1, ChatGLM style
-    // TODO: avoid creating a new tensor every time
+    //
+    // b is an int32 vector with size a->ne[2], it contains the positions
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx);
@@ -1226,7 +1262,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx);
@@ -1235,7 +1271,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1246,7 +1282,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1257,7 +1293,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             float                 base,
             bool                  down);
@@ -1267,7 +1303,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1866,39 +1902,39 @@ extern "C" {
 
     GGML_API const char * gguf_type_name(enum gguf_type type);
 
-    GGML_API int    gguf_get_version    (struct gguf_context * ctx);
-    GGML_API size_t gguf_get_alignment  (struct gguf_context * ctx);
-    GGML_API size_t gguf_get_data_offset(struct gguf_context * ctx);
-    GGML_API void * gguf_get_data       (struct gguf_context * ctx);
+    GGML_API int    gguf_get_version    (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_alignment  (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
+    GGML_API void * gguf_get_data       (const struct gguf_context * ctx);
 
-    GGML_API int          gguf_get_n_kv(struct gguf_context * ctx);
-    GGML_API int          gguf_find_key(struct gguf_context * ctx, const char * key);
-    GGML_API const char * gguf_get_key (struct gguf_context * ctx, int i);
+    GGML_API int          gguf_get_n_kv(const struct gguf_context * ctx);
+    GGML_API int          gguf_find_key(const struct gguf_context * ctx, const char * key);
+    GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int i);
 
-    GGML_API enum gguf_type gguf_get_kv_type (struct gguf_context * ctx, int i);
-    GGML_API enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i);
+    GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int i);
+    GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int i);
 
     // results are undefined if the wrong type is used for the key
-    GGML_API uint8_t      gguf_get_val_u8  (struct gguf_context * ctx, int i);
-    GGML_API int8_t       gguf_get_val_i8  (struct gguf_context * ctx, int i);
-    GGML_API uint16_t     gguf_get_val_u16 (struct gguf_context * ctx, int i);
-    GGML_API int16_t      gguf_get_val_i16 (struct gguf_context * ctx, int i);
-    GGML_API uint32_t     gguf_get_val_u32 (struct gguf_context * ctx, int i);
-    GGML_API int32_t      gguf_get_val_i32 (struct gguf_context * ctx, int i);
-    GGML_API float        gguf_get_val_f32 (struct gguf_context * ctx, int i);
-    GGML_API uint64_t     gguf_get_val_u64 (struct gguf_context * ctx, int i);
-    GGML_API int64_t      gguf_get_val_i64 (struct gguf_context * ctx, int i);
-    GGML_API double       gguf_get_val_f64 (struct gguf_context * ctx, int i);
-    GGML_API bool         gguf_get_val_bool(struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);
-    GGML_API int          gguf_get_arr_n   (struct gguf_context * ctx, int i);
-    GGML_API const void * gguf_get_arr_data(struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_arr_str (struct gguf_context * ctx, int key_id, int i);
+    GGML_API uint8_t      gguf_get_val_u8  (const struct gguf_context * ctx, int i);
+    GGML_API int8_t       gguf_get_val_i8  (const struct gguf_context * ctx, int i);
+    GGML_API uint16_t     gguf_get_val_u16 (const struct gguf_context * ctx, int i);
+    GGML_API int16_t      gguf_get_val_i16 (const struct gguf_context * ctx, int i);
+    GGML_API uint32_t     gguf_get_val_u32 (const struct gguf_context * ctx, int i);
+    GGML_API int32_t      gguf_get_val_i32 (const struct gguf_context * ctx, int i);
+    GGML_API float        gguf_get_val_f32 (const struct gguf_context * ctx, int i);
+    GGML_API uint64_t     gguf_get_val_u64 (const struct gguf_context * ctx, int i);
+    GGML_API int64_t      gguf_get_val_i64 (const struct gguf_context * ctx, int i);
+    GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int i);
+    GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int i);
+    GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int i);
+    GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int i);
+    GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int i);
+    GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
 
-    GGML_API int    gguf_get_n_tensors    (struct gguf_context * ctx);
-    GGML_API int    gguf_find_tensor      (struct gguf_context * ctx, const char * name);
-    GGML_API size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i);
-    GGML_API char * gguf_get_tensor_name  (struct gguf_context * ctx, int i);
+    GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
+    GGML_API int    gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
+    GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API char * gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
 
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
@@ -1943,11 +1979,11 @@ extern "C" {
     //
 
     // write the entire context to a binary file
-    GGML_API void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta);
+    GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
 
     // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
-    GGML_API size_t gguf_get_meta_size(struct gguf_context * ctx);
-    GGML_API void   gguf_get_meta_data(struct gguf_context * ctx, void * data);
+    GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
+    GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
     //
     // system info
@@ -1961,6 +1997,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_fma        (void);
     GGML_API int ggml_cpu_has_neon       (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
+    GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);
     GGML_API int ggml_cpu_has_fp16_va    (void);
     GGML_API int ggml_cpu_has_wasm_simd  (void);

gguf-py/README.md

@@ -27,8 +27,25 @@ In this case, upgrade Pip to the latest:
 pip install --upgrade pip
 ```
 
-## Publishing
-To publish the package, you need to have `twine` and `build` installed:
+## Automatic publishing with CI
+
+There's a GitHub workflow to make a release automatically upon creation of tags in a specified format.
+
+1. Bump the version in `pyproject.toml`.
+2. Create a tag named `gguf-vx.x.x` where `x.x.x` is the semantic version number.
+
+```sh
+git tag -a gguf-v1.0.0 -m "Version 1.0 release"
+```
+
+3. Push the tags.
+
+```sh
+git push origin --tags
+```
+
+## Manual publishing
+If you want to publish the package manually for any reason, you need to have `twine` and `build` installed:
 
 ```sh
 pip install build twine
@@ -36,7 +53,7 @@ pip install build twine
 
 Then, folow these steps to release a new version:
 
-1. Update the version in `pyproject.toml`.
+1. Bump the version in `pyproject.toml`.
 2. Build the package:
 
 ```sh

gguf-py/gguf/gguf.py

@@ -1,16 +1,18 @@
 #!/usr/bin/env python3
-import shutil
-import sys
-import struct
-import tempfile
-import numpy as np
+from __future__ import annotations
+
 import json
 import os
-from pathlib import Path
-
+import shutil
+import struct
+import sys
+import tempfile
 from enum import IntEnum, auto
 from io import BufferedWriter
-from typing import Any, BinaryIO, Callable, IO, Dict, List, Optional, Sequence, Tuple, Union
+from pathlib import Path
+from typing import IO, Any, BinaryIO, Callable, Sequence
+
+import numpy as np
 
 #
 # constants
@@ -75,12 +77,14 @@ KEY_TOKENIZER_RWKV       = "tokenizer.rwkv.world"
 
 
 class MODEL_ARCH(IntEnum):
-    LLAMA  : int = auto()
-    FALCON : int = auto()
-    GPT2   : int = auto()
-    GPTJ   : int = auto()
-    GPTNEOX: int = auto()
-    MPT    : int = auto()
+    LLAMA         : int = auto()
+    FALCON        : int = auto()
+    BAICHUAN      : int = auto()
+    GPT2          : int = auto()
+    GPTJ          : int = auto()
+    GPTNEOX       : int = auto()
+    MPT           : int = auto()
+    STARCODER     : int = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -103,16 +107,18 @@ class MODEL_TENSOR(IntEnum):
     FFN_NORM     : int = auto()
 
 
-MODEL_ARCH_NAMES: Dict[MODEL_ARCH, str] = {
-    MODEL_ARCH.LLAMA:   "llama",
-    MODEL_ARCH.FALCON:  "falcon",
-    MODEL_ARCH.GPT2:    "gpt2",
-    MODEL_ARCH.GPTJ:    "gptj",
-    MODEL_ARCH.GPTNEOX: "gptneox",
-    MODEL_ARCH.MPT:     "mpt",
+MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
+    MODEL_ARCH.LLAMA:          "llama",
+    MODEL_ARCH.FALCON:         "falcon",
+    MODEL_ARCH.BAICHUAN:       "baichuan",
+    MODEL_ARCH.GPT2:           "gpt2",
+    MODEL_ARCH.GPTJ:           "gptj",
+    MODEL_ARCH.GPTNEOX:        "gptneox",
+    MODEL_ARCH.MPT:            "mpt",
+    MODEL_ARCH.STARCODER:      "starcoder",
 }
 
-MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = {
+MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = {
     MODEL_ARCH.LLAMA: {
         MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
         MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
@@ -151,6 +157,34 @@ MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = {
         MODEL_TENSOR.FFN_DOWN:    "blk.{bid}.ffn_down",
         MODEL_TENSOR.FFN_UP:      "blk.{bid}.ffn_up",
     },
+    MODEL_ARCH.BAICHUAN: {
+        MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
+        MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
+        MODEL_TENSOR.OUTPUT:        "output",
+        MODEL_TENSOR.ROPE_FREQS:    "rope_freqs",
+        MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
+        MODEL_TENSOR.ATTN_Q:        "blk.{bid}.attn_q",
+        MODEL_TENSOR.ATTN_K:        "blk.{bid}.attn_k",
+        MODEL_TENSOR.ATTN_V:        "blk.{bid}.attn_v",
+        MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
+        MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
+        MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
+        MODEL_TENSOR.FFN_GATE:      "blk.{bid}.ffn_gate",
+        MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
+        MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
+    },
+    MODEL_ARCH.STARCODER: {
+        MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
+        MODEL_TENSOR.POS_EMBD:      "position_embd",
+        MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
+        MODEL_TENSOR.OUTPUT:        "output",
+        MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
+        MODEL_TENSOR.ATTN_QKV:      "blk.{bid}.attn_qkv",
+        MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
+        MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
+        MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
+        MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
+    },
     MODEL_ARCH.GPT2: {
         # TODO
     },
@@ -158,16 +192,20 @@ MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = {
 }
 
 # tensors that will not be serialized
-MODEL_TENSOR_SKIP: Dict[MODEL_ARCH, List[MODEL_TENSOR]] = {
+MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.BAICHUAN: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 
 class TensorNameMap:
-    mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = {
+    mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
         # Token embeddings
         MODEL_TENSOR.TOKEN_EMBD: (
             "gpt_neox.embed_in",           # gptneox
@@ -185,7 +223,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out", # gptneox
-            "lm_head",   # gpt2 mpt falcon llama-hf
+            "lm_head",   # gpt2 mpt falcon llama-hf baichuan
             "output",    # llama-pth
         ),
 
@@ -193,7 +231,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm", # gptneox
             "transformer.ln_f",          # gpt2 falcon
-            "model.norm",                # llama-hf
+            "model.norm",                # llama-hf baichuan
             "norm",                      # llama-pth
         ),
 
@@ -203,7 +241,7 @@ class TensorNameMap:
         ),
     }
 
-    block_mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = {
+    block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
         # Attention norm
         MODEL_TENSOR.ATTN_NORM: (
             "gpt_neox.layers.{bid}.input_layernorm", # gptneox
@@ -298,9 +336,9 @@ class TensorNameMap:
         ),
     }
 
-    mapping: Dict[str, Tuple[MODEL_TENSOR, str]]
+    mapping: dict[str, tuple[MODEL_TENSOR, str]]
 
-    tensor_names: Dict[MODEL_TENSOR, str]
+    tensor_names: dict[MODEL_TENSOR, str]
 
     def __init__(self, arch: MODEL_ARCH, n_blocks: int):
         mapping = self.mapping = {}
@@ -309,6 +347,7 @@ class TensorNameMap:
             tensor_name = tensor_names.get(tensor)
             if tensor_name is None:
                 continue
+            mapping[tensor_name] = (tensor, tensor_name)
             for key in keys:
                 mapping[key] = (tensor, tensor_name)
         for bid in range(n_blocks):
@@ -317,11 +356,12 @@ class TensorNameMap:
                 if tensor_name is None:
                     continue
                 tensor_name = tensor_name.format(bid = bid)
+                mapping[tensor_name] = (tensor, tensor_name)
                 for key in keys:
                     key = key.format(bid = bid)
                     mapping[key] = (tensor, tensor_name)
 
-    def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[Tuple[MODEL_TENSOR, str]]:
+    def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
         result = self.mapping.get(key)
         if result is not None:
             return result
@@ -332,13 +372,13 @@ class TensorNameMap:
                     return (result[0], result[1] + suffix)
         return None
 
-    def get_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[str]:
+    def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None:
         result = self.get_type_and_name(key, try_suffixes = try_suffixes)
         if result is None:
             return None
         return result[1]
 
-    def get_type(self, key: str, try_suffixes: Sequence[str]) -> Optional[MODEL_TENSOR]:
+    def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None:
         result = self.get_type_and_name(key, try_suffixes = try_suffixes)
         if result is None:
             return None
@@ -432,10 +472,10 @@ class GGUFWriter:
     ti_data = b""
     ti_data_count = 0
     use_temp_file: bool
-    temp_file: Optional[tempfile.SpooledTemporaryFile[bytes]] = None
-    tensors: List[Tuple[np.ndarray[Any, Any], int]]
+    temp_file: tempfile.SpooledTemporaryFile[bytes] | None = None
+    tensors: list[tuple[np.ndarray[Any, Any], int]]
 
-    def __init__(self, path: Union[os.PathLike[str], str], arch: str, use_temp_file = True):
+    def __init__(self, path: os.PathLike[str] | str, arch: str, use_temp_file = True):
         self.fout = open(path, "wb")
         self.arch = arch
         self.add_architecture()
@@ -531,7 +571,7 @@ class GGUFWriter:
         GGUFValueType.FLOAT64: "<d",
         GGUFValueType.BOOL:    "?" ,
     }
-    def add_val(self, val: Any, vtype: Optional[GGUFValueType] = None, add_vtype: bool = True):
+    def add_val(self, val: Any, vtype: GGUFValueType | None = None, add_vtype: bool = True):
         if vtype is None:
             vtype = GGUFValueType.get_type(val)
 
@@ -561,7 +601,7 @@ class GGUFWriter:
     def ggml_pad(x: int, n: int) -> int:
         return ((x + n - 1) // n) * n
 
-    def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: Union[np.dtype[np.float16], np.dtype[np.float32]], tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None):
+    def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32], tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None):
         assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
 
         encoded_name = name.encode("utf8")
@@ -580,7 +620,7 @@ class GGUFWriter:
         self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
         self.ti_data_count += 1
 
-    def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Optional[Sequence[int]] = None, raw_dtype: Optional[GGMLQuantizationType] = None):
+    def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, raw_dtype: GGMLQuantizationType | None = None):
         if self.use_temp_file and self.temp_file is None:
             fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024)
             fp.seek(0)
@@ -600,7 +640,7 @@ class GGUFWriter:
         if pad != 0:
             self.temp_file.write(bytes([0] * pad))
 
-    def write_padding(self, fp: BinaryIO, n: int, align: Optional[int] = None):
+    def write_padding(self, fp: BinaryIO, n: int, align: int | None = None):
         pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n
         if pad != 0:
             fp.write(bytes([0] * pad))
@@ -726,13 +766,13 @@ class GGUFWriter:
     def add_tokenizer_model(self, model: str):
         self.add_string(KEY_TOKENIZER_MODEL, model)
 
-    def add_token_list(self, tokens: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]):
+    def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]):
         self.add_array(KEY_TOKENIZER_LIST, tokens)
 
-    def add_token_merges(self, merges: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]):
+    def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]):
         self.add_array(KEY_TOKENIZER_MERGES, merges)
 
-    def add_token_types(self, types: Union[Sequence[TokenType], Sequence[int]]):
+    def add_token_types(self, types: Sequence[TokenType] | Sequence[int]):
         self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types)
 
     def add_token_scores(self, scores: Sequence[float]):
@@ -756,11 +796,11 @@ class GGUFWriter:
 
 class SpecialVocab:
     load_merges: bool = False
-    merges: List[str] = []
-    special_token_types: Tuple[str, ...] = tuple(('bos', 'eos', 'unk', 'sep', 'pad'))
-    special_token_ids: Dict[str, int] = {}
+    merges: list[str] = []
+    special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad')
+    special_token_ids: dict[str, int] = {}
 
-    def __init__(self, path: Path, load_merges: bool = False, special_token_types: Optional[Tuple[str, ...]] = None):
+    def __init__(self, path: Path, load_merges: bool = False, special_token_types: tuple[str, ...] | None = None):
         self.special_token_ids = {}
         self.load_merges = load_merges
         if special_token_types is not None:
@@ -799,7 +839,7 @@ class SpecialVocab:
             else:
                 continue
             for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content):
-                if isinstance(maybe_token_id, int):
+                if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
                     self.special_token_ids[typ] = maybe_token_id
                 break
         return True
@@ -812,7 +852,7 @@ class SpecialVocab:
             config = json.load(f)
         for typ in self.special_token_types:
             maybe_token_id = config.get(f'{typ}_token_id')
-            if isinstance(maybe_token_id, int):
+            if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
                 self.special_token_ids[typ] = maybe_token_id
         return True
 
@@ -821,7 +861,7 @@ class SpecialVocab:
             print(f'gguf: Adding {len(self.merges)} merge(s).')
             gw.add_token_merges(self.merges)
         for typ, tokid in self.special_token_ids.items():
-            handler: Optional[Callable[[int], None]] = getattr(gw, f'add_{typ}_token_id', None)
+            handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
             if handler is None:
                 print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping')
                 continue

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.2.1"
+version = "0.3.3"
 description = "Write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

grammars/c.gbnf

@@ -0,0 +1,42 @@
+root ::= (declaration)*
+
+declaration ::= dataType identifier "(" parameter? ")" "{" statement* "}"
+
+dataType  ::= "int" ws | "float" ws | "char" ws
+identifier ::= [a-zA-Z_] [a-zA-Z_0-9]*
+
+parameter ::= dataType identifier
+
+statement ::=
+    ( dataType identifier ws "=" ws expression ";" ) |
+    ( identifier ws "=" ws expression ";" ) |
+    ( identifier ws "(" argList? ")" ";" ) |
+    ( "return" ws expression ";" ) |
+    ( "while" "(" condition ")" "{" statement* "}" ) |
+    ( "for" "(" forInit ";" ws condition ";" ws forUpdate ")" "{" statement* "}" ) |
+    ( "if" "(" condition ")" "{" statement* "}" ("else" "{" statement* "}")? ) |
+    ( singleLineComment ) |
+    ( multiLineComment )
+
+forInit ::= dataType identifier ws "=" ws expression | identifier ws "=" ws expression
+forUpdate ::= identifier ws "=" ws expression
+
+condition ::= expression relationOperator expression
+relationOperator ::= ("<=" | "<" | "==" | "!=" | ">=" | ">")
+
+expression ::= term (("+" | "-") term)*
+term ::= factor(("*" | "/") factor)*
+
+factor ::= identifier | number | unaryTerm | funcCall | parenExpression
+unaryTerm ::= "-" factor
+funcCall ::= identifier "(" argList? ")"
+parenExpression ::= "(" ws expression ws ")"
+
+argList ::= expression ("," ws expression)*
+
+number ::= [0-9]+
+
+singleLineComment ::= "//" [^\n]* "\n"
+multiLineComment ::= "/*" ( [^*] | ("*" [^/]) )* "*/"
+
+ws ::= ([ \t\n]+)

grammars/json_arr.gbnf

@@ -0,0 +1,34 @@
+# This is the same as json.gbnf but we restrict whitespaces at the end of the root array
+# Useful for generating JSON arrays
+
+root   ::= arr
+value  ::= object | array | string | number | ("true" | "false" | "null") ws
+
+arr  ::=
+  "[\n" ws (
+            value
+    (",\n" ws value)*
+  )? "]"
+
+object ::=
+  "{" ws (
+            string ":" ws value
+    ("," ws string ":" ws value)*
+  )? "}" ws
+
+array  ::=
+  "[" ws (
+            value
+    ("," ws value)*
+  )? "]" ws
+
+string ::=
+  "\"" (
+    [^"\\] |
+    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
+  )* "\"" ws
+
+number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
+
+# Optional space: by convention, applied in this grammar after literal chars when allowed
+ws ::= ([ \t\n] ws)?

k_quants.c

@@ -13,6 +13,26 @@
 //
 #include <arm_neon.h>
 
+#if !defined(__aarch64__)
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+#endif
+
 #else
 
 #ifdef __wasm_simd128__
@@ -63,7 +83,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (!amax) { // all zero
+    if (amax < 1e-30f) { // all zero
         for (int i = 0; i < n; ++i) {
             L[i] = 0;
         }
@@ -183,13 +203,9 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
         int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
-    float sum_x = 0;
-    float sum_x2 = 0;
     for (int i = 1; i < n; ++i) {
         if (x[i] < min) min = x[i];
         if (x[i] > max) max = x[i];
-        sum_x += x[i];
-        sum_x2 += x[i]*x[i];
     }
     if (max == min) {
         for (int i = 0; i < n; ++i) L[i] = 0;
@@ -1070,6 +1086,13 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         }
 
+        if (!max_abs_scale) {
+            memset(&y[i], 0, sizeof(block_q6_K));
+            y[i].d = ggml_fp32_to_fp16(0.f);
+            x += QK_K;
+            continue;
+        }
+
         float iscale = -128.f/max_scale;
         y[i].d = ggml_fp32_to_fp16(1/iscale);
         for (int ib = 0; ib < QK_K/16; ++ib) {
@@ -1306,7 +1329,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
     const uint8x16_t m4 = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x2_t q2bytes;
     uint8_t aux[16];
@@ -1612,7 +1637,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q2bytes;
 
@@ -2060,7 +2087,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     __m256 acc = _mm256_setzero_ps();
 
-    uint32_t *aux;
+    const uint32_t *aux;
 
     for (int i = 0; i < nb; ++i) {
 
@@ -2070,7 +2097,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const int8_t  * restrict q8 = y[i].qs;
 
         // Set up scales
-        aux = (uint32_t *)x[i].scales;
+        aux = (const uint32_t *)x[i].scales;
         __m128i scales128 = _mm_set_epi32(
                 ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
                 ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
@@ -2582,7 +2609,10 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
         memcpy(utmp, x[i].scales, 12);
 
-        const uint32x2_t mins8 = {utmp[1] & kmask1, ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4)};
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
         utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
         utmp[0] &= kmask1;
 
@@ -2596,8 +2626,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        //int32x4_t isum = mzero;
-
         int32_t sumi1 = 0;
         int32_t sumi2 = 0;
 
@@ -3096,9 +3124,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mone = vdupq_n_u8(1);
     const uint8x16_t mtwo = vdupq_n_u8(2);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
 
@@ -3441,8 +3471,10 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mh = vdupq_n_u8(16);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
     uint8x16x4_t q5h;
@@ -3660,7 +3692,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
     //const int8x16_t  m32s = vdupq_n_s8(32);
 
     const uint8x16_t mone = vdupq_n_u8(3);
@@ -4049,8 +4083,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int32x4_t  vzero = vdupq_n_s32(0);
     const int8x16_t  m32s = vdupq_n_s8(32);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     const uint8x16_t mone = vdupq_n_u8(3);
 

llama.h

@@ -37,6 +37,8 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
+#define LLAMA_MAX_RNG_STATE (64*1024)
+
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
@@ -60,7 +62,9 @@ extern "C" {
     struct llama_model;
     struct llama_context;
 
-    typedef int llama_token;
+    typedef int32_t llama_pos;
+    typedef int32_t llama_token;
+    typedef int32_t llama_seq_id;
 
     enum llama_log_level {
         LLAMA_LOG_LEVEL_ERROR = 2,
@@ -86,24 +90,24 @@ extern "C" {
     // model file types
     enum llama_ftype {
         LLAMA_FTYPE_ALL_F32              = 0,
-        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_F16           = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4,  // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5,  // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6,  // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -122,6 +126,35 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
+    // Input data for llama_decode
+    // A llama_batch object can contain input about one or many sequences
+    // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
+    //
+    // - token  : the token ids of the input (used when embd is NULL)
+    // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
+    // - pos    : the positions of the respective token in the sequence
+    // - seq_id : the sequence to which the respective token belongs
+    // - logits : if zero, the logits for the respective token will not be output
+    //
+    typedef struct llama_batch {
+        int32_t n_tokens;
+
+        llama_token  * token;
+        float        * embd;
+        llama_pos    * pos;
+        llama_seq_id * seq_id;
+        int8_t       * logits;
+
+        // NOTE: helpers for smooth API transition - can be deprecated in the future
+        //       for future-proof code, use the above fields instead and ignore everything below
+        //
+        // pos[i] = all_pos_0 + i*all_pos_1
+        //
+        llama_pos    all_pos_0;  // used if pos == NULL
+        llama_pos    all_pos_1;  // used if pos == NULL
+        llama_seq_id all_seq_id; // used if seq_id == NULL
+    } llama_batch;
+
     struct llama_context_params {
         uint32_t seed;         // RNG seed, -1 for random
         int32_t  n_ctx;        // text context
@@ -164,6 +197,7 @@ extern "C" {
         enum llama_ftype ftype;      // quantize to this llama_ftype
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
     } llama_model_quantize_params;
 
     // grammar types
@@ -214,6 +248,7 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // Helpers for getting default parameters
     LLAMA_API struct llama_context_params llama_context_default_params(void);
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
 
@@ -244,20 +279,24 @@ extern "C" {
     LLAMA_API bool llama_mmap_supported (void);
     LLAMA_API bool llama_mlock_supported(void);
 
-    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
-    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
+    LLAMA_API int llama_n_vocab    (const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx_train(const struct llama_context * ctx);
+    LLAMA_API int llama_n_embd     (const struct llama_context * ctx);
 
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx);
 
-    LLAMA_API int llama_model_n_vocab(const struct llama_model * model);
-    LLAMA_API int llama_model_n_ctx  (const struct llama_model * model);
-    LLAMA_API int llama_model_n_embd (const struct llama_model * model);
+    LLAMA_API int llama_model_n_vocab    (const struct llama_model * model);
+    LLAMA_API int llama_model_n_ctx      (const struct llama_model * model);
+    LLAMA_API int llama_model_n_ctx_train(const struct llama_model * model);
+    LLAMA_API int llama_model_n_embd     (const struct llama_model * model);
 
     // Get a string describing the model type
     LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+
     // Returns the total size of all the tensors in the model in bytes
     LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+
     // Returns the total number of parameters in the model
     LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
 
@@ -278,7 +317,7 @@ extern "C" {
                       const char * path_lora,
                       const char * path_base_model,
                              int   n_threads),
-            "please use llama_model_apply_lora_from_file instead");
+            "use llama_model_apply_lora_from_file instead");
 
     LLAMA_API int llama_model_apply_lora_from_file(
             const struct llama_model * model,
@@ -286,11 +325,53 @@ extern "C" {
                           const char * path_base_model,
                                  int   n_threads);
 
-    // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    //
+    // KV cache
+    //
 
-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+    // Returns the number of tokens in the KV cache
+    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "avoid using this, it will be removed in the future, instead - count the tokens in user code");
+
+    // Remove all tokens data of cells in [c0, c1)
+    LLAMA_API void llama_kv_cache_tokens_rm(
+            struct llama_context * ctx,
+                         int32_t   c0,
+                         int32_t   c1);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    LLAMA_API void llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
+    LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    LLAMA_API void llama_kv_cache_seq_shift(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta);
+
+    //
+    // State / sessions
+    //
 
     // Returns the maximum size in bytes of the state (rng, logits, embedding
     // and kv_cache) - will often be smaller after compacting tokens
@@ -299,48 +380,100 @@ extern "C" {
     // Copies the state to the specified destination address.
     // Destination needs to have allocated enough memory.
     // Returns the number of bytes copied
-    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+    LLAMA_API size_t llama_copy_state_data(
+            struct llama_context * ctx,
+                         uint8_t * dst);
 
     // Set the state reading from the specified address
     // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+    LLAMA_API size_t llama_set_state_data(
+            struct llama_context * ctx,
+                         uint8_t * src);
 
     // Save/load session file
-    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
-    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    LLAMA_API bool llama_load_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+                     llama_token * tokens_out,
+                          size_t   n_token_capacity,
+                          size_t * n_token_count_out);
 
-    // Run the llama inference to obtain the logits and probabilities for the next token.
+    LLAMA_API bool llama_save_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+               const llama_token * tokens,
+                          size_t   n_token_count);
+
+    //
+    // Decoding
+    //
+
+    // Run the llama inference to obtain the logits and probabilities for the next token(s).
     // tokens + n_tokens is the provided batch of new tokens to process
     // n_past is the number of tokens to use from previous eval calls
     // Returns 0 on success
-    LLAMA_API int llama_eval(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval(
             struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
+                     llama_token * tokens,
+                         int32_t   n_tokens,
                              int   n_past,
-                             int   n_threads);
+                             int   n_threads),
+            "use llama_decode() instead");
 
     // Same as llama_eval, but use float matrix input directly.
-    LLAMA_API int llama_eval_embd(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval_embd(
             struct llama_context * ctx,
-                     const float * embd,
-                             int   n_tokens,
+                           float * embd,
+                         int32_t   n_tokens,
                              int   n_past,
-                             int   n_threads);
+                             int   n_threads),
+            "use llama_decode() instead");
 
-    // Export a static computation graph for context of 511 and batch size of 1
-    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
-    //       parameters here to keep things simple
-    // IMPORTANT: do not use for anything else other than debugging and testing!
-    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+    // Return batch for single sequence of tokens starting at pos_0
+    //
+    // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
+    //
+    LLAMA_API struct llama_batch llama_batch_get_one(
+                  llama_token * tokens,
+                      int32_t   n_tokens,
+                    llama_pos   pos_0,
+                 llama_seq_id   seq_id);
+
+    // Allocates a batch of tokens on the heap
+    // The batch has to be freed with llama_batch_free()
+    // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
+    // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
+    // The rest of the llama_batch members are allocated with size n_tokens
+    // All members are left uninitialized
+    LLAMA_API struct llama_batch llama_batch_init(
+            int32_t n_tokens,
+            int32_t embd);
+
+    // Frees a batch of tokens allocated with llama_batch_init()
+    LLAMA_API void llama_batch_free(struct llama_batch batch);
+
+    // Positive return values does not mean a fatal error, but rather a warning.
+    //   0 - success
+    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    // < 0 - error
+    LLAMA_API int llama_decode(
+            struct llama_context * ctx,
+              struct llama_batch   batch,
+                             int   n_threads);
 
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
     // Cols: n_vocab
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
+    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
+
     // Get the embeddings for the input
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
@@ -371,6 +504,7 @@ extern "C" {
     LLAMA_API int llama_tokenize(
             struct llama_context * ctx,
                       const char * text,
+                             int   text_len,
                      llama_token * tokens,
                              int   n_max_tokens,
                             bool   add_bos);
@@ -378,6 +512,7 @@ extern "C" {
     LLAMA_API int llama_tokenize_with_model(
         const struct llama_model * model,
                       const char * text,
+                             int   text_len,
                      llama_token * tokens,
                              int   n_max_tokens,
                             bool   add_bos);
@@ -409,15 +544,31 @@ extern "C" {
 
     LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
 
+    LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
+
     //
     // Sampling functions
     //
 
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
+    LLAMA_API void llama_sample_repetition_penalty(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   last_tokens_size,
+                          float    penalty);
 
     /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   last_tokens_size,
+                           float   alpha_frequency,
+                           float   alpha_presence);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
@@ -430,23 +581,54 @@ extern "C" {
                              float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API void llama_sample_softmax(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+    LLAMA_API void llama_sample_top_k(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                             int   k,
+                          size_t   min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_top_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
 
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+    LLAMA_API void llama_sample_tail_free(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   z,
+                          size_t   min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+    LLAMA_API void llama_sample_typical(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    LLAMA_API void llama_sample_temp(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   temp);
+
+    LLAMA_API DEPRECATED(void llama_sample_temperature(
+                struct llama_context * ctx,
+              llama_token_data_array * candidates,
+                               float   temp),
+            "use llama_sample_temp instead");
 
     /// @details Apply constraints from grammar
-    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+    LLAMA_API void llama_sample_grammar(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+      const struct llama_grammar * grammar);
 
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
@@ -454,23 +636,41 @@ extern "C" {
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                             int   m,
+                           float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                           float * mu);
 
     /// @details Selects the token with the highest probability.
-    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token_greedy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Randomly selects a token from the candidates based on their probabilities.
-    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Accepts the sampled token into the grammar
-    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+    LLAMA_API void llama_grammar_accept_token(
+            struct llama_context * ctx,
+            struct llama_grammar * grammar,
+                     llama_token   token);
 
     //
     // Beam search
@@ -478,9 +678,10 @@ extern "C" {
 
     struct llama_beam_view {
         const llama_token * tokens;
+
         size_t n_tokens;
-        float p;   // Cumulative beam probability (renormalized relative to all beams)
-        bool eob;  // Callback should set this to true when a beam is at end-of-beam.
+        float  p;        // Cumulative beam probability (renormalized relative to all beams)
+        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
     };
 
     // Passed to beam_search_callback function.
@@ -489,9 +690,10 @@ extern "C" {
     // These pointers are valid only during the synchronous callback, so should not be saved.
     struct llama_beams_state {
         struct llama_beam_view * beam_views;
+
         size_t n_beams;               // Number of elements in beam_views[].
         size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
-        bool last_call;               // True iff this is the last callback invocation.
+        bool   last_call;             // True iff this is the last callback invocation.
     };
 
     // Type of pointer to the beam_search_callback function.
@@ -507,10 +709,18 @@ extern "C" {
     /// @param n_past Number of tokens already evaluated.
     /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
     /// @param n_threads Number of threads as passed to llama_eval().
-    LLAMA_API void llama_beam_search(struct llama_context * ctx, llama_beam_search_callback_fn_t callback, void * callback_data, size_t n_beams, int n_past, int n_predict, int n_threads);
+    LLAMA_API void llama_beam_search(
+                   struct llama_context * ctx,
+        llama_beam_search_callback_fn_t   callback,
+                                   void * callback_data,
+                                 size_t   n_beams,
+                                    int   n_past,
+                                    int   n_predict,
+                                    int   n_threads);
 
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
     LLAMA_API void llama_reset_timings(struct llama_context * ctx);
 
@@ -535,7 +745,9 @@ extern "C" {
 
 struct ggml_tensor;
 
-const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+);
 
 #endif // LLAMA_API_INTERNAL
 

mypy.ini

@@ -0,0 +1,5 @@
+[mypy]
+strict = true
+allow_untyped_calls = true
+allow_untyped_defs = true
+allow_incomplete_defs = true

pocs/vdot/vdot.cpp

@@ -16,7 +16,7 @@
 
 constexpr int kVecSize = 1 << 18;
 
-float drawFromGaussianPdf(std::mt19937& rndm) {
+static float drawFromGaussianPdf(std::mt19937& rndm) {
     constexpr double kScale = 1./(1. + std::mt19937::max());
     constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale;
     static float lastX;
@@ -28,7 +28,8 @@ float drawFromGaussianPdf(std::mt19937& rndm) {
     haveX = true;
     return r*cos(phi);
 }
-void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
+
+static void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
     for (auto& v : values) v = mean + drawFromGaussianPdf(rndm);
 }
 

prompts/chat-with-baichuan.txt

@@ -0,0 +1,4 @@
+以下内容为人类用户与与一位智能助手的对话。
+
+用户:你好！
+助手:

run_with_preset.py

@@ -13,7 +13,7 @@ CLI_ARGS_MAIN_PERPLEXITY = [
     "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix", "instruct",
     "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base",
     "low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
-    "model", "mtest", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
+    "model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
     "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
     "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "random-prompt", "repeat-last-n",
     "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed",

scripts/LlamaConfig.cmake.in

@@ -0,0 +1,69 @@
+set(LLAMA_VERSION @LLAMA_INSTALL_VERSION@)
+set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
+set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
+set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
+set(LLAMA_BLAS @LLAMA_BLAS@)
+set(LLAMA_CUBLAS @LLAMA_CUBLAS@)
+set(LLAMA_METAL @LLAMA_METAL@)
+set(LLAMA_MPI @LLAMA_MPI@)
+set(LLAMA_CLBLAST @LLAMA_CLBLAST@)
+set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@)
+set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@)
+
+@PACKAGE_INIT@
+
+set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
+set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
+set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+
+# Ensure transient dependencies satisfied
+
+find_package(Threads REQUIRED)
+if (APPLE AND LLAMA_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+endif()
+
+if (LLAMA_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (LLAMA_CUBLAS)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (LLAMA_MPI)
+    find_package(MPI REQUIRED)
+endif()
+
+if (LLAMA_CLBLAST)
+    find_package(CLBlast REQUIRED)
+endif()
+
+if (LLAMA_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+find_library(llama_LIBRARY llama
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
+set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
+add_library(llama UNKNOWN IMPORTED)
+set_target_properties(llama
+    PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${llama_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
+
+check_required_components(Llama)

scripts/build-info.cmake

@@ -2,6 +2,8 @@ set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in")
 set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
 set(BUILD_NUMBER 0)
 set(BUILD_COMMIT "unknown")
+set(BUILD_COMPILER "unknown")
+set(BUILD_TARGET "unknown")
 
 # Look for git
 find_package(Git)
@@ -41,11 +43,45 @@ if(Git_FOUND)
     endif()
 endif()
 
+if(GIT_HEAD_RESULT EQUAL 0 AND GIT_COUNT_RESULT EQUAL 0)
+    set(BUILD_COMMIT ${HEAD})
+    set(BUILD_NUMBER ${COUNT})
+endif()
+
+execute_process(
+    COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
+    OUTPUT_VARIABLE OUT
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+    RESULT_VARIABLE RES
+)
+if (RES EQUAL 0)
+    set(BUILD_COMPILER ${OUT})
+endif()
+
+execute_process(
+    COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+    OUTPUT_VARIABLE OUT
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+    RESULT_VARIABLE RES
+)
+if (RES EQUAL 0)
+    set(BUILD_TARGET ${OUT})
+endif()
+
 # Only write the header if it's changed to prevent unnecessary recompilation
 if(EXISTS ${HEADER_FILE})
-    file(STRINGS ${HEADER_FILE} CONTENTS REGEX "BUILD_COMMIT \"([^\"]*)\"")
-    list(GET CONTENTS 0 EXISTING)
-    if(NOT EXISTING STREQUAL "#define BUILD_COMMIT \"${BUILD_COMMIT}\"")
+    file(READ ${HEADER_FILE} CONTENTS)
+    string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "BUILD_TARGET \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
         configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
     endif()
 else()

scripts/build-info.h.in

@@ -3,5 +3,7 @@
 
 #define BUILD_NUMBER @BUILD_NUMBER@
 #define BUILD_COMMIT "@BUILD_COMMIT@"
+#define BUILD_COMPILER "@BUILD_COMPILER@"
+#define BUILD_TARGET "@BUILD_TARGET@"
 
 #endif // BUILD_INFO_H

scripts/build-info.sh

@@ -1,23 +1,35 @@
 #!/bin/sh
 
-BUILD_NUMBER="0"
-BUILD_COMMIT="unknown"
+CC=$1
 
-REV_LIST=$(git rev-list --count HEAD)
-if [ $? -eq 0 ]; then
-  BUILD_NUMBER=$REV_LIST
+build_number="0"
+build_commit="unknown"
+build_compiler="unknown"
+build_target="unknown"
+
+if out=$(git rev-list --count HEAD); then
+  # git is broken on WSL so we need to strip extra newlines
+  build_number=$(printf '%s' "$out" | tr -d '\n')
 fi
 
-REV_PARSE=$(git rev-parse --short HEAD)
-if [ $? -eq 0 ]; then
-  BUILD_COMMIT=$REV_PARSE
+if out=$(git rev-parse --short HEAD); then
+  build_commit=$(printf '%s' "$out" | tr -d '\n')
+fi
+
+if out=$($CC --version | head -1); then
+  build_compiler=$out
+fi
+
+if out=$($CC -dumpmachine); then
+  build_target=$out
 fi
 
 echo "#ifndef BUILD_INFO_H"
 echo "#define BUILD_INFO_H"
-echo ""
-echo "#define BUILD_NUMBER $BUILD_NUMBER" | tr -d '\n'
-echo ""
-echo "#define BUILD_COMMIT \"$BUILD_COMMIT\"" | tr -d '\n'
-echo ""
+echo
+echo "#define BUILD_NUMBER $build_number"
+echo "#define BUILD_COMMIT \"$build_commit\""
+echo "#define BUILD_COMPILER \"$build_compiler\""
+echo "#define BUILD_TARGET \"$build_target\""
+echo
 echo "#endif // BUILD_INFO_H"

tests/CMakeLists.txt

@@ -29,15 +29,16 @@ llama_build_executable(test-tokenizer-0-llama.cpp)
 llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
 llama_build_executable(test-tokenizer-0-falcon.cpp)
 #llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_build_executable(test-tokenizer-1.cpp)
-# test-tokenizer-1 requires a BPE vocab. re-enable when we have one.
-#llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_build_executable(test-tokenizer-1-llama.cpp)
+llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
 #llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp) # SLOW
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
 
+llama_build_and_test_executable(test-rope.cpp)
+
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)
 add_executable(${TEST_TARGET} test-c.c)

tests/test-grad0.cpp

@@ -1404,6 +1404,11 @@ int main(int argc, const char ** argv) {
                     for (int n_past = 1; n_past < ne2[2]; ++n_past) {
                         x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
 
+                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
+                        for (int i = 0; i < ne2[2]; ++i) {
+                            ((int32_t *) p->data)[i] = n_past + i;
+                        }
+
                         ggml_set_param(ctx0, x[0]);
 
                         const bool skip_past = (mode & 1);
@@ -1415,7 +1420,7 @@ int main(int argc, const char ** argv) {
                             continue;
                         }
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode, 0));
 
                         GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                         check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
@@ -1438,6 +1443,11 @@ int main(int argc, const char ** argv) {
                     for (int n_past = 1; n_past < ne2[2]; ++n_past) {
                         x[0] = get_random_tensor_f16(ctx0, ndims, ne2, -1.0f, 1.0f);
 
+                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
+                        for (int i = 0; i < ne2[2]; ++i) {
+                            ((int32_t *) p->data)[i] = n_past + i;
+                        }
+
                         ggml_set_param(ctx0, x[0]);
 
                         const bool skip_past = (mode & 1);
@@ -1449,7 +1459,7 @@ int main(int argc, const char ** argv) {
                             continue;
                         }
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode, 0));
 
                         GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                         check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);

tests/test-opt.cpp

@@ -36,15 +36,15 @@
 #define GGML_PRINT(...) printf(__VA_ARGS__)
 
 
-float frand(void) {
+static float frand(void) {
     return (float)rand()/(float)RAND_MAX;
 }
 
-int irand(int n) {
+static int irand(int n) {
     return rand()%n;
 }
 
-void get_random_dims(int64_t * dims, int ndims) {
+static void get_random_dims(int64_t * dims, int ndims) {
     dims[0] = dims[1] = dims[2] = dims[3] = 1;
 
     for (int i = 0; i < ndims; i++) {
@@ -52,7 +52,7 @@ void get_random_dims(int64_t * dims, int ndims) {
     }
 }
 
-void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
+static void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
     dims[0] = dims[1] = dims[2] = dims[3] = 1;
 
     for (int i = 0; i < ndims; i++) {
@@ -61,12 +61,9 @@ void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
 }
 
 
-struct ggml_tensor * get_random_tensor(
-        struct ggml_context * ctx0,
-        int ndims,
-        int64_t ne[],
-        float fmin,
-        float fmax) {
+static struct ggml_tensor * get_random_tensor(
+    struct ggml_context * ctx0, int ndims, int64_t ne[], float fmin, float fmax
+) {
     struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
 
     switch (ndims) {
@@ -109,11 +106,11 @@ struct ggml_tensor * get_random_tensor(
     return result;
 }
 
-float get_element(const struct ggml_tensor * t, int idx) {
+static float get_element(const struct ggml_tensor * t, int idx) {
     return ((float *)t->data)[idx];
 }
 
-void set_element(struct ggml_tensor * t, int idx, float value) {
+static void set_element(struct ggml_tensor * t, int idx, float value) {
     ((float *)t->data)[idx] = value;
 }
 

tests/test-quantize-fns.cpp

@@ -13,24 +13,24 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
-const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
-const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
-const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
-const float MAX_DOT_PRODUCT_ERROR = 0.02f;
+constexpr float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
+constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f;
 
-const char* RESULT_STR[] = {"ok", "FAILED"};
+static const char* RESULT_STR[] = {"ok", "FAILED"};
 
 
 // Generate synthetic data
-void generate_data(float offset, size_t n, float * dst) {
+static void generate_data(float offset, size_t n, float * dst) {
     for (size_t i = 0; i < n; i++) {
         dst[i] = 0.1 + 2*cosf(i + offset);
     }
 }
 
 // Calculate RMSE between two float arrays
-float array_rmse(const float * a1, const float * a2, size_t n) {
+static float array_rmse(const float * a1, const float * a2, size_t n) {
     double sum = 0;
     for (size_t i = 0; i < n; i++) {
         double diff = a1[i] - a2[i];
@@ -40,7 +40,7 @@ float array_rmse(const float * a1, const float * a2, size_t n) {
 }
 
 // Total quantization error on test data
-float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
+static float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
 
@@ -50,7 +50,7 @@ float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, cons
 }
 
 // Total quantization error on test data
-float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
+static float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
     std::vector<float> tmp_out_ref(test_size);
@@ -64,7 +64,7 @@ float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size,
     return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
 }
 
-float dot_product(const float * a1, const float * a2, size_t test_size) {
+static float dot_product(const float * a1, const float * a2, size_t test_size) {
     double sum = 0;
     for (size_t i = 0; i < test_size; i++) {
         sum += a1[i] * a2[i];
@@ -73,7 +73,9 @@ float dot_product(const float * a1, const float * a2, size_t test_size) {
 }
 
 // Total dot product error
-float dot_product_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
+static float dot_product_error(
+    ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2
+) {
     std::vector<uint8_t> tmp_q1(2*test_size);
     std::vector<uint8_t> tmp_q2(2*test_size);
 

tests/test-quantize-perf.cpp

@@ -61,22 +61,22 @@ inline int64_t cpu_cycles() {
 
 
 // Generate synthetic data
-void generate_data(float offset, size_t n, float * dst) {
+static void generate_data(float offset, size_t n, float * dst) {
     for (size_t i = 0; i < n; i++) {
         dst[i] = 0.1 + 2*cosf(i + offset);
     }
 }
 
-float gigabytes_per_second(size_t bytes, int64_t usecs) {
+static float gigabytes_per_second(size_t bytes, int64_t usecs) {
     return bytes / (float) usecs * 1000000 / (1024*1024*1024);
 }
 
-void * align_with_offset(void * ptr, int offset) {
+static void * align_with_offset(void * ptr, int offset) {
     size_t dummy_size = MAX_ALIGNMENT * 4;
     return (char *) std::align(MAX_ALIGNMENT, MAX_ALIGNMENT, ptr, dummy_size) + offset;
 }
 
-void benchmark_function(size_t size, size_t q_size, int64_t iterations, std::function<size_t(void)> function) {
+static void benchmark_function(size_t size, size_t q_size, int64_t iterations, const std::function<size_t(void)> & function) {
     int64_t min_time_us = INT64_MAX;
     int64_t total_time_us = 0;
     int64_t min_time_cycles = INT64_MAX;
@@ -108,7 +108,7 @@ void benchmark_function(size_t size, size_t q_size, int64_t iterations, std::fun
     printf("      quantized throughput : %9.2f GB/s\n",  gigabytes_per_second(q_size * iterations, total_time_us));
 }
 
-void usage(char * argv[]) {
+static void usage(char * argv[]) {
     printf("Benchmark quantization specific functions on synthetic data\n");
     printf("\n");
     printf("usage: %s [options]\n", argv[0]);

tests/test-rope.cpp

@@ -0,0 +1,221 @@
+#include "ggml.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cassert>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wdouble-promotion"
+#endif
+
+#define MAX_NARGS 3
+
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+#define GGML_SILU_FP16
+
+//
+// logging
+//
+
+#if (GGML_DEBUG >= 1)
+#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG(...)
+#endif
+
+#if (GGML_DEBUG >= 5)
+#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_5(...)
+#endif
+
+#if (GGML_DEBUG >= 10)
+#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_10(...)
+#endif
+
+#define GGML_PRINT(...) printf(__VA_ARGS__)
+
+static float frand(void) {
+    return (float)rand()/(float)RAND_MAX;
+}
+
+static int irand(int n) {
+    if (n == 0) return 0;
+    return rand()%n;
+}
+
+static void get_random_dims(int64_t * dims, int ndims) {
+    dims[0] = dims[1] = dims[2] = dims[3] = 1;
+
+    for (int i = 0; i < ndims; i++) {
+        dims[i] = 1 + irand(4);
+    }
+}
+
+static struct ggml_tensor * get_random_tensor_f32(
+        struct ggml_context * ctx0,
+        int ndims,
+        const int64_t ne[],
+        float fmin,
+        float fmax) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
+
+    switch (ndims) {
+        case 1:
+            for (int i0 = 0; i0 < ne[0]; i0++) {
+                ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < ne[1]; i1++) {
+                for (int i0 = 0; i0 < ne[0]; i0++) {
+                    ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                }
+            }
+            break;
+        case 3:
+            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 *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                    }
+                }
+            }
+            break;
+        case 4:
+            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 *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            assert(false);
+    };
+
+    return result;
+}
+
+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) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    ggml_graph_compute(graph, &plan);
+}
+
+int main(int /*argc*/, const char ** /*argv*/) {
+    struct ggml_init_params params = {
+        /* .mem_size   = */ 128*1024*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ false,
+    };
+
+    std::vector<uint8_t> work_buffer;
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    struct ggml_tensor * x;
+
+    // rope f32
+    for (int m = 0; m < 3; ++m) {
+        const int ndims = 4;
+
+        const int64_t n_rot = 128;
+        const int64_t ne[4] = { 2*n_rot, 32, 73, 1 };
+
+        const int n_past_0 = 100;
+        const int n_past_2 = 33;
+
+        struct ggml_tensor * p0 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
+        struct ggml_tensor * p1 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
+        struct ggml_tensor * p2 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
+
+        for (int i = 0; i < ne[2]; ++i) {
+            ((int32_t *) p0->data)[i] = n_past_0 + i;
+            ((int32_t *) p1->data)[i] = n_past_2 - n_past_0;
+            ((int32_t *) p2->data)[i] = n_past_2 + i;
+        }
+
+        // test mode 0, 2, 4 (standard, GPT-NeoX, GLM)
+        const int mode = m == 0 ? 0 : m == 1 ? 2 : 4;
+
+        x = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+
+        // 100, 101, 102, ..., 172
+        struct ggml_tensor * r0 = ggml_rope(ctx0, x,  p0, n_rot, mode, 1024);
+        // -67, -67, -67, ..., -67
+        struct ggml_tensor * r1 = ggml_rope(ctx0, r0, p1, n_rot, mode, 1024); // "context swap", i.e. forget n_past_0 - n_past_2 tokens
+
+        //  33,  34,  35, ..., 105
+        struct ggml_tensor * r2 = ggml_rope(ctx0, x,  p2, n_rot, mode, 1024);
+
+        ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+        ggml_build_forward_expand(gf, r0);
+        ggml_build_forward_expand(gf, r1);
+        ggml_build_forward_expand(gf, r2);
+
+        ggml_graph_compute_helper(work_buffer, gf, 4);
+
+        // check that r1 and r2 are the same
+        {
+            double sum0 = 0.0f;
+            double sum1 = 0.0f;
+            double diff = 0.0f;
+
+            const float * r1_data = (float *) r1->data;
+            const float * r2_data = (float *) r2->data;
+
+            const int n_elements = ggml_nelements(r1);
+
+            for (int i = 0; i < n_elements; ++i) {
+                sum0 += fabs(r1_data[i]);
+                sum1 += fabs(r2_data[i]);
+                diff += fabs(r1_data[i] - r2_data[i]);
+                //if (fabs(r1_data[i] - r2_data[i]) > 0.0001f) {
+                //    printf("%d: %f %f\n", i, r1_data[i], r2_data[i]);
+                //    printf("diff: %f\n", fabs(r1_data[i] - r2_data[i]));
+                //}
+            }
+
+            //for (int i = 4096; i < 4096 + 128; ++i) {
+            //    printf("%f %f\n", r1_data[i], r2_data[i]);
+            //}
+
+            printf("mode: %d\n", mode);
+            printf("sum0: %f\n", sum0);
+            printf("sum1: %f\n", sum1);
+            printf("diff: %f\n", diff);
+            printf("rel err: %f\n", diff / sum0);
+            printf("rel err: %f\n", diff / sum1);
+
+            GGML_ASSERT(diff / sum0 < 0.0001f);
+            GGML_ASSERT(diff / sum1 < 0.0001f);
+        }
+    }
+
+    ggml_free(ctx0);
+
+    return 0;
+}
+

tests/test-sampling.cpp

@@ -12,7 +12,8 @@
 #include <vector>
 #include <algorithm>
 
-void dump(const llama_token_data_array * candidates) {
+
+static void dump(const llama_token_data_array * candidates) {
     for (size_t i = 0; i < candidates->size; i++) {
         printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit);
     }
@@ -21,9 +22,7 @@ void dump(const llama_token_data_array * candidates) {
 #define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0)
 
 
-void test_top_k(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                int k) {
+static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -45,10 +44,7 @@ void test_top_k(const std::vector<float> & probs,
 }
 
 
-void test_top_p(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                float p) {
-
+static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -70,9 +66,7 @@ void test_top_p(const std::vector<float> & probs,
 }
 
 
-void test_tfs(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                float z) {
+static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -93,9 +87,7 @@ void test_tfs(const std::vector<float> & probs,
 }
 
 
-void test_typical(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                float p) {
+static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -116,11 +108,10 @@ void test_typical(const std::vector<float> & probs,
 }
 
 
-void test_repetition_penalty(
-                const std::vector<float> & probs,
-                const std::vector<llama_token> & last_tokens,
-                const std::vector<float> & expected_probs,
-                float penalty) {
+static void test_repetition_penalty(
+    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
+    const std::vector<float> & expected_probs, float penalty
+) {
     assert(probs.size() == expected_probs.size());
 
     size_t n_vocab = probs.size();
@@ -145,11 +136,10 @@ void test_repetition_penalty(
 }
 
 
-void test_frequency_presence_penalty(
-                const std::vector<float> & probs,
-                const std::vector<llama_token> & last_tokens,
-                const std::vector<float> & expected_probs,
-                float alpha_frequency, float alpha_presence) {
+static void test_frequency_presence_penalty(
+    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
+    const std::vector<float> & expected_probs, float alpha_frequency, float alpha_presence
+) {
     assert(probs.size() == expected_probs.size());
 
     size_t n_vocab = probs.size();

tests/test-tokenizer-0-llama.cpp

@@ -1,5 +1,6 @@
 #include "llama.h"
 #include "common.h"
+#include "console.h"
 
 #include <cstdio>
 #include <string>
@@ -35,6 +36,7 @@ static const std::map<std::string, std::vector<llama_token>> & k_tests() {
         { "   Hello"              , {    1678,  15043, }, },
         { "    Hello"             , {     268,  15043, }, },
         { "    Hello\n    Hello"  , {     268,  15043,     13,   1678,  15043, }, },
+        { " ("                    , {   29871,  313, }, },
     };
 
     return _k_tests;
@@ -89,6 +91,12 @@ int main(int argc, char **argv) {
         return 2;
     }
 
+#ifdef _WIN32
+    // We need this for unicode console support
+    console::init(false, false);
+    atexit([]() { console::cleanup(); });
+#endif
+
     bool success = true;
 
     for (const auto & test_kv : k_tests()) {