ggml : fix arch check in bf16_to_fp32 (#10164)

* q6_k instruction reordering attempt

* better subtract method

* should be theoretically faster

small improvement with shuffle lut, likely because all loads are already done at that stage

* optimize bit fiddling

* handle -32 offset separately. bsums exists for a reason!

* use shift

* Update ggml-quants.c

* have to update ci macos version to 13 as 12 doesnt work now. 13 is still x86

.devops/full-musa.Dockerfile

@@ -0,0 +1,26 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release -j$(nproc) && \
+    cp build/bin/* .
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full-rocm.Dockerfile

@@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
+ARG ROCM_DOCKER_ARCH="\
     gfx803 \
     gfx900 \
     gfx906 \
@@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1030 \
     gfx1100 \
     gfx1101 \
-    gfx1102
+    gfx1102"
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -34,7 +34,7 @@ WORKDIR /app
 COPY . .
 
 # Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang

.devops/llama-cli-musa.Dockerfile

@@ -0,0 +1,30 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc)
+
+FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libgomp1
+
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-rocm.Dockerfile

@@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
+ARG ROCM_DOCKER_ARCH="\
     gfx803 \
     gfx900 \
     gfx906 \
@@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1030 \
     gfx1100 \
     gfx1101 \
-    gfx1102
+    gfx1102"
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -34,7 +34,7 @@ WORKDIR /app
 COPY . .
 
 # Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang

.devops/llama-server-musa.Dockerfile

@@ -0,0 +1,35 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-server -j$(nproc)
+
+FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-rocm.Dockerfile

@@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
+ARG ROCM_DOCKER_ARCH="\
     gfx803 \
     gfx900 \
     gfx906 \
@@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1030 \
     gfx1100 \
     gfx1101 \
-    gfx1102
+    gfx1102"
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -34,7 +34,7 @@ WORKDIR /app
 COPY . .
 
 # Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang

.dockerignore

@@ -1,7 +1,7 @@
 *.o
 *.a
 .cache/
-.git/
+# Do not ignore .git directory, otherwise the reported build number will always be 0
 .github/
 .gitignore
 .vs/

.github/workflows/bench.yml.disabled

@@ -27,10 +27,10 @@ on:
   push:
     branches:
       - master
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
   pull_request_target:
     types: [opened, synchronize, reopened]
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
   schedule:
     -  cron: '04 2 * * *'
 

.github/workflows/build.yml

@@ -19,6 +19,11 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  contents: write # for creating release
+
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
@@ -87,7 +92,7 @@ jobs:
           name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
-    runs-on: macos-12
+    runs-on: macos-13
 
     steps:
       - name: Clone
@@ -956,6 +961,7 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
@@ -967,6 +973,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-latest-cmake-hip:
+    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-latest
 
     steps:
@@ -994,8 +1001,72 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
-          cmake --build build --config Release
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+
+  windows-latest-cmake-hip-release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        gpu_target: [gfx1100, gfx1101, gfx1030]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+          md "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -1060,6 +1131,7 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cuda
+      - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
 

.github/workflows/close-issue.yml

@@ -3,6 +3,11 @@ on:
   schedule:
     - cron: "42 0 * * *"
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  issues: write
+
 jobs:
   close-issues:
     runs-on: ubuntu-latest

.github/workflows/docker.yml

@@ -15,11 +15,17 @@ on:
     branches:
       - master
     paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+  workflow_dispatch: # allows manual triggering, useful for debugging
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  packages: write
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
@@ -37,6 +43,9 @@ jobs:
           - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-musa", dockerfile: ".devops/llama-cli-musa.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-musa", dockerfile: ".devops/llama-server-musa.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-musa", dockerfile: ".devops/full-musa.Dockerfile", platforms: "linux/amd64" }
           # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
           #- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           #- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
@@ -46,6 +55,8 @@ jobs:
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
@@ -60,6 +71,34 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+
+          # determine tag name postfix (build number, commit hash)
+          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
+            TAG_POSTFIX="b${BUILD_NUMBER}"
+          else
+            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
+            TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
+          fi
+
+          # list all tags possible
+          TAGS=""
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
+
+          echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
+          echo "output_tags=$TAGS"  # print out for debugging
+        env:
+          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
       # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
       - name: Free Disk Space (Ubuntu)
         uses: jlumbroso/free-disk-space@main
@@ -77,25 +116,6 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Downcase github.repository_owner
-        run: |
-          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
       - name: Build and push Docker image (tagged + versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v6
@@ -103,5 +123,6 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          # tag list is generated from step above
+          tags: ${{ steps.tag.outputs.output_tags }}
           file: ${{ matrix.config.dockerfile }}

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

@@ -21,6 +21,13 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -12,6 +12,13 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-eval:
     strategy:

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

@@ -4,11 +4,13 @@ on:
   push:
     paths:
       - '.github/workflows/python-type-check.yml'
+      - 'pyrightconfig.json'
       - '**.py'
       - '**/requirements*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-type-check.yml'
+      - 'pyrightconfig.json'
       - '**.py'
       - '**/requirements*.txt'
 
@@ -33,6 +35,6 @@ jobs:
       - name: Type-check with Pyright
         uses: jakebailey/pyright-action@v2
         with:
-          version: 1.1.370
+          version: 1.1.382
           level: warning
           warnings: true

CMakeLists.txt

@@ -62,6 +62,9 @@ option(LLAMA_SANITIZE_THREAD    "llama: enable thread sanitizer"    OFF)
 option(LLAMA_SANITIZE_ADDRESS   "llama: enable address sanitizer"   OFF)
 option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 
+# utils
+option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})
+
 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
@@ -85,6 +88,10 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
+if (NOT DEFINED GGML_AMX)
+    set(GGML_AMX ON)
+endif()
+
 if (NOT DEFINED GGML_CUDA_GRAPHS)
     set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()
@@ -191,17 +198,19 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION lib/pkgconfig)
 
 #
-# programs, examples and tests
+# utils, programs, examples and tests
 #
 
-add_subdirectory(common)
+if (LLAMA_BUILD_COMMON)
+    add_subdirectory(common)
+endif()
 
-if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     include(CTest)
     add_subdirectory(tests)
-endif ()
+endif()
 
-if (LLAMA_BUILD_EXAMPLES)
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)
     add_subdirectory(pocs)
 endif()

CMakePresets.json

@@ -48,10 +48,23 @@
         }
     },
 
+    {
+        "name": "arm64-apple-clang", "hidden": true,
+        "architecture": { "value": "arm64",    "strategy": "external" },
+        "toolset":      { "value": "host=x64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
+        }
+    },
+
     { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
     { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg" ] },
     { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg", "static" ] },
 
+    { "name": "arm64-apple-clang-debug"  , "inherits": [ "base", "arm64-apple-clang",  "debug"   ] },
+    { "name": "arm64-apple-clang-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg"   ] },
+    { "name": "arm64-apple-clang+static-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
+
     { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },

CONTRIBUTING.md

@@ -1,24 +1,23 @@
 # Pull requests (for contributors)
 
 - Test your changes:
-  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
+  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the `ggml` library
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
-- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
-  - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
-- Consider allowing write access to your branch for faster review
+- Optionally rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs
+- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)
 
 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
-- Optionally, pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
+- Optionally pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
 
 # Coding guidelines
 
 - Avoid adding third-party dependencies, extra files, extra headers, etc.
 - Always consider cross-compatibility with other operating systems and architectures
-- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
+- Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
@@ -27,3 +26,8 @@
 
 ![matmul](media/matmul.png)
 
+# Resources
+
+The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
+
+https://github.com/ggerganov/llama.cpp/projects

Makefile

@@ -1,11 +1,9 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	libllava.a \
-	llama-baby-llama \
 	llama-batched \
 	llama-batched-bench \
 	llama-bench \
-	llama-benchmark-matmult \
 	llama-cli \
 	llama-convert-llama2c-to-ggml \
 	llama-embedding \
@@ -35,6 +33,7 @@ BUILD_TARGETS = \
 	llama-save-load-state \
 	llama-server \
 	llama-simple \
+	llama-simple-chat \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -56,7 +55,6 @@ TEST_TARGETS = \
 	tests/test-llama-grammar \
 	tests/test-log \
 	tests/test-model-load-cancel \
-	tests/test-opt \
 	tests/test-quantize-fns \
 	tests/test-quantize-perf \
 	tests/test-rope \
@@ -64,11 +62,12 @@ TEST_TARGETS = \
 	tests/test-tokenizer-0 \
 	tests/test-tokenizer-1-bpe \
 	tests/test-tokenizer-1-spm
+#	tests/test-opt \
 
 # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
 LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
-	retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm
+	retrieval speculative infill tokenize parallel export-lora lookahead lookup passkey gritlm
 
 # Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
 #  We don't want to clutter things too much, so we only build replacements for the most commonly used binaries.
@@ -94,11 +93,6 @@ GGML_METAL := 1
 DEPRECATE_WARNING := 1
 endif
 
-ifdef LLAMA_OPENMP
-GGML_OPENMP := 1
-DEPRECATE_WARNING := 1
-endif
-
 ifdef LLAMA_RPC
 GGML_RPC := 1
 DEPRECATE_WARNING := 1
@@ -585,6 +579,11 @@ ifndef GGML_NO_LLAMAFILE
 	OBJ_GGML    += ggml/src/llamafile/sgemm.o
 endif
 
+ifndef GGML_NO_AMX
+	MK_CPPFLAGS += -DGGML_USE_AMX
+	OBJ_GGML    += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o
+endif
+
 ifdef GGML_RPC
 	MK_CPPFLAGS += -DGGML_USE_RPC
 	OBJ_GGML    += ggml/src/ggml-rpc.o
@@ -611,7 +610,7 @@ ifdef GGML_CUDA
 
 		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
 		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
-		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
 	else
 		ifneq ('', '$(wildcard /opt/cuda)')
 			CUDA_PATH ?= /opt/cuda
@@ -916,6 +915,7 @@ endif # GGML_METAL
 
 OBJ_GGML += \
 	ggml/src/ggml.o \
+	ggml/src/ggml-cpu.o \
 	ggml/src/ggml-alloc.o \
 	ggml/src/ggml-backend.o \
 	ggml/src/ggml-quants.o \
@@ -936,7 +936,6 @@ OBJ_COMMON = \
 	common/console.o \
 	common/ngram-cache.o \
 	common/sampling.o \
-	common/train.o \
 	common/build-info.o \
 	common/json-schema-to-grammar.o
 
@@ -1048,6 +1047,12 @@ ggml/src/ggml.o: \
 	ggml/include/ggml.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
+ggml/src/ggml-cpu.o: \
+	ggml/src/ggml-cpu.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-common.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
 ggml/src/ggml-alloc.o: \
 	ggml/src/ggml-alloc.c \
 	ggml/include/ggml.h \
@@ -1055,10 +1060,11 @@ ggml/src/ggml-alloc.o: \
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
 ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.c \
+	ggml/src/ggml-backend.cpp \
+	ggml/src/ggml-backend-impl.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-backend.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ggml/src/ggml-quants.o: \
 	ggml/src/ggml-quants.c \
@@ -1087,6 +1093,19 @@ ggml/src/llamafile/sgemm.o: \
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # GGML_NO_LLAMAFILE
 
+ifndef GGML_NO_AMX
+ggml/src/ggml-amx.o: \
+	ggml/src/ggml-amx.cpp \
+	ggml/include/ggml-amx.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+ggml/src/ggml-amx/mmq.o: \
+	ggml/src/ggml-amx/mmq.cpp \
+	ggml/src/ggml-amx/mmq.h \
+	ggml/include/ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif
+
 ifdef GGML_RPC
 ggml/src/ggml-rpc.o: \
 	ggml/src/ggml-rpc.cpp \
@@ -1199,11 +1218,6 @@ common/json-schema-to-grammar.o: \
 	common/json-schema-to-grammar.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common/train.o: \
-	common/train.cpp \
-	common/train.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 common/ngram-cache.o: \
 	common/ngram-cache.cpp \
 	common/ngram-cache.h
@@ -1238,6 +1252,7 @@ clean:
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o
 	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
+	rm -vrf ggml/src/ggml-amx/*.o
 	rm -rvf $(BUILD_TARGETS)
 	rm -rvf $(TEST_TARGETS)
 	rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
@@ -1273,6 +1288,11 @@ llama-simple: examples/simple/simple.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-simple-chat: examples/simple-chat/simple-chat.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-tokenize: examples/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1370,11 +1390,6 @@ llama-bench: examples/llama-bench/llama-bench.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-baby-llama: examples/baby-llama/baby-llama.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 llama-export-lora: examples/export-lora/export-lora.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1523,16 +1538,6 @@ common/build-info.o: common/build-info.cpp
 
 tests: $(TEST_TARGETS)
 
-llama-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp \
-	$(OBJ_GGML) common/build-info.o
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-run-benchmark-matmult: llama-benchmark-matmult
-	./$@
-
-.PHONY: run-benchmark-matmult swift
-
 tests/test-arg-parser: tests/test-arg-parser.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

Package.swift

@@ -10,8 +10,9 @@ var sources = [
     "src/unicode.cpp",
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
+    "ggml/src/ggml-cpu.c",
     "ggml/src/ggml-alloc.c",
-    "ggml/src/ggml-backend.c",
+    "ggml/src/ggml-backend.cpp",
     "ggml/src/ggml-quants.c",
     "ggml/src/ggml-aarch64.c",
 ]

README.md

@@ -17,7 +17,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
+- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
+- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----
 
@@ -28,9 +30,9 @@ variety of hardware - locally and in the cloud.
 
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
-- AVX, AVX2 and AVX512 support for x86 architectures
+- AVX, AVX2, AVX512 and AMX support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
-- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
@@ -91,6 +93,8 @@ Typically finetunes of the base models below are supported as well.
 - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
+- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
+- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -112,6 +116,7 @@ Typically finetunes of the base models below are supported as well.
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
+- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
@@ -119,6 +124,7 @@ Typically finetunes of the base models below are supported as well.
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
+- C#/VB.NET (more features - community license): [LM-Kit.NET](https://docs.lm-kit.com/lm-kit-net/index.html)
 - 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)
@@ -127,6 +133,8 @@ Typically finetunes of the base models below are supported as well.
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 - PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
 - Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
+- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
+- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 
 **UI:**
 
@@ -166,12 +174,15 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
 - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
+- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [PocketPal AI - An iOS and Android App](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 - [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)
@@ -180,6 +191,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
+- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 
 **Games:**
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
@@ -408,7 +420,7 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [BLAS](./docs/build.md#blas-build) | All |
 | [BLIS](./docs/backend/BLIS.md) | All |
 | [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
+| [MUSA](./docs/build.md#musa) | Moore Threads MTT GPU |
 | [CUDA](./docs/build.md#cuda) | Nvidia GPU |
 | [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
 | [Vulkan](./docs/build.md#vulkan) | GPU |
@@ -440,7 +452,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - Contributors can open PRs
 - Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
-- Any help with managing issues and PRs is very appreciated!
+- Any help with managing issues, PRs and projects is very appreciated!
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
 - Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)

ci/run.sh

@@ -1,4 +1,4 @@
-#/bin/bash
+#!/bin/bash
 #
 # sample usage:
 #
@@ -53,7 +53,7 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
@@ -326,36 +326,36 @@ function gg_run_open_llama_7b_v2 {
     ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -460,34 +460,34 @@ function gg_run_pythia_1_4b {
     ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli --model ${model_f16}  -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -591,36 +591,36 @@ function gg_run_pythia_2_8b {
     ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -706,12 +706,88 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
 
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
+# rerank_tiny
+
+function gg_run_rerank_tiny {
+    cd ${SRC}
+
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+
+    gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/rerank-tiny"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+
+    # for this model, the SEP token is "</s>"
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+
+    # sample output
+    # rerank score 0:    0.029
+    # rerank score 1:    0.029
+    # rerank score 2:    0.135
+
+    # check that the score is in the range [$3, $4]
+    function check_score {
+        qnt="$1"
+        score=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$score < $3" | bc) -eq 1 ] || [ $(echo "$score > $4" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: score not in range [%s, %s])\n' "$qnt" "$score" "$3" "$4"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$score"
+        return 0
+    }
+
+    check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log
+
+    set +e
+}
+
+function gg_sum_rerank_tiny {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Rerank Tiny (Jina):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-rk-f16.log)"
+}
+
 function gg_check_build_requirements {
     if ! command -v cmake &> /dev/null; then
         gg_printf 'cmake not found, please install'
@@ -726,15 +802,6 @@ function gg_check_build_requirements {
     fi
 }
 
-function gg_sum_embd_bge_small {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'BGE Small (BERT):\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-}
-
 ## main
 
 export LLAMA_LOG_PREFIX=1
@@ -762,6 +829,7 @@ test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
+    test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
         test $ret -eq 0 && gg_run test_scripts_debug

cmake/arm64-apple-clang.cmake

@@ -0,0 +1,16 @@
+set( CMAKE_SYSTEM_NAME Darwin )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-apple-darwin-macho )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
+
+set( arch_c_flags "-march=armv8.4-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

common/CMakeLists.txt

@@ -66,8 +66,6 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     sampling.cpp
     sampling.h
-    train.cpp
-    train.h
     )
 
 if (BUILD_SHARED_LIBS)

common/arg.h

@@ -10,7 +10,7 @@
 // CLI argument parsing
 //
 
-struct llama_arg {
+struct common_arg {
     std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
     std::vector<const char *> args;
     const char * value_hint   = nullptr; // help text or example for arg value
@@ -18,60 +18,60 @@ struct llama_arg {
     const char * env          = nullptr;
     std::string help;
     bool is_sparam = false; // is current arg a sampling param?
-    void (*handler_void)   (gpt_params & params) = nullptr;
-    void (*handler_string) (gpt_params & params, const std::string &) = nullptr;
-    void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr;
-    void (*handler_int)    (gpt_params & params, int) = nullptr;
+    void (*handler_void)   (common_params & params) = nullptr;
+    void (*handler_string) (common_params & params, const std::string &) = nullptr;
+    void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
+    void (*handler_int)    (common_params & params, int) = nullptr;
 
-    llama_arg(
+    common_arg(
         const std::initializer_list<const char *> & args,
         const char * value_hint,
         const std::string & help,
-        void (*handler)(gpt_params & params, const std::string &)
+        void (*handler)(common_params & params, const std::string &)
     ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
 
-    llama_arg(
+    common_arg(
         const std::initializer_list<const char *> & args,
         const char * value_hint,
         const std::string & help,
-        void (*handler)(gpt_params & params, int)
+        void (*handler)(common_params & params, int)
     ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
 
-    llama_arg(
+    common_arg(
         const std::initializer_list<const char *> & args,
         const std::string & help,
-        void (*handler)(gpt_params & params)
+        void (*handler)(common_params & params)
     ) : args(args), help(help), handler_void(handler) {}
 
     // support 2 values for arg
-    llama_arg(
+    common_arg(
         const std::initializer_list<const char *> & args,
         const char * value_hint,
         const char * value_hint_2,
         const std::string & help,
-        void (*handler)(gpt_params & params, const std::string &, const std::string &)
+        void (*handler)(common_params & params, const std::string &, const std::string &)
     ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
 
-    llama_arg & set_examples(std::initializer_list<enum llama_example> examples);
-    llama_arg & set_env(const char * env);
-    llama_arg & set_sparam();
+    common_arg & set_examples(std::initializer_list<enum llama_example> examples);
+    common_arg & set_env(const char * env);
+    common_arg & set_sparam();
     bool in_example(enum llama_example ex);
     bool get_value_from_env(std::string & output);
     bool has_value_from_env();
     std::string to_string();
 };
 
-struct gpt_params_context {
+struct common_params_context {
     enum llama_example ex = LLAMA_EXAMPLE_COMMON;
-    gpt_params & params;
-    std::vector<llama_arg> options;
+    common_params & params;
+    std::vector<common_arg> options;
     void(*print_usage)(int, char **) = nullptr;
-    gpt_params_context(gpt_params & params) : params(params) {}
+    common_params_context(common_params & params) : params(params) {}
 };
 
 // parse input arguments from CLI
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
-bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
 // function to be used by test-arg-parser
-gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/common.cpp

@@ -12,6 +12,7 @@
 
 #include <algorithm>
 #include <cinttypes>
+#include <climits>
 #include <cmath>
 #include <codecvt>
 #include <cstdarg>
@@ -23,10 +24,10 @@
 #include <regex>
 #include <sstream>
 #include <string>
+#include <thread>
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
-#include <thread>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -362,10 +363,10 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
     return true;
 }
 
-void gpt_init() {
+void common_init() {
     llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
-        if (LOG_DEFAULT_LLAMA <= gpt_log_verbosity_thold) {
-            gpt_log_add(gpt_log_main(), level, "%s", text);
+        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+            common_log_add(common_log_main(), level, "%s", text);
         }
     }, NULL);
 
@@ -378,7 +379,7 @@ void gpt_init() {
     LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
 }
 
-std::string gpt_params_get_system_info(const gpt_params & params) {
+std::string common_params_get_system_info(const common_params & params) {
     std::ostringstream os;
 
     os << "system_info: n_threads = " << params.cpuparams.n_threads;
@@ -400,17 +401,19 @@ std::string gpt_params_get_system_info(const gpt_params & params) {
 // String utils
 //
 
-std::vector<std::string> string_split(std::string input, char separator) {
-    std::vector<std::string> parts;
-    size_t separator_pos = input.find(separator);
-    while (separator_pos != std::string::npos) {
-        std::string part = input.substr(0, separator_pos);
-        parts.emplace_back(part);
-        input = input.substr(separator_pos + 1);
-        separator_pos = input.find(separator);
-    }
-    parts.emplace_back(input);
-    return parts;
+std::string string_format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    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);
 }
 
 std::string string_strip(const std::string & str) {
@@ -493,7 +496,7 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
             first = false;
         }
 
-        auto detokenized = llama_token_to_piece(ctx, token);
+        auto detokenized = common_token_to_piece(ctx, token);
 
         detokenized.erase(
             std::remove_if(
@@ -524,7 +527,7 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
             first = false;
         }
 
-        auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
+        auto detokenized = common_token_to_piece(ctx, batch.token[i]);
 
         detokenized.erase(
                 std::remove_if(
@@ -819,16 +822,16 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
-    llama_init_result iparams;
-    auto mparams = llama_model_params_from_gpt_params(params);
+struct common_init_result common_init_from_params(common_params & params) {
+    common_init_result iparams;
+    auto mparams = common_model_params_to_llama(params);
 
     llama_model * model = nullptr;
 
     if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
+        model = common_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
     } else if (!params.model_url.empty()) {
-        model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
+        model = common_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
     } else {
         model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
@@ -838,7 +841,32 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         return iparams;
     }
 
-    auto cparams = llama_context_params_from_gpt_params(params);
+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free_model(model);
+
+            return iparams;
+        }
+    }
+
+    auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
@@ -851,10 +879,11 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
         if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
 
-        const auto cvec = llama_control_vector_load(params.control_vectors);
+        const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
+
             return iparams;
         }
 
@@ -867,13 +896,14 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
+
             return iparams;
         }
     }
 
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
-        llama_lora_adapter_container loaded_la;
+        common_lora_adapter_container loaded_la;
         loaded_la.path = la.path;
         loaded_la.scale = la.scale;
         loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
@@ -886,10 +916,10 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
     }
     if (!params.lora_init_without_apply) {
-        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
+        common_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
-    if (params.sparams.ignore_eos && llama_token_eos(model) == -1) {
+    if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
         LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sparams.ignore_eos = false;
     }
@@ -912,7 +942,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         }
 
         if (llama_model_has_encoder(model)) {
-            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
+            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
             if (decoder_start_token_id == -1) {
                 decoder_start_token_id = bos;
@@ -921,7 +951,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
             tmp.push_back(decoder_start_token_id);
         }
         if (llama_model_has_decoder(model)) {
-            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
@@ -930,10 +960,11 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
 
     iparams.model   = model;
     iparams.context = lctx;
+
     return iparams;
 }
 
-void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) {
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters) {
     llama_lora_adapter_clear(ctx);
     for (auto & la : lora_adapters) {
         if (la.scale != 0.0f) {
@@ -942,7 +973,7 @@ void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lor
     }
 }
 
-struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
+struct llama_model_params common_model_params_to_llama(const common_params & params) {
     auto mparams = llama_model_default_params();
 
     if (params.n_gpu_layers != -1) {
@@ -991,10 +1022,10 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
         return GGML_TYPE_Q5_1;
     }
 
-    throw std::runtime_error("Invalid cache type: " + s);
+    throw std::runtime_error("Unsupported cache type: " + s);
 }
 
-struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
+struct llama_context_params common_context_params_to_llama(const common_params & params) {
     auto cparams = llama_context_default_params();
 
     cparams.n_ctx             = params.n_ctx;
@@ -1003,7 +1034,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.cpuparams.n_threads;
     cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
-                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+                                params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
     cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
@@ -1023,6 +1054,11 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
 
+    if (params.reranking) {
+        cparams.embeddings    = true;
+        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
+    }
+
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
 
@@ -1079,7 +1115,7 @@ static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_
     return false;
 }
 
-static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
+static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
 
     // Initialize libcurl
     std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
@@ -1149,15 +1185,15 @@ static bool llama_download_file(const std::string & url, const std::string & pat
     }
 
     // Send a HEAD request to retrieve the etag and last-modified headers
-    struct llama_load_model_from_url_headers {
+    struct common_load_model_from_url_headers {
         std::string etag;
         std::string last_modified;
     };
-    llama_load_model_from_url_headers headers;
+    common_load_model_from_url_headers headers;
     {
         typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            llama_load_model_from_url_headers *headers = (llama_load_model_from_url_headers *) userdata;
+            common_load_model_from_url_headers *headers = (common_load_model_from_url_headers *) userdata;
 
             static std::regex header_regex("([^:]+): (.*)\r\n");
             static std::regex etag_regex("ETag", std::regex_constants::icase);
@@ -1293,7 +1329,7 @@ static bool llama_download_file(const std::string & url, const std::string & pat
     return true;
 }
 
-struct llama_model * llama_load_model_from_url(
+struct llama_model * common_load_model_from_url(
         const char * model_url,
         const char * path_model,
         const char * hf_token,
@@ -1304,7 +1340,7 @@ struct llama_model * llama_load_model_from_url(
         return NULL;
     }
 
-    if (!llama_download_file(model_url, path_model, hf_token)) {
+    if (!common_download_file(model_url, path_model, hf_token)) {
         return NULL;
     }
 
@@ -1357,7 +1393,7 @@ struct llama_model * llama_load_model_from_url(
                 char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
                 llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
 
-                return llama_download_file(split_url, split_path, hf_token);
+                return common_download_file(split_url, split_path, hf_token);
             }, idx));
         }
 
@@ -1372,7 +1408,7 @@ struct llama_model * llama_load_model_from_url(
     return llama_load_model_from_file(path_model, params);
 }
 
-struct llama_model * llama_load_model_from_hf(
+struct llama_model * common_load_model_from_hf(
         const char * repo,
         const char * model,
         const char * path_model,
@@ -1392,12 +1428,12 @@ struct llama_model * llama_load_model_from_hf(
     model_url += "/resolve/main/";
     model_url += model;
 
-    return llama_load_model_from_url(model_url.c_str(), path_model, hf_token, params);
+    return common_load_model_from_url(model_url.c_str(), path_model, hf_token, params);
 }
 
 #else
 
-struct llama_model * llama_load_model_from_url(
+struct llama_model * common_load_model_from_url(
         const char * /*model_url*/,
         const char * /*path_model*/,
         const char * /*hf_token*/,
@@ -1406,7 +1442,7 @@ struct llama_model * llama_load_model_from_url(
     return nullptr;
 }
 
-struct llama_model * llama_load_model_from_hf(
+struct llama_model * common_load_model_from_hf(
         const char * /*repo*/,
         const char * /*model*/,
         const char * /*path_model*/,
@@ -1422,16 +1458,18 @@ struct llama_model * llama_load_model_from_hf(
 // Batch utils
 //
 
-void llama_batch_clear(struct llama_batch & batch) {
+void common_batch_clear(struct llama_batch & batch) {
     batch.n_tokens = 0;
 }
 
-void llama_batch_add(
+void common_batch_add(
                  struct llama_batch & batch,
                         llama_token   id,
                           llama_pos   pos,
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits) {
+    GGML_ASSERT(batch.seq_id[batch.n_tokens] && "llama_batch size exceeded");
+
     batch.token   [batch.n_tokens] = id;
     batch.pos     [batch.n_tokens] = pos;
     batch.n_seq_id[batch.n_tokens] = seq_ids.size();
@@ -1447,15 +1485,15 @@ void llama_batch_add(
 // Vocab utils
 //
 
-std::vector<llama_token> llama_tokenize(
+std::vector<llama_token> common_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
-    return llama_tokenize(llama_get_model(ctx), text, add_special, parse_special);
+    return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
 }
 
-std::vector<llama_token> llama_tokenize(
+std::vector<llama_token> common_tokenize(
     const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
@@ -1474,7 +1512,7 @@ std::vector<llama_token> llama_tokenize(
     return result;
 }
 
-std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
+std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
     std::string piece;
     piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
     const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
@@ -1490,7 +1528,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
     return piece;
 }
 
-std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
+std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
     std::string text;
     text.resize(std::max(text.capacity(), tokens.size()));
     int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
@@ -1510,15 +1548,15 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
 // Chat template utils
 //
 
-bool llama_chat_verify_template(const std::string & tmpl) {
+bool common_chat_verify_template(const std::string & tmpl) {
     llama_chat_message chat[] = {{"user", "test"}};
     int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
     return res >= 0;
 }
 
-std::string llama_chat_apply_template(const struct llama_model * model,
+std::string common_chat_apply_template(const struct llama_model * model,
         const std::string & tmpl,
-        const std::vector<llama_chat_msg> & msgs,
+        const std::vector<common_chat_msg> & msgs,
         bool add_ass) {
     int alloc_size = 0;
     bool fallback = false; // indicate if we must fallback to default chatml
@@ -1560,42 +1598,42 @@ std::string llama_chat_apply_template(const struct llama_model * model,
     return formatted_chat;
 }
 
-std::string llama_chat_format_single(const struct llama_model * model,
+std::string common_chat_format_single(const struct llama_model * model,
         const std::string & tmpl,
-        const std::vector<llama_chat_msg> & past_msg,
-        const llama_chat_msg & new_msg,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
         bool add_ass) {
     std::ostringstream ss;
-    auto fmt_past_msg = past_msg.empty() ? "" : llama_chat_apply_template(model, tmpl, past_msg, false);
-    std::vector<llama_chat_msg> chat_new(past_msg);
+    auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(model, tmpl, past_msg, false);
+    std::vector<common_chat_msg> chat_new(past_msg);
     // if the past_msg ends with a newline, we must preserve it in the formatted version
     if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
         ss << "\n";
     };
     // format chat with new_msg
     chat_new.push_back(new_msg);
-    auto fmt_new_msg = llama_chat_apply_template(model, tmpl, chat_new, add_ass);
+    auto fmt_new_msg = common_chat_apply_template(model, tmpl, chat_new, add_ass);
     // get the diff part
     ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
     return ss.str();
 }
 
-std::string llama_chat_format_example(const struct llama_model * model,
+std::string common_chat_format_example(const struct llama_model * model,
         const std::string & tmpl) {
-    std::vector<llama_chat_msg> msgs = {
+    std::vector<common_chat_msg> msgs = {
         {"system",    "You are a helpful assistant"},
         {"user",      "Hello"},
         {"assistant", "Hi there"},
         {"user",      "How are you?"},
     };
-    return llama_chat_apply_template(model, tmpl, msgs, true);
+    return common_chat_apply_template(model, tmpl, msgs, true);
 }
 
 //
 // KV cache utils
 //
 
-void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
+void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
     static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
 
     printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
@@ -1618,7 +1656,7 @@ void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
     printf("\n=== Done dumping\n");
 }
 
-void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
+void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
     static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
 
     printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
@@ -1670,7 +1708,7 @@ void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_siz
 // Embedding utils
 //
 
-void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm) {
+void common_embd_normalize(const float * inp, float * out, int n, int embd_norm) {
     double sum = 0.0;
 
     switch (embd_norm) {
@@ -1704,7 +1742,7 @@ void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm)
     }
 }
 
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){
+float common_embd_similarity_cos(const float * embd1, const float * embd2, int n){
     double sum  = 0.0;
     double sum1 = 0.0;
     double sum2 = 0.0;
@@ -1730,8 +1768,8 @@ float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n)
 // Control vector utils
 //
 
-static llama_control_vector_data llama_control_vector_load_one(const llama_control_vector_load_info & load_info) {
-    llama_control_vector_data result = { -1, {} };
+static common_control_vector_data common_control_vector_load_one(const common_control_vector_load_info & load_info) {
+    common_control_vector_data result = { -1, {} };
 
     ggml_context * ctx = nullptr;
     struct gguf_init_params meta_gguf_params = {
@@ -1815,11 +1853,11 @@ static llama_control_vector_data llama_control_vector_load_one(const llama_contr
     return result;
 }
 
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos) {
-    llama_control_vector_data result = { -1, {} };
+common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos) {
+    common_control_vector_data result = { -1, {} };
 
     for (const auto & info : load_infos) {
-        auto cur = llama_control_vector_load_one(info);
+        auto cur = common_control_vector_load_one(info);
 
         if (cur.n_embd == -1) {
             result.n_embd = -1;
@@ -1911,8 +1949,10 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
     }
 }
 
-void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const llama_context * lctx,
+void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    ggml_cpu_init(); // some ARM features are detected at runtime
+
     const auto & sparams = params.sparams;
 
     fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
@@ -1968,6 +2008,10 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
+    fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
+    fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
+    fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
     fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
@@ -2048,11 +2092,12 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
     yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
 
-    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
     fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
+    fprintf(stream, "xtc_probability: %f # default: 0.0\n", sparams.xtc_probability);
+    fprintf(stream, "xtc_threshold: %f # default: 0.1\n", sparams.xtc_threshold);
     fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
     fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");

common/common.h

@@ -24,12 +24,12 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct llama_lora_adapter_info {
+struct common_lora_adapter_info {
     std::string path;
     float scale;
 };
 
-struct llama_lora_adapter_container : llama_lora_adapter_info {
+struct common_lora_adapter_container : common_lora_adapter_info {
     struct llama_lora_adapter * adapter;
 };
 
@@ -39,7 +39,7 @@ extern char const * LLAMA_COMMIT;
 extern char const * LLAMA_COMPILER;
 extern char const * LLAMA_BUILD_TARGET;
 
-struct llama_control_vector_load_info;
+struct common_control_vector_load_info;
 
 //
 // CPU utils
@@ -82,14 +82,17 @@ enum llama_example {
     LLAMA_EXAMPLE_COUNT,
 };
 
-enum gpt_sampler_type {
-    GPT_SAMPLER_TYPE_NONE        = 0,
-    GPT_SAMPLER_TYPE_TOP_K       = 1,
-    GPT_SAMPLER_TYPE_TOP_P       = 2,
-    GPT_SAMPLER_TYPE_MIN_P       = 3,
-    GPT_SAMPLER_TYPE_TFS_Z       = 4,
-    GPT_SAMPLER_TYPE_TYPICAL_P   = 5,
-    GPT_SAMPLER_TYPE_TEMPERATURE = 6,
+enum common_sampler_type {
+    COMMON_SAMPLER_TYPE_NONE        = 0,
+    COMMON_SAMPLER_TYPE_DRY         = 1,
+    COMMON_SAMPLER_TYPE_TOP_K       = 2,
+    COMMON_SAMPLER_TYPE_TOP_P       = 3,
+    COMMON_SAMPLER_TYPE_MIN_P       = 4,
+  //COMMON_SAMPLER_TYPE_TFS_Z       = 5,
+    COMMON_SAMPLER_TYPE_TYPICAL_P   = 6,
+    COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
+    COMMON_SAMPLER_TYPE_XTC         = 8,
+    COMMON_SAMPLER_TYPE_INFILL      = 9,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -99,38 +102,47 @@ enum dimre_method {
 };
 
 // sampler parameters
-struct gpt_sampler_params {
+struct common_sampler_params {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep          = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   min_p             = 0.05f; // 0.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
-    float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range    = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.00f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = false; // consider newlines as a repeatable token
-    bool    ignore_eos        = false;
-    bool    no_perf           = false; // disable performance metrics
+    int32_t n_prev             = 64;    // number of previous tokens to remember
+    int32_t n_probs            = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep           = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k              = 40;    // <= 0 to use vocab size
+    float   top_p              = 0.95f; // 1.0 = disabled
+    float   min_p              = 0.05f; // 0.0 = disabled
+    float   xtc_probability    = 0.00f; // 0.0 = disabled
+    float   xtc_threshold      = 0.10f; // > 0.5 disables XTC
+    float   typ_p              = 1.00f; // typical_p, 1.0 = disabled
+    float   temp               = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range     = 0.00f; // 0.0 = disabled
+    float   dynatemp_exponent  = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat     = 1.00f; // 1.0 = disabled
+    float   penalty_freq       = 0.00f; // 0.0 = disabled
+    float   penalty_present    = 0.00f; // 0.0 = disabled
+    float   dry_multiplier     = 0.0f;  // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
+    float   dry_base           = 1.75f; // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
+    int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
+    int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau       = 5.00f; // target entropy
+    float   mirostat_eta       = 0.10f; // learning rate
+    bool    penalize_nl        = false; // consider newlines as a repeatable token
+    bool    ignore_eos         = false;
+    bool    no_perf            = false; // disable performance metrics
 
-    std::vector<enum gpt_sampler_type> samplers = {
-        GPT_SAMPLER_TYPE_TOP_K,
-        GPT_SAMPLER_TYPE_TFS_Z,
-        GPT_SAMPLER_TYPE_TYPICAL_P,
-        GPT_SAMPLER_TYPE_TOP_P,
-        GPT_SAMPLER_TYPE_MIN_P,
-        GPT_SAMPLER_TYPE_TEMPERATURE
+    std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
+
+
+    std::vector<enum common_sampler_type> samplers = {
+        COMMON_SAMPLER_TYPE_DRY,
+        COMMON_SAMPLER_TYPE_TOP_K,
+        COMMON_SAMPLER_TYPE_TYPICAL_P,
+        COMMON_SAMPLER_TYPE_TOP_P,
+        COMMON_SAMPLER_TYPE_MIN_P,
+        COMMON_SAMPLER_TYPE_XTC,
+        COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
     std::string grammar; // optional BNF-like grammar to constrain sampling
@@ -141,9 +153,9 @@ struct gpt_sampler_params {
     std::string print() const;
 };
 
-struct gpt_params {
+struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =     0; // context size
+    int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -183,7 +195,7 @@ struct gpt_params {
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    struct gpt_sampler_params sparams;
+    struct common_sampler_params sparams;
 
     std::string model                = ""; // model path                                                    // NOLINT
     std::string model_draft          = ""; // draft model for speculative decoding                          // NOLINT
@@ -208,9 +220,9 @@ struct gpt_params {
     std::vector<llama_model_kv_override> kv_overrides;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
-    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
+    std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
-    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
+    std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
     int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
@@ -268,20 +280,21 @@ struct gpt_params {
 
     // embedding
     bool embedding         = false; // get only sentence embedding
-    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-    std::string embd_sep   = "\n";  // separator of embendings
+    std::string embd_sep   = "\n";  // separator of embeddings
+    bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    std::string system_prompt = "";                                                                         // NOLINT
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
@@ -289,7 +302,10 @@ struct gpt_params {
     std::string ssl_file_key  = "";                                                                         // NOLINT
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
-    bool endpoint_slots   = true;
+    // "advanced" endpoints are disabled by default for better security
+    bool webui            = true;
+    bool endpoint_slots   = false;
+    bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
     bool log_json = false;
@@ -344,20 +360,31 @@ struct gpt_params {
 
 // call once at the start of a program if it uses libcommon
 // initializes the logging system and prints info about the build
-void gpt_init();
+void common_init();
 
-std::string gpt_params_get_system_info(const gpt_params & params);
+std::string common_params_get_system_info(const common_params & params);
 
-bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
-bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
-void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
+bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
+bool parse_cpu_mask(const std::string & mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
+void postprocess_cpu_params(cpu_params & cpuparams, const cpu_params * role_model = nullptr);
 bool set_process_priority(enum ggml_sched_priority prio);
 
 //
 // String utils
 //
 
-std::vector<std::string> string_split(std::string input, char separator);
+#ifdef __GNUC__
+#ifdef __MINGW32__
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
+#endif
+
+LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
+std::string string_format(const char * fmt, ...);
 
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
@@ -366,6 +393,7 @@ void string_replace_all(std::string & s, const std::string & search, const std::
 
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
+    static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
     std::vector<T> values;
     std::istringstream str_stream(str);
     std::string token;
@@ -378,6 +406,22 @@ static std::vector<T> string_split(const std::string & str, char delim) {
     return values;
 }
 
+template<>
+std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
+{
+    std::vector<std::string> parts;
+    size_t begin_pos = 0;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(begin_pos, separator_pos - begin_pos);
+        parts.emplace_back(part);
+        begin_pos = separator_pos + 1;
+        separator_pos = input.find(separator, begin_pos);
+    }
+    parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
+    return parts;
+}
+
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
@@ -400,29 +444,29 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-struct llama_init_result {
+struct common_init_result {
     struct llama_model   * model   = nullptr;
     struct llama_context * context = nullptr;
-    std::vector<llama_lora_adapter_container> lora_adapters;
+    std::vector<common_lora_adapter_container> lora_adapters;
 };
 
-struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
+struct common_init_result     common_init_from_params(common_params & params);
 
-struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
-struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
+struct llama_model_params     common_model_params_to_llama  (const common_params & params);
+struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
-struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+struct llama_model * common_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+struct llama_model * common_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
-void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
 
 // Batch utils
 
-void llama_batch_clear(struct llama_batch & batch);
+void common_batch_clear(struct llama_batch & batch);
 
-void llama_batch_add(
+void common_batch_add(
                  struct llama_batch & batch,
                         llama_token   id,
                           llama_pos   pos,
@@ -435,13 +479,13 @@ void llama_batch_add(
 
 // tokenizes a string into a vector of tokens
 // should work similar to Python's `tokenizer.encode`
-std::vector<llama_token> llama_tokenize(
+std::vector<llama_token> common_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special = false);
 
-std::vector<llama_token> llama_tokenize(
+std::vector<llama_token> common_tokenize(
     const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
@@ -449,7 +493,7 @@ std::vector<llama_token> llama_tokenize(
 
 // tokenizes a token into a piece, optionally renders special/control tokens
 // should work similar to Python's `tokenizer.id_to_piece`
-std::string llama_token_to_piece(
+std::string common_token_to_piece(
         const struct llama_context * ctx,
                        llama_token   token,
                        bool          special = true);
@@ -457,7 +501,7 @@ std::string llama_token_to_piece(
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
 // optionally renders special/control tokens
-std::string llama_detokenize(
+std::string common_detokenize(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
@@ -467,31 +511,31 @@ std::string llama_detokenize(
 //
 
 // same with llama_chat_message, but uses std::string
-struct llama_chat_msg {
+struct common_chat_msg {
     std::string role;
     std::string content;
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-bool llama_chat_verify_template(const std::string & tmpl);
+bool common_chat_verify_template(const std::string & tmpl);
 
 // CPP wrapper for llama_chat_apply_template
 // If the built-in template is not supported, we default to chatml
 // If the custom "tmpl" is not supported, we throw an error
-std::string llama_chat_apply_template(const struct llama_model * model,
+std::string common_chat_apply_template(const struct llama_model * model,
         const std::string & tmpl,
-        const std::vector<llama_chat_msg> & chat,
+        const std::vector<common_chat_msg> & chat,
         bool add_ass);
 
 // Format single message, while taking into account the position of that message in chat history
-std::string llama_chat_format_single(const struct llama_model * model,
+std::string common_chat_format_single(const struct llama_model * model,
         const std::string & tmpl,
-        const std::vector<llama_chat_msg> & past_msg,
-        const llama_chat_msg & new_msg,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
         bool add_ass);
 
 // Returns an example of formatted chat
-std::string llama_chat_format_example(const struct llama_model * model,
+std::string common_chat_format_example(const struct llama_model * model,
         const std::string & tmpl);
 
 //
@@ -499,31 +543,31 @@ std::string llama_chat_format_example(const struct llama_model * model,
 //
 
 // Dump the KV cache view with the number of sequences per cell.
-void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
+void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
-void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
 
 //
 // Embedding utils
 //
 
-void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
+void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
 
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
+float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
 
 //
 // Control vector utils
 //
 
-struct llama_control_vector_data {
+struct common_control_vector_data {
     int n_embd;
 
     // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
     std::vector<float> data;
 };
 
-struct llama_control_vector_load_info {
+struct common_control_vector_load_info {
     float strength;
 
     std::string fname;
@@ -531,7 +575,7 @@ struct llama_control_vector_load_info {
 
 // Load control vectors, scale each by strength, and add them together.
 // On error, returns {-1, empty}
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
+common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos);
 
 //
 // Split utils
@@ -550,5 +594,5 @@ void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std
 void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
 
 void yaml_dump_non_result_info(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    FILE * stream, const common_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

common/console.cpp

@@ -94,6 +94,9 @@ namespace console {
                 simple_io = true;
             }
         }
+        if (simple_io) {
+            _setmode(_fileno(stdin), _O_U8TEXT);
+        }
 #else
         // POSIX-specific console initialization
         if (!simple_io) {

common/json-schema-to-grammar.cpp

@@ -611,7 +611,7 @@ private:
             }
             return join_seq();
         };
-        return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
+        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space");
     }
 
     /*

common/log.cpp

@@ -8,10 +8,10 @@
 #include <thread>
 #include <vector>
 
-int gpt_log_verbosity_thold = LOG_DEFAULT_LLAMA;
+int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
 
-void gpt_log_set_verbosity_thold(int verbosity) {
-    gpt_log_verbosity_thold = verbosity;
+void common_log_set_verbosity_thold(int verbosity) {
+    common_log_verbosity_thold = verbosity;
 }
 
 #define LOG_COL_DEFAULT "\033[0m"
@@ -29,16 +29,16 @@ static int64_t t_us() {
 }
 
 // colors
-enum gpt_log_col : int {
-    GPT_LOG_COL_DEFAULT = 0,
-    GPT_LOG_COL_BOLD,
-    GPT_LOG_COL_RED,
-    GPT_LOG_COL_GREEN,
-    GPT_LOG_COL_YELLOW,
-    GPT_LOG_COL_BLUE,
-    GPT_LOG_COL_MAGENTA,
-    GPT_LOG_COL_CYAN,
-    GPT_LOG_COL_WHITE,
+enum common_log_col : int {
+    COMMON_LOG_COL_DEFAULT = 0,
+    COMMON_LOG_COL_BOLD,
+    COMMON_LOG_COL_RED,
+    COMMON_LOG_COL_GREEN,
+    COMMON_LOG_COL_YELLOW,
+    COMMON_LOG_COL_BLUE,
+    COMMON_LOG_COL_MAGENTA,
+    COMMON_LOG_COL_CYAN,
+    COMMON_LOG_COL_WHITE,
 };
 
 // disable colors by default
@@ -54,7 +54,7 @@ static std::vector<const char *> g_col = {
     "",
 };
 
-struct gpt_log_entry {
+struct common_log_entry {
     enum ggml_log_level level;
 
     bool prefix;
@@ -71,7 +71,7 @@ struct gpt_log_entry {
         if (!fcur) {
             // stderr displays DBG messages only when their verbosity level is not higher than the threshold
             // these messages will still be logged to a file
-            if (level == GGML_LOG_LEVEL_DEBUG && gpt_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
+            if (level == GGML_LOG_LEVEL_DEBUG && common_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
                 return;
             }
 
@@ -82,23 +82,23 @@ struct gpt_log_entry {
             }
         }
 
-        if (level != GGML_LOG_LEVEL_NONE && prefix) {
+        if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
             if (timestamp) {
                 // [M.s.ms.us]
                 fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
-                        g_col[GPT_LOG_COL_BLUE],
+                        g_col[COMMON_LOG_COL_BLUE],
                         (int) (timestamp / 1000000 / 60),
                         (int) (timestamp / 1000000 % 60),
                         (int) (timestamp / 1000 % 1000),
                         (int) (timestamp % 1000),
-                        g_col[GPT_LOG_COL_DEFAULT]);
+                        g_col[COMMON_LOG_COL_DEFAULT]);
             }
 
             switch (level) {
-                case GGML_LOG_LEVEL_INFO:  fprintf(fcur, "%sI %s", g_col[GPT_LOG_COL_GREEN],   g_col[GPT_LOG_COL_DEFAULT]); break;
-                case GGML_LOG_LEVEL_WARN:  fprintf(fcur, "%sW %s", g_col[GPT_LOG_COL_MAGENTA], ""                        ); break;
-                case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[GPT_LOG_COL_RED],     ""                        ); break;
-                case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[GPT_LOG_COL_YELLOW],  ""                        ); break;
+                case GGML_LOG_LEVEL_INFO:  fprintf(fcur, "%sI %s", g_col[COMMON_LOG_COL_GREEN],   g_col[COMMON_LOG_COL_DEFAULT]); break;
+                case GGML_LOG_LEVEL_WARN:  fprintf(fcur, "%sW %s", g_col[COMMON_LOG_COL_MAGENTA], ""                        ); break;
+                case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[COMMON_LOG_COL_RED],     ""                        ); break;
+                case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[COMMON_LOG_COL_YELLOW],  ""                        ); break;
                 default:
                     break;
             }
@@ -107,18 +107,18 @@ struct gpt_log_entry {
         fprintf(fcur, "%s", msg.data());
 
         if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) {
-            fprintf(fcur, "%s", g_col[GPT_LOG_COL_DEFAULT]);
+            fprintf(fcur, "%s", g_col[COMMON_LOG_COL_DEFAULT]);
         }
 
         fflush(fcur);
     }
 };
 
-struct gpt_log {
+struct common_log {
     // default capacity - will be expanded if needed
-    gpt_log() : gpt_log(256) {}
+    common_log() : common_log(256) {}
 
-    gpt_log(size_t capacity) {
+    common_log(size_t capacity) {
         file = nullptr;
         prefix = false;
         timestamps = false;
@@ -137,7 +137,7 @@ struct gpt_log {
         resume();
     }
 
-    ~gpt_log() {
+    ~common_log() {
         pause();
         if (file) {
             fclose(file);
@@ -158,12 +158,12 @@ private:
     int64_t t_start;
 
     // ring buffer of entries
-    std::vector<gpt_log_entry> entries;
+    std::vector<common_log_entry> entries;
     size_t head;
     size_t tail;
 
     // worker thread copies into this
-    gpt_log_entry cur;
+    common_log_entry cur;
 
 public:
     void add(enum ggml_log_level level, const char * fmt, va_list args) {
@@ -219,7 +219,7 @@ public:
         tail = (tail + 1) % entries.size();
         if (tail == head) {
             // expand the buffer
-            std::vector<gpt_log_entry> new_entries(2*entries.size());
+            std::vector<common_log_entry> new_entries(2*entries.size());
 
             size_t new_tail = 0;
 
@@ -320,15 +320,15 @@ public:
         pause();
 
         if (colors) {
-            g_col[GPT_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
-            g_col[GPT_LOG_COL_BOLD]    = LOG_COL_BOLD;
-            g_col[GPT_LOG_COL_RED]     = LOG_COL_RED;
-            g_col[GPT_LOG_COL_GREEN]   = LOG_COL_GREEN;
-            g_col[GPT_LOG_COL_YELLOW]  = LOG_COL_YELLOW;
-            g_col[GPT_LOG_COL_BLUE]    = LOG_COL_BLUE;
-            g_col[GPT_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
-            g_col[GPT_LOG_COL_CYAN]    = LOG_COL_CYAN;
-            g_col[GPT_LOG_COL_WHITE]   = LOG_COL_WHITE;
+            g_col[COMMON_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
+            g_col[COMMON_LOG_COL_BOLD]    = LOG_COL_BOLD;
+            g_col[COMMON_LOG_COL_RED]     = LOG_COL_RED;
+            g_col[COMMON_LOG_COL_GREEN]   = LOG_COL_GREEN;
+            g_col[COMMON_LOG_COL_YELLOW]  = LOG_COL_YELLOW;
+            g_col[COMMON_LOG_COL_BLUE]    = LOG_COL_BLUE;
+            g_col[COMMON_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
+            g_col[COMMON_LOG_COL_CYAN]    = LOG_COL_CYAN;
+            g_col[COMMON_LOG_COL_WHITE]   = LOG_COL_WHITE;
         } else {
             for (size_t i = 0; i < g_col.size(); i++) {
                 g_col[i] = "";
@@ -355,47 +355,47 @@ public:
 // public API
 //
 
-struct gpt_log * gpt_log_init() {
-    return new gpt_log;
+struct common_log * common_log_init() {
+    return new common_log;
 }
 
-struct gpt_log * gpt_log_main() {
-    static struct gpt_log log;
+struct common_log * common_log_main() {
+    static struct common_log log;
 
     return &log;
 }
 
-void gpt_log_pause(struct gpt_log * log) {
+void common_log_pause(struct common_log * log) {
     log->pause();
 }
 
-void gpt_log_resume(struct gpt_log * log) {
+void common_log_resume(struct common_log * log) {
     log->resume();
 }
 
-void gpt_log_free(struct gpt_log * log) {
+void common_log_free(struct common_log * log) {
     delete log;
 }
 
-void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...) {
+void common_log_add(struct common_log * log, enum ggml_log_level level, const char * fmt, ...) {
     va_list args;
     va_start(args, fmt);
     log->add(level, fmt, args);
     va_end(args);
 }
 
-void gpt_log_set_file(struct gpt_log * log, const char * file) {
+void common_log_set_file(struct common_log * log, const char * file) {
     log->set_file(file);
 }
 
-void gpt_log_set_colors(struct gpt_log * log, bool colors) {
+void common_log_set_colors(struct common_log * log, bool colors) {
     log->set_colors(colors);
 }
 
-void gpt_log_set_prefix(struct gpt_log * log, bool prefix) {
+void common_log_set_prefix(struct common_log * log, bool prefix) {
     log->set_prefix(prefix);
 }
 
-void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps) {
+void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }

common/log.h

@@ -14,23 +14,23 @@
 #define LOG_DEFAULT_LLAMA 0
 
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
-// set via gpt_log_set_verbosity()
-extern int gpt_log_verbosity_thold;
+// set via common_log_set_verbosity()
+extern int common_log_verbosity_thold;
 
-void gpt_log_set_verbosity_thold(int verbosity); // not thread-safe
+void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
-// the gpt_log uses an internal worker thread to print/write log messages
+// the common_log uses an internal worker thread to print/write log messages
 // when the worker thread is paused, incoming log messages are discarded
-struct gpt_log;
+struct common_log;
 
-struct gpt_log * gpt_log_init();
-struct gpt_log * gpt_log_main(); // singleton, automatically destroys itself on exit
-void             gpt_log_pause (struct gpt_log * log); // pause  the worker thread, not thread-safe
-void             gpt_log_resume(struct gpt_log * log); // resume the worker thread, not thread-safe
-void             gpt_log_free  (struct gpt_log * log);
+struct common_log * common_log_init();
+struct common_log * common_log_main(); // singleton, automatically destroys itself on exit
+void                common_log_pause (struct common_log * log); // pause  the worker thread, not thread-safe
+void                common_log_resume(struct common_log * log); // resume the worker thread, not thread-safe
+void                common_log_free  (struct common_log * log);
 
 LOG_ATTRIBUTE_FORMAT(3, 4)
-void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...);
+void common_log_add(struct common_log * log, enum ggml_log_level level, const char * fmt, ...);
 
 // defaults: file = NULL, colors = false, prefix = false, timestamps = false
 //
@@ -54,10 +54,10 @@ void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * f
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
 //
 
-void gpt_log_set_file      (struct gpt_log * log, const char * file);       // not thread-safe
-void gpt_log_set_colors    (struct gpt_log * log,       bool   colors);     // not thread-safe
-void gpt_log_set_prefix    (struct gpt_log * log,       bool   prefix);     // whether to output prefix to each log
-void gpt_log_set_timestamps(struct gpt_log * log,       bool   timestamps); // whether to output timestamps in the prefix
+void common_log_set_file      (struct common_log * log, const char * file);       // not thread-safe
+void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
+void common_log_set_prefix    (struct common_log * log,       bool   prefix);     // whether to output prefix to each log
+void common_log_set_timestamps(struct common_log * log,       bool   timestamps); // whether to output timestamps in the prefix
 
 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
@@ -66,13 +66,13 @@ void gpt_log_set_timestamps(struct gpt_log * log,       bool   timestamps); // w
 //
 //   LOG_DBG("this is a debug message: %d\n", expensive_function());
 //
-// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > gpt_log_verbosity_thold
+// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > common_log_verbosity_thold
 //
 
 #define LOG_TMPL(level, verbosity, ...) \
     do { \
-        if ((verbosity) <= gpt_log_verbosity_thold) { \
-            gpt_log_add(gpt_log_main(), (level), __VA_ARGS__); \
+        if ((verbosity) <= common_log_verbosity_thold) { \
+            common_log_add(common_log_main(), (level), __VA_ARGS__); \
         } \
     } while (0)
 
@@ -83,8 +83,10 @@ void gpt_log_set_timestamps(struct gpt_log * log,       bool   timestamps); // w
 #define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
 #define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
 #define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
 
 #define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)
 #define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
 #define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
+#define LOG_CNTV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  verbosity, __VA_ARGS__)

common/ngram-cache.cpp

@@ -8,7 +8,7 @@
 #include <fstream>
 #include <thread>
 
-void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
+void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
                               std::vector<llama_token> & inp, int nnew, bool print_progress) {
     const int64_t t_start_ms = ggml_time_ms();
     const int64_t inp_size = inp.size();
@@ -20,16 +20,16 @@ void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, in
         const int64_t i_start = std::max(inp_size - nnew, ngram_size);
         for (int64_t i = i_start; i < inp_size; ++i) {
             const int64_t ngram_start = i - ngram_size;
-            llama_ngram ngram(&inp[ngram_start], ngram_size);
+            common_ngram ngram(&inp[ngram_start], ngram_size);
             const llama_token token = inp[i];
 
-            llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
+            common_ngram_cache::iterator part_it = ngram_cache.find(ngram);
             if (part_it == ngram_cache.end()) {
-                llama_ngram_cache_part part;
+                common_ngram_cache_part part;
                 part.emplace(token, 1);
                 ngram_cache.emplace(ngram, part);
             } else {
-                llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
+                common_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
                 if (token_count_it == part_it->second.end()) {
                     part_it->second.emplace(token, 1);
                 } else {
@@ -62,12 +62,12 @@ constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4,  3,  2,  2};
 constexpr int     draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
 
 // Helper function that tries to draft a token from only the static ngram cache:
-static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) {
-    llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
+    common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
     if (part_static_it == nc_static.end()) {
         return -1;
     }
-    const llama_ngram_cache_part part_static = part_static_it->second;
+    const common_ngram_cache_part part_static = part_static_it->second;
 
     int max_count_static  = 0;
     int sum_count_static  = 0;
@@ -95,19 +95,19 @@ static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ng
 
 // Try to draft a token from primary cache (context/dynamic), validate with static cache:
 static llama_token try_draft(
-    llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static,
+    common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
     const int * min_sample_size, const int * min_percent) {
 
     llama_token drafted_token = -1;
 
     for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
-        const llama_ngram ngram_primary = ngrams_primary[i];
+        const common_ngram ngram_primary = ngrams_primary[i];
 
-        llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
+        common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
         if (part_primary_it == nc_primary.end()) {
             continue;
         }
-        const llama_ngram_cache_part part_primary = part_primary_it->second;
+        const common_ngram_cache_part part_primary = part_primary_it->second;
 
         int max_count_primary = 0;
         int max_count_static  = 0;
@@ -117,7 +117,7 @@ static llama_token try_draft(
         for (std::pair<llama_token, int> token_count_primary : part_primary) {
             const llama_token token = token_count_primary.first;
 
-            llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
+            common_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
 
             const int32_t count_primary = token_count_primary.second;
             const int32_t count_static  = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
@@ -142,9 +142,9 @@ static llama_token try_draft(
     return drafted_token;
 }
 
-void llama_ngram_cache_draft(
+void common_ngram_cache_draft(
     std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
-    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static
+    common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static
 ) {
     GGML_ASSERT(draft.size() == 1);
     const int inp_size = inp.size();
@@ -157,21 +157,21 @@ void llama_ngram_cache_draft(
         llama_token drafted_token = -1;
 
         const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
-        llama_ngram ngram_static;
+        common_ngram ngram_static;
         for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
             ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
         }
-        llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
-        llama_ngram_cache_part part_static;
+        common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+        common_ngram_cache_part part_static;
         if (part_static_it != nc_static.end()) {
             part_static = part_static_it->second;
         }
 
         // cd = context + dynamic
-        std::vector<llama_ngram> ngrams_cd;
+        std::vector<common_ngram> ngrams_cd;
         for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
             const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
-            llama_ngram ngram_cd;
+            common_ngram ngram_cd;
             for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
                 ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
             }
@@ -196,16 +196,16 @@ void llama_ngram_cache_draft(
     }
 }
 
-void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) {
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) {
     std::ofstream file_out(filename, std::ios::binary);
-    for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) {
-        const llama_ngram      ngram        = item.first;
-        llama_ngram_cache_part token_counts = item.second;
+    for (std::pair<common_ngram, common_ngram_cache_part> item : ngram_cache) {
+        const common_ngram      ngram        = item.first;
+        common_ngram_cache_part token_counts = item.second;
         GGML_ASSERT(!token_counts.empty());
         const int32_t ntokens = token_counts.size();
         GGML_ASSERT(ntokens > 0);
 
-        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(llama_ngram));
+        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(common_ngram));
         file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
         for (std::pair<llama_token, int32_t> item2 : token_counts) {
             const llama_token token = item2.first;
@@ -219,14 +219,14 @@ void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filen
 
 }
 
-llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
+common_ngram_cache common_ngram_cache_load(std::string & filename) {
     std::ifstream hashmap_file(filename, std::ios::binary);
     if (!hashmap_file) {
         throw std::ifstream::failure("Unable to open file " + filename);
     }
-    llama_ngram_cache ngram_cache;
+    common_ngram_cache ngram_cache;
 
-    llama_ngram ngram;
+    common_ngram ngram;
     int32_t     ntokens;
     llama_token token;
     int32_t     count;
@@ -235,11 +235,11 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
     char * ntokensc = reinterpret_cast<char*>(&ntokens);
     char * tokenc   = reinterpret_cast<char*>(&token);
     char * countc   = reinterpret_cast<char*>(&count);
-    while(hashmap_file.read(ngramc, sizeof(llama_ngram))) {
+    while(hashmap_file.read(ngramc, sizeof(common_ngram))) {
         GGML_ASSERT(!hashmap_file.eof());
         GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
         GGML_ASSERT(ntokens > 0);
-        llama_ngram_cache_part token_counts;
+        common_ngram_cache_part token_counts;
 
         for (int i = 0; i < ntokens; ++i) {
             GGML_ASSERT(!hashmap_file.eof());
@@ -257,12 +257,12 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
     return ngram_cache;
 }
 
-void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) {
-    for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) {
-        const llama_ngram      ngram = ngram_part.first;
-        llama_ngram_cache_part  part = ngram_part.second;
+void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add) {
+    for (std::pair<common_ngram, common_ngram_cache_part> ngram_part : ngram_cache_add) {
+        const common_ngram      ngram = ngram_part.first;
+        common_ngram_cache_part  part = ngram_part.second;
 
-        llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
+        common_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
         if (part_merged_it == ngram_cache_target.end()) {
             ngram_cache_target.emplace(ngram, part);
             continue;
@@ -273,7 +273,7 @@ void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram
             const int32_t     count = token_count.second;
             GGML_ASSERT(count > 0);
 
-            llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
+            common_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
             if (token_count_merged_it == part_merged_it->second.end()) {
                 part_merged_it->second.emplace(token, count);
                 continue;

common/ngram-cache.h

@@ -12,22 +12,22 @@
 
 // Data structures to map n-grams to empirical token probabilities:
 
-struct llama_ngram {
+struct common_ngram {
     llama_token tokens[LLAMA_NGRAM_MAX];
 
-    llama_ngram() {
+    common_ngram() {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
             tokens[i] = -1;
         }
     }
 
-    llama_ngram(const llama_token * input, const int ngram_size) {
+    common_ngram(const llama_token * input, const int ngram_size) {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
             tokens[i] = i < ngram_size ? input[i] : -1;
         }
     }
 
-    bool operator==(const llama_ngram & other) const {
+    bool operator==(const common_ngram & other) const {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
             if (tokens[i] != other.tokens[i]) {
                 return false;
@@ -37,28 +37,28 @@ struct llama_ngram {
     }
 };
 
-struct llama_token_hash_function {
+struct common_token_hash_function {
     size_t operator()(const llama_token token) const {
         // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
         return token * 11400714819323198485llu;
     }
 };
 
-struct llama_ngram_hash_function {
-    size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
+struct common_ngram_hash_function {
+    size_t operator()(const common_ngram & ngram) const {
+        size_t hash = common_token_hash_function{}(ngram.tokens[0]);
         for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= llama_token_hash_function{}(ngram.tokens[i]);
+            hash ^= common_token_hash_function{}(ngram.tokens[i]);
         }
         return hash;
     }
 };
 
 // token -> number of times token has been seen
-typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part;
+typedef std::unordered_map<llama_token, int32_t> common_ngram_cache_part;
 
 // n-gram -> empirical distribution of following tokens
-typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
+typedef std::unordered_map<common_ngram, common_ngram_cache_part, common_ngram_hash_function> common_ngram_cache;
 
 
 // Update an ngram cache with tokens.
@@ -70,8 +70,8 @@ typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash
 //
 // In order to get correct results inp_data can ONLY BE APPENDED TO.
 // Changes in the middle need a complete rebuild.
-void llama_ngram_cache_update(
-    llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
+void common_ngram_cache_update(
+    common_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
 
 // Try to draft tokens from ngram caches.
 // inp:                the tokens generated so far.
@@ -81,21 +81,21 @@ void llama_ngram_cache_update(
 // nc_context:         ngram cache based on current context.
 // nc_dynamic:         ngram cache based on previous user generations.
 // nc_static:          ngram cache generated from a large text corpus, used for validation.
-void llama_ngram_cache_draft(
+void common_ngram_cache_draft(
     std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
-    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static);
+    common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static);
 
 // Save an ngram cache to a file.
 // ngram_cache: the ngram cache to save.
 // filename:    the path under which to save the ngram cache.
-void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename);
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
 
-// Load an ngram cache saved with llama_ngram_cache_save.
+// Load an ngram cache saved with common_ngram_cache_save.
 // filename: the path from which to load the ngram cache.
 // returns:  an ngram cache containing the information saved to filename.
-llama_ngram_cache llama_ngram_cache_load(std::string & filename);
+common_ngram_cache common_ngram_cache_load(std::string & filename);
 
 // Merge two ngram caches.
 // ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
 // ngram_cache_add:    the ngram cache to add to ngram_cache_target.
-void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add);
+void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add);

common/sampling.cpp

@@ -98,8 +98,8 @@ struct ring_buffer {
     std::vector<T> data;
 };
 
-struct gpt_sampler {
-    gpt_sampler_params params;
+struct common_sampler {
+    common_sampler_params params;
 
     struct llama_sampler * grmr;
     struct llama_sampler * chain;
@@ -125,26 +125,28 @@ struct gpt_sampler {
     }
 };
 
-std::string gpt_sampler_params::print() const {
+std::string common_sampler_params::print() const {
     char result[1024];
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
-            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
+            "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
-            top_k, tfs_z, top_p, min_p, typ_p, temp,
+            dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
+            top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
 }
 
-struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {
+struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params) {
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
 
-    auto * result = new gpt_sampler {
+    auto * result = new common_sampler {
         /* .params = */ params,
         /* .grmr   = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
         /* .chain  = */ llama_sampler_chain_init(lparams),
@@ -171,52 +173,60 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
                 params.penalize_nl,
                 params.ignore_eos));
 
-    if (params.temp > 0.0f) {
-        if (params.mirostat == 0) {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case GPT_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+    if (params.mirostat == 0) {
+        for (const auto & cnstr : params.samplers) {
+            switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                    {
+                        std::vector<const char*> c_breakers;
+                        c_breakers.reserve(params.dry_sequence_breakers.size());
+                        for (const auto& str : params.dry_sequence_breakers) {
+                            c_breakers.push_back(str.c_str());
+                        }
+
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                    }
                         break;
-                    case GPT_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_TFS_Z:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+                case COMMON_SAMPLER_TYPE_TOP_K:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_MIN_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_XTC:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    break;
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    break;
+                case COMMON_SAMPLER_TYPE_INFILL:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
+                    break;
+                default:
+                    GGML_ASSERT(false && "unknown sampler type");
             }
-            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-            llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
-        } else if (params.mirostat == 1) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
-        } else if (params.mirostat == 2) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
-        } else {
-            GGML_ASSERT(false && "unknown mirostat version");
         }
+        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
+    } else if (params.mirostat == 1) {
+        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+    } else if (params.mirostat == 2) {
+        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
     } else {
-        llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
+        GGML_ASSERT(false && "unknown mirostat version");
     }
 
     return result;
 }
 
-void gpt_sampler_free(struct gpt_sampler * gsmpl) {
+void common_sampler_free(struct common_sampler * gsmpl) {
     if (gsmpl) {
         llama_sampler_free(gsmpl->grmr);
 
@@ -226,7 +236,7 @@ void gpt_sampler_free(struct gpt_sampler * gsmpl) {
     }
 }
 
-void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar) {
+void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
     if (accept_grammar) {
         llama_sampler_accept(gsmpl->grmr, token);
     }
@@ -236,14 +246,14 @@ void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool acce
     gsmpl->prev.push_back(token);
 }
 
-void gpt_sampler_reset(struct gpt_sampler * gsmpl) {
+void common_sampler_reset(struct common_sampler * gsmpl) {
     llama_sampler_reset(gsmpl->grmr);
 
     llama_sampler_reset(gsmpl->chain);
 }
 
-struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) {
-    return new gpt_sampler {
+struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
+    return new common_sampler {
         /* .params = */ gsmpl->params,
         /* .grmr   = */ llama_sampler_clone(gsmpl->grmr),
         /* .chain  = */ llama_sampler_clone(gsmpl->chain),
@@ -253,7 +263,7 @@ struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) {
     };
 }
 
-void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl) {
+void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
     // TODO: measure grammar performance
 
     if (gsmpl) {
@@ -264,7 +274,7 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
     }
 }
 
-llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
     gsmpl->set_logits(ctx, idx);
 
     auto & grmr  = gsmpl->grmr;
@@ -310,21 +320,21 @@ llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context
     return cur_p.data[cur_p.selected].id;
 }
 
-uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) {
+uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
     return llama_sampler_get_seed(gsmpl->chain);
 }
 
 // helpers
 
-llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) {
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
     return &gsmpl->cur_p;
 }
 
-llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) {
+llama_token common_sampler_last(const struct common_sampler * gsmpl) {
     return gsmpl->prev.rat(0);
 }
 
-std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
+std::string common_sampler_print(const struct common_sampler * gsmpl) {
     std::string result = "logits ";
 
     for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
@@ -335,7 +345,7 @@ std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
     return result;
 }
 
-std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main, int n) {
+std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_main, int n) {
     n = std::min(n, (int) gsmpl->prev.size());
 
     if (n <= 0) {
@@ -350,63 +360,67 @@ std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main,
 
         GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen");
 
-        result += llama_token_to_piece(ctx_main, id);
+        result += common_token_to_piece(ctx_main, id);
     }
 
     return result;
 }
 
-char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr) {
+char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
     switch (cnstr) {
-        case GPT_SAMPLER_TYPE_TOP_K:       return 'k';
-        case GPT_SAMPLER_TYPE_TFS_Z:       return 'f';
-        case GPT_SAMPLER_TYPE_TYPICAL_P:   return 'y';
-        case GPT_SAMPLER_TYPE_TOP_P:       return 'p';
-        case GPT_SAMPLER_TYPE_MIN_P:       return 'm';
-        case GPT_SAMPLER_TYPE_TEMPERATURE: return 't';
+        case COMMON_SAMPLER_TYPE_DRY:         return 'd';
+        case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
+        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
+        case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
+        case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
+        case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
+        case COMMON_SAMPLER_TYPE_XTC:         return 'x';
+        case COMMON_SAMPLER_TYPE_INFILL:      return 'i';
         default : return '?';
     }
 }
 
-std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr) {
+std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
     switch (cnstr) {
-        case GPT_SAMPLER_TYPE_TOP_K:       return "top_k";
-        case GPT_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
-        case GPT_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
-        case GPT_SAMPLER_TYPE_TOP_P:       return "top_p";
-        case GPT_SAMPLER_TYPE_MIN_P:       return "min_p";
-        case GPT_SAMPLER_TYPE_TEMPERATURE: return "temperature";
+        case COMMON_SAMPLER_TYPE_DRY:         return "dry";
+        case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
+        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
+        case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
+        case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
+        case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
+        case COMMON_SAMPLER_TYPE_INFILL:      return "infill";
         default : return "";
     }
 }
 
-std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
-    std::unordered_map<std::string, gpt_sampler_type> sampler_canonical_name_map {
-        { "top_k",       GPT_SAMPLER_TYPE_TOP_K },
-        { "top_p",       GPT_SAMPLER_TYPE_TOP_P },
-        { "typ_p",       GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "min_p",       GPT_SAMPLER_TYPE_MIN_P },
-        { "tfs_z",       GPT_SAMPLER_TYPE_TFS_Z },
-        { "temperature", GPT_SAMPLER_TYPE_TEMPERATURE },
+std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
+        { "dry",         COMMON_SAMPLER_TYPE_DRY },
+        { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
+        { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
+        { "xtc",         COMMON_SAMPLER_TYPE_XTC },
+        { "infill",      COMMON_SAMPLER_TYPE_INFILL },
     };
 
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, gpt_sampler_type> sampler_alt_name_map {
-        { "top-k",       GPT_SAMPLER_TYPE_TOP_K },
-        { "top-p",       GPT_SAMPLER_TYPE_TOP_P },
-        { "nucleus",     GPT_SAMPLER_TYPE_TOP_P },
-        { "typical-p",   GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "typical",     GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "typ-p",       GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "typ",         GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "min-p",       GPT_SAMPLER_TYPE_MIN_P },
-        { "tfs-z",       GPT_SAMPLER_TYPE_TFS_Z },
-        { "tfs",         GPT_SAMPLER_TYPE_TFS_Z },
-        { "temp",        GPT_SAMPLER_TYPE_TEMPERATURE },
+    std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
+        { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
+        { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
+        { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "typ-p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "temp",        COMMON_SAMPLER_TYPE_TEMPERATURE },
     };
 
-    std::vector<gpt_sampler_type> samplers;
+    std::vector<common_sampler_type> samplers;
     samplers.reserve(names.size());
 
     for (const auto & name : names) {
@@ -426,17 +440,19 @@ std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std
     return samplers;
 }
 
-std::vector<gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars) {
-    std::unordered_map<char, gpt_sampler_type> sampler_name_map = {
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_K),       GPT_SAMPLER_TYPE_TOP_K },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TFS_Z),       GPT_SAMPLER_TYPE_TFS_Z },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P),   GPT_SAMPLER_TYPE_TYPICAL_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_P),       GPT_SAMPLER_TYPE_TOP_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_MIN_P),       GPT_SAMPLER_TYPE_MIN_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TEMPERATURE), GPT_SAMPLER_TYPE_TEMPERATURE }
+std::vector<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
+    std::unordered_map<char, common_sampler_type> sampler_name_map = {
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_DRY),         COMMON_SAMPLER_TYPE_DRY },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL),      COMMON_SAMPLER_TYPE_INFILL },
     };
 
-    std::vector<gpt_sampler_type> samplers;
+    std::vector<common_sampler_type> samplers;
     samplers.reserve(chars.size());
 
     for (const auto & c : chars) {

common/sampling.h

@@ -7,7 +7,7 @@
 #include <string>
 #include <vector>
 
-// gpt_sampler extends llama_sampler with additional functionality:
+// common_sampler extends llama_sampler with additional functionality:
 //
 //  - grammar support
 //  - custom sampler logic based on the parameters
@@ -23,30 +23,30 @@
 // token in order to verify if it fits the grammar. And only if the token doesn't fit the grammar, the
 // grammar constraints are applied to the full vocabulary and the token is resampled.
 //
-// The gpt_sampler also maintains a container with the last accepted tokens. In the future, this can
+// The common_sampler also maintains a container with the last accepted tokens. In the future, this can
 // be moved into the core llama library.
 //
-// For convenience, the gpt_sampler also maintains a container with the current candidate tokens.
+// For convenience, the common_sampler also maintains a container with the current candidate tokens.
 // This can be used to access the probabilities of the rest of the non-sampled tokens.
 //
 // TODO: measure grammar performance
 //
 
-struct gpt_sampler;
+struct common_sampler;
 
 // llama_sampler API overloads
 
-struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params);
+struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params);
 
-void gpt_sampler_free(struct gpt_sampler * gsmpl);
+void common_sampler_free(struct common_sampler * gsmpl);
 
 // if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
-void                 gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar);
-void                 gpt_sampler_reset (struct gpt_sampler * gsmpl);
-struct gpt_sampler * gpt_sampler_clone (struct gpt_sampler * gsmpl);
+void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar);
+void                    common_sampler_reset (struct common_sampler * gsmpl);
+struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 
 // arguments can be nullptr to skip printing
-void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl);
+void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
 
 // extended sampling implementation:
 //
@@ -58,26 +58,26 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
 // if grammar_first is true, the grammar is applied before the samplers (slower)
 // useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
 //
-llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
-uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl);
+uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 
 // helpers
 
 // access the internal list of current candidate tokens
-llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl);
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
 
 // get the last accepted token
-llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl);
+llama_token common_sampler_last(const struct common_sampler * gsmpl);
 
 // print the sampler chain into a string
-std::string gpt_sampler_print(const struct gpt_sampler * gsmpl);
+std::string common_sampler_print(const struct common_sampler * gsmpl);
 
 // get a string representation of the last accepted tokens
-std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx, int n);
+std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx, int n);
 
-char        gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr);
-std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr);
+char        common_sampler_type_to_chr(enum common_sampler_type cnstr);
+std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
 
-std::vector<enum gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
-std::vector<enum gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars);
+std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);

common/train.h

@@ -1,233 +0,0 @@
-// Various helper functions and utilities for training
-
-#pragma once
-
-#include <string>
-#include <random>
-#include <vector>
-
-#include "ggml.h"
-#include "llama.h"
-
-#define LLAMA_TRAIN_MAX_NODES 16384
-
-typedef std::string mt19937_state;
-
-struct train_state {
-    struct ggml_opt_context * opt;
-
-    uint64_t train_its;
-    uint64_t train_samples;
-    uint64_t train_tokens;
-    uint64_t train_epochs;
-
-    size_t        shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
-    mt19937_state shuffle_rng_state_current;
-    mt19937_state shuffle_rng_state_next;
-    size_t        shuffle_sample_count;
-    size_t        shuffle_next_sample;
-};
-
-struct train_params_common {
-    const char * fn_train_data;
-    const char * fn_checkpoint_in;
-    const char * fn_checkpoint_out;
-    const char * pattern_fn_it;
-    const char * fn_latest;
-
-    bool print_usage;
-
-    int save_every;
-
-    uint32_t seed;
-
-    int n_ctx;
-    int n_threads;
-    int n_batch;
-    int n_gradient_accumulation;
-    int n_epochs;
-    int n_gpu_layers;
-
-    bool custom_n_ctx;
-
-    bool use_flash;
-    bool use_checkpointing;
-
-    std::string sample_start;
-    bool include_sample_start;
-    bool escape;
-    bool overlapping_samples;
-    bool fill_with_next_samples;
-    bool separate_with_eos;
-    bool separate_with_bos;
-    bool sample_random_offsets;
-
-    bool force_reshuffle;
-
-    int   warmup;
-    int   cos_decay_steps;
-    float cos_decay_restart;
-    float cos_decay_min;
-    bool  enable_restart;
-
-    int   opt_past;
-    float opt_delta;
-    int   opt_max_no_improvement;
-
-    int   adam_n_iter;
-    float adam_alpha;
-    float adam_min_alpha;
-    float adam_decay;
-    int   adam_decay_min_ndim;
-    float adam_beta1;
-    float adam_beta2;
-    float adam_gclip;
-    float adam_eps_f;
-};
-
-typedef void (*save_train_files_callback)(void * data, struct train_state * train);
-
-struct train_opt_callback_data {
-    struct train_params_common * params;
-    struct train_state         * train;
-    save_train_files_callback    save_cb;
-    void                       * save_data;
-    struct llama_context       * lctx;
-    int                          last_save_iter;
-    llama_token                * tokens_data;
-    size_t                       tokens_size;
-    size_t                     * samples_begin;
-    size_t                     * samples_size;
-    size_t                     * shuffled_samples_offs;
-    size_t                     * shuffled_samples_begin;
-    size_t                     * shuffled_samples_size;
-    size_t                       samples_count;
-    struct ggml_tensor         * tokens_input;
-    struct ggml_tensor         * target_probs;
-    int                          first_iter;
-    int                          first_epoch;
-    int                          iter_at_last_epoch;
-    int64_t                      last_time;
-    double                       millis_per_iter;
-};
-
-struct train_state * init_train_state();
-void free_train_state(struct train_state  * state);
-
-struct train_params_common get_default_train_params_common();
-void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
-
-bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
-void finish_processing_train_args(struct train_params_common * params);
-
-struct random_normal_distribution;
-struct random_uniform_distribution;
-
-struct random_normal_distribution  * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
-struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
-
-void free_random_normal_distribution (struct random_normal_distribution  * rnd);
-void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
-
-struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
-struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
-
-// generate random float in interval [0,1)
-float frand();
-float frand_normal (struct random_normal_distribution * rnd);
-float frand_uniform(struct random_uniform_distribution * rnd);
-
-int   clamp (const int v, const int min, const int max);
-float fclamp(const float v, const float min, const float max);
-
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
-
-size_t tokenize_file(
-        struct llama_context     * lctx,
-        const char               * filename,
-        const std::string        & sample_start,
-        bool                       include_sample_start,
-        bool                       overlapping_samples,
-        unsigned                   context_length,
-        std::vector<llama_token> & out_tokens,
-        std::vector<size_t>      & out_samples_begin,
-        std::vector<size_t>      & out_samples_size);
-
-int64_t get_example_targets_batch(
-        struct llama_context * lctx,
-        struct ggml_tensor   * tokens_input,
-        struct ggml_tensor   * target_probs,
-        int64_t                example_id,
-        const size_t         * samples_offs,
-        const size_t         * samples_begin,
-        const size_t         * samples_size,
-              size_t           samples_count,
-        const llama_token    * train_data,
-        size_t                 n_train_data,
-        bool                   separate_with_eos,
-        bool                   separate_with_bos,
-        bool                   fill_with_next_samples,
-        bool                   sample_random_offsets);
-
-
-void          mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
-mt19937_state mt19937_get_state(const std::mt19937& rng);
-mt19937_state mt19937_seed_to_state(unsigned seed);
-
-mt19937_state shuffle_samples(
-        const mt19937_state & rng_state,
-        size_t              * shuffled_offs,
-        size_t              * shuffled_begins,
-        size_t              * shuffled_sizes,
-        const size_t        * begins,
-        const size_t        * sizes,
-        size_t                count);
-
-size_t hash_combine(size_t h1, size_t h2);
-
-size_t compute_samples_hash(
-    const char* fn,
-    const size_t* samples_begin,
-    const size_t* samples_size,
-    size_t sample_count);
-
-
-std::string replace_str(const char * s, const char * needle, const char * replacement);
-
-void print_duration(double milliseconds);
-
-float cosine_decay(
-    int64_t step,
-    int64_t decay_steps,
-    float   minimum);
-
-float cosine_decay_restart(
-    int64_t step,
-    int64_t decay_steps,
-    float   minimum,
-    float   restart_step_mult);
-
-float learning_schedule(
-    int64_t step,
-    int64_t warmup_steps,
-    int64_t decay_steps,
-    float   learning_rate,
-    float   overall_minimum,
-    float   cos_decay_minimum,
-    float   cos_decay_restart_step_mult,
-    bool    enable_restart);
-
-void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
-
-void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
-void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
-
-bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
-void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
-
-std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
-
-void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);

convert_hf_to_gguf.py

@@ -15,6 +15,7 @@ from enum import IntEnum
 from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
+from itertools import chain
 
 import math
 import numpy as np
@@ -64,7 +65,6 @@ class Model:
     model_name: str | None
     metadata_override: Path | None
     dir_model_card: Path
-    is_lora: bool
 
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
@@ -72,7 +72,8 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
-                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -87,14 +88,13 @@ class Model:
         self.is_safetensors = len(self.part_names) > 0
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
-        self.hparams = Model.load_hparams(self.dir_model)
+        self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         self.metadata_override = metadata_override
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
-        self.is_lora = is_lora  # true if model is used inside convert_lora_to_gguf.py
 
         # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
         if self.ftype == gguf.LlamaFileType.GUESSED:
@@ -270,10 +270,14 @@ class Model:
 
         return False
 
+    # some models need extra generated tensors (like rope_freqs)
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        return ()
+
     def prepare_tensors(self):
         max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
 
-        for name, data_torch in self.get_tensors():
+        for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()):
             # we don't need these
             if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
                 continue
@@ -291,8 +295,13 @@ class Model:
                     bid = int(part)
                     break
 
-            for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
-                data: np.ndarray  # type hint
+            for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
+                data = data_torch.squeeze().numpy()
+
+                # if data ends up empty, it means data_torch was a scalar tensor -> restore
+                if len(data.shape) == 0:
+                    data = data_torch.numpy()
+
                 n_dims = len(data.shape)
                 data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
@@ -565,6 +574,9 @@ class Model:
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
+        if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
+            # ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
+            res = "bert-bge-large"
         if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
             # ref: https://huggingface.co/mosaicml/mpt-7b
             res = "mpt"
@@ -592,6 +604,9 @@ class Model:
         if chkhsh == "a8594e3edff7c29c003940395316294b2c623e09894deebbc65f33f1515df79e":
             # ref: https://huggingface.co/databricks/dbrx-base
             res = "dbrx"
+        if chkhsh == "c7699093ba4255a91e702aa38a596aa81669f3525dae06c2953267dde580f448":
+            # ref: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+            res = "jina-v1-en"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
             res = "jina-v2-en"
@@ -640,6 +655,9 @@ class Model:
         if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
             # ref: https://huggingface.co/microsoft/phi-2
             res = "phi-2"
+        if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
+            # ref: https://huggingface.co/facebook/chameleon-7b
+            res = "chameleon"
 
         if res is None:
             logger.warning("\n")
@@ -1524,6 +1542,17 @@ class LlamaModel(Model):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -1540,17 +1569,6 @@ class LlamaModel(Model):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
-        # Apply to granite small models only
-        if self.hparams.get("vocab_size", 32000) == 49152:
-            self.gguf_writer.add_add_bos_token(False)
-
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
         if n_head_kv is not None and n_head != n_head_kv:
@@ -1606,7 +1624,7 @@ class LlamaModel(Model):
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def prepare_tensors(self):
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
@@ -1633,9 +1651,9 @@ class LlamaModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
-                if not self.is_lora:
-                    self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
 
+    def prepare_tensors(self):
         super().prepare_tensors()
 
         if self._experts is not None:
@@ -1859,8 +1877,6 @@ class MiniCPM3Model(Model):
     def set_gguf_parameters(self):
         hparams = self.hparams
 
-        rope_dims = hparams["qk_rope_head_dim"]
-
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
@@ -1876,24 +1892,25 @@ class MiniCPM3Model(Model):
         self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         rope_scaling = self.find_hparam(['rope_scaling'], True)
-        if rope_scaling is None:
-            return
+        if rope_scaling is not None:
+            rope_dims = self.hparams["qk_rope_head_dim"]
 
-        long_factors = rope_scaling.get('long_factor', None)
-        short_factors = rope_scaling.get('short_factor', None)
+            long_factors = rope_scaling.get('long_factor', None)
+            short_factors = rope_scaling.get('short_factor', None)
 
-        if long_factors is None or short_factors is None:
-            raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
+            if long_factors is None or short_factors is None:
+                raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
 
-        if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
-            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
+            if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
+                raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
 
-        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
-        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
     def set_vocab(self):
-        self._set_vocab_llama_hf()
+        self._set_vocab_sentencepiece()
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -2205,6 +2222,13 @@ class Phi3MiniModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
 
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
+        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
+        rope_dims = n_embd // n_head
+
         # write rope scaling for long context (128k) model
         rope_scaling = self.find_hparam(['rope_scaling'], True)
         if rope_scaling is None:
@@ -2234,9 +2258,8 @@ class Phi3MiniModel(Model):
         if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
             raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
 
-        if not self.is_lora:
-            self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
-            self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
 
 @Model.register("PlamoForCausalLM")
@@ -2598,7 +2621,7 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
-@Model.register("XLMRobertaModel")
+@Model.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -2696,6 +2719,11 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_add_eos_token(True)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
         # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
         if name == "embeddings.position_embeddings.weight":
             if self._position_offset is not None:
@@ -2840,6 +2868,9 @@ class Rwkv6Model(Model):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.chat_template = "rwkv-world"
+        # hack: Add '\n\n' as the EOT token to make it chat normally
+        special_vocab._set_special_token("eot", 261)
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -3107,6 +3138,14 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "bert.", remove the prefix
+        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+        if name.startswith("bert."):
+            name = name[5:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
 
 @Model.register("OpenELMForCausalLM")
 class OpenELMModel(Model):
@@ -4047,7 +4086,7 @@ class ExaoneModel(Model):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
 
-    def prepare_tensors(self):
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
@@ -4074,14 +4113,112 @@ class ExaoneModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
-                if not self.is_lora:
-                    self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
 
-        super().prepare_tensors()
+
+@Model.register("GraniteForCausalLM")
+class GraniteModel(LlamaModel):
+    """Conversion for IBM's GraniteForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE
+
+    def set_gguf_parameters(self):
+        """Granite uses standard llama parameters with the following differences:
+
+        - No head_dim support
+        - New multiplier params:
+            - attention_scale
+            - embedding_scale
+            - residual_scale
+        - logits_scaling
+        """
+        if head_dim := self.hparams.pop("head_dim", None):
+            logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim)
+        super().set_gguf_parameters()
+        # NOTE: Convert _multiplier params to _scale params for naming
+        #   consistency
+        if attention_scale := self.hparams.get("attention_multiplier"):
+            self.gguf_writer.add_attention_scale(attention_scale)
+            logger.info("gguf: (granite) attention_scale = %s", attention_scale)
+        if embedding_scale := self.hparams.get("embedding_multiplier"):
+            self.gguf_writer.add_embedding_scale(embedding_scale)
+            logger.info("gguf: (granite) embedding_scale = %s", embedding_scale)
+        if residual_scale := self.hparams.get("residual_multiplier"):
+            self.gguf_writer.add_residual_scale(residual_scale)
+            logger.info("gguf: (granite) residual_scale = %s", residual_scale)
+        if logits_scale := self.hparams.get("logits_scaling"):
+            self.gguf_writer.add_logit_scale(logits_scale)
+            logger.info("gguf: (granite) logits_scale = %s", logits_scale)
+
+
+@Model.register("GraniteMoeForCausalLM")
+class GraniteMoeModel(GraniteModel):
+    """Conversion for IBM's GraniteMoeForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_MOE
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        """In modeling_granitemoe, the JetMoe implementation of parallel experts
+        is used. This essentially merges w1 and w3 into a single tensor with 2x
+        the hidden size that is then split during forward. To keep compatibility
+        with existing mixtral support, we pull them apart here.
+        """
+
+        if name.endswith("block_sparse_moe.input_linear.weight"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
+            gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
+            ]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@Model.register("ChameleonForConditionalGeneration")
+@Model.register("ChameleonForCausalLM")  # obsolete
+class ChameleonModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHAMELEON
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # ignore image tokenizer for now
+        # TODO: remove this once image support is implemented for Chameleon
+        if name.startswith("model.vqmodel"):
+            return []
+
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        hidden_dim = self.hparams.get("hidden_size")
+
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+        if name.endswith(("q_norm.weight", "q_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
+        if name.endswith(("k_norm.weight", "k_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
+    @staticmethod
+    def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
+        head_dim = hidden_dim // n_heads
+        data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
+        data_torch = data_torch.repeat_interleave(n_heads, 0)
+        return data_torch
 
 
 ###### CONVERSION LOGIC ######
 
+
 # tree of lazy tensors
 class LazyTorchTensor(gguf.LazyBase):
     _tensor_type = torch.Tensor

convert_hf_to_gguf_update.py

@@ -72,6 +72,7 @@ models = [
     {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
     {"name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
     {"name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
+    {"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
@@ -81,6 +82,7 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-v1-en",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
@@ -99,6 +101,7 @@ models = [
     {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
     {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
     {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
+    {"name": "chameleon",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
 ]
 
 

convert_lora_to_gguf.py

@@ -12,6 +12,7 @@ import json
 from math import prod
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
+from transformers import AutoConfig
 
 import torch
 
@@ -230,7 +231,7 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
-        description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
+        description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
     parser.add_argument(
         "--outfile", type=Path,
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
@@ -256,17 +257,23 @@ def parse_args() -> argparse.Namespace:
         help="only print out what will be done, without writing any new files",
     )
     parser.add_argument(
-        "--base", type=Path, required=True,
-        help="directory containing base model file",
+        "--base", type=Path,
+        help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
     )
     parser.add_argument(
         "lora_path", type=Path,
-        help="directory containing LoRA adapter file",
+        help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
     )
 
     return parser.parse_args()
 
 
+def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
+    # normally, adapter does not come with base model config, we need to load it from AutoConfig
+    config = AutoConfig.from_pretrained(hf_model_id)
+    return config.to_dict()
+
+
 if __name__ == '__main__':
     args = parse_args()
     logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
@@ -281,7 +288,7 @@ if __name__ == '__main__':
 
     ftype = ftype_map[args.outtype]
 
-    dir_base_model: Path = args.base
+    dir_base_model: Path | None = args.base
     dir_lora: Path = args.lora_path
     lora_config = dir_lora / "adapter_config.json"
     input_model = dir_lora / "adapter_model.safetensors"
@@ -301,9 +308,29 @@ if __name__ == '__main__':
         input_model = os.path.join(dir_lora, "adapter_model.bin")
         lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
 
+    # load LoRA config
+    with open(lora_config, "r") as f:
+        lparams: dict[str, Any] = json.load(f)
+
     # load base model
-    logger.info(f"Loading base model: {dir_base_model.name}")
-    hparams = Model.load_hparams(dir_base_model)
+    if dir_base_model is None:
+        if "base_model_name_or_path" in lparams:
+            model_id = lparams["base_model_name_or_path"]
+            logger.info(f"Loading base model from Hugging Face: {model_id}")
+            try:
+                hparams = load_hparams_from_hf(model_id)
+            except OSError as e:
+                logger.error(f"Failed to load base model config: {e}")
+                logger.error("Please try downloading the base model and add its path to --base")
+                sys.exit(1)
+        else:
+            logger.error("'base_model_name_or_path' is not found in adapter_config.json")
+            logger.error("Base model config is required. Please download the base model and add its path to --base")
+            sys.exit(1)
+    else:
+        logger.info(f"Loading base model: {dir_base_model.name}")
+        hparams = Model.load_hparams(dir_base_model)
+
     with torch.inference_mode():
         try:
             model_class = Model.from_model_architecture(hparams["architectures"][0])
@@ -323,13 +350,19 @@ if __name__ == '__main__':
                 self.dir_model_card = dir_lora_model
                 self.lora_alpha = float(lora_alpha)
 
+            def set_vocab(self):
+                pass
+
             def set_type(self):
                 self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
                 self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
 
             def set_gguf_parameters(self):
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
-                super().set_gguf_parameters()
+
+            def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+                # Never add extra tensors (e.g. rope_freqs) for LoRA adapters
+                return ()
 
             def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
                 tensor_map: dict[str, PartialLoraTensor] = {}
@@ -344,6 +377,9 @@ if __name__ == '__main__':
                         if ".base_layer.weight" in name:
                             continue
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
+                        if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
+                            logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
+                            logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
                         sys.exit(1)
 
                     if base_name in tensor_map:
@@ -377,9 +413,6 @@ if __name__ == '__main__':
                     yield (dest_name + ".lora_a", lora_a)
                     yield (dest_name + ".lora_b", lora_b)
 
-        with open(lora_config, "r") as f:
-            lparams: dict[str, Any] = json.load(f)
-
         alpha: float = lparams["lora_alpha"]
 
         model_instance = LoraModel(
@@ -392,7 +425,7 @@ if __name__ == '__main__':
             dry_run=args.dry_run,
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
-            is_lora=True,
+            hparams=hparams,
         )
 
         logger.info("Exporting model...")

docs/android.md

@@ -2,55 +2,82 @@
 # Android
 
 ## Build on Android using Termux
-[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
-```
-apt update && apt upgrade -y
-apt install git make cmake
-```
 
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
+[Termux](https://termux.dev/en/) is an Android terminal emulator and Linux environment app (no root required). As of writing, Termux is available experimentally in the Google Play Store; otherwise, it may be obtained directly from the project repo or on F-Droid.
 
-[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
-
-## Building the Project using Android NDK
-Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
-
-Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
-```
-$ mkdir build-android
-$ cd build-android
-$ export NDK=<your_ndk_directory>
-$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
-$ make
-```
-
-Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
-
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+With Termux, you can install and run `llama.cpp` as if the environment were Linux. Once in the Termux shell:
 
 ```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$ apt update && apt upgrade -y
+$ apt install git cmake
 ```
 
-Now, you can start chatting:
+Then, follow the [build instructions](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md), specifically for CMake.
+
+Once the binaries are built, download your model of choice (e.g., from Hugging Face). It's recommended to place it in the `~/` directory for best performance:
+
 ```
-$cd /data/data/com.termux/files/home/bin
-$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+$ curl -L {model-url} -o ~/{model}.gguf
 ```
 
-Here's a demo of an interactive session running on Pixel 5 phone:
+Then, if you are not already in the repo directory, `cd` into `llama.cpp` and:
+
+```
+$ ./build/bin/llama-simple -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
+```
+
+Here, we show `llama-simple`, but any of the executables under `examples` should work, in theory. Be sure to set `context-size` to a reasonable number (say, 4096) to start with; otherwise, memory could spike and kill your terminal.
+
+To see what it might look like visually, here's an old demo of an interactive session running on a Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
+
+## Cross-compile using Android NDK
+It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
+
+Once you're ready and have cloned `llama.cpp`, invoke the following in the project directory:
+
+```
+$ cmake \
+  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DCMAKE_C_FLAGS="-march=armv8.7a" \
+  -DCMAKE_CXX_FLAGS="-march=armv8.7a" \
+  -DGGML_OPENMP=OFF \
+  -DGGML_LLAMAFILE=OFF \
+  -B build-android
+```
+
+Notes:
+  - While later versions of Android NDK ship with OpenMP, it must still be installed by CMake as a dependency, which is not supported at this time
+  - `llamafile` does not appear to support Android devices (see: https://github.com/Mozilla-Ocho/llamafile/issues/325)
+
+The above command should configure `llama.cpp` with the most performant options for modern devices. Even if your device is not running `armv8.7a`, `llama.cpp` includes runtime checks for available CPU features it can use.
+
+Feel free to adjust the Android ABI for your target. Once the project is configured:
+
+```
+$ cmake --build build-android --config Release -j{n}
+$ cmake --install build-android --prefix {install-dir} --config Release
+```
+
+After installing, go ahead and download the model of your choice to your host system. Then:
+
+```
+$ adb shell "mkdir /data/local/tmp/llama.cpp"
+$ adb push {install-dir} /data/local/tmp/llama.cpp/
+$ adb push {model}.gguf /data/local/tmp/llama.cpp/
+$ adb shell
+```
+
+In the `adb shell`:
+
+```
+$ cd /data/local/tmp/llama.cpp
+$ LD_LIBRARY_PATH=lib ./bin/llama-simple -m {model}.gguf -c {context-size} -p "{your-prompt}"
+```
+
+That's it!
+
+Be aware that Android will not find the library path `lib` on its own, so we must specify `LD_LIBRARY_PATH` in order to run the installed executables. Android does support `RPATH` in later API levels, so this could change in the future. Refer to the previous section for information about `context-size` (very important!) and running other `examples`.

docs/backend/SYCL.md

@@ -26,7 +26,7 @@
 
 ### Llama.cpp + SYCL
 
-The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*).
+The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.
 
 ## Recommended Release
 
@@ -111,10 +111,18 @@ SYCL backend supports Intel GPU Family:
 
 **Verified devices**
 
-| Nvidia GPU               | Status  | Verified Model |
-|--------------------------|---------|----------------|
-| Ampere Series            | Support | A100, A4000    |
-| Ampere Series *(Mobile)* | Support | RTX 40 Series  |
+| Nvidia GPU               | Status    | Verified Model |
+|--------------------------|-----------|----------------|
+| Ampere Series            | Supported | A100, A4000    |
+| Ampere Series *(Mobile)* | Supported | RTX 40 Series  |
+
+| AMD GPU                  | Status       | Verified Model |
+|--------------------------|--------------|----------------|
+| Radeon Pro               | Experimental | W6800          |
+| Radeon RX                | Experimental | 6700 XT        |
+
+Note: AMD GPU support is highly experimental and is incompatible with F16.
+Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
 
 ## Docker
 The docker build option is currently limited to *intel GPU* targets.
@@ -186,6 +194,10 @@ Platform #0: Intel(R) OpenCL HD Graphics
 
 In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
 
+- **AMD GPU**
+
+To target AMD GPUs with SYCL, the ROCm stack must be installed first.
+
 2. **Install Intel® oneAPI Base toolkit**
 
 - **For Intel GPU**
@@ -212,6 +224,19 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB
 cmake --build buildWithCublas --config Release
 ```
 
+- **Adding support to AMD GPUs**
+
+**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
+
+**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
+
+```sh
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+# Find your HIPTARGET with rocminfo, under the key 'Name:'
+cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
+cmake --build buildWithrocBLAS --config Release
+```
 
 3. **Verify installation and environment**
 
@@ -223,22 +248,32 @@ sycl-ls
 
 - **Intel GPU**
 
-When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`ext_oneapi_level_zero:gpu:0`] in the sample output below:
+When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:
 
 ```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
-[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
-[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
-[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
+[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu][opencl:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
 ```
 
 - **Nvidia GPU**
 
-Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
+Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below:
+
 ```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
-[opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
-[ext_oneapi_cuda:gpu:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.2]
+[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
+[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
+[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5]
+```
+
+- **AMD GPU**
+
+For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
+
+```
+[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000]
+[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9]
 ```
 
 ### II. Build llama.cpp
@@ -266,6 +301,7 @@ cmake --build build --config Release -j -v
 ```
 
 #### Nvidia GPU
+
 ```sh
 # Export relevant ENV variables
 export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
@@ -283,7 +319,25 @@ cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -
 
 # build all binary
 cmake --build build --config Release -j -v
+```
 
+#### AMD GPU
+
+```sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
+
+# Build LLAMA with rocBLAS acceleration through SYCL
+
+## AMD
+# Use FP32, FP16 is not supported
+# Find your GGML_SYCL_HIP_TARGET with rocminfo, under the key 'Name:'
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_HIP_TARGET=${GGML_SYCL_HIP_TARGET} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# build all binary
+cmake --build build --config Release -j -v
 ```
 
 ### III. Run the inference
@@ -586,11 +640,11 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 #### Build
 
-| Name               | Value                             | Function                                    |
-|--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
-| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
-| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
+| Name               | Value                                 | Function                                    |
+|--------------------|---------------------------------------|---------------------------------------------|
+| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
@@ -636,6 +690,14 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   It's same for other projects including llama.cpp SYCL backend.
 
+- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+
+  Device Memory is not enough.
+
+  |Reason|Solution|
+  |-|-|
+  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
+  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
 
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.

docs/build.md

@@ -198,6 +198,8 @@ The following compilation options are also available to tweak performance:
 
 ### MUSA
 
+This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa).
+
 - Using `make`:
   ```bash
   make GGML_MUSA=1
@@ -209,6 +211,12 @@ The following compilation options are also available to tweak performance:
   cmake --build build --config Release
   ```
 
+The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used.
+
+The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted.
+
+Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
+
 ### hipBLAS
 
 This provides BLAS acceleration on HIP-supported AMD GPUs.

docs/docker.md

@@ -19,8 +19,11 @@ Additionally, there the following images, similar to the above:
 - `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`)
 - `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
 
-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 GitHub 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).
+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 GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 
 ## Usage
 
@@ -84,3 +87,37 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
+
+## Docker With MUSA
+
+Assuming one has the [mt-container-toolkit](https://developer.mthreads.com/musa/native) properly installed on Linux, `muBLAS` should be accessible inside the container.
+
+## Building Docker locally
+
+```bash
+docker build -t local/llama.cpp:full-musa -f .devops/full-musa.Dockerfile .
+docker build -t local/llama.cpp:light-musa -f .devops/llama-cli-musa.Dockerfile .
+docker build -t local/llama.cpp:server-musa -f .devops/llama-server-musa.Dockerfile .
+```
+
+You may want to pass in some different `ARGS`, depending on the MUSA environment supported by your container host, as well as the GPU architecture.
+
+The defaults are:
+
+- `MUSA_VERSION` set to `rc3.1.0`
+
+The resulting images, are essentially the same as the non-MUSA images:
+
+1. `local/llama.cpp:full-musa`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-musa`: This image only includes the main executable file.
+3. `local/llama.cpp:server-musa`: This image only includes the server executable file.
+
+## Usage
+
+After building locally, Usage is similar to the non-MUSA examples, but you'll need to set `mthreads` as default Docker runtime. This can be done by executing `(cd /usr/bin/musa && sudo ./docker setup $PWD)` and verifying the changes by executing `docker info | grep mthreads` on the host machine. You will also want to use the `--n-gpu-layers` flag.
+
+```bash
+docker run -v /path/to/models:/models local/llama.cpp:full-musa --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run -v /path/to/models:/models local/llama.cpp:light-musa -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run -v /path/to/models:/models local/llama.cpp:server-musa -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
+```

examples/CMakeLists.txt

@@ -13,10 +13,8 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 if (EMSCRIPTEN)
 else()
     add_subdirectory(cvector-generator)
-    add_subdirectory(baby-llama)
     add_subdirectory(batched-bench)
     add_subdirectory(batched)
-    add_subdirectory(benchmark)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
@@ -50,6 +48,7 @@ else()
     endif()
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(simple-chat)
     add_subdirectory(speculative)
     add_subdirectory(tokenize)
 endif()

examples/baby-llama/CMakeLists.txt

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

examples/batched-bench/batched-bench.cpp

@@ -15,13 +15,13 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     int is_pp_shared = params.is_pp_shared;
 
@@ -36,7 +36,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = common_model_params_to_llama(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -45,7 +45,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = common_context_params_to_llama(params);
 
     // ensure enough sequences are available
     ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
@@ -74,7 +74,6 @@ int main(int argc, char ** argv) {
                 batch.n_seq_id + i,
                 batch.seq_id   + i,
                 batch.logits   + i,
-                0, 0, 0, // unused
             };
 
             const int ret = llama_decode(ctx, batch_view);
@@ -92,7 +91,7 @@ int main(int argc, char ** argv) {
     // warm up
     {
         for (int i = 0; i < 16; ++i) {
-            llama_batch_add(batch, 0, i, { 0 }, false);
+            common_batch_add(batch, 0, i, { 0 }, false);
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -122,11 +121,11 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                llama_batch_clear(batch);
+                common_batch_clear(batch);
 
                 for (int i = 0; i < pp; ++i) {
                     for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                        llama_batch_add(batch, 0, i, { j }, false);
+                        common_batch_add(batch, 0, i, { j }, false);
                     }
                 }
                 batch.logits[batch.n_tokens - 1] = true;
@@ -151,10 +150,10 @@ int main(int argc, char ** argv) {
                 const auto t_tg_start = ggml_time_us();
 
                 for (int i = 0; i < tg; ++i) {
-                    llama_batch_clear(batch);
+                    common_batch_clear(batch);
 
                     for (int j = 0; j < pl; ++j) {
-                        llama_batch_add(batch, 0, pp + i, { j }, true);
+                        common_batch_add(batch, 0, pp + i, { j }, true);
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {

examples/batched/batched.cpp

@@ -15,16 +15,16 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     // number of parallel batches
     int n_parallel = params.n_parallel;
@@ -39,7 +39,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = common_model_params_to_llama(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -51,13 +51,13 @@ int main(int argc, char ** argv) {
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
-    tokens_list = ::llama_tokenize(model, params.prompt, true);
+    tokens_list = common_tokenize(model, params.prompt, true);
 
     const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = common_context_params_to_llama(params);
 
     ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_predict, n_parallel);
@@ -94,7 +94,7 @@ int main(int argc, char ** argv) {
     LOG("\n");
 
     for (auto id : tokens_list) {
-        LOG("%s", llama_token_to_piece(ctx, id).c_str());
+        LOG("%s", common_token_to_piece(ctx, id).c_str());
     }
 
     // create a llama_batch
@@ -108,7 +108,7 @@ int main(int argc, char ** argv) {
 
     // evaluate the initial prompt
     for (size_t i = 0; i < tokens_list.size(); ++i) {
-        llama_batch_add(batch, tokens_list[i], i, seq_ids, false);
+        common_batch_add(batch, tokens_list[i], i, seq_ids, false);
     }
     GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
@@ -123,8 +123,8 @@ int main(int argc, char ** argv) {
             decoder_start_token_id = llama_token_bos(model);
         }
 
-        llama_batch_clear(batch);
-        llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
+        common_batch_clear(batch);
+        common_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
     }
 
     // llama_decode will output logits only for the last token of the prompt
@@ -161,7 +161,7 @@ int main(int argc, char ** argv) {
 
     while (n_cur <= n_predict) {
         // prepare the next batch
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -185,15 +185,15 @@ int main(int argc, char ** argv) {
 
             // if there is only one stream, we print immediately to stdout
             if (n_parallel == 1) {
-                LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
             }
 
-            streams[i] += llama_token_to_piece(ctx, new_token_id);
+            streams[i] += common_token_to_piece(ctx, new_token_id);
 
             i_batch[i] = batch.n_tokens;
 
             // push this new token for next evaluation
-            llama_batch_add(batch, new_token_id, n_cur, { i }, true);
+            common_batch_add(batch, new_token_id, n_cur, { i }, true);
 
             n_decode += 1;
         }

examples/benchmark/CMakeLists.txt

@@ -1,6 +0,0 @@
-set(TARGET llama-bench-matmult)
-add_executable(${TARGET} benchmark-matmult.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/benchmark/benchmark-matmult.cpp

@@ -1,275 +0,0 @@
-#include "common.h"
-#include "ggml.h"
-
-#include <locale.h>
-#include <assert.h>
-#include <math.h>
-#include <cstring>
-#include <cstdio>
-#include <cinttypes>
-#include <unordered_map>
-#include <queue>
-#include <string.h>
-#include <cassert>
-#include <fstream>
-#include <string>
-#include <iterator>
-#include <algorithm>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-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, nullptr);
-
-    if (plan.work_size > 0) {
-        buf.resize(plan.work_size);
-        plan.work_data = buf.data();
-    }
-
-    ggml_graph_compute(graph, &plan);
-}
-
-static float tensor_sum_elements(const ggml_tensor * tensor) {
-    double sum = 0;
-    if (tensor->type == GGML_TYPE_F32) {
-        for (int j = 0; j < tensor->ne[1]; j++) {
-            for (int k = 0; k < tensor->ne[0]; k++) {
-                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
-            }
-        }
-    }
-    return sum;
-}
-
-static void tensor_dump(const ggml_tensor * tensor, const char * name) {
-    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
-        tensor->type, ggml_type_name(tensor->type),
-        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
-    float sum = tensor_sum_elements(tensor);
-    printf("Sum of tensor %s is %6.2f\n", name, sum);
-}
-
-#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
-
-struct benchmark_params_struct {
-    int     n_threads     = 1;
-    int32_t n_iterations  = 10;
-};
-
-static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -i N, --iter N     number of iterations to use during computation (default: %d)\n", params.n_iterations);
-    fprintf(stderr, "\n");
-}
-
-int main(int argc, char ** argv)  {
-    struct benchmark_params_struct benchmark_params;
-
-    bool invalid_param = false;
-    std::string arg;
-    for (int i = 1; i < argc; i++) {
-        arg = argv[i];
-
-        if (arg == "-t" || arg == "--threads") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            benchmark_params.n_threads = std::stoi(argv[i]);
-        } else if (arg == "-i" || arg == "--iter") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            benchmark_params.n_iterations = std::stoi(argv[i]);
-        }  else if (arg == "-h" || arg == "--help") {
-            print_usage(argc, argv, benchmark_params);
-            exit(0);
-        }
-    }
-    if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        print_usage(argc, argv, benchmark_params);
-        exit(1);
-    }
-
-    print_build_info();
-    printf("Starting Test\n");
-
-    // create the ggml context
-    struct ggml_context * ctx;
-    //const int sizex = 4096;
-    //const int sizey = 11008;
-
-#undef VERBOSE_DEBUGGING
-#ifndef VERBOSE_DEBUGGING
-    const int sizey = 4096;
-    const int sizex = 11008;
-    const int sizez = 128;
-#else
-    /* Working - let's increase size */
-    const int sizey = 1;
-    const int sizex = (8*32);
-    const int sizez = 1;
-
-    /*const int sizey = 1;
-    const int sizex = 3*(8*32);
-    const int sizez = 1;*/
-#endif
-
-    //printf("Memsize required = %i\n", sizex*sizex);
-
-    // TODO: perform the bench for all types or for a user specified type
-    const ggml_type qtype = GGML_TYPE_Q4_1;
-
-    size_t ctx_size = 0;
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += 1024*1024*16;
-
-    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
-
-    struct ggml_init_params params = {
-        /*.mem_size   =*/ ctx_size,
-        /*.mem_buffer =*/ NULL,
-        /* no_alloc   =*/ 0
-    };
-
-    ctx = ggml_init(params);
-    if (!ctx) {
-        fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-        return 1;
-    }
-
-
-    printf("Creating new tensors\n");
-    // printf("Creating new tensor m1\n");
-    struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
-    ggml_set_f32(m11, 1.0f);
-
-    // printf("Creating new tensor m1\n");
-    struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
-    ggml_set_f32(m12, 1.5f);
-
-    // printf("Creating new tensor m2\n");
-    struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
-    ggml_set_f32(m2, 2.0f);
-
-    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
-    // printf("Creating new tensor m11xm2\n");
-    struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
-
-    // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf, m11xm2);
-
-    printf("n_threads=%i\n", benchmark_params.n_threads);
-
-    TENSOR_DUMP(m11);
-    TENSOR_DUMP(m2);
-
-    std::vector<uint8_t> work_buffer;
-
-    ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
-
-    TENSOR_DUMP(ggml_graph_node(gf, 0));
-
-    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
-
-    int32_t nelements = sizex*sizey;
-
-    // Set up a the benchmark matrices
-    // printf("Creating new tensor q11 & Running quantize\n");
-    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
-
-    // Set up a the compute graph
-    // printf("Creating new tensor q31\n");
-    struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
-
-    // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf31, q31);
-
-    // Set up a second graph computation to make sure we override the CPU cache lines
-    // printf("Creating new tensor q12 & Running quantize\n");
-    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
-
-    // printf("Creating new tensor q32\n");
-    struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
-
-    //printf("Creating compute graph\n");
-    struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf32, q32);
-    printf("n_threads=%i\n", benchmark_params.n_threads);
-
-    const int dimx = sizex;
-    const int dimy = sizey;
-    const int dimz = sizez;
-    long long int flops_per_dot_product = dimy + dimy;
-    long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
-    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
-
-
-    // Let's use the F32 result from above as a reference for the quantized multiplication
-    float sum_of_F32_reference = tensor_sum_elements(ggml_graph_node(gf, 0));
-
-    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
-    printf("=====================================================================================\n");
-
-    double  gflops_sum = 0;
-    for (int i=0;i<benchmark_params.n_iterations ;i++) {
-
-        long long int start = ggml_time_us();
-        //printf("Running ggml_graph_compute\n");
-        ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads);
-
-        long long int stop = ggml_time_us();
-        long long int usec = stop-start;
-        double gflops = (double)(flops_per_matrix)/usec/1000.0;
-        gflops_sum += gflops;
-        printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
-            i,
-            benchmark_params.n_threads,
-            sizex, sizey, sizez, flops_per_matrix,
-            usec,gflops);
-
-#ifdef VERBOSE_DEBUGGING
-        TENSOR_DUMP("res",gf31.nodes[0])
-#endif
-
-        // Check that the matrix multiplication result is in the right ballpark
-        // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
-        float sum_of_Q4_result = tensor_sum_elements(ggml_graph_node(gf31, 0));
-        float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
-        float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
-
-        if (delta > allowed_delta)  {
-            printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
-                sum_of_F32_reference,
-                sum_of_Q4_result,
-                delta,
-                allowed_delta
-            );
-            exit(0);
-        }
-
-        // Running a different graph computation to make sure we override the CPU cache lines
-        ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads);
-    }
-    printf("\n");
-    printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
-    printf("=====================================================================================\n");
-}

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

@@ -201,7 +201,7 @@ static void print_sample_weights(TransformerWeights *w){
 
 //////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
 
-struct llama_vocab {
+struct my_llama_vocab {
     using id    = int32_t;
     using token = std::string;
     using ttype = llama_token_type;
@@ -525,7 +525,7 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
+static void load_vocab(const char * filename, const Config * config, struct my_llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
         LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
@@ -583,13 +583,13 @@ static void load_vocab(const char * filename, const Config * config, struct llam
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (llama_vocab::id id=0; id<n_vocab; ++id) {
+        for (my_llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
 
             unsigned char byte_val;
-            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            my_llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
             if (id == UNKNOWN_TOKEN_ID) {
                 text = "<unk>";
                 type = LLAMA_TOKEN_TYPE_UNKNOWN;
@@ -631,7 +631,7 @@ static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const floa
 }
 
 static void save_as_llama_model(
-    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+    struct my_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
@@ -671,7 +671,7 @@ static void save_as_llama_model(
     std::vector<const char*> tokens;
     std::vector<float> scores;
     std::vector<llama_token_type> token_types;
-    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+    for (const my_llama_vocab::token_data & token_data : vocab->id_to_token) {
         tokens.push_back(token_data.text.c_str());
         scores.push_back(token_data.score);
         token_types.push_back(token_data.type);
@@ -872,7 +872,7 @@ static std::string basename(const std::string &path) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_init();
+    common_init();
 
     struct train_params params = get_default_train_params();
     if (!params_parse(argc, argv, &params)) {
@@ -905,7 +905,7 @@ int main(int argc, char ** argv) {
         fclose(file);
     }
 
-    struct llama_vocab vocab;
+    struct my_llama_vocab vocab;
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;

examples/cvector-generator/cvector-generator.cpp

@@ -31,7 +31,7 @@ template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     std::string ret;
     for (; begin != end; ++begin) {
-        ret += llama_token_to_piece(ctx, *begin);
+        ret += common_token_to_piece(ctx, *begin);
     }
 
     return ret;
@@ -272,8 +272,8 @@ struct tokenized_prompt {
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
         const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-        tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
-        tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
+        tokens_pos = common_tokenize(ctx, pos, add_bos, true);
+        tokens_neg = common_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
         padding_seq(ctx, tokens_pos, max_seq_len);
         padding_seq(ctx, tokens_neg, max_seq_len);
@@ -281,7 +281,7 @@ struct tokenized_prompt {
 
     void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
         // TODO: customize padding token
-        std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
+        std::vector<llama_token> pad_tokens = common_tokenize(ctx, " ", false);
         llama_token pad_tok = pad_tokens.back();
         while (tokens.size() < len) {
             tokens.push_back(pad_tok);
@@ -339,7 +339,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
     llama_kv_cache_clear(ctx);
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
     }
@@ -370,7 +370,7 @@ static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const
  * Load prompt files and completion file.
  * Then format each pair of prompt + completion to make an entry.
  */
-static int prepare_entries(gpt_params & params, train_context & ctx_train) {
+static int prepare_entries(common_params & params, train_context & ctx_train) {
     // load prompts
     std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
     std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
@@ -388,9 +388,9 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
         return 1;
     }
 
@@ -413,7 +413,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;

examples/cvector-generator/pca.hpp

@@ -204,13 +204,6 @@ static ggml_status compute_piter(
         ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
     }
 
-// TODO: enable GPU support when support for GGML_OP_SQRT is added
-//#ifdef GGML_USE_METAL
-//    if (ggml_backend_is_metal(model.backend)) {
-//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
-//    }
-//#endif
-
     ggml_status res = ggml_backend_graph_compute(model.backend, gf);
     if (res == GGML_STATUS_SUCCESS) {
         auto extract_i = [](std::string prefix, std::string str) -> int {

examples/embedding/embedding.cpp

@@ -28,7 +28,7 @@ static std::vector<std::string> split_lines(const std::string & s, const std::st
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
     size_t n_tokens = tokens.size();
     for (size_t i = 0; i < n_tokens; i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, true);
+        common_batch_add(batch, tokens[i], i, { seq_id }, true);
     }
 }
 
@@ -74,18 +74,18 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         }
 
         float * out = output + embd_pos * n_embd;
-        llama_embd_normalize(embd, out, n_embd, embd_norm);
+        common_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     params.embedding = true;
     // For non-causal models, batch size must be equal to ubatch size
@@ -95,7 +95,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -122,7 +122,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
     // split the prompt into lines
@@ -135,7 +135,7 @@ int main(int argc, char ** argv) {
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true, false);
+        auto inp = common_tokenize(ctx, prompt, true, true);
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -159,7 +159,7 @@ int main(int argc, char ** argv) {
             LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
             LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
             for (int j = 0; j < (int) inputs[i].size(); j++) {
-                LOG("%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+                LOG("%6d -> '%s'\n", inputs[i][j], common_token_to_piece(ctx, inputs[i][j]).c_str());
             }
             LOG("\n\n");
         }
@@ -199,7 +199,7 @@ int main(int argc, char ** argv) {
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
             e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
             s = 0;
-            llama_batch_clear(batch);
+            common_batch_clear(batch);
         }
 
         // add to batch
@@ -234,6 +234,11 @@ int main(int argc, char ** argv) {
                 }
                 LOG("\n");
             }
+        } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            for (int j = 0; j < n_embd_count; j++) {
+                // NOTE: if you change this log - update the tests in ci/run.sh
+                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+            }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding
             for (int j = 0; j < n_prompts; j++) {
@@ -258,7 +263,7 @@ int main(int argc, char ** argv) {
                 LOG("\n");
                 for (int i = 0; i < n_prompts; i++) {
                     for (int j = 0; j < n_prompts; j++) {
-                        float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                        float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                         LOG("%6.2f ", sim);
                     }
                     LOG("%1.10s", prompts[i].c_str());
@@ -291,7 +296,7 @@ int main(int argc, char ** argv) {
             for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
                 LOG("    [");
                 for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                    float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                     LOG("%6.2f", sim);
                     j++;
                     if (j < n_embd_count) LOG(", "); else break;

examples/eval-callback/eval-callback.cpp

@@ -126,12 +126,12 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
     return true;
 }
 
-static bool run(llama_context * ctx, const gpt_params & params) {
+static bool run(llama_context * ctx, const common_params & params) {
     const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;
     }
@@ -142,13 +142,13 @@ static bool run(llama_context * ctx, const gpt_params & params) {
 int main(int argc, char ** argv) {
     callback_data cb_data;
 
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -160,7 +160,7 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -172,7 +172,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
         LOG_INF("\n");
     }
 

examples/export-lora/export-lora.cpp

@@ -128,7 +128,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<llama_lora_adapter_info> & lora_files,
+            std::vector<common_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -314,9 +314,9 @@ struct lora_merge_ctx {
             // optionally dequantize it
             printf("%s :   + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
             auto nels = ggml_nelements(inp_base);
-            ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
+            const auto * qtype = ggml_get_type_traits(base->type);
             std::vector<uint8_t> dequant_buf(nels * sizeof(float));
-            qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
+            qtype->to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
             ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
         } else {
             ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
@@ -400,9 +400,9 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
         return 1;
     }
 

examples/gen-docs/gen-docs.cpp

@@ -6,42 +6,73 @@
 
 // Export usage message (-h) to markdown format
 
+static void write_table_header(std::ofstream & file) {
+    file << "| Argument | Explanation |\n";
+    file << "| -------- | ----------- |\n";
+}
+
+static void write_table_entry(std::ofstream & file, const common_arg & opt) {
+    file << "| `";
+    // args
+    for (const auto & arg : opt.args) {
+    if (arg == opt.args.front()) {
+            file << arg;
+            if (opt.args.size() > 1) file << ", ";
+        } else {
+            file << arg << (arg != opt.args.back() ? ", " : "");
+        }
+    }
+    // value hint
+    if (opt.value_hint) {
+        std::string md_value_hint(opt.value_hint);
+        string_replace_all(md_value_hint, "|", "\\|");
+        file << " " << md_value_hint;
+    }
+    if (opt.value_hint_2) {
+        std::string md_value_hint_2(opt.value_hint_2);
+        string_replace_all(md_value_hint_2, "|", "\\|");
+        file << " " << md_value_hint_2;
+    }
+    // help text
+    std::string md_help(opt.help);
+    string_replace_all(md_help, "\n", "<br/>");
+    string_replace_all(md_help, "|", "\\|");
+    file << "` | " << md_help << " |\n";
+}
+
+static void write_table(std::ofstream & file, std::vector<common_arg *> & opts) {
+    write_table_header(file);
+    for (const auto & opt : opts) {
+        write_table_entry(file, *opt);
+    }
+}
+
 static void export_md(std::string fname, llama_example ex) {
     std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
 
-    gpt_params params;
-    auto ctx_arg = gpt_params_parser_init(params, ex);
+    common_params params;
+    auto ctx_arg = common_params_parser_init(params, ex);
 
-    file << "| Argument | Explanation |\n";
-    file << "| -------- | ----------- |\n";
+    std::vector<common_arg *> common_options;
+    std::vector<common_arg *> sparam_options;
+    std::vector<common_arg *> specific_options;
     for (auto & opt : ctx_arg.options) {
-        file << "| `";
-        // args
-        for (const auto & arg : opt.args) {
-        if (arg == opt.args.front()) {
-                file << arg;
-                if (opt.args.size() > 1) file << ", ";
-            } else {
-                file << arg << (arg != opt.args.back() ? ", " : "");
-            }
+        // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
+        if (opt.is_sparam) {
+            sparam_options.push_back(&opt);
+        } else if (opt.in_example(ctx_arg.ex)) {
+            specific_options.push_back(&opt);
+        } else {
+            common_options.push_back(&opt);
         }
-        // value hint
-        if (opt.value_hint) {
-            std::string md_value_hint(opt.value_hint);
-            string_replace_all(md_value_hint, "|", "\\|");
-            file << " " << md_value_hint;
-        }
-        if (opt.value_hint_2) {
-            std::string md_value_hint_2(opt.value_hint_2);
-            string_replace_all(md_value_hint_2, "|", "\\|");
-            file << " " << md_value_hint_2;
-        }
-        // help text
-        std::string md_help(opt.help);
-        string_replace_all(md_help, "\n", "<br/>");
-        string_replace_all(md_help, "|", "\\|");
-        file << "` | " << md_help << " |\n";
     }
+
+    file << "**Common params**\n\n";
+    write_table(file, common_options);
+    file << "\n\n**Sampling params**\n\n";
+    write_table(file, sparam_options);
+    file << "\n\n**Example-specific params**\n\n";
+    write_table(file, specific_options);
 }
 
 int main(int, char **) {

examples/gguf-split/gguf-split.cpp

@@ -22,12 +22,20 @@
 #endif
 
 enum split_operation : uint8_t {
-    SPLIT_OP_SPLIT,
-    SPLIT_OP_MERGE,
+    OP_NONE,
+    OP_SPLIT,
+    OP_MERGE,
+};
+
+enum split_mode : uint8_t {
+    MODE_NONE,
+    MODE_TENSOR,
+    MODE_SIZE,
 };
 
 struct split_params {
-    split_operation operation = SPLIT_OP_SPLIT;
+    split_operation operation = OP_NONE;
+    split_mode mode = MODE_NONE;
     size_t n_bytes_split = 0;
     int n_split_tensors = 128;
     std::string input;
@@ -87,59 +95,52 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
         }
 
         bool arg_found = false;
-        bool is_op_set = false;
-        bool is_mode_set = false;
         if (arg == "-h" || arg == "--help") {
             split_print_usage(argv[0]);
             exit(0);
-        }
-        if (arg == "--version") {
+        } else if (arg == "--version") {
             fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
             fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
             exit(0);
-        }
-        if (arg == "--dry-run") {
+        } else if (arg == "--dry-run") {
             arg_found = true;
             params.dry_run = true;
-        }
-        if (arg == "--no-tensor-first-split") {
+        } else if (arg == "--no-tensor-first-split") {
             arg_found = true;
             params.no_tensor_first_split = true;
-        }
-
-        if (is_op_set) {
-            throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
-        }
-        if (arg == "--merge") {
+        } else if (arg == "--merge") {
             arg_found = true;
-            is_op_set = true;
-            params.operation = SPLIT_OP_MERGE;
-        }
-        if (arg == "--split") {
+            if (params.operation != OP_NONE && params.operation != OP_MERGE) {
+                throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+            }
+            params.operation = OP_MERGE;
+        } else if (arg == "--split") {
             arg_found = true;
-            is_op_set = true;
-            params.operation = SPLIT_OP_SPLIT;
-        }
-
-        if (is_mode_set) {
-            throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
-        }
-        if (arg == "--split-max-tensors") {
+            if (params.operation != OP_NONE && params.operation != OP_SPLIT) {
+                throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+            }
+            params.operation = OP_SPLIT;
+        } else if (arg == "--split-max-tensors") {
             if (++arg_idx >= argc) {
                 invalid_param = true;
                 break;
             }
             arg_found = true;
-            is_mode_set = true;
+            if (params.mode != MODE_NONE && params.mode != MODE_TENSOR) {
+                throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+            }
+            params.mode = MODE_TENSOR;
             params.n_split_tensors = atoi(argv[arg_idx]);
-        }
-        if (arg == "--split-max-size") {
+        } else if (arg == "--split-max-size") {
             if (++arg_idx >= argc) {
                 invalid_param = true;
                 break;
             }
             arg_found = true;
-            is_mode_set = true;
+            if (params.mode != MODE_NONE && params.mode != MODE_SIZE) {
+                throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+            }
+            params.mode = MODE_SIZE;
             params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
         }
 
@@ -148,6 +149,15 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
         }
     }
 
+    // the operation is split if not specified
+    if (params.operation == OP_NONE) {
+        params.operation = OP_SPLIT;
+    }
+    // the split mode is by tensor if not specified
+    if (params.mode == MODE_NONE) {
+        params.mode = MODE_TENSOR;
+    }
+
     if (invalid_param) {
         throw std::invalid_argument("error: invalid parameter for argument: " + arg);
     }
@@ -265,13 +275,15 @@ struct split_strategy {
     }
 
     bool should_split(int i_tensor, size_t next_size) {
-        if (params.n_bytes_split > 0) {
+        if (params.mode == MODE_SIZE) {
             // split by max size per file
             return next_size > params.n_bytes_split;
-        } else {
+        } else if (params.mode == MODE_TENSOR) {
             // split by number of tensors per file
             return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
         }
+        // should never happen
+        GGML_ABORT("invalid mode");
     }
 
     void print_info() {
@@ -559,9 +571,9 @@ int main(int argc, const char ** argv) {
     split_params_parse(argc, argv, params);
 
     switch (params.operation) {
-        case SPLIT_OP_SPLIT: gguf_split(params);
+        case OP_SPLIT: gguf_split(params);
             break;
-        case SPLIT_OP_MERGE: gguf_merge(params);
+        case OP_MERGE: gguf_merge(params);
             break;
         default: split_print_usage(argv[0]);
             exit(EXIT_FAILURE);

examples/gritlm/gritlm.cpp

@@ -15,11 +15,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
     llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
     for (uint64_t i = 0; i < sentences.size(); i++) {
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         const std::string input_string = instruction + sentences[i];
 
-        std::vector<llama_token> inputs = llama_tokenize(model, input_string, true, false);
+        std::vector<llama_token> inputs = common_tokenize(model, input_string, true, false);
 
         const int32_t n_toks = inputs.size();
 
@@ -28,7 +28,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         // inputs.push_back(llama_token_eos(model));
 
         // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(model, instruction, true, false).size();
+        const int32_t n_inst = common_tokenize(model, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -40,7 +40,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
@@ -75,7 +75,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         std::vector<float> emb_norm(emb_unorm.size());
-        llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
+        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
         result.push_back(emb_norm);
 
 #ifdef GRIT_DEBUG
@@ -105,16 +105,16 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
-    std::vector<llama_token> inputs = llama_tokenize(model, prompt, false, true);
+    std::vector<llama_token> inputs = common_tokenize(model, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
-        llama_batch_clear(bat);
+        common_batch_clear(bat);
         {
             const int32_t n_inputs = inputs.size();
 
             for (int32_t i = 0; i < n_inputs; i++) {
-                llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
+                common_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
             }
         }
         inputs.clear();
@@ -127,7 +127,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
             break;
         }
 
-        std::string piece = llama_token_to_piece(ctx, token);
+        std::string piece = common_token_to_piece(ctx, token);
         if (stream) {
             std::printf("%s", piece.c_str());
             std::fflush(stdout);
@@ -152,16 +152,16 @@ static std::string gritlm_instruction(const std::string & instruction) {
 }
 
 int main(int argc, char * argv[]) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
+    llama_model_params mparams = common_model_params_to_llama(params);
+    llama_context_params cparams = common_context_params_to_llama(params);
 
     llama_backend_init();
 
@@ -199,10 +199,10 @@ int main(int argc, char * argv[]) {
 
         const int n_embd = llama_n_embd(model);
 
-        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
-        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q1_d0 = common_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q1_d1 = common_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
 
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);

examples/imatrix/imatrix.cpp

@@ -37,13 +37,13 @@ struct Stats {
 class IMatrixCollector {
 public:
     IMatrixCollector() = default;
-    void set_params(gpt_params params) { m_params = std::move(params); }
+    void set_params(common_params params) { m_params = std::move(params); }
     bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
     void save_imatrix(int ncall = -1) const;
     bool load_imatrix(const char * file_name);
 private:
     std::unordered_map<std::string, Stats> m_stats;
-    gpt_params                             m_params;
+    common_params                          m_params;
     std::mutex                             m_mutex;
     int                                    m_last_call = 0;
     std::vector<float>                     m_src1_data;
@@ -428,7 +428,7 @@ static void process_logits(
     }
 }
 
-static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
+static bool compute_imatrix(llama_context * ctx, const common_params & params) {
     const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
     GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
@@ -436,7 +436,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
 
     auto tim2 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
@@ -496,6 +496,8 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
+        llama_batch batch = llama_batch_init(n_batch, 0, 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);
@@ -508,9 +510,14 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            // TODO: use batch.logits to save computations instead of relying on logits_all == true
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            common_batch_clear(batch);
+            for (int i = 0; i < batch_size; i++) {
+                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            }
+
+            if (llama_decode(ctx, batch)) {
                 LOG_ERR("%s : failed to eval\n", __func__);
+                llama_batch_free(batch);
                 return false;
             }
 
@@ -523,6 +530,8 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
             }
         }
 
+        llama_batch_free(batch);
+
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -568,16 +577,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.n_ctx = 512;
     params.logits_all = true;
+    params.escape = false;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
@@ -606,7 +616,7 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -624,7 +634,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
     if (!compute_imatrix(ctx, params)) {

examples/infill/infill.cpp

@@ -35,8 +35,8 @@
 
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
-static gpt_sampler             ** g_smpl;
-static gpt_params               * g_params;
+static common_sampler          ** g_smpl;
+static common_params            * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
@@ -44,7 +44,7 @@ static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;
 
 static void write_logfile(
-    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const llama_context * ctx, const common_params & params, const llama_model * model,
     const std::vector<llama_token> & input_tokens, const std::string & output,
     const std::vector<llama_token> & output_tokens
 ) {
@@ -95,8 +95,13 @@ static void sigint_handler(int signo) {
         } else {
             console::cleanup();
             LOG("\n");
-            gpt_perf_print(*g_ctx, *g_smpl);
+            common_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+
+            // make sure all logs are flushed
+            LOG("Interrupted by user\n");
+            common_log_pause(common_log_main());
+
             _exit(130);
         }
     }
@@ -104,14 +109,14 @@ static void sigint_handler(int signo) {
 #endif
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
     g_params = &params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     auto & sparams = params.sparams;
 
@@ -161,7 +166,7 @@ int main(int argc, char ** argv) {
 
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
-    gpt_sampler  * smpl = nullptr;
+    common_sampler * smpl = nullptr;
 
     g_model = &model;
     g_ctx = &ctx;
@@ -169,7 +174,7 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     model = llama_init.model;
     ctx = llama_init.context;
@@ -190,21 +195,21 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
     const bool add_bos = llama_add_bos_token(model);
     GGML_ASSERT(!llama_add_eos_token(model));
 
     std::vector<llama_token> embd_inp;
     std::vector<llama_token> embd_end;
-    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
-    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+    std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
+    std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-    GGML_ASSERT(llama_token_prefix(model) >= 0);
-    GGML_ASSERT(llama_token_suffix(model) >= 0);
+    GGML_ASSERT(llama_token_fim_pre(model) >= 0);
+    GGML_ASSERT(llama_token_fim_suf(model) >= 0);
 
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
 
     embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
     embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@@ -213,7 +218,7 @@ int main(int argc, char ** argv) {
     }
     embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
 
-    const llama_token middle_token = llama_token_middle(model);
+    const llama_token middle_token = llama_token_fim_mid(model);
     if (middle_token >= 0) {
         embd_inp.push_back(middle_token);
     }
@@ -252,15 +257,15 @@ int main(int argc, char ** argv) {
         LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
         LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (params.n_keep > 0) {
         LOG_INF("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                LOG("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            LOG("'\n");
+            LOG_CNT("'\n");
         }
         LOG_INF("\n");
     }
@@ -293,16 +298,16 @@ int main(int argc, char ** argv) {
             LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    smpl = gpt_sampler_init(model, sparams);
+    smpl = common_sampler_init(model, sparams);
 
-    LOG_INF("sampler seed: %u\n",     gpt_sampler_get_seed(smpl));
+    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
     LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
-    LOG_INF("sampler chain: %s\n",    gpt_sampler_print(smpl).c_str());
+    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
 
     LOG_INF("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("\n");
-    LOG("\n#####  Infill mode  #####\n\n");
+    LOG_INF("\n");
+    LOG_INF("\n#####  Infill mode  #####\n\n");
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
@@ -313,11 +318,11 @@ int main(int argc, char ** argv) {
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
-        LOG("== Running in interactive mode. ==\n");
+        LOG_INF("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-        LOG(       " - Press Ctrl+C to interject at any time.\n");
+        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-        LOG(       "%s\n", control_message);
+        LOG_INF(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -391,7 +396,7 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -406,9 +411,9 @@ int main(int argc, char ** argv) {
         embd.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
+            const llama_token id = common_sampler_sample(smpl, ctx, -1);
 
-            gpt_sampler_accept(smpl, id, true);
+            common_sampler_accept(smpl, id, true);
 
             // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
 
@@ -429,7 +434,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                gpt_sampler_accept(smpl, embd_inp[n_consumed], false);
+                common_sampler_accept(smpl, embd_inp[n_consumed], false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -441,7 +446,7 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo) {
             for (auto id : embd) {
-                const std::string token_str = llama_token_to_piece(ctx, id);
+                const std::string token_str = common_token_to_piece(ctx, id);
                 LOG("%s", token_str.c_str());
 
                 if (embd.size() > 1) {
@@ -460,10 +465,10 @@ int main(int argc, char ** argv) {
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
             // deal with eot token in infill mode
-            if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
+            if ((common_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if (is_interacting && !params.interactive_first) {
                     // print an eot token
-                    LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+                    LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
                 LOG("\n");
                 console::set_display(console::user_input);
@@ -500,11 +505,11 @@ int main(int argc, char ** argv) {
                 }
 
                 // tokenize new prefix and suffix
-                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
-                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+                std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
+                std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
 
                 embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
                 embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@@ -524,7 +529,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of generation tokens in interactive mode
-            else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
+            else if (llama_token_is_eog(model, common_sampler_last(smpl))) {
                 LOG_DBG("found EOS token\n");
 
                 if (params.interactive) {
@@ -574,7 +579,7 @@ int main(int argc, char ** argv) {
 
                     const size_t original_size = embd_inp.size();
 
-                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    const auto line_inp = common_tokenize(ctx, buffer, false);
                     LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
 
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
@@ -582,7 +587,7 @@ int main(int argc, char ** argv) {
                     for (size_t i = original_size; i < embd_inp.size(); ++i) {
                         const llama_token token = embd_inp[i];
                         output_tokens.push_back(token);
-                        output_ss << llama_token_to_piece(ctx, token);
+                        output_ss << common_token_to_piece(ctx, token);
                     }
 
                     n_remain -= line_inp.size();
@@ -596,7 +601,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    gpt_sampler_reset(smpl);
+                    common_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -615,17 +620,17 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+        LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
     }
 
     LOG("\n");
-    gpt_perf_print(ctx, smpl);
+    common_perf_print(ctx, smpl);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     llama_free(ctx);
     llama_free_model(model);
 
-    gpt_sampler_free(smpl);
+    common_sampler_free(smpl);
     llama_backend_free();
 
     return 0;

examples/json_schema_to_grammar.py

@@ -540,7 +540,7 @@ class SchemaConverter:
         return self._add_rule(
             name,
             to_rule(transform()) if self._raw_pattern \
-                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
+                else "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space")
 
 
     def _resolve_ref(self, ref):

examples/llama-bench/llama-bench.cpp

@@ -21,12 +21,6 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
-#include "ggml-cuda.h"
-#include "ggml-sycl.h"
-
-#ifdef GGML_USE_CANN
-#include "ggml-cann.h"
-#endif
 
 #ifdef _WIN32
 #define WIN32_LEAN_AND_MEAN
@@ -82,95 +76,27 @@ static T stdev(const std::vector<T> & v) {
 }
 
 static std::string get_cpu_info() {
-    std::string id;
-#ifdef __linux__
-    FILE * f = fopen("/proc/cpuinfo", "r");
-    if (f) {
-        char buf[1024];
-        while (fgets(buf, sizeof(buf), f)) {
-            if (strncmp(buf, "model name", 10) == 0) {
-                char * p = strchr(buf, ':');
-                if (p) {
-                    p++;
-                    while (std::isspace(*p)) {
-                        p++;
-                    }
-                    while (std::isspace(p[strlen(p) - 1])) {
-                        p[strlen(p) - 1] = '\0';
-                    }
-                    id = p;
-                    break;
-                }
-            }
-        }
-        fclose(f);
-    }
-#elif defined(_WIN32)
-    HKEY hKey;
-    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
-                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
-                     0,
-                     KEY_READ,
-                     &hKey) != ERROR_SUCCESS) {
-        // fail to open registry key
-        return "";
-    }
-    char cpu_brand[256];
-    DWORD cpu_brand_size = sizeof(cpu_brand);
-    if (RegQueryValueExA(hKey,
-                        TEXT("ProcessorNameString"),
-                        NULL,
-                        NULL,
-                        (LPBYTE)cpu_brand,
-                        &cpu_brand_size) == ERROR_SUCCESS) {
-        id.assign(cpu_brand, cpu_brand_size);
-        if (id.find('\0') != std::string::npos) {
-            id.resize(id.find('\0'));
+    std::vector<std::string> cpu_list;
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        auto * dev = ggml_backend_dev_get(i);
+        auto dev_type = ggml_backend_dev_type(dev);
+        if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU || dev_type == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
+            cpu_list.push_back(ggml_backend_dev_description(dev));
         }
     }
-    RegCloseKey(hKey);
-#endif
-    // TODO: other platforms
-    return id;
+    return join(cpu_list, ", ");
 }
 
 static std::string get_gpu_info() {
-    std::string id;
-#ifdef GGML_USE_CUDA
-    int count = ggml_backend_cuda_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
+    std::vector<std::string> gpu_list;
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        auto * dev = ggml_backend_dev_get(i);
+        auto dev_type = ggml_backend_dev_type(dev);
+        if (dev_type == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            gpu_list.push_back(ggml_backend_dev_description(dev));
         }
     }
-#endif
-#ifdef GGML_USE_SYCL
-    int count = ggml_backend_sycl_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_sycl_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
-        }
-    }
-#endif
-#ifdef GGML_USE_CANN
-    uint32_t count = ggml_backend_cann_get_device_count();
-    for (uint32_t i = 0; i < count; i++) {
-        char buf[128];
-        ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
-        }
-    }
-#endif
-    // TODO: other backends
-    return id;
+    return join(gpu_list, ", ");
 }
 
 // command line params
@@ -304,9 +230,9 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  --cpu-strict <0|1>                        (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str());
     printf("  --poll <0...100>                          (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>                  (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-#ifdef GGML_USE_RPC
-    printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
-#endif
+    if (llama_supports_rpc()) {
+        printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
+    }
     printf("  -sm, --split-mode <none|layer|row>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
     printf("  -mg, --main-gpu <i>                       (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>              (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
@@ -439,6 +365,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
+            if (invalid_param) {
+                break;
+            }
             params.type_k.insert(params.type_k.end(), types.begin(), types.end());
         } else if (arg == "-ctv" || arg == "--cache-type-v") {
             if (++i >= argc) {
@@ -455,6 +384,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
+            if (invalid_param) {
+                break;
+            }
             params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
@@ -491,14 +423,12 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
-#ifdef GGML_USE_RPC
-        } else if (arg == "-rpc" || arg == "--rpc") {
+        } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rpc_servers.push_back(argv[i]);
-#endif
         } else if (arg == "-sm" || arg == "--split-mode") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -520,6 +450,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 modes.push_back(mode);
             }
+            if (invalid_param) {
+                break;
+            }
             params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {
@@ -931,29 +864,15 @@ struct test {
     }
 
     static std::string get_backend() {
-        if (cuda) {
-            return GGML_CUDA_NAME;
+        std::vector<std::string> backends;
+        for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+            auto * reg = ggml_backend_reg_get(i);
+            std::string name = ggml_backend_reg_name(reg);
+            if (name != "CPU") {
+                backends.push_back(ggml_backend_reg_name(reg));
+            }
         }
-        if (vulkan) {
-            return "Vulkan";
-        }
-        if (kompute) {
-            return "Kompute";
-        }
-        if (metal) {
-            return "Metal";
-        }
-        if (sycl) {
-            return GGML_SYCL_NAME;
-        }
-        if (gpu_blas) {
-            return "GPU BLAS";
-        }
-        if (blas) {
-            return "BLAS";
-        }
-
-        return "CPU";
+        return backends.empty() ? "CPU" : join(backends, ",");
     }
 
     static const std::vector<std::string> & get_fields() {
@@ -1421,7 +1340,7 @@ struct sql_printer : public printer {
     }
 };
 
-static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
+static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model = llama_get_model(ctx);
@@ -1437,14 +1356,14 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
-        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens));
         n_processed += n_tokens;
     }
 
     llama_synchronize(ctx);
 }
 
-static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
+static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model = llama_get_model(ctx);
@@ -1453,7 +1372,7 @@ static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads)
     llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
 
     for (int i = 0; i < n_gen; i++) {
-        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
+        llama_decode(ctx, llama_batch_get_one(&token, 1));
         llama_synchronize(ctx);
         token = std::rand() % n_vocab;
     }
@@ -1589,13 +1508,13 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
             }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             if (params.progress) {
                 fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
             }
-            test_gen(ctx, 1, 0, t.n_threads);
+            test_gen(ctx, 1, t.n_threads);
         }
 
         for (int i = 0; i < params.reps; i++) {
@@ -1607,13 +1526,13 @@ int main(int argc, char ** argv) {
                 if (params.progress) {
                     fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps);
                 }
-                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+                test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
                 if (params.progress) {
                     fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps);
                 }
-                test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
+                test_gen(ctx, t.n_gen, t.n_threads);
             }
 
             uint64_t t_ns = get_time_ns() - t_start;

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

@@ -18,6 +18,7 @@ android {
         }
         externalNativeBuild {
             cmake {
+                arguments += "-DLLAMA_BUILD_COMMON=ON"
                 arguments += "-DCMAKE_BUILD_TYPE=Release"
                 cppFlags += listOf()
                 arguments += listOf()

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

@@ -186,11 +186,11 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
     for (nri = 0; nri < nr; nri++) {
         LOGi("Benchmark prompt processing (pp)");
 
-        llama_batch_clear(*batch);
+        common_batch_clear(*batch);
 
         const int n_tokens = pp;
         for (i = 0; i < n_tokens; i++) {
-            llama_batch_add(*batch, 0, i, { 0 }, false);
+            common_batch_add(*batch, 0, i, { 0 }, false);
         }
 
         batch->logits[batch->n_tokens - 1] = true;
@@ -210,9 +210,9 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
-            llama_batch_clear(*batch);
+            common_batch_clear(*batch);
             for (j = 0; j < pl; j++) {
-                llama_batch_add(*batch, 0, i, { j }, true);
+                common_batch_add(*batch, 0, i, { j }, true);
             }
 
             LOGi("llama_decode() text generation: %d", i);
@@ -283,9 +283,6 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
         nullptr,
         nullptr,
         nullptr,
-        0,
-        0,
-        0,
     };
 
     if (embd) {
@@ -357,7 +354,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
 
-    const auto tokens_list = llama_tokenize(context, text, 1);
+    const auto tokens_list = common_tokenize(context, text, 1);
 
     auto n_ctx = llama_n_ctx(context);
     auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
@@ -369,14 +366,14 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     }
 
     for (auto id : tokens_list) {
-        LOGi("%s", llama_token_to_piece(context, id).c_str());
+        LOGi("%s", common_token_to_piece(context, id).c_str());
     }
 
-    llama_batch_clear(*batch);
+    common_batch_clear(*batch);
 
     // evaluate the initial prompt
     for (auto i = 0; i < tokens_list.size(); i++) {
-        llama_batch_add(*batch, tokens_list[i], i, { 0 }, false);
+        common_batch_add(*batch, tokens_list[i], i, { 0 }, false);
     }
 
     // llama_decode will output logits only for the last token of the prompt
@@ -419,7 +416,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         return nullptr;
     }
 
-    auto new_token_chars = llama_token_to_piece(context, new_token_id);
+    auto new_token_chars = common_token_to_piece(context, new_token_id);
     cached_token_chars += new_token_chars;
 
     jstring new_token = nullptr;
@@ -431,8 +428,8 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         new_token = env->NewStringUTF("");
     }
 
-    llama_batch_clear(*batch);
-    llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
+    common_batch_clear(*batch);
+    common_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
 
     env->CallVoidMethod(intvar_ncur, la_int_var_inc);
 

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

@@ -46,7 +46,6 @@ actor LlamaContext {
         let sparams = llama_sampler_chain_default_params()
         self.sampling = llama_sampler_chain_init(sparams)
         llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
-        llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
         llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
     }
 

examples/llama.vim

@@ -1,135 +1,783 @@
-" Requires an already running llama.cpp server
-" To install either copy or symlink to ~/.vim/autoload/llama.vim
-" Then start with either :call llama#doLlamaGen(),
-" or add a keybind to your vimrc such as
-" nnoremap Z :call llama#doLlamaGen()<CR>
-" Similarly, you could add an insert mode keybind with
-" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
+" LLM-based text completion using llama.cpp
 "
-" g:llama_api_url, g:llama_api_key and g:llama_overrides can be configured in your .vimrc
-" let g:llama_api_url = "192.168.1.10:8080"
-" llama_overrides can also be set through buffer/window scopes. For instance
-" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
-" Could be added to your .vimrc to automatically set a lower temperature when
-" editing a python script
-" Additionally, an override dict can be stored at the top of a file
-" !*{"stop": ["User:"]}
-" Could be added to the start of your chatlog.txt to set the stopping token
-" These parameter dicts are merged together from lowest to highest priority:
-" server default -> g:llama_overrides -> w:llama_overrides ->
-" b:llama_overrides -> in file (!*) overrides
+" requires:
+"
+"   - neovim or vim
+"   - curl
+"   - llama.cpp server instance
+"   - FIM-compatible model
+"
+" sample config:
+"
+"   - Tab       - accept the current suggestion
+"   - Shift+Tab - accept just the first line of the suggestion
+"   - Ctrl+F    - toggle FIM completion manually
+"
+" make symlink or copy this file to ~/.config/nvim/autoload/llama.vim
+"
+" start the llama.cpp server with a FIM-compatible model. for example:
+"
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"
+"   --batch-size [512, model max context]
+"
+"     adjust the batch size to control how much of the provided local context will be used during the inference
+"     lower values will use smaller part of the context around the cursor, which will result in faster processing
+"
+"   --ubatch-size [64, 2048]
+"
+"     chunks the batch into smaller chunks for faster processing
+"     depends on the specific hardware. use llama-bench to profile and determine the best size
+"
+"   --cache-reuse (ge:llama_config.n_predict, 1024]
+"
+"     this should be either 0 (disabled) or strictly larger than g:llama_config.n_predict
+"     using non-zero value enables context reuse on the server side which dramatically improves the performance at
+"     large contexts. a value of 256 should be good for all cases
+"
+" run this once to initialise llama.vim:
+"
+"   :call llama#init()
+"
+" more info: https://github.com/ggerganov/llama.cpp/pull/9787
 "
-" Sublists (like logit_bias and stop) are overridden, not merged
-" Example override:
-" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
-if !exists("g:llama_api_url")
-    let g:llama_api_url= "127.0.0.1:8080"
-endif
-if !exists("g:llama_overrides")
-   let g:llama_overrides = {}
-endif
-const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
-const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
-let s:linedict = {}
 
-func s:callbackHandler(bufn, channel, msg)
-   if len(a:msg) < 3
-      return
-   elseif a:msg[0] == "d"
-      let l:msg = a:msg[6:-1]
-   else
-      let l:msg = a:msg
-   endif
-   let l:decoded_msg = json_decode(l:msg)
-   let l:newtext = split(l:decoded_msg['content'], "\n", 1)
-   if len(l:newtext) > 0
-      call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
-   else
-      echo "nothing genned"
-   endif
-   if len(newtext) > 1
-      let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
-      let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
-   endif
-   if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
-       echo "Finished generation"
-   endif
-endfunction
+" colors (adjust to your liking)
+highlight llama_hl_hint guifg=#ff772f ctermfg=202
+highlight llama_hl_info guifg=#77ff2f ctermfg=119
 
-func llama#doLlamaGen()
-   if exists("b:job")
-      if job_status(b:job) == "run"
-         call job_stop(b:job)
-         return
-      endif
-   endif
+" general parameters:
+"
+"   endpoint:         llama.cpp server endpoint
+"   n_prefix:         number of lines before the cursor location to include in the local prefix
+"   n_suffix:         number of lines after  the cursor location to include in the local suffix
+"   n_predict:        max number of tokens to predict
+"   t_max_prompt_ms:  max alloted time for the prompt processing (TODO: not yet supported)
+"   t_max_predict_ms: max alloted time for the prediction
+"   show_info:        show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
+"   auto_fim:         trigger FIM completion automatically on cursor movement
+"   max_line_suffix:  do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
+"
+" ring buffer of chunks, accumulated with time upon:
+"
+"  - completion request
+"  - yank
+"  - entering a buffer
+"  - leaving a buffer
+"  - writing a file
+"
+" parameters for the ring-buffer with extra context:
+"
+"   ring_n_chunks:    max number of chunks to pass as extra context to the server (0 to disable)
+"   ring_chunk_size:  max size of the chunks (in number of lines)
+"                     note: adjust these numbers so that you don't overrun your context
+"                           at ring_n_chunks = 64 and ring_chunk_size = 64 you need ~32k context
+"   ring_scope:       the range around the cursor position (in number of lines) for gathering chunks after FIM
+"   ring_update_ms:   how often to process queued chunks in normal mode
+"
+let s:default_config = {
+    \ 'endpoint':         'http://127.0.0.1:8012/infill',
+    \ 'n_prefix':         256,
+    \ 'n_suffix':         64,
+    \ 'n_predict':        128,
+    \ 't_max_prompt_ms':  500,
+    \ 't_max_predict_ms': 3000,
+    \ 'show_info':        2,
+    \ 'auto_fim':         v:true,
+    \ 'max_line_suffix':  8,
+    \ 'ring_n_chunks':    64,
+    \ 'ring_chunk_size':  64,
+    \ 'ring_scope':       1024,
+    \ 'ring_update_ms':   1000,
+    \ }
 
-   let l:cbuffer = bufnr("%")
-   let s:linedict[l:cbuffer] = line('$')
-   let l:buflines = getbufline(l:cbuffer, 1, 1000)
-   let l:querydata = copy(s:querydata)
-   call extend(l:querydata, g:llama_overrides)
-   if exists("w:llama_overrides")
-      call extend(l:querydata, w:llama_overrides)
-   endif
-   if exists("b:llama_overrides")
-      call extend(l:querydata, b:llama_overrides)
-   endif
-   if l:buflines[0][0:1] == '!*'
-      let l:userdata = json_decode(l:buflines[0][2:-1])
-      call extend(l:querydata, l:userdata)
-      let l:buflines = l:buflines[1:-1]
-   endif
-   let l:querydata.prompt = join(l:buflines, "\n")
-   let l:curlcommand = copy(s:curlcommand)
-   if exists("g:llama_api_key")
-       call extend(l:curlcommand, ['--header', 'Authorization: Bearer ' .. g:llama_api_key])
-   endif
-   let l:curlcommand[2] = json_encode(l:querydata)
-   let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
-endfunction
+let g:llama_config = get(g:, 'llama_config', s:default_config)
 
-" Echos the tokkenization of the provided string , or cursor to end of word
-" Onus is placed on the user to include the preceding space
-func llama#tokenizeWord(...)
-    if (a:0 > 0)
-        let l:input = a:1
-    else
-        exe "normal \"*ye"
-        let l:input = @*
-    endif
-    let l:querydata = {"content": l:input}
-    let l:curlcommand = copy(s:curlcommand)
-    let l:curlcommand[2] = json_encode(l:querydata)
-    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
-   let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
-endfunction
-
-func s:tokenizeWordCallback(plaintext, channel, msg)
-    echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
-endfunction
-
-
-" Echos the token count of the entire buffer (or provided string)
-" Example usage :echo llama#tokenCount()
-func llama#tokenCount(...)
-    if (a:0 > 0)
-        let l:buflines = a:1
-    else
-        let l:buflines = getline(1,1000)
-        if l:buflines[0][0:1] == '!*'
-            let l:buflines = l:buflines[1:-1]
+function! s:get_indent(str)
+    let l:count = 0
+    for i in range(len(a:str))
+        if a:str[i] == "\t"
+            let l:count += &tabstop - 1
+        else
+            break
         endif
-        let l:buflines = join(l:buflines, "\n")
-    endif
-    let l:querydata = {"content": l:buflines}
-    let l:curlcommand = copy(s:curlcommand)
-    let l:curlcommand[2] = json_encode(l:querydata)
-    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
-   let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
+    endfor
+    return l:count
 endfunction
 
-func s:tokenCountCallback(channel, msg)
-    let resp = json_decode(a:msg)
-    echo len(resp.tokens)
+function! s:rand(i0, i1) abort
+    return a:i0 + rand() % (a:i1 - a:i0 + 1)
+endfunction
+
+function! llama#init()
+    if !executable('curl')
+        echohl WarningMsg
+        echo 'llama.vim requires the "curl" command to be available'
+        echohl None
+        return
+    endif
+
+    let s:pos_x = 0 " cursor position upon start of completion
+    let s:pos_y = 0
+
+    let s:line_cur = ''
+
+    let s:line_cur_prefix = ''
+    let s:line_cur_suffix = ''
+
+    let s:ring_chunks = [] " current set of chunks used as extra context
+    let s:ring_queued = [] " chunks that are queued to be sent for processing
+    let s:ring_n_evict = 0
+
+    let s:hint_shown = v:false
+    let s:pos_y_pick = -9999 " last y where we picked a chunk
+    let s:pos_dx = 0
+    let s:content = []
+    let s:can_accept = v:false
+
+    let s:timer_fim = -1
+    let s:t_fim_start = reltime() " used to measure total FIM time
+    let s:t_last_move = reltime() " last time the cursor moved
+
+    let s:current_job = v:null
+
+    let s:ghost_text_nvim = exists('*nvim_buf_get_mark')
+    let s:ghost_text_vim = has('textprop')
+
+    if s:ghost_text_vim
+        let s:hlgroup_hint = 'llama_hl_hint'
+        let s:hlgroup_info = 'llama_hl_info'
+
+        if empty(prop_type_get(s:hlgroup_hint))
+            call prop_type_add(s:hlgroup_hint, {'highlight': s:hlgroup_hint})
+        endif
+        if empty(prop_type_get(s:hlgroup_info))
+            call prop_type_add(s:hlgroup_info, {'highlight': s:hlgroup_info})
+        endif
+    endif
+
+    augroup llama
+        autocmd!
+        autocmd InsertEnter     * inoremap <expr> <silent> <C-F> llama#fim_inline(v:false)
+        autocmd InsertLeavePre  * call llama#fim_cancel()
+
+        autocmd CursorMoved     * call s:on_move()
+        autocmd CursorMovedI    * call s:on_move()
+        autocmd CompleteChanged * call llama#fim_cancel()
+
+        if g:llama_config.auto_fim
+            autocmd CursorMovedI * call llama#fim(v:true)
+        endif
+
+        " gather chunks upon yanking
+        autocmd TextYankPost    * if v:event.operator ==# 'y' | call s:pick_chunk(v:event.regcontents, v:false, v:true) | endif
+
+        " gather chunks upon entering/leaving a buffer
+        autocmd BufEnter        * call timer_start(100, {-> s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)})
+        autocmd BufLeave        * call                      s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
+
+        " gather chunk upon saving the file
+        autocmd BufWritePost    * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
+    augroup END
+
+    silent! call llama#fim_cancel()
+
+    " init background update of the ring buffer
+    if g:llama_config.ring_n_chunks > 0
+        call s:ring_update()
+    endif
+endfunction
+
+" compute how similar two chunks of text are
+" 0 - no similarity, 1 - high similarity
+" TODO: figure out something better
+function! s:chunk_sim(c0, c1)
+    let l:lines0 = len(a:c0)
+    let l:lines1 = len(a:c1)
+
+    let l:common = 0
+
+    for l:line0 in a:c0
+        for l:line1 in a:c1
+            if l:line0 == l:line1
+                let l:common += 1
+                break
+            endif
+        endfor
+    endfor
+
+    return 2.0 * l:common / (l:lines0 + l:lines1)
+endfunction
+
+" pick a random chunk of size g:llama_config.ring_chunk_size from the provided text and queue it for processing
+"
+" no_mod   - do not pick chunks from buffers with pending changes
+" do_evict - evict chunks that are very similar to the new one
+"
+function! s:pick_chunk(text, no_mod, do_evict)
+    " do not pick chunks from buffers with pending changes or buffers that are not files
+    if a:no_mod && (getbufvar(bufnr('%'), '&modified') || !buflisted(bufnr('%')) || !filereadable(expand('%')))
+        return
+    endif
+
+    " if the extra context option is disabled - do nothing
+    if g:llama_config.ring_n_chunks <= 0
+        return
+    endif
+
+    " don't pick very small chunks
+    if len(a:text) < 3
+        return
+    endif
+
+    if len(a:text) + 1 < g:llama_config.ring_chunk_size
+        let l:chunk = a:text
+    else
+        let l:l0 = s:rand(0, max([0, len(a:text) - g:llama_config.ring_chunk_size/2]))
+        let l:l1 = min([l:l0 + g:llama_config.ring_chunk_size/2, len(a:text)])
+
+        let l:chunk = a:text[l:l0:l:l1]
+    endif
+
+    let l:chunk_str = join(l:chunk, "\n") . "\n"
+
+    " check if this chunk is already added
+    let l:exist = v:false
+
+    for i in range(len(s:ring_chunks))
+        if s:ring_chunks[i].data == l:chunk
+            let l:exist = v:true
+            break
+        endif
+    endfor
+
+    for i in range(len(s:ring_queued))
+        if s:ring_queued[i].data == l:chunk
+            let l:exist = v:true
+            break
+        endif
+    endfor
+
+    if l:exist
+        return
+    endif
+
+    " evict queued chunks that are very similar to the new one
+    for i in range(len(s:ring_queued) - 1, 0, -1)
+        if s:chunk_sim(s:ring_queued[i].data, l:chunk) > 0.9
+            if a:do_evict
+                call remove(s:ring_queued, i)
+                let s:ring_n_evict += 1
+            else
+                return
+            endif
+        endif
+    endfor
+
+    " also from s:ring_chunks
+    for i in range(len(s:ring_chunks) - 1, 0, -1)
+        if s:chunk_sim(s:ring_chunks[i].data, l:chunk) > 0.9
+            if a:do_evict
+                call remove(s:ring_chunks, i)
+                let s:ring_n_evict += 1
+            else
+                return
+            endif
+        endif
+    endfor
+
+    " TODO: become parameter ?
+    if len(s:ring_queued) == 16
+        call remove(s:ring_queued, 0)
+    endif
+
+    call add(s:ring_queued, {'data': l:chunk, 'str': l:chunk_str, 'time': reltime(), 'filename': expand('%')})
+
+    "let &statusline = 'extra context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
+endfunction
+
+" picks a queued chunk, sends it for processing and adds it to s:ring_chunks
+" called every g:llama_config.ring_update_ms
+function! s:ring_update()
+    call timer_start(g:llama_config.ring_update_ms, {-> s:ring_update()})
+
+    " update only if in normal mode or if the cursor hasn't moved for a while
+    if mode() !=# 'n' && reltimefloat(reltime(s:t_last_move)) < 3.0
+        return
+    endif
+
+    if len(s:ring_queued) == 0
+        return
+    endif
+
+    " move the first queued chunk to the ring buffer
+    if len(s:ring_chunks) == g:llama_config.ring_n_chunks
+        call remove(s:ring_chunks, 0)
+    endif
+
+    call add(s:ring_chunks, remove(s:ring_queued, 0))
+
+    "let &statusline = 'updated context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
+
+    " send asynchronous job with the new extra context so that it is ready for the next FIM
+    let l:extra_context = []
+    for l:chunk in s:ring_chunks
+        call add(l:extra_context, {
+            \ 'text':     l:chunk.str,
+            \ 'time':     l:chunk.time,
+            \ 'filename': l:chunk.filename
+            \ })
+    endfor
+
+    " no samplers needed here
+    let l:request = json_encode({
+        \ 'input_prefix':     "",
+        \ 'input_suffix':     "",
+        \ 'input_extra':      l:extra_context,
+        \ 'prompt':           "",
+        \ 'n_predict':        1,
+        \ 'temperature':      0.0,
+        \ 'stream':           v:false,
+        \ 'samplers':         ["temperature"],
+        \ 'cache_prompt':     v:true,
+        \ 't_max_prompt_ms':  1,
+        \ 't_max_predict_ms': 1
+        \ })
+
+    let l:curl_command = [
+        \ "curl",
+        \ "--silent",
+        \ "--no-buffer",
+        \ "--request", "POST",
+        \ "--url", g:llama_config.endpoint,
+        \ "--header", "Content-Type: application/json",
+        \ "--data", l:request
+        \ ]
+
+    " no callbacks because we don't need to process the response
+    if s:ghost_text_nvim
+        call jobstart(l:curl_command, {})
+    elseif s:ghost_text_vim
+        call job_start(l:curl_command, {})
+    endif
+endfunction
+
+" necessary for 'inoremap <expr>'
+function! llama#fim_inline(is_auto) abort
+    call llama#fim(a:is_auto)
+    return ''
+endfunction
+
+" the main FIM call
+" takes local context around the cursor and sends it together with the extra context to the server for completion
+function! llama#fim(is_auto) abort
+    " we already have a suggestion for the current cursor position
+    if s:hint_shown && !a:is_auto
+        call llama#fim_cancel()
+        return
+    endif
+
+    call llama#fim_cancel()
+
+    " avoid sending repeated requests too fast
+    if reltimefloat(reltime(s:t_fim_start)) < 0.6
+        if s:timer_fim != -1
+            call timer_stop(s:timer_fim)
+            let s:timer_fim = -1
+        endif
+
+        let s:t_fim_start = reltime()
+        let s:timer_fim = timer_start(600, {-> llama#fim(v:true)})
+        return
+    endif
+
+    let s:t_fim_start = reltime()
+
+    let s:content = []
+    let s:can_accept = v:false
+
+    let s:pos_x = col('.') - 1
+    let s:pos_y = line('.')
+    let l:max_y = line('$')
+
+    let l:lines_prefix = getline(max([1, s:pos_y - g:llama_config.n_prefix]), s:pos_y - 1)
+    let l:lines_suffix = getline(s:pos_y + 1, min([l:max_y, s:pos_y + g:llama_config.n_suffix]))
+
+    let s:line_cur = getline('.')
+
+    let s:line_cur_prefix = strpart(s:line_cur, 0, s:pos_x)
+    let s:line_cur_suffix = strpart(s:line_cur, s:pos_x)
+
+    if a:is_auto && len(s:line_cur_suffix) > g:llama_config.max_line_suffix
+        return
+    endif
+
+    let l:prefix = ""
+        \ . join(l:lines_prefix, "\n")
+        \ . "\n"
+
+    let l:prompt = ""
+        \ . s:line_cur_prefix
+
+    let l:suffix = ""
+        \ . s:line_cur_suffix
+        \ . "\n"
+        \ . join(l:lines_suffix, "\n")
+        \ . "\n"
+
+    " prepare the extra context data
+    let l:extra_context = []
+    for l:chunk in s:ring_chunks
+        call add(l:extra_context, {
+            \ 'text':     l:chunk.str,
+            \ 'time':     l:chunk.time,
+            \ 'filename': l:chunk.filename
+            \ })
+    endfor
+
+    " the indentation of the current line
+    let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
+
+    let l:request = json_encode({
+        \ 'input_prefix':     l:prefix,
+        \ 'input_suffix':     l:suffix,
+        \ 'input_extra':      l:extra_context,
+        \ 'prompt':           l:prompt,
+        \ 'n_predict':        g:llama_config.n_predict,
+        \ 'n_indent':         l:indent,
+        \ 'top_k':            40,
+        \ 'top_p':            0.99,
+        \ 'stream':           v:false,
+        \ 'samplers':         ["top_k", "top_p", "infill"],
+        \ 'cache_prompt':     v:true,
+        \ 't_max_prompt_ms':  g:llama_config.t_max_prompt_ms,
+        \ 't_max_predict_ms': g:llama_config.t_max_predict_ms
+        \ })
+
+    let l:curl_command = [
+        \ "curl",
+        \ "--silent",
+        \ "--no-buffer",
+        \ "--request", "POST",
+        \ "--url", g:llama_config.endpoint,
+        \ "--header", "Content-Type: application/json",
+        \ "--data", l:request
+        \ ]
+
+    if s:current_job != v:null
+        if s:ghost_text_nvim
+            call jobstop(s:current_job)
+        elseif s:ghost_text_vim
+            call job_stop(s:current_job)
+        endif
+    endif
+
+    " send the request asynchronously
+    if s:ghost_text_nvim
+        let s:current_job = jobstart(l:curl_command, {
+            \ 'on_stdout': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
+            \ 'on_exit':   function('s:fim_on_exit'),
+            \ 'stdout_buffered': v:true
+            \ })
+    elseif s:ghost_text_vim
+        let s:current_job = job_start(l:curl_command, {
+            \ 'out_cb': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
+            \ 'exit_cb':   function('s:fim_on_exit')
+            \ })
+    endif
+
+    " TODO: per-file location
+    let l:delta_y = abs(s:pos_y - s:pos_y_pick)
+
+    " gather some extra context nearby and process it in the background
+    " only gather chunks if the cursor has moved a lot
+    " TODO: something more clever? reranking?
+    if a:is_auto && l:delta_y > 32
+        " expand the prefix even further
+        call s:pick_chunk(getline(max([1,       s:pos_y - g:llama_config.ring_scope]), max([1,       s:pos_y - g:llama_config.n_prefix])), v:false, v:false)
+
+        " pick a suffix chunk
+        call s:pick_chunk(getline(min([l:max_y, s:pos_y + g:llama_config.n_suffix]),   min([l:max_y, s:pos_y + g:llama_config.n_suffix + g:llama_config.ring_chunk_size])), v:false, v:false)
+
+        let s:pos_y_pick = s:pos_y
+    endif
+endfunction
+
+" if first_line == v:true accept only the first line of the response
+function! llama#fim_accept(first_line)
+    " insert the suggestion at the cursor location
+    if s:can_accept && len(s:content) > 0
+        call setline(s:pos_y, s:line_cur[:(s:pos_x - 1)] . s:content[0])
+        if len(s:content) > 1
+            if !a:first_line
+                call append(s:pos_y, s:content[1:-1])
+            endif
+        endif
+
+        " move the cursor to the end of the accepted text
+        if !a:first_line && len(s:content) > 1
+            call cursor(s:pos_y + len(s:content) - 1, s:pos_x + s:pos_dx + 1)
+        else
+            call cursor(s:pos_y, s:pos_x + len(s:content[0]))
+        endif
+    endif
+
+    call llama#fim_cancel()
+endfunction
+
+function! llama#fim_cancel()
+    let s:hint_shown = v:false
+
+    " clear the virtual text
+    let l:bufnr = bufnr('%')
+
+    if s:ghost_text_nvim
+        let l:id_vt_fim = nvim_create_namespace('vt_fim')
+        call nvim_buf_clear_namespace(l:bufnr, l:id_vt_fim,  0, -1)
+    elseif s:ghost_text_vim
+        call prop_remove({'type': s:hlgroup_hint, 'all': v:true})
+        call prop_remove({'type': s:hlgroup_info, 'all': v:true})
+    endif
+
+    " remove the mappings
+    silent! iunmap <buffer> <Tab>
+    silent! iunmap <buffer> <S-Tab>
+    silent! iunmap <buffer> <Esc>
+endfunction
+
+function! s:on_move()
+    let s:t_last_move = reltime()
+
+    call llama#fim_cancel()
+endfunction
+
+" callback that processes the FIM result from the server and displays the suggestion
+function! s:fim_on_stdout(pos_x, pos_y, is_auto, job_id, data, event = v:null)
+    if s:ghost_text_nvim
+        let l:raw = join(a:data, "\n")
+    elseif s:ghost_text_vim
+        let l:raw = a:data
+    endif
+
+    if len(l:raw) == 0
+        return
+    endif
+
+    if a:pos_x != col('.') - 1 || a:pos_y != line('.')
+        return
+    endif
+
+    " show the suggestion only in insert mode
+    if mode() !=# 'i'
+        return
+    endif
+
+    let s:pos_x = a:pos_x
+    let s:pos_y = a:pos_y
+
+    let s:can_accept = v:true
+    let l:has_info   = v:false
+
+    if s:can_accept && v:shell_error
+        if !a:is_auto
+            call add(s:content, "<| curl error: is the server on? |>")
+        endif
+        let s:can_accept = v:false
+    endif
+
+    let l:n_prompt    = 0
+    let l:t_prompt_ms = 1.0
+    let l:s_prompt    = 0
+
+    let l:n_predict    = 0
+    let l:t_predict_ms = 1.0
+    let l:s_predict    = 0
+
+    " get the generated suggestion
+    if s:can_accept
+        let l:response = json_decode(l:raw)
+
+        for l:part in split(get(l:response, 'content', ''), "\n", 1)
+            call add(s:content, l:part)
+        endfor
+
+        " remove trailing new lines
+        while len(s:content) > 0 && s:content[-1] == ""
+            call remove(s:content, -1)
+        endwhile
+
+        let l:generation_settings = get(l:response, 'generation_settings', {})
+        let l:n_ctx = get(l:generation_settings, 'n_ctx', 0)
+
+        let l:n_cached  = get(l:response, 'tokens_cached', 0)
+        let l:truncated = get(l:response, 'truncated', v:false)
+
+        " if response.timings is available
+        if len(get(l:response, 'timings', {})) > 0
+            let l:has_info = v:true
+            let l:timings  = get(l:response, 'timings', {})
+
+            let l:n_prompt    = get(l:timings, 'prompt_n', 0)
+            let l:t_prompt_ms = get(l:timings, 'prompt_ms', 1)
+            let l:s_prompt    = get(l:timings, 'prompt_per_second', 0)
+
+            let l:n_predict    = get(l:timings, 'predicted_n', 0)
+            let l:t_predict_ms = get(l:timings, 'predicted_ms', 1)
+            let l:s_predict    = get(l:timings, 'predicted_per_second', 0)
+        endif
+    endif
+
+    if len(s:content) == 0
+        call add(s:content, "")
+        let s:can_accept = v:false
+    endif
+
+    if len(s:content) == 0
+        return
+    endif
+
+    " NOTE: the following is logic for discarding predictions that repeat existing text
+    "       the code is quite ugly and there is very likely a simpler and more canonical way to implement this
+    "
+    "       still, I wonder if there is some better way that avoids having to do these special hacks?
+    "       on one hand, the LLM 'sees' the contents of the file before we start editing, so it is normal that it would
+    "       start generating whatever we have given it via the extra context. but on the other hand, it's not very
+    "       helpful to re-generate the same code that is already there
+
+    " truncate the suggestion if the first line is empty
+    if len(s:content) == 1 && s:content[0] == ""
+        let s:content = [""]
+    endif
+
+    " ... and the next lines are repeated
+    if len(s:content) > 1 && s:content[0] == "" && s:content[1:] == getline(s:pos_y + 1, s:pos_y + len(s:content) - 1)
+        let s:content = [""]
+    endif
+
+    " truncate the suggestion if it repeats the suffix
+    if len(s:content) == 1 && s:content[0] == s:line_cur_suffix
+        let s:content = [""]
+    endif
+
+    " find the first non-empty line (strip whitespace)
+    let l:cmp_y = s:pos_y + 1
+    while l:cmp_y < line('$') && getline(l:cmp_y) =~? '^\s*$'
+        let l:cmp_y += 1
+    endwhile
+
+    if (s:line_cur_prefix . s:content[0]) == getline(l:cmp_y)
+        " truncate the suggestion if it repeats the next line
+        if len(s:content) == 1
+            let s:content = [""]
+        endif
+
+        " ... or if the second line of the suggestion is the prefix of line l:cmp_y + 1
+        if len(s:content) == 2 && s:content[-1] == getline(l:cmp_y + 1)[:len(s:content[-1]) - 1]
+            let s:content = [""]
+        endif
+
+        " ... or if the middle chunk of lines of the suggestion is the same as [l:cmp_y + 1, l:cmp_y + len(s:content) - 1)
+        if len(s:content) > 2 && join(s:content[1:-1], "\n") == join(getline(l:cmp_y + 1, l:cmp_y + len(s:content) - 1), "\n")
+            let s:content = [""]
+        endif
+    endif
+
+    " keep only lines that have the same or larger whitespace prefix as s:line_cur_prefix
+    "let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
+    "for i in range(1, len(s:content) - 1)
+    "    if strlen(matchstr(s:content[i], '^\s*')) < l:indent
+    "        let s:content = s:content[:i - 1]
+    "        break
+    "    endif
+    "endfor
+
+    let s:pos_dx = len(s:content[-1])
+
+    let s:content[-1] .= s:line_cur_suffix
+
+    call llama#fim_cancel()
+
+    " display virtual text with the suggestion
+    let l:bufnr = bufnr('%')
+
+    if s:ghost_text_nvim
+        let l:id_vt_fim = nvim_create_namespace('vt_fim')
+    endif
+
+    " construct the info message
+    if g:llama_config.show_info > 0 && l:has_info
+        let l:prefix = '   '
+
+        if l:truncated
+            let l:info = printf("%s | WARNING: the context is full: %d / %d, increase the server context size or reduce g:llama_config.ring_n_chunks",
+                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
+                \ l:n_cached, l:n_ctx
+                \ )
+        else
+            let l:info = printf("%s | c: %d / %d, r: %d / %d, e: %d, q: %d / 16 | p: %d (%.2f ms, %.2f t/s) | g: %d (%.2f ms, %.2f t/s) | t: %.2f ms",
+                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
+                \ l:n_cached,  l:n_ctx, len(s:ring_chunks), g:llama_config.ring_n_chunks, s:ring_n_evict, len(s:ring_queued),
+                \ l:n_prompt,  l:t_prompt_ms,  l:s_prompt,
+                \ l:n_predict, l:t_predict_ms, l:s_predict,
+                \ 1000.0 * reltimefloat(reltime(s:t_fim_start))
+                \ )
+        endif
+
+        if g:llama_config.show_info == 1
+            " display the info in the statusline
+            let &statusline = l:info
+            let l:info = ''
+        endif
+    endif
+
+    " display the suggestion and append the info to the end of the first line
+    if s:ghost_text_nvim
+        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, s:pos_x - 1, {
+            \ 'virt_text': [[s:content[0], 'llama_hl_hint'], [l:info, 'llama_hl_info']],
+            \ 'virt_text_win_col': virtcol('.') - 1
+            \ })
+
+        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, 0, {
+            \ 'virt_lines': map(s:content[1:], {idx, val -> [[val, 'llama_hl_hint']]}),
+            \ 'virt_text_win_col': virtcol('.')
+            \ })
+    elseif s:ghost_text_vim
+        let l:new_suffix = s:content[0]
+        if !empty(l:new_suffix)
+            call prop_add(s:pos_y, s:pos_x + 1, {
+                        \ 'type': s:hlgroup_hint,
+                        \ 'text': l:new_suffix
+                        \ })
+        endif
+        for line in s:content[1:]
+            call prop_add(s:pos_y, 0, {
+                        \ 'type': s:hlgroup_hint,
+                        \ 'text': line,
+                        \ 'text_padding_left': s:get_indent(line),
+                        \ 'text_align': 'below'
+                        \ })
+        endfor
+        if !empty(l:info)
+            call prop_add(s:pos_y, 0, {
+                        \ 'type': s:hlgroup_info,
+                        \ 'text': l:info,
+                        \ 'text_padding_left': col('$'),
+                        \ 'text_wrap': 'truncate'
+                        \ })
+        endif
+    endif
+
+    " setup accept shortcuts
+    inoremap <buffer> <Tab>   <C-O>:call llama#fim_accept(v:false)<CR>
+    inoremap <buffer> <S-Tab> <C-O>:call llama#fim_accept(v:true)<CR>
+
+    let s:hint_shown = v:true
+endfunction
+
+function! s:fim_on_exit(job_id, exit_code, event = v:null)
+    if a:exit_code != 0
+        echom "Job failed with exit code: " . a:exit_code
+    endif
+
+    let s:current_job = v:null
 endfunction

examples/llava/clip.cpp

@@ -4,6 +4,7 @@
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
 #include "ggml.h"
+#include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
@@ -2444,12 +2445,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
     }
 
-#ifdef GGML_USE_METAL
-    if (ggml_backend_is_metal(ctx->backend)) {
-        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
-    }
-#endif
-
     ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor

examples/llava/convert_image_encoder_to_gguf.py

@@ -274,7 +274,7 @@ fout.add_bool("clip.use_gelu", use_gelu)
 
 
 if has_llava_projector:
-    model.vision_model.encoder.layers.pop(-1)  # pyright: ignore[reportAttributeAccessIssue]
+    model.vision_model.encoder.layers.pop(-1)
     projector = torch.load(args.llava_projector)
     for name, data in projector.items():
         name = get_tensor_name(name)
@@ -288,7 +288,7 @@ if has_llava_projector:
 
     print("Projector tensors added\n")
 
-state_dict = model.state_dict()  # pyright: ignore[reportAttributeAccessIssue]
+state_dict = model.state_dict()
 for name, data in state_dict.items():
     if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
         # we don't need this

examples/llava/llava-cli.cpp

@@ -20,7 +20,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
@@ -37,21 +37,21 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
 
 static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
     std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
     eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
     return true;
 }
 
-static const char * sample(struct gpt_sampler * smpl,
+static const char * sample(struct common_sampler * smpl,
                            struct llama_context * ctx_llama,
                            int * n_past) {
-    const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
-    gpt_sampler_accept(smpl, id, true);
+    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
+    common_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
-        ret = llama_token_to_piece(ctx_llama, id);
+        ret = common_token_to_piece(ctx_llama, id);
     }
     eval_id(ctx_llama, id, n_past);
     return ret.c_str();
@@ -120,7 +120,7 @@ static void print_usage(int, char ** argv) {
     LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
-static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) {
+static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
 
     // load and preprocess the image
     llava_image_embed * embed = NULL;
@@ -146,7 +146,7 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
     return embed;
 }
 
-static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) {
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
     int n_past = 0;
 
     const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
@@ -159,16 +159,16 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
         LOG_INF("system_prompt: %s\n", system_prompt.c_str());
         if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
         LOG_INF("user_prompt: %s\n", user_prompt.c_str());
         if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     } else {
@@ -176,9 +176,9 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
         if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     }
@@ -191,7 +191,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     LOG("\n");
 
-    struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
     if (!smpl) {
         LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);
@@ -211,15 +211,15 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         fflush(stdout);
     }
 
-    gpt_sampler_free(smpl);
+    common_sampler_free(smpl);
     LOG("\n");
 }
 
-static struct llama_model * llava_init(gpt_params * params) {
+static struct llama_model * llava_init(common_params * params) {
     llama_backend_init();
     llama_numa_init(params->numa);
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+    llama_model_params model_params = common_model_params_to_llama(*params);
 
     llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
@@ -229,7 +229,7 @@ static struct llama_model * llava_init(gpt_params * params) {
     return model;
 }
 
-static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
     const char * clip_path = params->mmproj.c_str();
 
     auto prompt = params->prompt;
@@ -240,7 +240,7 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    llama_context_params ctx_params = common_context_params_to_llama(*params);
     ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
 
     llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
@@ -272,13 +272,13 @@ static void llava_free(struct llava_context * ctx_llava) {
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         print_usage(argc, argv);

examples/llava/llava.cpp

@@ -401,6 +401,39 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     return true;
 }
 
+struct llava_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
 bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
     int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
 
@@ -409,8 +442,9 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
-        if (llama_decode(ctx_llama, batch)) {
+        float * embd = image_embed->embed+i*n_embd;
+        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
+        if (llama_decode(ctx_llama, llava_batch.batch)) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
@@ -432,7 +466,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
     if (!image_embed_result) {
         clip_image_u8_free(img);
-        LOG_ERR("%s: coulnd't embed the image\n", __func__);
+        LOG_ERR("%s: couldn't embed the image\n", __func__);
         return NULL;
     }
 

examples/llava/minicpmv-cli.cpp

@@ -25,11 +25,11 @@ static void show_additional_info(int /*argc*/, char ** argv) {
     LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
-static struct llama_model * llava_init(gpt_params * params) {
+static struct llama_model * llava_init(common_params * params) {
     llama_backend_init();
     llama_numa_init(params->numa);
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+    llama_model_params model_params = common_model_params_to_llama(*params);
 
     llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
@@ -39,13 +39,13 @@ static struct llama_model * llava_init(gpt_params * params) {
     return model;
 }
 
-static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
     auto prompt = params->prompt;
     if (prompt.empty()) {
         prompt = "describe the image in detail.";
     }
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    llama_context_params ctx_params = common_context_params_to_llama(*params);
     if (params->n_ctx < 2048) {
         // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
         LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
@@ -79,7 +79,7 @@ static void llava_free(struct llava_context * ctx_llava) {
     llama_backend_free();
 }
 
-static struct clip_ctx * clip_init_context(gpt_params * params) {
+static struct clip_ctx * clip_init_context(common_params * params) {
     const char * clip_path = params->mmproj.c_str();
 
     auto prompt = params->prompt;
@@ -97,7 +97,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
@@ -114,7 +114,7 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
 
 static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
     std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
     return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
 }
 
@@ -129,7 +129,7 @@ static void process_eval_image_embed(struct llava_context * ctx_llava, const str
     llava_image_embed_free(slice_embed);
 }
 
-static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) {
+static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
     std::string system_prompt;
     int idx = 0;
     int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
@@ -162,22 +162,22 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
     LOG_INF("%s: image token past: %d\n", __func__, n_past);
 }
 
-static const char * sample(struct gpt_sampler * smpl,
+static const char * sample(struct common_sampler * smpl,
                            struct llama_context * ctx_llama,
                            int * n_past) {
-    const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
-    gpt_sampler_accept(smpl, id, true);
+    const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
+    common_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
-        ret = llama_token_to_piece(ctx_llama, id);
+        ret = common_token_to_piece(ctx_llama, id);
     }
     eval_id(ctx_llama, id, n_past);
     return ret.c_str();
 }
 
-static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
+static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
     auto * ctx_clip = clip_init_context(params);
     auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
     if (!embeds) {
@@ -213,7 +213,7 @@ static struct llava_context * minicpmv_init(gpt_params * params, const std::stri
     return ctx_llava;
 }
 
-static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, const std::string & prompt, int & n_past, bool is_first = false){
+static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
     std::string user_prompt = prompt;
     int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
     if (!is_first) {
@@ -237,11 +237,11 @@ static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_par
 
     LOG_INF("\n");
 
-    struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
     return smpl;
 }
 
-static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampler * smpl, int &n_past){
+static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
 
     const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
     return tmp;
@@ -250,13 +250,13 @@ static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampl
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     if (params.mmproj.empty() || (params.image.empty())) {
         show_additional_info(argc, argv);
@@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
 
                 fflush(stdout);
             }
-            gpt_sampler_free(smpl);
+            common_sampler_free(smpl);
         }else {
             while (true) {
                 LOG("<user>");
@@ -309,7 +309,7 @@ int main(int argc, char ** argv) {
                     if (strstr(response.c_str(), "<user>")) break; // minicpm-v
                     fflush(stdout);
                 }
-                gpt_sampler_free(smpl);
+                common_sampler_free(smpl);
             }
         }
         printf("\n");

examples/lookahead/lookahead.cpp

@@ -37,13 +37,13 @@ struct ngram_container {
 };
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     const int W = 15; // lookahead window
     const int N = 5;  // n-gram size
@@ -56,7 +56,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the target model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -65,7 +65,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> inp;
     std::vector<llama_token> all;
 
-    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
     all = inp;
 
     const int max_context_size     = llama_n_ctx(ctx);
@@ -79,7 +79,7 @@ int main(int argc, char ** argv) {
     LOG("\n\n");
 
     for (auto id : inp) {
-        LOG("%s", llama_token_to_piece(ctx, id).c_str());
+        LOG("%s", common_token_to_piece(ctx, id).c_str());
     }
 
     fflush(stderr);
@@ -89,8 +89,8 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // eval the prompt
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
         llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
@@ -115,7 +115,7 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
 
     // target model sampling context
-    struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
 
     // verification n-grams
     std::vector<ngram_data> ngrams_cur(G);
@@ -156,12 +156,12 @@ int main(int argc, char ** argv) {
 
     // sample first token
     {
-        id = gpt_sampler_sample(smpl, ctx, 0);
+        id = common_sampler_sample(smpl, ctx, 0);
 
-        gpt_sampler_accept(smpl, id, true);
+        common_sampler_accept(smpl, id, true);
 
         {
-            const std::string token_str = llama_token_to_piece(ctx, id);
+            const std::string token_str = common_token_to_piece(ctx, id);
 
             LOG("%s", token_str.c_str());
             fflush(stdout);
@@ -172,7 +172,7 @@ int main(int argc, char ** argv) {
         // debug
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            llama_kv_cache_dump_view_seqs(kvc_view, 40);
+            common_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
@@ -201,10 +201,10 @@ int main(int argc, char ** argv) {
         //                                                      V  V  V  V  V  V
         //                                                             id
         {
-            llama_batch_clear(batch);
+            common_batch_clear(batch);
 
             // current token - first token of the first level
-            llama_batch_add(batch, id, n_past, seq_id_all, true);
+            common_batch_add(batch, id, n_past, seq_id_all, true);
 
             // verification n-grams - queue this before the lookahead tokens for less KV cache fragmentation
             {
@@ -229,7 +229,7 @@ int main(int argc, char ** argv) {
                         ngrams_cur[g].tokens [j + 1] = t;
                         ngrams_cur[g].i_batch[j + 1] = batch.n_tokens;
 
-                        llama_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
+                        common_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
                     }
                 }
             }
@@ -241,13 +241,13 @@ int main(int argc, char ** argv) {
                     seq_id_look[j] = i + j + 1;
                 }
 
-                llama_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
+                common_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
             }
 
             // fill the rest of the levels
             for (int j = 1; j < N - 1; j++) {
                 for (int i = 0; i < W; i++) {
-                    llama_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
+                    common_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
                 }
             }
         }
@@ -281,13 +281,13 @@ int main(int argc, char ** argv) {
             }
 
             // sample the next token
-            id = gpt_sampler_sample(smpl, ctx, i_batch);
+            id = common_sampler_sample(smpl, ctx, i_batch);
 
-            gpt_sampler_accept(smpl, id, true);
+            common_sampler_accept(smpl, id, true);
 
             // print
             {
-                const std::string token_str = llama_token_to_piece(ctx, id);
+                const std::string token_str = common_token_to_piece(ctx, id);
 
                 if (v == 0) {
                     LOG("%s", token_str.c_str());
@@ -327,7 +327,7 @@ int main(int argc, char ** argv) {
             // print known n-grams starting with token id (debug)
             if (0 && v == 0) {
                 if (ngrams_observed.cnt[id] > 0) {
-                    LOG("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], llama_token_to_piece(ctx, id).c_str());
+                    LOG("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], common_token_to_piece(ctx, id).c_str());
                 }
 
                 for (int i = 0; i < ngrams_observed.cnt[id]; i++) {
@@ -336,7 +336,7 @@ int main(int argc, char ** argv) {
                     const int idx = id*(N - 1)*G + i*(N - 1);
 
                     for (int j = 0; j < N - 1; j++) {
-                        const std::string token_str = llama_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
+                        const std::string token_str = common_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
 
                         LOG("%s", token_str.c_str());
                     }
@@ -358,7 +358,7 @@ int main(int argc, char ** argv) {
                 if (v == 0) {
                     // sample from the last level
                     for (int i = 0; i < W; i++) {
-                        tokens_j[N - 2][i] = gpt_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
+                        tokens_j[N - 2][i] = common_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
                     }
                 } else {
                     for (int i = 0; i < W; i++) {
@@ -466,9 +466,9 @@ int main(int argc, char ** argv) {
     LOG_INF("n_accept  = %d\n", n_accept);
 
     LOG_INF("\n");
-    gpt_perf_print(ctx, smpl);
+    common_perf_print(ctx, smpl);
 
-    gpt_sampler_free(smpl);
+    common_sampler_free(smpl);
 
     llama_kv_cache_view_free(&kvc_view);
 

examples/lookup/lookup-create.cpp

@@ -12,9 +12,9 @@
 #include <vector>
 
 int main(int argc, char ** argv){
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
@@ -23,7 +23,7 @@ int main(int argc, char ** argv){
     llama_numa_init(params.numa);
 
     // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -31,15 +31,15 @@ int main(int argc, char ** argv){
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
     fprintf(stderr, "%s: tokenization done\n", __func__);
 
 
-    llama_ngram_cache ngram_cache;
-    llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
+    common_ngram_cache ngram_cache;
+    common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
     fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
 
-    llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+    common_ngram_cache_save(ngram_cache, params.lookup_cache_static);
 
     return 0;
 }

examples/lookup/lookup-merge.cpp

@@ -33,15 +33,15 @@ int main(int argc, char ** argv){
     }
 
     fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
-    llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]);
+    common_ngram_cache ngram_cache_merged = common_ngram_cache_load(args[0]);
 
     for (size_t i = 1; i < args.size()-1; ++i) {
         fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
-        llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]);
+        common_ngram_cache ngram_cache = common_ngram_cache_load(args[i]);
 
-        llama_ngram_cache_merge(ngram_cache_merged, ngram_cache);
+        common_ngram_cache_merge(ngram_cache_merged, ngram_cache);
     }
 
     fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
-    llama_ngram_cache_save(ngram_cache_merged, args.back());
+    common_ngram_cache_save(ngram_cache_merged, args.back());
 }

examples/lookup/lookup-stats.cpp

@@ -13,13 +13,13 @@
 #include <vector>
 
 int main(int argc, char ** argv){
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     const int n_draft = params.n_draft;
 
@@ -28,18 +28,18 @@ int main(int argc, char ** argv){
     llama_numa_init(params.numa);
 
     // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
 
-    llama_ngram_cache ngram_cache_context;
-    llama_ngram_cache ngram_cache_dynamic;
-    llama_ngram_cache ngram_cache_static;
+    common_ngram_cache ngram_cache_context;
+    common_ngram_cache ngram_cache_dynamic;
+    common_ngram_cache ngram_cache_static;
     int64_t t_draft_flat_us = 0;
     int64_t t_draft_us = 0;
 
@@ -48,7 +48,7 @@ int main(int argc, char ** argv){
 
         if (!params.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
                 LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
                 exit(1);
@@ -57,7 +57,7 @@ int main(int argc, char ** argv){
 
         if (!params.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
@@ -86,7 +86,7 @@ int main(int argc, char ** argv){
 
             {
                 const int64_t t_start_draft_us = ggml_time_us();
-                llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+                common_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
                 t_draft_us += ggml_time_us() - t_start_draft_us;
             }
 
@@ -105,7 +105,7 @@ int main(int argc, char ** argv){
 
                 {
                     const int64_t t_start_draft_us = ggml_time_us();
-                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
                     t_draft_us += ggml_time_us() - t_start_draft_us;
                 }
             }
@@ -115,7 +115,7 @@ int main(int argc, char ** argv){
                 pseudo_output.push_back(inp_slice[pseudo_output.size()]);
                 {
                     const int64_t t_start_draft_us = ggml_time_us();
-                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
                     t_draft_us += ggml_time_us() - t_start_draft_us;
                 }
             }
@@ -133,7 +133,7 @@ int main(int argc, char ** argv){
         }
 
         // After each chunk, update the dynamic ngram cache with the context ngram cache:
-        llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+        common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
         ngram_cache_context.clear();
     }
 

examples/lookup/lookup.cpp

@@ -13,13 +13,13 @@
 #include <vector>
 
 int main(int argc, char ** argv){
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     // max. number of additional tokens to draft if match is found
     const int n_draft = params.n_draft;
@@ -31,29 +31,29 @@ int main(int argc, char ** argv){
     llama_numa_init(params.numa);
 
     // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
 
-    llama_ngram_cache ngram_cache_context;
-    llama_ngram_cache ngram_cache_dynamic;
-    llama_ngram_cache ngram_cache_static;
+    common_ngram_cache ngram_cache_context;
+    common_ngram_cache ngram_cache_dynamic;
+    common_ngram_cache ngram_cache_static;
     int64_t t_draft_flat_us = 0;
     int64_t t_draft_us = 0;
 
     {
         // Fill up context ngram cache with tokens from user input:
         const int64_t t_start_draft_us = ggml_time_us();
-        llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
+        common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
 
         if (!params.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
                 LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
                 exit(1);
@@ -62,7 +62,7 @@ int main(int argc, char ** argv){
 
         if (!params.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
@@ -80,7 +80,7 @@ int main(int argc, char ** argv){
     LOG("\n\n");
 
     for (auto id : inp) {
-        LOG("%s", llama_token_to_piece(ctx, id).c_str());
+        LOG("%s", common_token_to_piece(ctx, id).c_str());
     }
 
     fflush(stderr);
@@ -89,8 +89,8 @@ int main(int argc, char ** argv){
 
     const auto t_enc_start = ggml_time_us();
 
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     const auto t_enc_end = ggml_time_us();
 
@@ -102,7 +102,7 @@ int main(int argc, char ** argv){
 
     bool has_eos = false;
 
-    struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
 
     std::vector<llama_token> draft;
 
@@ -117,7 +117,7 @@ int main(int argc, char ** argv){
         // debug
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            llama_kv_cache_dump_view_seqs(kvc_view, 40);
+            common_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         // print current draft sequence
@@ -126,11 +126,11 @@ int main(int argc, char ** argv){
         int i_dft = 0;
         while (true) {
             // sample from the target model
-            llama_token id = gpt_sampler_sample(smpl, ctx, i_dft);
+            llama_token id = common_sampler_sample(smpl, ctx, i_dft);
 
-            gpt_sampler_accept(smpl, id, true);
+            common_sampler_accept(smpl, id, true);
 
-            const std::string token_str = llama_token_to_piece(ctx, id);
+            const std::string token_str = common_token_to_piece(ctx, id);
 
             if (!params.use_color) {
                 LOG("%s", token_str.c_str());
@@ -152,7 +152,7 @@ int main(int argc, char ** argv){
                 {
                     // Update context ngram cache with the newly accepted token:
                     const int64_t t_start_draft_us = ggml_time_us();
-                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                    common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
                     t_draft_us += ggml_time_us() - t_start_draft_us;
                 }
 
@@ -178,7 +178,7 @@ int main(int argc, char ** argv){
             {
                 // Update context ngram cache with the newly accepted token:
                 const int64_t t_start_draft_us = ggml_time_us();
-                llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
                 t_draft_us += ggml_time_us() - t_start_draft_us;
             }
             break;
@@ -192,18 +192,18 @@ int main(int argc, char ** argv){
         // clean the cache of draft tokens that weren't accepted
         llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
 
-        llama_batch_clear(batch_tgt);
-        llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
+        common_batch_clear(batch_tgt);
+        common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
 
         // Draft already contains a single token sampled from the model:
         GGML_ASSERT(draft.size() == 1);
         GGML_ASSERT(draft[0] == inp.back());
         const int64_t t_start_draft_us = ggml_time_us();
 
-        llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+        common_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
 
         for (size_t i = 1; i < draft.size(); ++i) {
-            llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+            common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
         }
 
         t_draft_us += ggml_time_us() - t_start_draft_us;
@@ -218,8 +218,8 @@ int main(int argc, char ** argv){
     auto t_dec_end = ggml_time_us();
 
     // Update dynamic ngram cache with context ngram cache and save it to disk:
-    llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-    llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+    common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+    common_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
 
     LOG("\n\n");
 
@@ -237,9 +237,9 @@ int main(int argc, char ** argv){
     LOG_INF("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_INF("\ntarget:\n\n");
-    gpt_perf_print(ctx, smpl);
+    common_perf_print(ctx, smpl);
 
-    gpt_sampler_free(smpl);
+    common_sampler_free(smpl);
 
     llama_batch_free(batch_tgt);
 

examples/main/README.md

@@ -69,7 +69,7 @@ In this section, we cover the most commonly used options for running the `llama-
 -   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 -   `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\'
 -   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has.
--   -   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 
 ## Input Prompts
 
@@ -187,6 +187,30 @@ Use the `--no-penalize-nl` option to disable newline penalization when applying
 
 Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
 
+### DRY Repetition Penalty
+
+DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
+
+- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
+- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
+- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
+- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
+- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
+
+The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
+
+The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
+
+The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
+
+The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
+
+The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
+
+DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
+
+Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
+
 ### Top-K Sampling
 
 -   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
@@ -211,14 +235,6 @@ The Min-P sampling method was designed as an alternative to Top-P, and aims to e
 
 Example usage: `--min-p 0.05`
 
-### Tail-Free Sampling (TFS)
-
--   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
-
-Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
-
-Example usage: `--tfs 0.95`
-
 ### Locally Typical Sampling
 
 -   `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
@@ -241,6 +257,19 @@ The `--mirostat-ent` option sets the Mirostat target entropy (tau), which repres
 
 Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
 
+### XTC Sampling
+
+-   `--xtc-probability N`: Sets the chance for token removal (checked once on sampler start) (default: 0.0).
+-   `--xtc-threshold N`: Sets a minimum probability threshold for tokens to be removed (default: 0.1).
+
+Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
+
+By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
+
+Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
+
+Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
+
 ### Logit Bias
 
 -   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
@@ -284,10 +313,6 @@ These options help improve the performance and memory usage of the LLaMA models.
 
  These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
-### Memory Float 32
-
--   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
-
 ### Batch Size
 
 -   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
@@ -308,6 +333,15 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
 
+## LoRA (Low-Rank Adaptation) adapters
+
+-   `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
+-   `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
+
+You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
+
+LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
+
 ## Additional Options
 
 These options provide extra functionality and customization when running the LLaMA models:
@@ -316,6 +350,4 @@ These options provide extra functionality and customization when running the LLa
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
--   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 -   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

examples/main/main.cpp

@@ -33,8 +33,8 @@
 
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
-static gpt_sampler             ** g_smpl;
-static gpt_params               * g_params;
+static common_sampler          ** g_smpl;
+static common_params            * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
@@ -63,7 +63,7 @@ static bool file_is_empty(const std::string & path) {
 }
 
 static void write_logfile(
-    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const llama_context * ctx, const common_params & params, const llama_model * model,
     const std::vector<llama_token> & input_tokens, const std::string & output,
     const std::vector<llama_token> & output_tokens
 ) {
@@ -114,30 +114,35 @@ static void sigint_handler(int signo) {
         } else {
             console::cleanup();
             LOG("\n");
-            gpt_perf_print(*g_ctx, *g_smpl);
+            common_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+
+            // make sure all logs are flushed
+            LOG("Interrupted by user\n");
+            common_log_pause(common_log_main());
+
             _exit(130);
         }
     }
 }
 #endif
 
-static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, const std::string & role, const std::string & content) {
-    llama_chat_msg new_msg{role, content};
-    auto formatted = llama_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
+static std::string chat_add_and_format(struct llama_model * model, std::vector<common_chat_msg> & chat_msgs, const std::string & role, const std::string & content) {
+    common_chat_msg new_msg{role, content};
+    auto formatted = common_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
     chat_msgs.push_back({role, content});
     LOG_DBG("formatted: '%s'\n", formatted.c_str());
     return formatted;
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
     g_params = &params;
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     auto & sparams = params.sparams;
 
@@ -182,9 +187,9 @@ int main(int argc, char ** argv) {
 
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
-    gpt_sampler * smpl = nullptr;
+    common_sampler * smpl = nullptr;
 
-    std::vector<llama_chat_msg> chat_msgs;
+    std::vector<common_chat_msg> chat_msgs;
 
     g_model = &model;
     g_ctx = &ctx;
@@ -192,7 +197,7 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     model = llama_init.model;
     ctx = llama_init.context;
@@ -241,7 +246,7 @@ int main(int argc, char ** argv) {
     // print chat template example in conversation mode
     if (params.conversation) {
         if (params.enable_chat_template) {
-            LOG_INF("%s: chat template example:\n%s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
+            LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(model, params.chat_template).c_str());
         } else {
             LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
         }
@@ -250,7 +255,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
         LOG_INF("\n");
     }
 
@@ -291,7 +296,7 @@ int main(int argc, char ** argv) {
             : params.prompt;
         if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
             LOG_DBG("tokenize the prompt\n");
-            embd_inp = ::llama_tokenize(ctx, prompt, true, true);
+            embd_inp = common_tokenize(ctx, prompt, true, true);
         } else {
             LOG_DBG("use session tokens\n");
             embd_inp = session_tokens;
@@ -374,15 +379,15 @@ int main(int argc, char ** argv) {
         LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
         LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (params.n_keep > add_bos) {
             LOG_INF("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                LOG("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            LOG("'\n");
+            LOG_CNT("'\n");
         }
         LOG_INF("\n");
     }
@@ -404,54 +409,54 @@ int main(int argc, char ** argv) {
     }
 
     if (params.interactive) {
-        LOG("%s: interactive mode on.\n", __func__);
+        LOG_INF("%s: interactive mode on.\n", __func__);
 
         if (!params.antiprompt.empty()) {
             for (const auto & antiprompt : params.antiprompt) {
-                LOG("Reverse prompt: '%s'\n", antiprompt.c_str());
+                LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str());
                 if (params.verbose_prompt) {
-                    auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
+                    auto tmp = common_tokenize(ctx, antiprompt, false, true);
                     for (int i = 0; i < (int) tmp.size(); i++) {
-                        LOG("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                        LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
                     }
                 }
             }
         }
 
         if (params.input_prefix_bos) {
-            LOG("Input prefix with BOS\n");
+            LOG_INF("Input prefix with BOS\n");
         }
 
         if (!params.input_prefix.empty()) {
-            LOG("Input prefix: '%s'\n", params.input_prefix.c_str());
+            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
             if (params.verbose_prompt) {
-                auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
+                auto tmp = common_tokenize(ctx, params.input_prefix, true, true);
                 for (int i = 0; i < (int) tmp.size(); i++) {
-                    LOG("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                    LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
                 }
             }
         }
 
         if (!params.input_suffix.empty()) {
-            LOG("Input suffix: '%s'\n", params.input_suffix.c_str());
+            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
             if (params.verbose_prompt) {
-                auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
+                auto tmp = common_tokenize(ctx, params.input_suffix, false, true);
                 for (int i = 0; i < (int) tmp.size(); i++) {
-                    LOG("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                    LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
                 }
             }
         }
     }
 
-    smpl = gpt_sampler_init(model, sparams);
+    smpl = common_sampler_init(model, sparams);
     if (!smpl) {
         LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
         return 1;
     }
 
-    LOG_INF("sampler seed: %u\n",     gpt_sampler_get_seed(smpl));
+    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
     LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
-    LOG_INF("sampler chain: %s\n",    gpt_sampler_print(smpl).c_str());
+    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
 
     LOG_INF("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);
 
@@ -469,7 +474,7 @@ int main(int argc, char ** argv) {
       //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
         LOG_INF("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
     }
-    LOG("\n");
+    LOG_INF("\n");
 
     if (params.interactive) {
         const char * control_message;
@@ -481,11 +486,11 @@ int main(int argc, char ** argv) {
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
-        LOG("== Running in interactive mode. ==\n");
+        LOG_INF("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-        LOG(       " - Press Ctrl+C to interject at any time.\n");
+        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-        LOG(       "%s\n", control_message);
+        LOG_INF(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -516,14 +521,14 @@ int main(int argc, char ** argv) {
 
     antiprompt_ids.reserve(params.antiprompt.size());
     for (const std::string & antiprompt : params.antiprompt) {
-        antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
+        antiprompt_ids.emplace_back(::common_tokenize(ctx, antiprompt, false, true));
     }
 
     if (llama_model_has_encoder(model)) {
         int enc_input_size = embd_inp.size();
         llama_token * enc_input_buf = embd_inp.data();
 
-        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
+        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return 1;
         }
@@ -564,30 +569,30 @@ int main(int argc, char ** argv) {
                     if (!params.ctx_shift){
                         LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
                         break;
-                    } else {
-                        if (params.n_predict == -2) {
-                            LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                            break;
-                        }
-
-                        const int n_left    = n_past - params.n_keep;
-                        const int n_discard = n_left/2;
-
-                        LOG_DBG("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            , params.n_keep + n_discard);
-                        llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
-
-                        n_past -= n_discard;
-
-                        LOG_DBG("after swap: n_past = %d\n", n_past);
-
-                        LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
-
-                        LOG_DBG("clear session path\n");
-                        path_session.clear();
                     }
+
+                    if (params.n_predict == -2) {
+                        LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                        break;
+                    }
+
+                    const int n_left    = n_past - params.n_keep;
+                    const int n_discard = n_left/2;
+
+                    LOG_DBG("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            , params.n_keep + n_discard);
+                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+
+                    n_past -= n_discard;
+
+                    LOG_DBG("after swap: n_past = %d\n", n_past);
+
+                    LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
+
+                    LOG_DBG("clear session path\n");
+                    path_session.clear();
                 }
             } else {
                 // context extension via Self-Extend
@@ -643,7 +648,7 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -674,9 +679,9 @@ int main(int argc, char ** argv) {
                 LOG_DBG("saved session to %s\n", path_session.c_str());
             }
 
-            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
+            const llama_token id = common_sampler_sample(smpl, ctx, -1);
 
-            gpt_sampler_accept(smpl, id, /* accept_grammar= */ true);
+            common_sampler_accept(smpl, id, /* accept_grammar= */ true);
 
             // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
 
@@ -697,7 +702,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                gpt_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
+                common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -709,7 +714,7 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo && display) {
             for (auto id : embd) {
-                const std::string token_str = llama_token_to_piece(ctx, id, params.special);
+                const std::string token_str = common_token_to_piece(ctx, id, params.special);
 
                 // Console/Stream Output
                 LOG("%s", token_str.c_str());
@@ -738,7 +743,7 @@ int main(int argc, char ** argv) {
             // check for reverse prompt in the last n_prev tokens
             if (!params.antiprompt.empty()) {
                 const int n_prev = 32;
-                const std::string last_output = gpt_sampler_prev_str(smpl, ctx, n_prev);
+                const std::string last_output = common_sampler_prev_str(smpl, ctx, n_prev);
 
                 is_antiprompt = false;
                 // Check if each of the reverse prompts appears at the end of the output.
@@ -760,7 +765,7 @@ int main(int argc, char ** argv) {
                 }
 
                 // check for reverse prompt using special tokens
-                llama_token last_token = gpt_sampler_last(smpl);
+                llama_token last_token = common_sampler_last(smpl);
                 for (std::vector<llama_token> ids : antiprompt_ids) {
                     if (ids.size() == 1 && last_token == ids[0]) {
                         if (params.interactive) {
@@ -777,13 +782,13 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of generation tokens in interactive mode
-            if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
+            if (llama_token_is_eog(model, common_sampler_last(smpl))) {
                 LOG_DBG("found an EOG token\n");
 
                 if (params.interactive) {
                     if (!params.antiprompt.empty()) {
                         // tokenize and inject first reverse prompt
-                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
+                        const auto first_antiprompt = common_tokenize(ctx, params.antiprompt.front(), false, true);
                         embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
                         is_antiprompt = true;
                     }
@@ -798,8 +803,8 @@ int main(int argc, char ** argv) {
 
             // if current token is not EOG, we add it to current assistant message
             if (params.conversation) {
-                const auto id = gpt_sampler_last(smpl);
-                assistant_ss << llama_token_to_piece(ctx, id, false);
+                const auto id = common_sampler_last(smpl);
+                assistant_ss << common_token_to_piece(ctx, id, false);
             }
 
             if (n_past > 0 && is_interacting) {
@@ -857,9 +862,9 @@ int main(int argc, char ** argv) {
                         ? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
                         : std::move(buffer);
                     // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
-                    const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, user_inp,            false, format_chat);
-                    const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
+                    const auto line_pfx = common_tokenize(ctx, params.input_prefix, false, true);
+                    const auto line_inp = common_tokenize(ctx, user_inp,            false, format_chat);
+                    const auto line_sfx = common_tokenize(ctx, params.input_suffix, false, true);
 
                     LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
 
@@ -877,7 +882,7 @@ int main(int argc, char ** argv) {
                     for (size_t i = original_size; i < embd_inp.size(); ++i) {
                         const llama_token token = embd_inp[i];
                         output_tokens.push_back(token);
-                        output_ss << llama_token_to_piece(ctx, token);
+                        output_ss << common_token_to_piece(ctx, token);
                     }
 
                     // reset assistant message
@@ -894,7 +899,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    gpt_sampler_reset(smpl);
+                    common_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -920,10 +925,10 @@ int main(int argc, char ** argv) {
     }
 
     LOG("\n\n");
-    gpt_perf_print(ctx, smpl);
+    common_perf_print(ctx, smpl);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
-    gpt_sampler_free(smpl);
+    common_sampler_free(smpl);
 
     llama_free(ctx);
     llama_free_model(model);

examples/parallel/parallel.cpp

@@ -54,7 +54,7 @@ static std::vector<std::string> k_prompts = {
 struct client {
     ~client() {
         if (smpl) {
-            gpt_sampler_free(smpl);
+            common_sampler_free(smpl);
         }
     }
 
@@ -75,7 +75,7 @@ struct client {
     std::string prompt;
     std::string response;
 
-    struct gpt_sampler * smpl = nullptr;
+    struct common_sampler * smpl = nullptr;
 };
 
 static void print_date_time() {
@@ -103,13 +103,13 @@ static std::vector<std::string> split_string(const std::string& input, char deli
 int main(int argc, char ** argv) {
     srand(1234);
 
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     // number of simultaneous "clients" to simulate
     const int32_t n_clients = params.n_parallel;
@@ -130,7 +130,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the target model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -160,11 +160,11 @@ int main(int argc, char ** argv) {
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.smpl = gpt_sampler_init(model, params.sparams);
+        client.smpl = common_sampler_init(model, params.sparams);
     }
 
     std::vector<llama_token> tokens_system;
-    tokens_system = ::llama_tokenize(ctx, k_system, true);
+    tokens_system = common_tokenize(ctx, k_system, true);
     const int32_t n_tokens_system = tokens_system.size();
 
     llama_seq_id g_seq_id = 0;
@@ -189,7 +189,7 @@ int main(int argc, char ** argv) {
         LOG_INF("%s: Evaluating the system prompt ...\n", __func__);
 
         for (int32_t i = 0; i < n_tokens_system; ++i) {
-            llama_batch_add(batch, tokens_system[i], i, { 0 }, false);
+            common_batch_add(batch, tokens_system[i], i, { 0 }, false);
         }
 
         if (llama_decode(ctx, batch) != 0) {
@@ -210,10 +210,10 @@ int main(int argc, char ** argv) {
     while (true) {
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            llama_kv_cache_dump_view_seqs(kvc_view, 40);
+            common_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         // decode any currently ongoing sequences
         for (auto & client : clients) {
@@ -223,7 +223,7 @@ int main(int argc, char ** argv) {
 
             client.i_batch = batch.n_tokens;
 
-            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
+            common_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
 
             client.n_decoded += 1;
         }
@@ -252,14 +252,14 @@ int main(int argc, char ** argv) {
                     client.prompt   = client.input + "\nAssistant:";
                     client.response = "";
 
-                    gpt_sampler_reset(client.smpl);
+                    common_sampler_reset(client.smpl);
 
                     // do not prepend BOS because we have a system prompt!
                     std::vector<llama_token> tokens_prompt;
-                    tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
+                    tokens_prompt = common_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
+                        common_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
                     }
 
                     // extract the logits only for the last token
@@ -308,7 +308,6 @@ int main(int argc, char ** argv) {
                 batch.n_seq_id + i,
                 batch.seq_id   + i,
                 batch.logits   + i,
-                0, 0, 0, // unused
             };
 
             const int ret = llama_decode(ctx, batch_view);
@@ -340,9 +339,9 @@ int main(int argc, char ** argv) {
                 //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 = gpt_sampler_sample(client.smpl, ctx, client.i_batch - i);
+                const llama_token id = common_sampler_sample(client.smpl, ctx, client.i_batch - i);
 
-                gpt_sampler_accept(client.smpl, id, true);
+                common_sampler_accept(client.smpl, id, true);
 
                 if (client.n_decoded == 1) {
                     // start measuring generation time after the first token to make sure all concurrent clients
@@ -350,7 +349,7 @@ int main(int argc, char ** argv) {
                     client.t_start_gen = ggml_time_us();
                 }
 
-                const std::string token_str = llama_token_to_piece(ctx, id);
+                const std::string token_str = common_token_to_piece(ctx, id);
 
                 client.response += token_str;
                 client.sampled = id;

examples/passkey/passkey.cpp

@@ -15,17 +15,17 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.n_junk = 250;
     params.n_keep = 32;
     params.i_pos  = -1;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     int n_junk = params.n_junk;
     int n_keep = params.n_keep;
@@ -61,7 +61,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = common_model_params_to_llama(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -72,7 +72,7 @@ int main(int argc, char ** argv) {
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = common_context_params_to_llama(params);
 
     ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;
 
@@ -92,10 +92,10 @@ int main(int argc, char ** argv) {
 
     // tokenize the prompt
     std::vector<llama_token> tokens_list;
-    tokens_list = ::llama_tokenize(ctx, params.prompt, true);
+    tokens_list = common_tokenize(ctx, params.prompt, true);
 
     // tokenize the prefix and use it as a sink
-    const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size();
+    const int n_tokens_prefix = common_tokenize(ctx, prompt_prefix, true).size();
 
     const int n_tokens_all = tokens_list.size();
 
@@ -137,10 +137,10 @@ int main(int argc, char ** argv) {
             n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
 
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
-            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+            common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
         }
 
         if (i + n_batch >= n_tokens_all) {
@@ -171,10 +171,10 @@ int main(int argc, char ** argv) {
 
         n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
 
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
-            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+            common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
         }
 
         if (i + n_batch >= n_tokens_all) {
@@ -229,15 +229,15 @@ int main(int argc, char ** argv) {
                 break;
             }
 
-            LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
 
             n_decode += 1;
 
             // prepare the next batch
-            llama_batch_clear(batch);
+            common_batch_clear(batch);
 
             // push this new token for next evaluation
-            llama_batch_add(batch, new_token_id, n_past++, { 0 }, true);
+            common_batch_add(batch, new_token_id, n_past++, { 0 }, true);
         }
 
         n_cur += 1;

examples/perplexity/perplexity.cpp

@@ -35,7 +35,7 @@ struct results_log_softmax {
 };
 
 static void write_logfile(
-    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const llama_context * ctx, const common_params & params, const llama_model * model,
     const struct results_perplexity & results
 ) {
     if (params.logdir.empty()) {
@@ -169,7 +169,7 @@ static void process_logits(
                 break;
             }
             lock.unlock();
-            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            const results_log_softmax results = log_softmax(n_vocab, logits + size_t(i)*n_vocab, tokens[i+1]);
             const double v = -results.log_softmax;
             local_nll += v;
             local_nll2 += v*v;
@@ -203,7 +203,7 @@ static void process_logits(std::ostream& out, int n_vocab, const float * logits,
                 break;
             }
             lock.unlock();
-            const double v = log_softmax(n_vocab, logits + i*n_vocab, log_probs.data() + i*nv, tokens[i+1]);
+            const double v = log_softmax(n_vocab, logits + size_t(i)*n_vocab, log_probs.data() + i*nv, tokens[i+1]);
             local_nll += v;
             local_nll2 += v*v;
         }
@@ -281,7 +281,9 @@ static std::pair<double, float> log_softmax(int n_vocab, const float * logits, c
     kld.sum_kld  += sum;
     kld.sum_kld2 += sum*sum;
     ++kld.count;
-    if (imax == imax_base) ++kld.n_same_top;
+    if (imax == imax_base) {
+        ++kld.n_same_top;
+    }
 
     const float p_base = expf(-nll_base);
     const float p = expf(-nll);
@@ -323,7 +325,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
                 break;
             }
             lock.unlock();
-            std::pair<double, float> v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld);
+            std::pair<double, float> v = log_softmax(n_vocab, logits + size_t(i)*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld);
             kld_values[i]    = (float)v.first;
             p_diff_values[i] = v.second;
         }
@@ -337,7 +339,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
     }
 }
 
-static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
+static results_perplexity perplexity_v2(llama_context * ctx, const common_params & params) {
     // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
@@ -348,7 +350,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
 
     const int n_ctx = llama_n_ctx(ctx);
 
@@ -383,9 +385,10 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
     int count = 0;
     double nll = 0.0;
 
@@ -405,14 +408,21 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
+        llama_batch batch = llama_batch_init(n_batch, 0, 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);
 
+            common_batch_clear(batch);
+            for (int i = 0; i < batch_size; i++) {
+                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            }
+
             //LOG_DBG("    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            // TODO: use llama_batch.logits instead of relying on logits_all == true
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            if (llama_decode(ctx, batch)) {
                 //LOG_ERR("%s : failed to eval\n", __func__);
+                llama_batch_free(batch);
                 return {tokens, -1, logit_history, prob_history};
             }
 
@@ -424,14 +434,16 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            const auto batch_logits = llama_get_logits(ctx);
-            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+            const auto * batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + size_t(batch_size) * n_vocab);
 
             if (j == 0) {
                 tokens[batch_start] = token_org;
             }
         }
 
+        llama_batch_free(batch);
+
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -444,15 +456,13 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
             }
             LOG("%.2f minutes\n", total_seconds / 60.0);
         }
-        LOG("\n");
 
         //LOG_DBG("%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
         for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {
-
             // Calculate probability of next token, given the previous ones.
             const std::vector<float> tok_logits(
-                logits.begin() + (j + 0) * n_vocab,
-                logits.begin() + (j + 1) * n_vocab);
+                logits.begin() + size_t(j + 0) * n_vocab,
+                logits.begin() + size_t(j + 1) * n_vocab);
 
             const float prob = softmax(tok_logits)[tokens[start + j + 1]];
             logit_history[start + j + 1] = tok_logits[tokens[start + j + 1]];
@@ -473,7 +483,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     return {tokens, std::exp(nll / count), logit_history, prob_history};
 }
 
-static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) {
+static results_perplexity perplexity(llama_context * ctx, const common_params & params, const int32_t n_ctx) {
     if (params.ppl_stride > 0) {
         return perplexity_v2(ctx, params);
     }
@@ -501,7 +511,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
 
     auto tim2 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
@@ -522,9 +532,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     const int n_chunk_max = tokens.size() / n_ctx;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
     int count = 0;
     double nll = 0.0;
     double nll2 = 0.0;
@@ -539,7 +550,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
     std::vector<float> logits;
     if (num_batches > 1) {
-        logits.reserve((size_t)n_ctx * n_vocab);
+        logits.reserve(size_t(n_ctx) * n_vocab);
     }
 
     LOG_INF("%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
@@ -621,7 +632,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
             if (num_batches > 1 && n_outputs > 0) {
                 const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + n_outputs * n_vocab);
+                logits.insert(logits.end(), batch_logits, batch_logits + size_t(n_outputs) * n_vocab);
             }
         }
 
@@ -638,7 +649,6 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             }
             LOG("%.2f minutes\n", total_seconds / 60.0);
         }
-        LOG("\n");
 
         for (int seq = 0; seq < n_seq_batch; seq++) {
             const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx + first);
@@ -663,7 +673,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             } else {
                 double av = nll/count;
                 double av2 = nll2/count - av*av;
-                if (av2 > 0) av2 = sqrt(av2/(count-1));
+                if (av2 > 0) {
+                    av2 = sqrt(av2/(count-1));
+                }
                 LOG("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
             }
         }
@@ -688,10 +700,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     return {tokens, ppl, logit_history, prob_history};
 }
 
-static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int32_t n_batch, int32_t n_vocab) {
+static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int n_batch, int n_vocab) {
     int prev_outputs = 0;
-    for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
-        const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+    for (int i = 0; i < (int) batch.n_tokens; i += n_batch) {
+        const int n_tokens = std::min<int>(n_batch, batch.n_tokens - i);
 
         llama_batch batch_view = {
             n_tokens,
@@ -701,7 +713,6 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             batch.n_seq_id + i,
             batch.seq_id   + i,
             batch.logits   + i,
-            0, 0, 0, // unused
         };
 
         const int ret = llama_decode(ctx, batch_view);
@@ -715,7 +726,7 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             n_outputs += batch_view.logits[i] != 0;
         }
 
-        memcpy(batch_logits.data() + prev_outputs*n_vocab, llama_get_logits(ctx), n_outputs*n_vocab*sizeof(float));
+        memcpy(batch_logits.data() + size_t(prev_outputs)*n_vocab, llama_get_logits(ctx), size_t(n_outputs)*n_vocab*sizeof(float));
 
         prev_outputs += n_outputs;
     }
@@ -730,7 +741,9 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
     if (eval_results.size() != eval_pairs.size()) {
         eval_results.resize(eval_pairs.size());
     }
-    if (eval_pairs.empty()) return;
+    if (eval_pairs.empty()) {
+        return;
+    }
 
     size_t max_threads = std::min((eval_pairs.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK, workers.size());
 
@@ -738,11 +751,13 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
     auto compute = [&counter, &eval_pairs, &eval_results, batch_logits, n_vocab] () {
         float local_logprobs[K_TOKEN_CHUNK];
         while (true) {
-            size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed);
-            if (first >= eval_results.size()) break;
-            size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size());
+            const size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed);
+            if (first >= eval_results.size()) {
+                break;
+            }
+            const size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size());
             for (size_t i = first; i < last; ++i) {
-                auto logits = batch_logits + eval_pairs[i].first * n_vocab;
+                const auto * logits = batch_logits + eval_pairs[i].first * n_vocab;
                 float max_logit = logits[0];
                 for (int j = 1; j < n_vocab; ++j) {
                     max_logit = std::max(max_logit, logits[j]);
@@ -765,7 +780,7 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
     }
 }
 
-static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
+static void hellaswag_score(llama_context * ctx, const common_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
@@ -846,7 +861,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         hs_cur.gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
         for (size_t j = 0; j < 4; j++) {
             hs_cur.ending[j] = prompt_lines[idx*6+2+j];
-            hs_cur.seq_tokens[j] = ::llama_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], true);
+            hs_cur.seq_tokens[j] = common_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], true);
         }
 
         // determine the common prefix of the endings
@@ -879,10 +894,11 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
 
     double acc = 0.0f;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_ctx   = llama_n_ctx(ctx);
     const int n_batch = params.n_batch;
 
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
@@ -890,7 +906,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
 
     std::vector<float> tok_logits(n_vocab);
     // TODO: this could be made smaller; it's currently the worst-case size
-    std::vector<float> batch_logits(n_vocab*n_ctx);
+    std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
     std::vector<float> eval_results;
@@ -902,7 +918,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         size_t i1 = i0;
         size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
         // each task has 4 unique sequence ids - one for each ending
@@ -918,7 +934,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             }
 
             for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
-                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
+                common_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
             n_logits += 1;
@@ -928,7 +944,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
                 // TODO: don't evaluate the last token of each sequence
                 for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
                     const bool needs_logits = i < seq_tokens_size - 1;
-                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    common_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
                     n_logits += needs_logits;
                 }
             }
@@ -977,7 +993,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             auto & hs_cur = hs_data[i];
 
             // get the logits of the last token of the common prefix
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*hs_cur.i_logits, n_vocab*sizeof(float));
+            std::memcpy(tok_logits.data(), batch_logits.data() + hs_cur.i_logits*n_vocab, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -1104,7 +1120,7 @@ static std::vector<winogrande_entry> load_winogrande_from_csv(const std::string
  *    0,Sarah was a much better surgeon than Maria so _ always got the easier cases.,Sarah,Maria,2
  *
  */
-static void winogrande_score(llama_context * ctx, const gpt_params & params) {
+static void winogrande_score(llama_context * ctx, const common_params & params) {
 
     constexpr int k_min_trailing_ctx = 3;
 
@@ -1138,8 +1154,8 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     LOG_INF("%s : tokenizing selected tasks\n", __func__);
 
     for (auto & task : data) {
-        task.seq_tokens[0] = ::llama_tokenize(ctx, task.first + task.choices[0] + task.second, true);
-        task.seq_tokens[1] = ::llama_tokenize(ctx, task.first + task.choices[1] + task.second, true);
+        task.seq_tokens[0] = common_tokenize(ctx, task.first + task.choices[0] + task.second, true);
+        task.seq_tokens[1] = common_tokenize(ctx, task.first + task.choices[1] + task.second, true);
 
         task.common_prefix = 0;
         for (size_t k = 0; k < task.seq_tokens[0].size(); k++) {
@@ -1154,16 +1170,17 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             task.seq_tokens[0].size() - task.common_prefix +
             task.seq_tokens[1].size() - task.common_prefix;
 
-        task.n_base1 = ::llama_tokenize(ctx, task.first + task.choices[0], true).size();
-        task.n_base2 = ::llama_tokenize(ctx, task.first + task.choices[1], true).size();
+        task.n_base1 = common_tokenize(ctx, task.first + task.choices[0], true).size();
+        task.n_base2 = common_tokenize(ctx, task.first + task.choices[1], true).size();
     }
 
     LOG_INF("%s : calculating winogrande score over selected tasks.\n", __func__);
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_ctx   = llama_n_ctx(ctx);
     const int n_batch = params.n_batch;
 
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
     const int max_tasks_per_batch = 128;
     const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
@@ -1171,7 +1188,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
 
     std::vector<float> tok_logits(n_vocab);
     // TODO: this could be made smaller; it's currently the worst-case size
-    std::vector<float> batch_logits(n_vocab*n_ctx);
+    std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
     std::vector<float> eval_results;
@@ -1186,7 +1203,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
         size_t i1 = i0;
         size_t i_logits = 0;
 
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
             int n_logits = 0;
@@ -1196,7 +1213,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             }
 
             for (size_t i = 0; i < data[i1].common_prefix; ++i) {
-                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
+                common_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true;
             n_logits += 1;
@@ -1204,7 +1221,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             for (int s = 0; s < 2; ++s) {
                 // TODO: end before the last token, no need to predict past the end of the sequences
                 for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
+                    common_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
                     n_logits += 1;
                 }
             }
@@ -1361,7 +1378,7 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
             }
             return false;
         }
-        task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, true));
+        task.seq_tokens.emplace_back(::common_tokenize(ctx, task.question + " " + answer, true));
     }
     auto min_len = task.seq_tokens.front().size();
     for (auto& seq : task.seq_tokens) {
@@ -1405,7 +1422,7 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
 //     git@hf.co:datasets/Stevross/mmlu
 //     https://huggingface.co/datasets/truthful_qa
 //
-static void multiple_choice_score(llama_context * ctx, const gpt_params & params) {
+static void multiple_choice_score(llama_context * ctx, const common_params & params) {
 
     std::istringstream strstream(params.prompt);
     uint32_t n_task;
@@ -1511,17 +1528,18 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
 
     LOG("\ntask\tacc_norm\n");
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_ctx   = llama_n_ctx(ctx);
     const int n_batch = params.n_batch;
 
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
     std::vector<float> tok_logits(n_vocab);
-    std::vector<float> batch_logits(n_vocab*n_ctx);
+    std::vector<float> batch_logits(size_t(n_ctx)*n_vocab);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
     std::vector<float> eval_results;
@@ -1538,7 +1556,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         size_t i1 = i0;
         size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
         // each task has 4 unique sequence ids - one for each ending
@@ -1561,7 +1579,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
 
             for (size_t i = 0; i < cur_task.common_prefix; ++i) {
                 //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
-                llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
+                common_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
             n_logits += 1;
@@ -1571,7 +1589,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
                 // TODO: don't evaluate the last token of each sequence
                 for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
                     const bool needs_logits = i < seq_tokens_size - 1;
-                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    common_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
                     n_logits += needs_logits;
                 }
             }
@@ -1629,7 +1647,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             //LOG("\n    common_prefix: %zu\n", cur_task.common_prefix);
 
             // get the logits of the last token of the common prefix
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*cur_task.i_logits, n_vocab*sizeof(float));
+            std::memcpy(tok_logits.data(), batch_logits.data() + cur_task.i_logits*n_vocab, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -1685,7 +1703,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     LOG_INF("\n");
 }
 
-static void kl_divergence(llama_context * ctx, const gpt_params & params) {
+static void kl_divergence(llama_context * ctx, const common_params & params) {
     if (params.logits_file.empty()) {
         LOG_ERR("%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
         return;
@@ -1711,7 +1729,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
                 __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
     }
 
-    int n_vocab, n_chunk;
+    int n_vocab;
+    int n_chunk;
     in.read((char *)&n_vocab, sizeof(n_vocab));
     in.read((char *)&n_chunk, sizeof(n_chunk));
     if (in.fail()) {
@@ -1722,7 +1741,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
         LOG_ERR("%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
     }
 
-    std::vector<llama_token> tokens(n_ctx * n_chunk);
+    std::vector<llama_token> tokens(size_t(n_ctx) * n_chunk);
     if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) {
         LOG_ERR("%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
         return;
@@ -1739,7 +1758,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     std::vector<float> p_diff_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
     std::vector<float> logits;
     if (num_batches > 1) {
-        logits.reserve(n_ctx * n_vocab);
+        logits.reserve(size_t(n_ctx) * n_vocab);
     }
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
@@ -1780,6 +1799,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
+        llama_batch batch = llama_batch_init(n_batch, 0, 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);
@@ -1792,9 +1813,14 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            // TODO: use llama_batch.logits instead of relying on logits_all == true
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            common_batch_clear(batch);
+            for (int i = 0; i < batch_size; i++) {
+                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            }
+
+            if (llama_decode(ctx, batch)) {
                 LOG_ERR("%s : failed to eval\n", __func__);
+                llama_batch_free(batch);
                 return;
             }
 
@@ -1803,10 +1829,12 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
 
             if (num_batches > 1) {
                 const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+                logits.insert(logits.end(), batch_logits, batch_logits + size_t(batch_size) * n_vocab);
             }
         }
 
+        llama_batch_free(batch);
+
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -1824,7 +1852,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
 
         const int first = n_ctx/2;
         const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
-        process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+        process_logits(n_vocab, all_logits + size_t(first)*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
                 workers, log_probs_uint16, kld, kld_ptr, p_diff_ptr);
         p_diff_ptr += n_ctx - 1 - first;
         kld_ptr    += n_ctx - 1 - first;
@@ -1957,16 +1985,17 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.n_ctx = 512;
     params.logits_all = true;
+    params.escape = false;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     const int32_t n_ctx = params.n_ctx;
 
@@ -2005,7 +2034,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model and apply lora adapter, if any
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -2024,7 +2053,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
     struct results_perplexity results;

examples/quantize-stats/quantize-stats.cpp

@@ -142,7 +142,7 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
 }
 
 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,
+    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, bool use_reference,
     float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
 ) {
     if (layer->type == GGML_TYPE_F16) {
@@ -166,7 +166,7 @@ static void test_roundtrip_on_chunk(
 
 // Run quantization function for a single layer and update error stats
 static void test_roundtrip_on_layer(
-    std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
+    std::string & name, bool print_layer_stats, const ggml_type_traits & 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
 ) {
@@ -371,8 +371,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
-        if (qfns.from_float && qfns.to_float) {
+        const auto *  qfns = ggml_get_type_traits(type);
+        if (qfns->from_float && qfns->to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
             }
@@ -393,7 +393,7 @@ int main(int argc, char ** argv) {
                 test_roundtrip_on_layer(
                         layer_name,
                         params.per_layer_stats,
-                        qfns,
+                        *qfns,
                         params.reference,
                         kv_tensor.second,
                         input_scratch,

examples/quantize/quantize.cpp

@@ -63,6 +63,16 @@ static const char * const LLM_KV_QUANTIZE_IMATRIX_DATASET    = "quantize.imatrix
 static const char * const LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES  = "quantize.imatrix.entries_count";
 static const char * const LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS   = "quantize.imatrix.chunks_count";
 
+static bool striequals(const char * a, const char * b) {
+    while (*a && *b) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+        a++; b++;
+    }
+    return *a == *b;
+}
+
 static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
     std::string ftype_str;
 
@@ -70,7 +80,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
         ftype_str.push_back(std::toupper(ch));
     }
     for (auto & it : QUANT_OPTIONS) {
-        if (it.name == ftype_str) {
+        if (striequals(it.name.c_str(), ftype_str.c_str())) {
             ftype = it.ftype;
             ftype_str_out = it.name;
             return true;
@@ -225,15 +235,15 @@ static int prepare_imatrix(const std::string & imatrix_file,
 }
 
 static ggml_type parse_ggml_type(const char * arg) {
-    ggml_type result = GGML_TYPE_COUNT;
-    for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
-        auto type = ggml_type(j);
+    for (int i = 0; i < GGML_TYPE_COUNT; ++i) {
+        auto type = (ggml_type)i;
         const auto * name = ggml_type_name(type);
-        if (name && strcmp(arg, name) == 0) {
-            result = type; break;
+        if (name && striequals(name, arg)) {
+            return type;
         }
     }
-    return result;
+    fprintf(stderr, "%s: invalid ggml_type '%s'\n", __func__, arg);
+    return GGML_TYPE_COUNT;
 }
 
 int main(int argc, char ** argv) {
@@ -254,12 +264,18 @@ int main(int argc, char ** argv) {
         } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
             if (arg_idx < argc-1) {
                 params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+                if (params.output_tensor_type == GGML_TYPE_COUNT) {
+                    usage(argv[0]);
+                }
             } else {
                 usage(argv[0]);
             }
         } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
             if (arg_idx < argc-1) {
                 params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+                if (params.token_embedding_type == GGML_TYPE_COUNT) {
+                    usage(argv[0]);
+                }
             } else {
                 usage(argv[0]);
             }

examples/retrieval/retrieval.cpp

@@ -77,7 +77,7 @@ static std::vector<chunk> chunk_file(const std::string & filename, int chunk_siz
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
     size_t n_tokens = tokens.size();
     for (size_t i = 0; i < n_tokens; i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, true);
+        common_batch_add(batch, tokens[i], i, { seq_id }, true);
     }
 }
 
@@ -107,18 +107,18 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         }
 
         float * out = output + batch.seq_id[i][0] * n_embd;
-        llama_embd_normalize(embd, out, n_embd);
+        common_embd_normalize(embd, out, n_embd);
     }
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
         return 1;
     }
 
-    gpt_init();
+    common_init();
 
     // For BERT models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
@@ -149,7 +149,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -176,7 +176,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
     // max batch size
@@ -185,7 +185,7 @@ int main(int argc, char ** argv) {
 
     // tokenize the prompts and trim
     for (auto & chunk : chunks) {
-        auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false);
+        auto inp = common_tokenize(ctx, chunk.textdata, true, false);
         if (inp.size() > n_batch) {
             LOG_ERR("%s: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -204,7 +204,7 @@ int main(int argc, char ** argv) {
             LOG_INF("%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
             LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
             for (int j = 0; j < (int) chunks[i].tokens.size(); j++) {
-                LOG_INF("%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
+                LOG_INF("%6d -> '%s'\n", chunks[i].tokens[j], common_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
             }
             LOG_INF("\n\n");
         }
@@ -232,7 +232,7 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens + n_toks > n_batch) {
             float * out = emb + p * n_embd;
             batch_decode(ctx, batch, out, s, n_embd);
-            llama_batch_clear(batch);
+            common_batch_clear(batch);
             p += s;
             s = 0;
         }
@@ -260,20 +260,20 @@ int main(int argc, char ** argv) {
     while (true) {
         LOG("Enter query: ");
         std::getline(std::cin, query);
-        std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);
+        std::vector<int32_t> query_tokens = common_tokenize(ctx, query, true);
 
         batch_add_seq(query_batch, query_tokens, 0);
 
         std::vector<float> query_emb(n_embd, 0);
         batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
 
-        llama_batch_clear(query_batch);
+        common_batch_clear(query_batch);
 
         // compute cosine similarities
         {
             std::vector<std::pair<int, float>> similarities;
             for (int i = 0; i < n_chunks; i++) {
-                float sim = llama_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
+                float sim = common_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
                 similarities.push_back(std::make_pair(i, sim));
             }
 

examples/rpc/rpc-server.cpp

@@ -1,3 +1,5 @@
+#include "ggml-cpu.h"
+
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif
@@ -6,6 +8,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #include "ggml-rpc.h"
 #ifdef _WIN32
 #  include <windows.h>
@@ -79,6 +85,12 @@ static ggml_backend_t create_backend() {
     if (!backend) {
         fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
     }
+#elif GGML_USE_VULKAN
+    fprintf(stderr, "%s: using Vulkan backend\n", __func__);
+    backend = ggml_backend_vk_init(0); // init device 0
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__);
+    }
 #endif
 
     // if there aren't GPU Backends fallback to CPU backend
@@ -92,6 +104,8 @@ static ggml_backend_t create_backend() {
 static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
 #ifdef GGML_USE_CUDA
     ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
+#elif GGML_USE_VULKAN
+    ggml_backend_vk_get_device_memory(0, free_mem, total_mem);
 #else
     #ifdef _WIN32
         MEMORYSTATUSEX status;
@@ -139,7 +153,7 @@ int main(int argc, char * argv[]) {
         get_backend_memory(&free_mem, &total_mem);
     }
     printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
-    start_rpc_server(backend, endpoint.c_str(), free_mem, total_mem);
+    ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
     ggml_backend_free(backend);
     return 0;
 }

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

@@ -6,12 +6,12 @@
 #include <cstdio>
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.prompt = "The quick brown fox";
     params.sparams.seed = 1234;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -28,7 +28,7 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -42,15 +42,21 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
 
     // tokenize prompt
-    auto tokens = llama_tokenize(ctx, params.prompt, true);
+    auto tokens = common_tokenize(ctx, params.prompt, true);
+
+    // prepare the batch
+    llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
+    for (size_t i = 0; i < tokens.size(); i++) {
+        common_batch_add(batch, tokens[i], i, {0}, false);
+    }
+    batch.logits[batch.n_tokens - 1] = true; // generate next token
 
     // evaluate prompt
-    llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
-    n_past += tokens.size();
+    llama_decode(ctx, batch);
+    n_past += batch.n_tokens;
 
     // save state (rng, logits, embedding and kv_cache) to file
     {
@@ -72,13 +78,17 @@ int main(int argc, char ** argv) {
 
     for (auto i = 0; i < params.n_predict; i++) {
         auto next_token     = llama_sampler_sample(smpl, ctx, -1);
-        auto next_token_str = llama_token_to_piece(ctx, next_token);
+        auto next_token_str = common_token_to_piece(ctx, next_token);
 
         printf("%s", next_token_str.c_str());
         result0 += next_token_str;
 
-        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
+        common_batch_clear(batch);
+        common_batch_add(batch, next_token, n_past, {0}, true);
+
+        if (llama_decode(ctx, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_batch_free(batch);
             llama_free(ctx);
             llama_free_model(model);
             return 1;
@@ -92,11 +102,10 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
 
     // make new context
-    auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+    auto * ctx2 = llama_new_context_with_model(model, common_context_params_to_llama(params));
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl2, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsecond run: %s", params.prompt.c_str());
@@ -128,13 +137,17 @@ int main(int argc, char ** argv) {
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
         auto next_token     = llama_sampler_sample(smpl2, ctx2, -1);
-        auto next_token_str = llama_token_to_piece(ctx2, next_token);
+        auto next_token_str = common_token_to_piece(ctx2, next_token);
 
         printf("%s", next_token_str.c_str());
         result1 += next_token_str;
 
-        if (llama_decode(ctx2, llama_batch_get_one(&next_token, 1, n_past, 0))) {
+        common_batch_clear(batch);
+        common_batch_add(batch, next_token, n_past, {0}, true);
+
+        if (llama_decode(ctx2, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_batch_free(batch);
             llama_free(ctx2);
             llama_free_model(model);
             return 1;
@@ -152,11 +165,10 @@ int main(int argc, char ** argv) {
     }
 
     // make new context
-    auto * ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+    auto * ctx3 = llama_new_context_with_model(model, common_context_params_to_llama(params));
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl3, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());
@@ -216,13 +228,17 @@ int main(int argc, char ** argv) {
     // third run with seq 1 instead of 0
     for (auto i = 0; i < params.n_predict; i++) {
         auto next_token     = llama_sampler_sample(smpl3, ctx3, -1);
-        auto next_token_str = llama_token_to_piece(ctx3, next_token);
+        auto next_token_str = common_token_to_piece(ctx3, next_token);
 
         printf("%s", next_token_str.c_str());
         result2 += next_token_str;
 
-        if (llama_decode(ctx3, llama_batch_get_one(&next_token, 1, n_past, 1))) {
+        common_batch_clear(batch);
+        common_batch_add(batch, next_token, n_past, {1}, true);
+
+        if (llama_decode(ctx3, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_batch_free(batch);
             llama_free(ctx3);
             llama_free_model(model);
             return 1;
@@ -236,6 +252,7 @@ int main(int argc, char ** argv) {
     llama_sampler_free(smpl2);
     llama_sampler_free(smpl3);
 
+    llama_batch_free(batch);
     llama_free(ctx3);
     llama_free_model(model);
 

examples/server/README.md

@@ -7,6 +7,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
  * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
+ * Reranking endoint (WIP: https://github.com/ggerganov/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
  * Continuous batching
  * Multimodal (wip)
@@ -17,12 +18,15 @@ The project is under active development, and we are [looking for feedback and co
 
 ## Usage
 
+<!-- Note for contributors: The list below is generated by llama-gen-docs -->
+
+**Common params**
+
 | Argument | Explanation |
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
-| `-v, --verbose` | print verbose information |
-| `--verbosity N` | set specific verbosity level (default: 0) |
+| `--verbose-prompt` | print a verbose prompt before generation (default: false) |
 | `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
 | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
 | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
@@ -42,61 +46,33 @@ The project is under active development, and we are [looking for feedback and co
 | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
 | `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `-p, --prompt PROMPT` | prompt to start generation with |
+| `--no-perf` | disable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_NO_PERF) |
 | `-f, --file FNAME` | a file containing the prompt (default: none) |
 | `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
 | `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
 | `--no-escape` | do not process escape sequences |
-| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
-| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
-| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) |
-| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
-| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--penalize-nl` | penalize newline tokens (default: false) |
-| `--temp N` | temperature (default: 0.8) |
-| `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--tfs N` | tail free sampling, parameter z (default: 1.0, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
-| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
-| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
-| `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
-| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
-| `--grammar-file FNAME` | file to read grammar from |
-| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
-| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model |
-| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N |
-| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model) |
-| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N |
-| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0) |
-| `-gan, --grp-attn-n N` | group-attention factor (default: 1) |
-| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0) |
+| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model<br/>(env: LLAMA_ARG_ROPE_SCALING_TYPE) |
+| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N<br/>(env: LLAMA_ARG_ROPE_SCALE) |
+| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
+| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
+| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
-| `-nkvo, --no-kv-offload` | disable KV offload |
-| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16) |
-| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) |
+| `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
-| `-np, --parallel N` | number of parallel sequences to decode (default: 1) |
-| `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
-| `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
-| `--mlock` | force system to keep model in RAM rather than swapping or compressing |
-| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock) |
-| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437 |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
+| `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
+| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
+| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
-| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs |
-| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1 |
-| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0) |
+| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
+| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
+| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
 | `--check-tensors` | check model tensor data for invalid values (default: false) |
 | `--override-kv KEY=TYPE:VALUE` | advanced option to override model metadata by key. may be specified multiple times.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false |
 | `--lora FNAME` | path to LoRA adapter (can be repeated to use multiple adapters) |
@@ -104,36 +80,87 @@ The project is under active development, and we are [looking for feedback and co
 | `--control-vector FNAME` | add a control vector<br/>note: this argument can be repeated to add multiple control vectors |
 | `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE<br/>note: this argument can be repeated to add multiple scaled control vectors |
 | `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
-| `-a, --alias STRING` | set alias for model name (to be used by REST API) |
 | `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)<br/>(env: LLAMA_ARG_MODEL) |
 | `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
 | `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
 | `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
 | `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
+| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
+| `--log-disable` | Log disable |
+| `--log-file FNAME` | Log to file |
+| `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |
+| `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) |
+| `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.<br/>(env: LLAMA_LOG_VERBOSITY) |
+| `--log-prefix` | Enable prefx in log messages<br/>(env: LLAMA_LOG_PREFIX) |
+| `--log-timestamps` | Enable timestamps in log messages<br/>(env: LLAMA_LOG_TIMESTAMPS) |
+
+
+**Sampling params**
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;typ_p;top_p;min_p;temperature) |
+| `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
+| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
+| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
+| `--penalize-nl` | penalize newline tokens (default: false) |
+| `--temp N` | temperature (default: 0.8) |
+| `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--dry-base N` | DRY sampling base value (default: 1.75) |
+| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
+| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
+| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
+| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
+| `--grammar-file FNAME` | file to read grammar from |
+| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+
+
+**Example-specific params**
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
+| `-sp, --special` | special tokens output enabled (default: false) |
+| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
+| `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
+| `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
+| `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
+| `-a, --alias STRING` | set alias for model name (to be used by REST API)<br/>(env: LLAMA_ARG_ALIAS) |
 | `--host HOST` | ip address to listen (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
 | `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
-| `--path PATH` | path to serve static files from (default: ) |
+| `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
+| `--reranking, --rerank` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
 | `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
 | `--api-key-file FNAME` | path to file containing API keys (default: none) |
-| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key |
-| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate |
-| `-to, --timeout N` | server read/write timeout in seconds (default: 600) |
+| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
+| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
+| `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
-| `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications |
+| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
-| `--no-slots` | disables slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
+| `--slots` | enable slots monitoring endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
+| `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
+| `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
 | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted:<br/>https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
-| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
-| `--log-test` | Log test |
-| `--log-disable` | Log disable |
-| `--log-enable` | Log enable |
-| `--log-new` | Log new |
-| `--log-append` | Log append |
-| `--log-file FNAME` | Log file |
+
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
 
@@ -295,7 +322,18 @@ node index.js
 
       - The prompt is a string or an array with the first element given as a string
       - The model's `tokenizer.ggml.add_bos_token` metadata is `true`
-      - The system prompt is empty
+
+    These input shapes and data type are allowed for `prompt`:
+
+      - Single string: `"string"`
+      - Single sequence of tokens: `[12, 34, 56]`
+      - Mixed tokens and strings: `[12, 34, "string", 56, 78]`
+
+    Multiple prompts are also supported. In this case, the completion result will be an array.
+
+      - Only strings: `["string1", "string2"]`
+      - Strings and sequences of tokens: `["string1", [12, 34, 56]]`
+      - Mixed types: `[[12, 34, "string", 56, 78], [12, 34, 56], "string"]`
 
     `temperature`: Adjust the randomness of the generated text. Default: `0.8`
 
@@ -311,6 +349,8 @@ node index.js
 
     `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.
 
+    `n_indent`: Specify the minimum line indentation for the generated text in number of whitespace characters. Useful for code completion tasks. Default: `0`
+
     `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token.
     By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
 
@@ -319,8 +359,6 @@ node index.js
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`
 
-    `tfs_z`: Enable tail free sampling with parameter z. Default: `1.0`, which is disabled.
-
     `typical_p`: Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.
 
     `repeat_penalty`: Control the repetition of token sequences in the generated text. Default: `1.1`
@@ -333,6 +371,16 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
+    `dry_multiplier`: Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.
+
+    `dry_base`: Set the DRY repetition penalty base value. Default: `1.75`
+
+    `dry_allowed_length`: Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`
+
+    `dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.
+
+    `dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
+
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@@ -353,15 +401,15 @@ node index.js
 
     `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum. Default: `0`
 
+    `t_max_predict_ms`: Set a time limit in milliseconds for the prediction (a.k.a. text-generation) phase. The timeout will trigger if the generation takes more than the specified time (measured since the first token was generated) and if a new-line character has already been generated. Useful for FIM applications. Default: `0`, which is disabled.
+
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
     `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
 
-    `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
-
-    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
 **Response format**
 
@@ -461,38 +509,99 @@ The same as [the embedding example](../embedding) does.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
+### POST `/reranking`: Rerank documents according to a given query
+
+Similar to https://jina.ai/reranker/ but might change in the future.
+Requires a reranker model (such as [bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3)) and the `--embedding --pooling rank` options.
+
+    *Options:*
+
+    `query`: The query against which the documents will be ranked.
+
+    `documents`: An array strings representing the documents to be ranked.
+
+    *Aliases:*
+      - `/rerank`
+      - `/v1/rerank`
+      - `/v1/reranking`
+
+    *Examples:*
+
+    ```shell
+    curl http://127.0.0.1:8012/v1/rerank \
+        -H "Content-Type: application/json" \
+        -d '{
+            "model": "some-model",
+                "query": "What is panda?",
+                "top_n": 3,
+                "documents": [
+                    "hi",
+                "it is a bear",
+                "The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China."
+                ]
+        }' | jq
+    ```
+
 ### POST `/infill`: For code infilling.
 
 Takes a prefix and a suffix and returns the predicted completion as stream.
 
-    *Options:*
+*Options:*
 
-    `input_prefix`: Set the prefix of the code to infill.
+- `input_prefix`: Set the prefix of the code to infill.
+- `input_suffix`: Set the suffix of the code to infill.
+- `input_extra`:  Additional context inserted before the FIM prefix.
+- `prompt`:       Added after the `FIM_MID` token
 
-    `input_suffix`: Set the suffix of the code to infill.
+`input_extra` is array of `{"filename": string, "text": string}` objects.
 
-    It also accepts all the options of `/completion` except `stream` and `prompt`.
+The endpoint also accepts all the options of `/completion`.
 
-- **GET** `/props`: Return current server settings.
+If the model has `FIM_REPO` and `FIM_FILE_SEP` tokens, the [repo-level pattern](https://arxiv.org/pdf/2409.12186) is used:
+
+```txt
+<FIM_REP>myproject
+<FIM_SEP>{chunk 0 filename}
+{chunk 0 text}
+<FIM_SEP>{chunk 1 filename}
+{chunk 1 text}
+...
+<FIM_SEP>filename
+<FIM_PRE>[input_prefix]<FIM_SUF>[input_suffix]<FIM_MID>[prompt]
+```
+
+If the tokens are missing, then the extra context is simply prefixed at the start:
+
+```txt
+[input_extra]<FIM_PRE>[input_prefix]<FIM_SUF>[input_suffix]<FIM_MID>[prompt]
+```
+
+### **GET** `/props`: Get server global properties.
+
+This endpoint is public (no API key check). By default, it is read-only. To make POST request to change global properties, you need to start server with `--props`
 
 **Response format**
 
 ```json
 {
-  "assistant_name": "",
-  "user_name": "",
   "default_generation_settings": { ... },
   "total_slots": 1,
   "chat_template": ""
 }
 ```
 
-- `assistant_name` - the required assistant name to generate the prompt in case you have specified a system prompt for all slots.
-- `user_name` - the required anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
 - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `chat_template` - the model's original Jinja2 prompt template
 
+### POST `/props`: Change server global properties.
+
+To use this endpoint with POST method, you need to start server with `--props`
+
+*Options:*
+
+- None yet
+
 ### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
 
 Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
@@ -501,7 +610,7 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
 
     See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
 
-    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}`), similar to other OpenAI-inspired API providers.
+    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
 
     *Examples:*
 
@@ -583,7 +692,10 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
 
 ### GET `/slots`: Returns the current slots processing state
 
-This endpoint can be disabled with `--no-slots`
+> [!WARNING]
+> This endpoint is intended for debugging and may be modified in future versions. For security reasons, we strongly advise against enabling it in production environments.
+
+This endpoint is disabled by default and can be enabled with `--slots`
 
 If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots.
 
@@ -600,6 +712,7 @@ Example:
         "grammar": "",
         "id": 0,
         "ignore_eos": false,
+        "is_processing": false,
         "logit_bias": [],
         "min_p": 0.05000000074505806,
         "mirostat": 0,
@@ -626,21 +739,18 @@ Example:
         "repeat_penalty": 1.100000023841858,
         "samplers": [
             "top_k",
-            "tfs_z",
             "typical_p",
             "top_p",
             "min_p",
             "temperature"
         ],
         "seed": 42,
-        "state": 1,
         "stop": [
             "\n"
         ],
         "stream": false,
         "task_id": 0,
         "temperature": 0.0,
-        "tfs_z": 1.0,
         "top_k": 40,
         "top_p": 0.949999988079071,
         "typical_p": 1.0
@@ -648,10 +758,6 @@ Example:
 ]
 ```
 
-Possible values for `slot[i].state` are:
-- `0`: SLOT_STATE_IDLE
-- `1`: SLOT_STATE_PROCESSING
-
 ### GET `/metrics`: Prometheus compatible metrics exporter
 
 This endpoint is only accessible if `--metrics` is set.
@@ -759,28 +865,6 @@ To know the `id` of the adapter, use GET `/lora-adapters`
 
 ## More examples
 
-### Change system prompt on runtime
-
-To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt`. This only needs to be used once.
-
-`prompt`: Specify a context that you want all connecting clients to respect.
-
-`anti_prompt`: Specify the word you want to use to instruct the model to stop. This must be sent to each client through the `/props` endpoint.
-
-`assistant_name`: The bot's name is necessary for each customer to generate the prompt. This must be sent to each client through the `/props` endpoint.
-
-```json
-{
-    "system_prompt": {
-        "prompt": "Transcript of a never ending dialog, where the User interacts with an Assistant.\nThe Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.\nUser: Recommend a nice restaurant in the area.\nAssistant: 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.\nUser: Who is Richard Feynman?\nAssistant: 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?\".\nUser:",
-        "anti_prompt": "User:",
-        "assistant_name": "Assistant:"
-    }
-}
-```
-
-**NOTE**: You can do this automatically when starting the server by simply creating a .json file with these options and using the CLI option `-spf FNAME` or `--system-prompt-file FNAME`.
-
 ### Interactive mode
 
 Check the sample in [chat.mjs](chat.mjs).

examples/server/public/index-new.html

@@ -40,10 +40,15 @@
       repeat_last_n: 0, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.0, // 1.0 = disabled
       penalize_nl: false, // true only useful for infinite completion
+      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
+      dry_base: 1.75,     // 0.0 = disabled
+      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
+      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
       top_k: 0, // <= 0 to use vocab size
       top_p: 1.0, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled; recommended for non-english: ~ 0.4
-      tfs_z: 1.0, // 1.0 = disabled
+      xtc_probability: 0.0, // 0 = disabled;
+      xtc_threshold: 0.1, // > 0.5 disables XTC;
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -831,11 +836,16 @@ return html`
         <fieldset class="params">
           ${IntField({ label: "Top-K", title: "Limits the selection of the next token to the K most probable tokens. 1 means no randomness = greedy sampling. If set to 0, it means the entire vocabulary size is considered.", max: 100, min: 0, step: 1, name: "top_k", value: params.value.top_k })}
           ${IntField({ label: "Penalize Last N", title: "The last n tokens that are taken into account to penalise repetitions. A value of 0 means that this function is deactivated and -1 means that the entire size of the context is taken into account.", max: 2048, min: 0, step: 16, name: "repeat_last_n", value: params.value.repeat_last_n })}
-          ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
           ${FloatField({ label: "Presence Penalty", title: "A penalty that is applied if certain tokens appear repeatedly in the generated text. A higher value leads to fewer repetitions.", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
-          ${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
           ${FloatField({ label: "Frequency Penalty", title: "A penalty that is applied based on the frequency with which certain tokens occur in the training data set. A higher value results in rare tokens being favoured.", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
+          ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
           ${FloatField({ label: "Typical-P", title: "Activates local typical sampling, a method used to limit the prediction of tokens that are atypical in the current context. The parameter p controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
+          ${FloatField({ label: "XTC probability", title: "Sets the chance for token removal (checked once on sampler start)", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
+          ${FloatField({ label: "XTC threshold", title: "Sets a minimum probability threshold for tokens to be removed", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
+          ${FloatField({ label: "DRY Penalty Multiplier", title: "Set the DRY repetition penalty multiplier. Default is 0.0, which disables DRY.", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
+          ${FloatField({ label: "DRY Base", title: "Set the DRY repetition penalty base value. Default is 1.75", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
+          ${IntField({ label: "DRY Allowed Length", title: "Tokens that extend repetition beyond this receive exponentially increasing penalty. Default is 2", max: 10, min: 1, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
+          ${IntField({ label: "DRY Penalty Last N", title: "How many tokens to scan for repetitions. Default is -1, where 0 is disabled and -1 is context size", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
           ${IntField({ label: "Min Keep", title: "If greater than 0, samplers are forced to return N possible tokens at minimum. Default is 0", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
         </fieldset>
 
@@ -1132,12 +1142,15 @@ document.addEventListener('DOMContentLoaded', (event) => {
   const snapSettings = {
     temperature: { snapValue: 1.0, snapRangeMultiplier: 6 },
     min_p: { snapValue: 0.05, snapRangeMultiplier: 2 },
+    xtc_probability: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    xtc_threshold: { snapValue: 0.5, snapRangeMultiplier: 4 },
     top_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
-    tfs_z: { snapValue: 1.0, snapRangeMultiplier: 4 },
     typical_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
     repeat_penalty: { snapValue: 1.0, snapRangeMultiplier: 4 },
     presence_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
     frequency_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    dry_multiplier: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    dry_base: { snapValue: 1.75, snapRangeMultiplier: 4 },
   };
   // add an event listener for each slider
   Object.keys(snapSettings).forEach(sliderName => {

examples/server/public/index.html

@@ -304,10 +304,15 @@
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
       penalize_nl: false,
+      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
+      dry_base: 1.75,     // 0.0 = disabled
+      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
+      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
-      tfs_z: 1.0, // 1.0 = disabled
+      xtc_probability: 0.0, // 0 = disabled;
+      xtc_threshold: 0.1, // > 0.5 disables XTC;
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -1009,10 +1014,15 @@
           <details>
             <summary>More options</summary>
             <fieldset class="two">
-              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
               ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
               ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
               ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
+              ${FloatField({ label: "DRY Penalty Multiplier", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
+              ${FloatField({ label: "DRY Base", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
+              ${IntField({ label: "DRY Allowed Length", max: 10, min: 2, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
+              ${IntField({ label: "DRY Penalty Last N", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
+              ${FloatField({ label: "XTC probability", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
+              ${FloatField({ label: "XTC threshold", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
             </fieldset>
             <hr />
             <fieldset class="three">

examples/server/public/json-schema-to-grammar.mjs

@@ -529,7 +529,7 @@ export class SchemaConverter {
       return joinSeq();
     };
 
-    return this._addRule(name, "\"\\\"\" " + toRule(transform()) + " \"\\\"\" space")
+    return this._addRule(name, "\"\\\"\" (" + toRule(transform()) + ") \"\\\"\" space")
   }
 
   _notStrings(strings) {

examples/server/tests/features/ctx_shift.feature

@@ -0,0 +1,66 @@
+@llama.cpp
+@ctx_shift
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
+    And   a model file test-model.gguf
+    And   a model alias tinyllama-2
+    And   BOS token is 1
+    And   42 as server seed
+    And   256 KV cache size
+    And   32 as batch size
+    And   2 slots
+
+    # the prompt is 301 tokens
+    # the slot context is 256/2 = 128 tokens
+    # the prompt is truncated to keep the last 109 tokens
+    # 64 tokens are generated thanks to shifting the context when it gets full
+  Scenario: Inference with context shift
+    And   64 server max tokens to predict
+    Then  the server is starting
+    Then  the server is healthy
+    Given a prompt:
+    """
+    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
+    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
+    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
+    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching fun|Annaks|popcorns|pictry|bowl
+    And   the completion is  truncated
+    And   109 prompt tokens are processed
+
+  Scenario Outline: Inference without context shift
+    And   <n_predict> server max tokens to predict
+    And   disable context shifting
+    Then  the server is starting
+    Then  the server is healthy
+    Given a prompt:
+    """
+    Hi how are you
+    """
+    And   a completion request with no api error
+    Then  <n_token_output> tokens are predicted matching twind|Anna
+    And   the completion is <truncated> truncated
+    And   8 prompt tokens are processed
+    Examples:
+      | n_predict | n_token_output | truncated |
+      | 64        | 64             | not       |
+      | -1        | 120            |           |
+
+  Scenario: Inference without context shift (expected error: prompt too long)
+    And   disable context shifting
+    Then  the server is starting
+    Then  the server is healthy
+    Given a prompt:
+    """
+    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
+    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
+    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
+    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+    """
+    And   a completion request with 400 api error
+

examples/server/tests/features/embeddings.feature

@@ -10,12 +10,12 @@ Feature: llama.cpp server
     And   42 as server seed
     And   2 slots
     # the bert-bge-small model has context size of 512
-    # since the generated prompts are as big as the batch size, we need to set the batch size to 512
+    # since the generated prompts are as big as the batch size, we need to set the batch size to <= 512
     # ref: https://huggingface.co/BAAI/bge-small-en-v1.5/blob/5c38ec7c405ec4b44b94cc5a9bb96e735b38267a/config.json#L20
-    And   512 as batch size
-    And   512 as ubatch size
-    And   2048 KV cache size
-    And   embeddings extraction
+    And   128 as batch size
+    And   128 as ubatch size
+    And   512 KV cache size
+    And   enable embeddings endpoint
     Then  the server is starting
     Then  the server is healthy
 
@@ -26,6 +26,20 @@ Feature: llama.cpp server
     """
     Then embeddings are generated
 
+  Scenario: Embedding (error: prompt too long)
+    When embeddings are computed for:
+    """
+    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
+    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
+    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
+    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
+    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
+    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
+    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+    """
+    And  embeddings request with 500 api error
+
   Scenario: OAI Embeddings compatibility
     Given a model bert-bge-small
     When an OAI compatible embeddings computation request for:

examples/server/tests/features/infill.feature

@@ -0,0 +1,36 @@
+@llama.cpp
+@infill
+Feature: llama.cpp server
+
+  # The current model is made by adding FIM tokens to the existing stories260K
+  # We may want to use a better model in the future, maybe something like SmolLM 360M
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model file tinyllamas/stories260K-infill.gguf from HF repo ggml-org/models
+    And   a model file test-model-infill.gguf
+    And   a model alias tinyllama-infill
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Infill without input_extra
+    Given a prompt "Complete this"
+    And   an infill input extra none none
+    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
+    And   an infill input suffix "}\n"
+    And   an infill request with no api error
+    Then  64 tokens are predicted matching One|day|she|saw|big|scary|bird
+
+  Scenario: Infill with input_extra
+    Given a prompt "Complete this"
+    And   an infill input extra "llama.h" "LLAMA_API int32_t llama_n_threads();\n"
+    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
+    And   an infill input suffix "}\n"
+    And   an infill request with no api error
+    Then  64 tokens are predicted matching cuts|Jimmy|mom|came|into|the|room"

examples/server/tests/features/rerank.feature

@@ -0,0 +1,42 @@
+@llama.cpp
+@rerank
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/models/resolve/main/jina-reranker-v1-tiny-en/ggml-model-f16.gguf
+    And   a model file jina-reranker-v1-tiny-en.gguf
+    And   a model alias jina-reranker-v1-tiny-en
+    And   42 as server seed
+    And   2 slots
+    And   512 as batch size
+    And   512 as ubatch size
+    And   512 KV cache size
+    And   enable reranking endpoint
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Rerank
+    Given a rerank query:
+      """
+      Machine learning is
+      """
+    And   a rerank document:
+      """
+      A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.
+      """
+    And   a rerank document:
+      """
+      Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.
+      """
+    And   a rerank document:
+      """
+      Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.
+      """
+    And   a rerank document:
+      """
+      Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine.
+      """
+    When  reranking request
+    Then  reranking results are returned
+    Then  reranking highest score is index 2 and lowest score is index 3

examples/server/tests/features/security.feature

@@ -5,7 +5,7 @@ Feature: Security
   Background: Server startup with an api key defined
     Given a server listening on localhost:8080
     And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a server api key llama.cpp
+    And   a server api key THIS_IS_THE_KEY
     Then  the server is starting
     Then  the server is healthy
 
@@ -16,11 +16,11 @@ Feature: Security
     And   a completion request with <api_error> api error
 
     Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackeme   | raised    |
-      |           | raised    |
+      | api_key         | api_error |
+      | THIS_IS_THE_KEY | no        |
+      | THIS_IS_THE_KEY | no        |
+      | hackeme         | raised    |
+      |                 | raised    |
 
   Scenario Outline: OAI Compatibility
     Given a system prompt test
@@ -32,10 +32,10 @@ Feature: Security
     Given an OAI compatible chat completions request with <api_error> api error
 
     Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackme    | raised    |
+      | api_key         | api_error |
+      | THIS_IS_THE_KEY | no        |
+      | THIS_IS_THE_KEY | no        |
+      | hackme          | raised    |
 
   Scenario Outline: OAI Compatibility (invalid response formats)
     Given a system prompt test
@@ -55,7 +55,7 @@ Feature: Security
 
 
   Scenario Outline: CORS Options
-    Given a user api key llama.cpp
+    Given a user api key THIS_IS_THE_KEY
     When  an OPTIONS request is sent from <origin>
     Then  CORS header <cors_header> is set to <cors_header_value>