llamamodel: trigger CUDA OOM early so we can fall back

Signed-off-by: Jared Van Bortel <jared@nomic.ai>

.circleci/config.yml

@@ -11,9 +11,10 @@ workflows:
           base-revision: main
           config-path: .circleci/continue_config.yml
           mapping: |
+            .circleci/.* run-all-workflows true
+            gpt4all-backend/.* run-all-workflows true
             gpt4all-bindings/python/.* run-python-workflow true
             gpt4all-bindings/typescript/.* run-ts-workflow true
             gpt4all-bindings/csharp/.* run-csharp-workflow true
-            gpt4all-backend/.* run-chat-workflow true
             gpt4all-chat/.* run-chat-workflow true
             .* run-default-workflow true

.circleci/continue_config.yml

@@ -5,6 +5,9 @@ orbs:
   node: circleci/node@5.1
 
 parameters:
+  run-all-workflows:
+    type: boolean
+    default: false
   run-default-workflow:
     type: boolean
     default: false
@@ -27,8 +30,56 @@ jobs:
       - image: circleci/python:3.7
     steps:
       - run: echo "CircleCI pipeline triggered"
-
-  build-gpt4all-chat-linux:
+  build-offline-chat-installer-macos:
+    macos:
+      xcode: 14.0.0
+    steps:
+      - checkout
+      - run:
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
+      - restore_cache:  # this is the new step to restore cache
+          keys:
+            - macos-qt-cache-v3
+      - run:
+          name: Installing Qt
+          command: |
+            if [ ! -d ~/Qt ]; then
+              curl -o qt-unified-macOS-x64-4.6.0-online.dmg https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg
+              hdiutil attach qt-unified-macOS-x64-4.6.0-online.dmg
+              /Volumes/qt-unified-macOS-x64-4.6.0-online/qt-unified-macOS-x64-4.6.0-online.app/Contents/MacOS/qt-unified-macOS-x64-4.6.0-online --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+              hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
+            fi
+      - save_cache:  # this is the new step to save cache
+          key: macos-qt-cache-v3
+          paths:
+            - ~/Qt
+      - run:
+          name: Build
+          command: |
+            mkdir build
+            cd build
+            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.7/bin
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
+              -DCMAKE_GENERATOR:STRING=Ninja \
+              -DBUILD_UNIVERSAL=ON \
+              -DMACDEPLOYQT=~/Qt/6.5.1/macos/bin/macdeployqt \
+              -DGPT4ALL_OFFLINE_INSTALLER=ON \
+              -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.5.1/macos/lib/cmake/Qt6 \
+              -DCMAKE_MAKE_PROGRAM:FILEPATH=~/Qt/Tools/Ninja/ninja \
+              -S ../gpt4all-chat \
+              -B .
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target all
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target install
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target package
+            mkdir upload
+            cp gpt4all-installer-* upload
+      - store_artifacts:
+          path: build/upload
+  build-offline-chat-installer-linux:
     machine:
       image: ubuntu-2204:2023.04.2
     steps:
@@ -40,35 +91,50 @@ jobs:
             git submodule update --init --recursive
       - restore_cache:  # this is the new step to restore cache
           keys:
-            - linux-qt-cache
+            - linux-qt-cache-v2
       - run:
           name: Setup Linux and Dependencies
           command: |
             wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
             sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-            sudo apt update && sudo apt install -y libfontconfig1 libfreetype6 libx11-6 libx11-xcb1 libxext6 libxfixes3 libxi6 libxrender1 libxcb1 libxcb-cursor0 libxcb-glx0 libxcb-keysyms1 libxcb-image0 libxcb-shm0 libxcb-icccm4 libxcb-sync1 libxcb-xfixes0 libxcb-shape0 libxcb-randr0 libxcb-render-util0 libxcb-util1 libxcb-xinerama0 libxcb-xkb1 libxkbcommon0 libxkbcommon-x11-0 bison build-essential flex gperf python3 gcc g++ libgl1-mesa-dev libwayland-dev vulkan-sdk
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
+            sudo apt update && sudo apt install -y libfontconfig1 libfreetype6 libx11-6 libx11-xcb1 libxext6 libxfixes3 libxi6 libxrender1 libxcb1 libxcb-cursor0 libxcb-glx0 libxcb-keysyms1 libxcb-image0 libxcb-shm0 libxcb-icccm4 libxcb-sync1 libxcb-xfixes0 libxcb-shape0 libxcb-randr0 libxcb-render-util0 libxcb-util1 libxcb-xinerama0 libxcb-xkb1 libxkbcommon0 libxkbcommon-x11-0 bison build-essential flex gperf python3 gcc g++ libgl1-mesa-dev libwayland-dev vulkan-sdk patchelf cuda-compiler-12-4 libcublas-dev-12-4 libnvidia-compute-550-server libmysqlclient21 libodbc2 libpq5
       - run:
           name: Installing Qt
           command: |
             if [ ! -d ~/Qt ]; then
               wget https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run
               chmod +x qt-unified-linux-x64-4.6.0-online.run
-              ./qt-unified-linux-x64-4.6.0-online.run --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.46 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver qt.qt6.651.qtwaylandcompositor
+              ./qt-unified-linux-x64-4.6.0-online.run --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver qt.qt6.651.qtwaylandcompositor
             fi
       - save_cache:  # this is the new step to save cache
-          key: linux-qt-cache
+          key: linux-qt-cache-v2
           paths:
             - ~/Qt
+      - run:
+          name: Build linuxdeployqt
+          command: |
+            git clone https://github.com/nomic-ai/linuxdeployqt
+            cd linuxdeployqt && qmake && sudo make install
       - run:
           name: Build
           command: |
+            set -eo pipefail
             export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
+            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.7/bin
+            export PATH=$PATH:/usr/local/cuda/bin
             mkdir build
             cd build
-            ~/Qt/Tools/CMake/bin/cmake -DCMAKE_BUILD_TYPE=Release -S ../gpt4all-chat -B .
+            mkdir upload
+            ~/Qt/Tools/CMake/bin/cmake -DGPT4ALL_OFFLINE_INSTALLER=ON -DCMAKE_BUILD_TYPE=Release -S ../gpt4all-chat -B .
             ~/Qt/Tools/CMake/bin/cmake --build . --target all
-
-  build-gpt4all-chat-windows:
+            ~/Qt/Tools/CMake/bin/cmake --build . --target install
+            ~/Qt/Tools/CMake/bin/cmake --build . --target package
+            cp gpt4all-installer-* upload
+      - store_artifacts:
+          path: build/upload
+  build-offline-chat-installer-windows:
     machine:
       image: 'windows-server-2019-vs2019:2022.08.1'
       resource_class: windows.large
@@ -82,16 +148,16 @@ jobs:
             git submodule update --init --recursive
       - restore_cache:  # this is the new step to restore cache
           keys:
-            - windows-qt-cache
+            - windows-qt-cache-v2
       - run:
           name: Installing Qt
           command: |
             if (-not (Test-Path C:\Qt)) {
               Invoke-WebRequest -Uri https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe -OutFile qt-unified-windows-x64-4.6.0-online.exe
-              & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.46 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+              & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
             }
       - save_cache:  # this is the new step to save cache
-          key: windows-qt-cache
+          key: windows-qt-cache-v2
           paths:
             - C:\Qt
       - run:
@@ -99,6 +165,11 @@ jobs:
           command: |
             Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
+      - run:
+          name: Install CUDA Toolkit
+          command: |
+            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/12.4.1/network_installers/cuda_12.4.1_windows_network.exe -OutFile cuda_12.4.1_windows_network.exe
+            .\cuda_12.4.1_windows_network.exe -s cudart_12.4 nvcc_12.4 cublas_12.4 cublas_dev_12.4
       - run:
           name: Build
           command: |
@@ -106,6 +177,7 @@ jobs:
             $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Windows Kits\10\bin\10.0.22000.0\x64"
             $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX64\x64"
             $Env:PATH = "${Env:PATH};C:\VulkanSDK\1.3.261.1\bin"
+            $Env:PATH = "${Env:PATH};C:\Qt\Tools\QtInstallerFramework\4.7\bin"
             $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\ucrt\x64"
             $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\um\x64"
             $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\lib\x64"
@@ -126,9 +198,122 @@ jobs:
               "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_64" `
               "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
               "-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON" `
+              "-DGPT4ALL_OFFLINE_INSTALLER=ON" `
               "-S ..\gpt4all-chat" `
               "-B ."
             & "C:\Qt\Tools\Ninja\ninja.exe"
+            & "C:\Qt\Tools\Ninja\ninja.exe" install
+            & "C:\Qt\Tools\Ninja\ninja.exe" package
+            mkdir upload
+            copy gpt4all-installer-win64.exe upload
+      - store_artifacts:
+          path: build/upload
+  build-gpt4all-chat-linux:
+    machine:
+      image: ubuntu-2204:2023.04.2
+    steps:
+      - checkout
+      - run:
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
+      - restore_cache:  # this is the new step to restore cache
+          keys:
+            - linux-qt-cache-v2
+      - run:
+          name: Setup Linux and Dependencies
+          command: |
+            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
+            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
+            sudo apt update && sudo apt install -y libfontconfig1 libfreetype6 libx11-6 libx11-xcb1 libxext6 libxfixes3 libxi6 libxrender1 libxcb1 libxcb-cursor0 libxcb-glx0 libxcb-keysyms1 libxcb-image0 libxcb-shm0 libxcb-icccm4 libxcb-sync1 libxcb-xfixes0 libxcb-shape0 libxcb-randr0 libxcb-render-util0 libxcb-util1 libxcb-xinerama0 libxcb-xkb1 libxkbcommon0 libxkbcommon-x11-0 bison build-essential flex gperf python3 gcc g++ libgl1-mesa-dev libwayland-dev vulkan-sdk cuda-compiler-12-4 libcublas-dev-12-4 libnvidia-compute-550-server libmysqlclient21 libodbc2 libpq5
+      - run:
+          name: Installing Qt
+          command: |
+            if [ ! -d ~/Qt ]; then
+              wget https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run
+              chmod +x qt-unified-linux-x64-4.6.0-online.run
+              ./qt-unified-linux-x64-4.6.0-online.run --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver qt.qt6.651.qtwaylandcompositor
+            fi
+      - save_cache:  # this is the new step to save cache
+          key: linux-qt-cache-v2
+          paths:
+            - ~/Qt
+      - run:
+          name: Build
+          command: |
+            export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
+            export PATH=$PATH:/usr/local/cuda/bin
+            ~/Qt/Tools/CMake/bin/cmake -DCMAKE_BUILD_TYPE=Release -S gpt4all-chat -B build
+            ~/Qt/Tools/CMake/bin/cmake --build build --target all
+
+  build-gpt4all-chat-windows:
+    machine:
+      image: 'windows-server-2019-vs2019:2022.08.1'
+      resource_class: windows.large
+      shell: powershell.exe -ExecutionPolicy Bypass
+    steps:
+      - checkout
+      - run:
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
+      - restore_cache:  # this is the new step to restore cache
+          keys:
+            - windows-qt-cache-v2
+      - run:
+          name: Installing Qt
+          command: |
+            if (-not (Test-Path C:\Qt)) {
+              Invoke-WebRequest -Uri https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe -OutFile qt-unified-windows-x64-4.6.0-online.exe
+              & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            }
+      - save_cache:  # this is the new step to save cache
+          key: windows-qt-cache-v2
+          paths:
+            - C:\Qt
+      - run:
+          name: Install VulkanSDK
+          command: |
+            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
+      - run:
+          name: Install CUDA Toolkit
+          command: |
+            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/12.4.1/network_installers/cuda_12.4.1_windows_network.exe -OutFile cuda_12.4.1_windows_network.exe
+            .\cuda_12.4.1_windows_network.exe -s cudart_12.4 nvcc_12.4 cublas_12.4 cublas_dev_12.4
+      - run:
+          name: Build
+          command: |
+            $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Windows Kits\10\bin\x64"
+            $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Windows Kits\10\bin\10.0.22000.0\x64"
+            $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX64\x64"
+            $Env:PATH = "${Env:PATH};C:\VulkanSDK\1.3.261.1\bin"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\ucrt\x64"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\um\x64"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\lib\x64"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\lib\x64"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\ucrt"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\um"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\shared"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\winrt"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\cppwinrt"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\VS\include"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include"
+            $Env:VULKAN_SDK = "C:\VulkanSDK\1.3.261.1"
+            & "C:\Qt\Tools\CMake_64\bin\cmake.exe" `
+              "-DCMAKE_GENERATOR:STRING=Ninja" `
+              "-DCMAKE_BUILD_TYPE=Release" `
+              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_64" `
+              "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
+              "-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON" `
+              "-S gpt4all-chat" `
+              "-B build"
+            & "C:\Qt\Tools\Ninja\ninja.exe" -C build
 
   build-gpt4all-chat-macos:
     macos:
@@ -142,51 +327,50 @@ jobs:
             git submodule update --init --recursive
       - restore_cache:  # this is the new step to restore cache
           keys:
-            - macos-qt-cache_v2
+            - macos-qt-cache-v3
       - run:
           name: Installing Qt
           command: |
             if [ ! -d ~/Qt ]; then
               curl -o qt-unified-macOS-x64-4.6.0-online.dmg https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg
               hdiutil attach qt-unified-macOS-x64-4.6.0-online.dmg
-              /Volumes/qt-unified-macOS-x64-4.6.0-online/qt-unified-macOS-x64-4.6.0-online.app/Contents/MacOS/qt-unified-macOS-x64-4.6.0-online --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.46 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+              /Volumes/qt-unified-macOS-x64-4.6.0-online/qt-unified-macOS-x64-4.6.0-online.app/Contents/MacOS/qt-unified-macOS-x64-4.6.0-online --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
               hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
             fi
       - save_cache:  # this is the new step to save cache
-          key: macos-qt-cache_v2
+          key: macos-qt-cache-v3
           paths:
             - ~/Qt
       - run:
           name: Build
           command: |
-            mkdir build
-            cd build
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
               -DCMAKE_GENERATOR:STRING=Ninja \
+              -DBUILD_UNIVERSAL=ON \
               -DCMAKE_BUILD_TYPE=Release \
               -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.5.1/macos/lib/cmake/Qt6 \
               -DCMAKE_MAKE_PROGRAM:FILEPATH=~/Qt/Tools/Ninja/ninja \
-              -S ../gpt4all-chat \
-              -B .
-            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target all
+              -S gpt4all-chat \
+              -B build
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build build --target all
   build-ts-docs: 
     docker: 
       - image: cimg/base:stable
     steps:
       - checkout
       - node/install:
-          install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - node/install-packages:
-          pkg-manager: yarn
+          pkg-manager: npm
           app-dir: gpt4all-bindings/typescript
-          override-ci-command: yarn install
+          override-ci-command: npm install --ignore-scripts
       - run:
           name: build docs ts yo
           command: |
             cd gpt4all-bindings/typescript
-            yarn docs:build
+            npm run docs:build
   build-py-docs:
     docker:
       - image: circleci/python:3.8
@@ -202,13 +386,13 @@ jobs:
       - run:
           name: Make Documentation
           command: |
-            cd gpt4all-bindings/python/
+            cd gpt4all-bindings/python
             mkdocs build
       - run:
           name: Deploy Documentation
           command: |
-            cd gpt4all-bindings/python/
-            aws s3 cp ./site s3://docs.gpt4all.io/ --recursive | cat
+            cd gpt4all-bindings/python
+            aws s3 sync --delete site/ s3://docs.gpt4all.io/
       - run:
           name: Invalidate docs.gpt4all.io cloudfront
           command: aws cloudfront create-invalidation --distribution-id E1STQOW63QL2OH --paths "/*"
@@ -226,24 +410,26 @@ jobs:
           command: |
             wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
             sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
             sudo apt-get update
-            sudo apt-get install -y cmake build-essential vulkan-sdk
+            sudo apt-get install -y cmake build-essential vulkan-sdk cuda-compiler-12-4 libcublas-dev-12-4 libnvidia-compute-550-server libmysqlclient21 libodbc2 libpq5
             pip install setuptools wheel cmake
       - run:
           name: Build C library
           command: |
-            git submodule init
-            git submodule update
+            export PATH=$PATH:/usr/local/cuda/bin
+            git submodule update --init --recursive
             cd gpt4all-backend
-            mkdir build
-            cd build
-            cmake ..
-            cmake --build . --parallel
+            cmake -B build -DCMAKE_BUILD_TYPE=Release
+            cmake --build build --parallel
       - run:
           name: Build wheel
           command: |
             cd gpt4all-bindings/python/
             python setup.py bdist_wheel --plat-name=manylinux1_x86_64
+      - store_artifacts:
+          path: gpt4all-bindings/python/dist
       - persist_to_workspace:
           root: gpt4all-bindings/python/dist
           paths:
@@ -263,69 +449,84 @@ jobs:
       - run:
           name: Build C library
           command: |
-            git submodule init
-            git submodule update
+            git submodule update --init  # don't use --recursive because macOS doesn't use Kompute
             cd gpt4all-backend
-            mkdir build
-            cd build
-            cmake .. -DCMAKE_OSX_ARCHITECTURES="x86_64;arm64"
-            cmake --build . --parallel
+            cmake -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_OSX_ARCHITECTURES="x86_64;arm64"
+            cmake --build build --parallel
       - run:
           name: Build wheel
           command: |
             cd gpt4all-bindings/python
-            python setup.py bdist_wheel --plat-name=macosx_10_9_universal2
+            python setup.py bdist_wheel --plat-name=macosx_10_15_universal2
+      - store_artifacts:
+          path: gpt4all-bindings/python/dist
       - persist_to_workspace:
           root: gpt4all-bindings/python/dist
           paths:
             - "*.whl"
 
   build-py-windows:
-    executor:
-      name: win/default
+    machine:
+      image: 'windows-server-2019-vs2019:2022.08.1'
+      resource_class: windows.large
+      shell: powershell.exe -ExecutionPolicy Bypass
     steps:
       - checkout
       - run:
-          name: Install MinGW64
-          command: choco install -y mingw --force --no-progress
-      - run:
-          name: Add MinGW64 to PATH
-          command: $env:Path += ";C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin"
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
       - run:
           name: Install VulkanSDK
           command: |
             Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
+      - run:
+          name: Install CUDA Toolkit
+          command: |
+            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/12.4.1/network_installers/cuda_12.4.1_windows_network.exe -OutFile cuda_12.4.1_windows_network.exe
+            .\cuda_12.4.1_windows_network.exe -s cudart_12.4 nvcc_12.4 cublas_12.4 cublas_dev_12.4
       - run:
           name: Install dependencies
           command:
-            choco install -y cmake --installargs 'ADD_CMAKE_TO_PATH=System'
+            choco install -y cmake ninja --installargs 'ADD_CMAKE_TO_PATH=System'
       - run:
           name: Install Python dependencies
           command: pip install setuptools wheel cmake
       - run:
           name: Build C library
           command: |
-            git submodule init
-            git submodule update
+            # Visual Studio setup
+            # I would use Enter-VsDevShell but it causes cudafe++ to segfault
+            $Env:PATH += ";C:\Program Files (x86)\Windows Kits\10\bin\x64"
+            $Env:PATH += ";C:\Program Files (x86)\Windows Kits\10\bin\10.0.22000.0\x64"
+            $Env:PATH += ";C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX64\x64"
+            $Env:LIB  = "C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\ucrt\x64"
+            $Env:LIB += ";C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\um\x64"
+            $Env:LIB += ";C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\lib\x64"
+            $Env:LIB += ";C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\lib\x64"
+            $Env:INCLUDE  = "C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\ucrt"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\um"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\shared"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\winrt"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\cppwinrt"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\VS\include"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include"
+            $Env:INCLUDE += ";C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include"
+
+            $Env:PATH += ";C:\VulkanSDK\1.3.261.1\bin"
+            $Env:VULKAN_SDK = "C:\VulkanSDK\1.3.261.1"
             cd gpt4all-backend
-            mkdir build
-            cd build
-            $env:Path += ";C:\VulkanSDK\1.3.261.1\bin"
-            cmake -G "MinGW Makefiles" .. -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=OFF
-            cmake --build . --parallel
+            cmake -G Ninja -B build -DCMAKE_BUILD_TYPE=Release -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON
+            cmake --build build --parallel
       - run:
           name: Build wheel
-          # TODO: As part of this task, we need to move mingw64 binaries into package.
-          # This is terrible and needs a more robust solution eventually.
           command: |
             cd gpt4all-bindings/python
-            cd gpt4all
-            mkdir llmodel_DO_NOT_MODIFY
-            mkdir llmodel_DO_NOT_MODIFY/build/
-            cp 'C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin\*dll' 'llmodel_DO_NOT_MODIFY/build/'
-            cd ..
             python setup.py bdist_wheel --plat-name=win_amd64
+      - store_artifacts:
+          path: gpt4all-bindings/python/dist
       - persist_to_workspace:
           root: gpt4all-bindings/python/dist
           paths:
@@ -366,11 +567,14 @@ jobs:
           command: |
             wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
             sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list            
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
             sudo apt-get update
-            sudo apt-get install -y cmake build-essential vulkan-sdk
+            sudo apt-get install -y cmake build-essential vulkan-sdk cuda-compiler-12-4 libcublas-dev-12-4 libnvidia-compute-550-server libmysqlclient21 libodbc2 libpq5
       - run:
           name: Build Libraries
           command: |
+            export PATH=$PATH:/usr/local/cuda/bin
             cd gpt4all-backend
             mkdir -p runtimes/build
             cd runtimes/build
@@ -435,6 +639,11 @@ jobs:
           command: |
             Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
+      - run:
+          name: Install CUDA Toolkit
+          command: |
+            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/12.4.1/network_installers/cuda_12.4.1_windows_network.exe -OutFile cuda_12.4.1_windows_network.exe
+            .\cuda_12.4.1_windows_network.exe -s cudart_12.4 nvcc_12.4 cublas_12.4 cublas_dev_12.4
       - run:
           name: Install dependencies
           command: |
@@ -442,10 +651,11 @@ jobs:
       - run:
           name: Build Libraries
           command: |
-            $MinGWBin = "C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin"
+            $MinGWBin = "C:\ProgramData\mingw64\mingw64\bin"
             $Env:Path += ";$MinGwBin"
             $Env:Path += ";C:\Program Files\CMake\bin"
             $Env:Path += ";C:\VulkanSDK\1.3.261.1\bin"
+            $Env:VULKAN_SDK = "C:\VulkanSDK\1.3.261.1"
             cd gpt4all-backend
             mkdir runtimes/win-x64
             cd runtimes/win-x64
@@ -477,6 +687,11 @@ jobs:
           command: |
             Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
+      - run:
+          name: Install CUDA Toolkit
+          command: |
+            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/12.4.1/network_installers/cuda_12.4.1_windows_network.exe -OutFile cuda_12.4.1_windows_network.exe
+            .\cuda_12.4.1_windows_network.exe -s cudart_12.4 nvcc_12.4 cublas_12.4 cublas_dev_12.4
       - run:
           name: Install dependencies
           command: |
@@ -486,6 +701,7 @@ jobs:
           command: |
             $Env:Path += ";C:\Program Files\CMake\bin"
             $Env:Path += ";C:\VulkanSDK\1.3.261.1\bin"
+            $Env:VULKAN_SDK = "C:\VulkanSDK\1.3.261.1"
             cd gpt4all-backend
             mkdir runtimes/win-x64_msvc
             cd runtimes/win-x64_msvc
@@ -499,7 +715,7 @@ jobs:
 
   build-csharp-linux:
     docker:
-      - image: mcr.microsoft.com/dotnet/sdk:7.0-jammy # Ubuntu 22.04
+      - image: mcr.microsoft.com/dotnet/sdk:8.0
     steps:
       - checkout
       - attach_workspace:
@@ -555,6 +771,10 @@ jobs:
             - gpt4all-csharp-nuget-packages-win
       - attach_workspace:
           at: C:\Users\circleci\workspace
+      - run:
+          name: "Install .NET"
+          command: |
+            choco install -y dotnet-8.0-sdk
       - run:
           name: "Prepare Native Libs"
           command: |
@@ -602,7 +822,8 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install --cask dotnet-sdk
+            brew tap isen-ng/dotnet-sdk-versions
+            brew install --cask dotnet-sdk8-0-100
       - attach_workspace:
           at: /tmp/workspace
       - run:
@@ -644,7 +865,7 @@ jobs:
 
   store-and-upload-nupkgs:
     docker:
-      - image: mcr.microsoft.com/dotnet/sdk:6.0-jammy # Ubuntu 22.04
+      - image: mcr.microsoft.com/dotnet/sdk:8.0
     steps:
       - attach_workspace:
           at: /tmp/workspace
@@ -660,9 +881,9 @@ jobs:
             cp /tmp/workspace/runtimes/linux-x64/*.so runtimes/linux-x64/native/
             mkdir -p runtimes/win-x64/native
             cp /tmp/workspace/runtimes/win-x64/*.dll runtimes/win-x64/native/
-            mkdir -p runtimes/osx/native
-            cp /tmp/workspace/runtimes/osx-x64/*.dylib runtimes/osx/native/
-            cp /tmp/workspace/runtimes/osx-x64/*.metal runtimes/osx/native/
+            #mkdir -p runtimes/osx/native
+            #cp /tmp/workspace/runtimes/osx-x64/*.dylib runtimes/osx/native/
+            #cp /tmp/workspace/runtimes/osx-x64/*.metal runtimes/osx/native/
             dotnet pack ./Gpt4All/Gpt4All.csproj -p:IncludeSymbols=true -p:SymbolPackageFormat=snupkg -c Release
             dotnet nuget push ./Gpt4All/bin/Release/Gpt4All.*.nupkg -s $NUGET_URL -k $NUGET_TOKEN --skip-duplicate
       - store_artifacts:
@@ -679,9 +900,11 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - node/install-packages:
           app-dir: gpt4all-bindings/typescript
           pkg-manager: yarn
+          override-ci-command: yarn install
       - run:
           command: | 
             cd gpt4all-bindings/typescript
@@ -708,9 +931,11 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - node/install-packages:
           app-dir: gpt4all-bindings/typescript
           pkg-manager: yarn
+          override-ci-command: yarn install
       - run:
           command: | 
             cd gpt4all-bindings/typescript
@@ -719,14 +944,14 @@ jobs:
           name: "Persisting all necessary things to workspace"
           command: |  
             mkdir -p gpt4all-backend/prebuilds/darwin-x64
-            mkdir -p gpt4all-backend/runtimes/darwin-x64
-            cp /tmp/gpt4all-backend/runtimes/osx-x64/*-*.* gpt4all-backend/runtimes/darwin-x64
+            mkdir -p gpt4all-backend/runtimes/darwin
+            cp /tmp/gpt4all-backend/runtimes/osx-x64/*-*.* gpt4all-backend/runtimes/darwin
             cp gpt4all-bindings/typescript/prebuilds/darwin-x64/*.node gpt4all-backend/prebuilds/darwin-x64
       - persist_to_workspace:
           root: gpt4all-backend
           paths:
             - prebuilds/darwin-x64/*.node
-            - runtimes/darwin-x64/*-*.*
+            - runtimes/darwin/*-*.*
 
   build-nodejs-windows: 
     executor:
@@ -748,6 +973,7 @@ jobs:
               nvm install 18.16.0
               nvm use 18.16.0
       - run: node --version 
+      - run: corepack enable
       - run:           
           command: |
             npm install -g yarn
@@ -781,6 +1007,7 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - run: 
           command: |
             cd gpt4all-bindings/typescript
@@ -795,9 +1022,12 @@ jobs:
             cp /tmp/gpt4all-backend/runtimes/linux-x64/*-*.so runtimes/linux-x64/native/ 
             cp /tmp/gpt4all-backend/prebuilds/linux-x64/*.node prebuilds/linux-x64/    
 
-            mkdir -p runtimes/darwin-x64/native
+            # darwin has univeral runtime libraries
+            mkdir -p runtimes/darwin/native
             mkdir -p prebuilds/darwin-x64/
-            cp /tmp/gpt4all-backend/runtimes/darwin-x64/*-*.* runtimes/darwin-x64/native/
+
+            cp /tmp/gpt4all-backend/runtimes/darwin/*-*.* runtimes/darwin/native/
+
             cp /tmp/gpt4all-backend/prebuilds/darwin-x64/*.node prebuilds/darwin-x64/    
             
             # Fallback build if user is not on above prebuilds
@@ -820,16 +1050,39 @@ jobs:
           command: |
             cd gpt4all-bindings/typescript
             npm set //registry.npmjs.org/:_authToken=$NPM_TOKEN
-            npm publish --access public --tag alpha
+            npm publish
 
 workflows:
   version: 2
   default:
-    when: << pipeline.parameters.run-default-workflow >>
+    when:
+      or:
+       - << pipeline.parameters.run-all-workflows >>
+       - << pipeline.parameters.run-default-workflow >>
     jobs:
       - default-job
+  build-chat-offline-installers:
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-chat-workflow >>
+    jobs:
+      - hold:
+          type: approval
+      - build-offline-chat-installer-macos:
+          requires:
+            - hold
+      - build-offline-chat-installer-windows:
+          requires:
+            - hold
+      - build-offline-chat-installer-linux:
+          requires:
+            - hold
   build-and-test-gpt4all-chat:
-    when: << pipeline.parameters.run-chat-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-chat-workflow >>
     jobs:
       - hold:
           type: approval
@@ -843,7 +1096,10 @@ workflows:
           requires:
             - hold
   deploy-docs:
-    when: << pipeline.parameters.run-python-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-python-workflow >>
     jobs:
       - build-ts-docs:
           filters:
@@ -856,7 +1112,10 @@ workflows:
               only:
                 - main
   build-py-deploy:
-    when: << pipeline.parameters.run-python-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-python-workflow >>
     jobs:
       - pypi-hold:
           type: approval
@@ -891,15 +1150,20 @@ workflows:
             - build-py-macos
   build-bindings:
     when: 
-        or:
-         - << pipeline.parameters.run-python-workflow >>
-         - << pipeline.parameters.run-csharp-workflow >>
-         - <<   pipeline.parameters.run-ts-workflow   >>
+      or:
+       - << pipeline.parameters.run-all-workflows >>
+       - << pipeline.parameters.run-python-workflow >>
+       - << pipeline.parameters.run-csharp-workflow >>
+       - << pipeline.parameters.run-ts-workflow >>
     jobs:
       - hold:
           type: approval
+      - csharp-hold:
+          type: approval
       - nuget-hold:
           type: approval
+      - nodejs-hold:
+          type: approval
       - npm-hold:
           type: approval
       - build-bindings-backend-linux:
@@ -942,21 +1206,21 @@ workflows:
             branches:
               only:
           requires:
-            - npm-hold
+            - nodejs-hold
             - build-bindings-backend-linux
       - build-nodejs-windows:
           filters:
             branches:
               only:
           requires:
-            - npm-hold
+            - nodejs-hold
             - build-bindings-backend-windows-msvc
       - build-nodejs-macos:
           filters:
             branches:
               only:
           requires:
-            - npm-hold
+            - nodejs-hold
             - build-bindings-backend-macos
 
 
@@ -966,21 +1230,21 @@ workflows:
             branches:
               only:
           requires:
-            - nuget-hold
+            - csharp-hold
             - build-bindings-backend-linux
       - build-csharp-windows:
           filters:
             branches:
               only:
           requires:
-            - nuget-hold
+            - csharp-hold
             - build-bindings-backend-windows
       - build-csharp-macos:
           filters:
             branches:
               only:
           requires:
-            - nuget-hold
+            - csharp-hold
             - build-bindings-backend-macos
       - store-and-upload-nupkgs:
           filters:
@@ -990,4 +1254,4 @@ workflows:
             - nuget-hold
             - build-csharp-windows
             - build-csharp-linux
-            - build-csharp-macos
+            #- build-csharp-macos

.codespellrc

@@ -1,3 +1,3 @@
 [codespell]
-ignore-words-list = blong, belong, afterall, som
+ignore-words-list = blong, afterall, som, assistent, crasher
 skip = .git,*.pdf,*.svg,*.lock

.github/ISSUE_TEMPLATE/bindings-bug.md

@@ -0,0 +1,35 @@
+---
+name: "\U0001F6E0 Bindings Bug Report"
+about: A bug report for the GPT4All Bindings
+labels: ["bindings", "bug-unconfirmed"]
+---
+
+<!-- Before creating a new issue, please make sure to take a few moments to check the issue tracker for existing issues about the bug. -->
+
+### Bug Report
+
+<!-- A clear and concise description of what the bug is. -->
+
+### Example Code
+
+<!-- Please provide a minimal code example that can be used to experience this issue. Delete this section if it does not apply. -->
+
+### Steps to Reproduce
+
+<!-- List the steps that should be taken to experience this issue. -->
+
+1.
+2.
+3.
+
+### Expected Behavior
+
+<!-- In a few words, what did you expect to happen? -->
+
+### Your Environment
+
+- Bindings version (e.g. "Version" from `pip show gpt4all`):
+- Operating System:
+- Chat model used (if applicable):
+
+<!-- You can freely edit this text, please remove all the lines you believe are unnecessary. -->

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,70 +0,0 @@
-name: "\U0001F41B Bug Report"
-description: Submit a bug report to help us improve GPT4All
-labels: ["02 Bug Report"]
-body:
-  - type: markdown
-    attributes:
-      value: >
-        Thank you for taking the time to file a bug report. Before creating a new
-        issue, please make sure to take a few moments to check the issue tracker
-        for existing issues about the bug.
-
-  - type: textarea
-    id: system-info
-    attributes:
-      label: System Info
-      description: Please share your system info with us.
-      placeholder: GPT4All version, platform, python version, etc...
-    validations:
-      required: true
-
-  - type: checkboxes
-    id: information-scripts-examples
-    attributes:
-      label: Information
-      description: "The problem arises when using:"
-      options:
-        - label: "The official example notebooks/scripts"
-        - label: "My own modified scripts"
-
-  - type: checkboxes
-    id: related-components
-    attributes:
-      label: Related Components
-      description: "Select the components related to the issue (if applicable):"
-      options:
-        - label: "backend"
-        - label: "bindings"
-        - label: "python-bindings"
-        - label: "chat-ui"
-        - label: "models"
-        - label: "circleci"
-        - label: "docker"
-        - label: "api"
-
-  - type: textarea
-    id: reproduction
-    validations:
-      required: true
-    attributes:
-      label: Reproduction
-      description: |
-        Please provide a [code sample](https://stackoverflow.com/help/minimal-reproducible-example) that reproduces the problem you ran into. It can be a Colab link or just a code snippet.
-        If you have code snippets, error messages, stack traces please provide them here as well.
-        Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting
-        Avoid screenshots when possible, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.
-
-      placeholder: |
-        Steps to reproduce the behavior:
-
-          1.
-          2.
-          3.
-
-  - type: textarea
-    id: expected-behavior
-    validations:
-      required: true
-    attributes:
-      label: Expected behavior
-      description: "A clear and concise description of what you would expect to happen."

.github/ISSUE_TEMPLATE/chat-bug.md

@@ -0,0 +1,31 @@
+---
+name: "\U0001F4AC GPT4All Bug Report"
+about: A bug report for GPT4All Chat
+labels: ["chat", "bug-unconfirmed"]
+---
+
+<!-- Before creating a new issue, please make sure to take a few moments to check the issue tracker for existing issues about the bug. -->
+
+### Bug Report
+
+<!-- A clear and concise description of what the bug is. -->
+
+### Steps to Reproduce
+
+<!-- List the steps that should be taken to experience this issue. Provide any relevant information about your configuration, and describe anything that was unexpected. -->
+
+1.
+2.
+3.
+
+### Expected Behavior
+
+<!-- In a few words, what did you expect to happen? -->
+
+### Your Environment
+
+- GPT4All version:
+- Operating System:
+- Chat model used (if applicable):
+
+<!-- You can freely edit this text, please remove all the lines you believe are unnecessary. -->

.github/ISSUE_TEMPLATE/config.yml

@@ -1,2 +1 @@
-blank_issues_enabled: false
-version: 2.1
+version: 2.1

.github/ISSUE_TEMPLATE/documentation.md

@@ -0,0 +1,9 @@
+---
+name: "\U0001F4C4 Documentation"
+about: An issue related to the GPT4All documentation
+labels: ["documentation"]
+---
+
+### Documentation
+
+<!-- Please describe the issue with the documentation as clearly as possible. -->

.github/ISSUE_TEMPLATE/documentation.yml

@@ -1,19 +0,0 @@
-name: Documentation
-description: Report an issue related to the GPT4All documentation.
-title: "DOC: <Please write a comprehensive title after the 'DOC: ' prefix>"
-labels: [03 - Documentation]
-
-body:
-- type: textarea
-  attributes: 
-    label: "Issue with current documentation:"
-    description: >
-      Please make sure to leave a reference to the document/code you're
-      referring to.
-
-- type: textarea
-  attributes:
-    label: "Idea or request for content:"
-    description: >
-      Please describe as clearly as possible what topics you think are missing
-      from the current documentation.

.github/ISSUE_TEMPLATE/feature-request.md

@@ -0,0 +1,10 @@
+---
+name: "\U0001F680 Feature Request"
+about: Submit a proposal/request for a new GPT4All feature
+title: "[Feature] Feature request title..."
+labels: ["enhancement"]
+---
+
+### Feature Request
+
+<!-- A clear and concise description of the feature proposal. -->

.github/ISSUE_TEMPLATE/feature-request.yml

@@ -1,30 +0,0 @@
-name: "\U0001F680 Feature Request"
-description: Submit a proposal/request for a new GPT4All feature
-labels: ["02 Feature Request"]
-body:
-  - type: textarea
-    id: feature-request
-    validations:
-      required: true
-    attributes:
-      label: Feature request
-      description: |
-        A clear and concise description of the feature proposal. Please provide links to any relevant GitHub repos, papers, or other resources if relevant.
-
-  - type: textarea
-    id: motivation
-    validations:
-      required: true
-    attributes:
-      label: Motivation
-      description: |
-        Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too.
-
-  - type: textarea
-    id: contribution
-    validations:
-      required: true
-    attributes:
-      label: Your contribution
-      description: |
-        Is there any way that you could help, e.g. by submitting a PR? Make sure to read the CONTRIBUTING.MD [readme](https://github.com/nomic-ai/gpt4all/blob/main/CONTRIBUTING.md)

.github/ISSUE_TEMPLATE/other-bug.md

@@ -0,0 +1,32 @@
+---
+name: "\U0001F41B Other Bug Report"
+about: A bug in another component of GPT4All
+labels: ["bug-unconfirmed"]
+---
+
+<!-- Before creating a new issue, please make sure to take a few moments to check the issue tracker for existing issues about the bug. -->
+
+### Bug Report
+
+<!-- A clear and concise description of what the bug is. -->
+
+### Steps to Reproduce
+
+<!-- List the steps that should be taken to experience this issue. Provide any relevant information about your configuration, and describe anything that was unexpected. If this bug involves original code, please provide a minimal version that can reproduce the issue. -->
+
+1.
+2.
+3.
+
+### Expected Behavior
+
+<!-- In a few words, what did you expect to happen? -->
+
+### Your Environment
+
+- GPT4All version (if applicable):
+- Operating System:
+- Chat model used (if applicable):
+
+<!-- You can freely edit this text, please remove all the lines you believe are unnecessary. -->
+

.github/ISSUE_TEMPLATE/other.yml

@@ -1,18 +0,0 @@
-name: Other Issue
-description: Raise an issue that wouldn't be covered by the other templates.
-title: "Issue: <Please write a comprehensive title after the 'Issue: ' prefix>"
-labels: [04 - Other]
-
-body:
-  - type: textarea
-    attributes:
-      label: "Issue you'd like to raise."
-      description: >
-        Please describe the issue you'd like to raise as clearly as possible.
-        Make sure to include any relevant links or references.
-
-  - type: textarea
-    attributes:
-      label: "Suggestion:"
-      description: >
-        Please outline a suggestion to improve the issue here.

.gitignore

@@ -183,4 +183,7 @@ build_*
 build-*
 
 # IntelliJ
-.idea/
+.idea/
+
+# LLM models
+*.gguf

.gitmodules

@@ -1,9 +1,4 @@
-[submodule "llama.cpp-230519"]
-	path = gpt4all-backend/llama.cpp-230519
-	url = https://github.com/ggerganov/llama.cpp.git
-[submodule "llama.cpp-230511"]
-	path = gpt4all-backend/llama.cpp-230511
-	url = https://github.com/nomic-ai/llama.cpp
 [submodule "llama.cpp-mainline"]
 	path = gpt4all-backend/llama.cpp-mainline
 	url = https://github.com/nomic-ai/llama.cpp.git
+	branch = master

LICENSE_SOM.txt

@@ -1,30 +0,0 @@
-Software for Open Models License (SOM)
-Version 1.0 dated August 30th, 2023
-
-This license governs use of the accompanying Software. If you use the Software, you accept this license. If you do not accept the license, do not use the Software.
-
-This license is intended to encourage open release of models created, modified, processed, or otherwise used via the Software under open licensing terms, and should be interpreted in light of that intent.
-
-1. Definitions
-The “Licensor” is the person or entity who is making the Software available under this license. “Software” is the software made available by Licensor under this license.
-A “Model” is the output of a machine learning algorithm, and excludes the Software.
-“Model Source Materials” must include the Model and model weights, and may include any input data, input data descriptions, documentation or training descriptions for the Model.
-“Open Licensing Terms” means: (a) any open source license approved by the Open Source Initiative, or (b) any other terms that make the Model Source Materials publicly available free of charge, and allow recipients to use, modify and distribute the Model Source Materials. Terms described in (b) may include reasonable restrictions such as non-commercial or non-production limitations, or require use in compliance with law.
-
-2. Grant of Rights. Subject to the conditions and limitations in section 3:
-(A) Copyright Grant. Licensor grants you a non-exclusive, worldwide, royalty-free copyright license to copy, modify, and distribute the Software and any modifications of the Software you create under this license. The foregoing license includes without limitation the right to create, modify, and use Models using this Software.
-
-(B) Patent Grant. Licensor grants you a non-exclusive, worldwide, royalty-free license, under any patents owned or controlled by Licensor, to make, have made, use, sell, offer for sale, import, or otherwise exploit the Software.  No license is granted to patent rights that are not embodied in the operation of the Software in the form provided by Licensor.
-
-3. Conditions and Limitations
-(A) Model Licensing and Access. If you use the Software to create, modify, process, or otherwise use any Model, including usage to create inferences with a Model, whether or not you make the Model available to others, you must make that Model Source Materials publicly available under Open Licensing Terms. 
-
-(B) No Re-Licensing. If you redistribute the Software, or modifications to the Software made under the license granted above, you must make it available only under the terms of this license. You may offer additional terms such as warranties, maintenance and support, but You, and not Licensor, are responsible for performing such terms.
-
-(C) No Trademark License. This license does not grant you rights to use the Licensor’s name, logo, or trademarks.
-
-(D) If you assert in writing a claim against any person or entity alleging that the use of the Software infringes any patent, all of your licenses to the Software under Section 2 end automatically as of the date you asserted the claim.
-
-(E) If you distribute any portion of the Software, you must retain all copyright, patent, trademark, and attribution notices that are present in the Software, and you must include a copy of this license.
-
-(F) The Software is licensed “as-is.” You bear the entire risk of using it. Licensor gives You no express warranties, guarantees or conditions. You may have additional consumer rights under your local laws that this license cannot change. To the extent permitted under your local laws, the Licensor disclaims and excludes the implied warranties of merchantability, fitness for a particular purpose and non-infringement. To the extent this disclaimer is unlawful, you, and not Licensor, are responsible for any liability.

README.md

@@ -1,72 +1,73 @@
 <h1 align="center">GPT4All</h1>
-
-<p align="center">Open-source assistant-style large language models that run locally on your CPU</p>
-
+<p align="center">Privacy-oriented software for chatting with large language models that run on your own computer.</p>
 <p align="center">
-<a href="https://gpt4all.io">GPT4All Website</a>
+  <a href="https://gpt4all.io">Official Website</a> &bull; <a href="https://docs.gpt4all.io">Documentation</a> &bull; <a href="https://discord.gg/mGZE39AS3e">Discord</a>
 </p>
-
 <p align="center">
-<a href="https://docs.gpt4all.io">GPT4All Documentation</a>
+  Official Download Links: <a href="https://gpt4all.io/installers/gpt4all-installer-win64.exe">Windows</a> &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-darwin.dmg">macOS</a> &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-linux.run">Ubuntu</a>
 </p>
-
 <p align="center">
-<a href="https://discord.gg/mGZE39AS3e">Discord</a>
+  <b>NEW:</b> <a href="https://forms.nomic.ai/gpt4all-release-notes-signup">Subscribe to our mailing list</a> for updates and news!
 </p>
-
-<p align="center">
-<a href="https://python.langchain.com/en/latest/modules/models/llms/integrations/gpt4all.html">🦜️🔗 Official Langchain Backend</a> 
-</p>
-
 <p align="center">
 GPT4All is made possible by our compute partner <a href="https://www.paperspace.com/">Paperspace</a>.
 </p>
-
 <p align="center">
-  <img width="600" height="365" src="https://user-images.githubusercontent.com/13879686/231876409-e3de1934-93bb-4b4b-9013-b491a969ebbc.gif">
-</p>
-<p align="center">
-Run on an M1 macOS Device (not sped up!)
+ <a href="https://www.phorm.ai/query?projectId=755eecd3-24ad-49cc-abf4-0ab84caacf63"><img src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg" alt="phorm.ai"></a>
 </p>
 
-## GPT4All: An ecosystem of open-source on-edge large language models.
-GPT4All is an ecosystem to train and deploy **powerful** and **customized** large language models that run locally on consumer grade CPUs. Note that your CPU needs to support [AVX or AVX2 instructions](https://en.wikipedia.org/wiki/Advanced_Vector_Extensions).
+<p align="center">
+  <img width="auto" height="400" src="https://github.com/nomic-ai/gpt4all/assets/14168726/495fce3e-769b-4e5a-a394-99f072ac4d29">
+</p>
+<p align="center">
+Run on an M2 MacBook Pro (not sped up!)
+</p>
+
+
+## About GPT4All
+
+GPT4All is an ecosystem to run **powerful** and **customized** large language models that work locally on consumer grade CPUs and NVIDIA and AMD GPUs. Note that your CPU needs to support [AVX instructions](https://en.wikipedia.org/wiki/Advanced_Vector_Extensions).
 
 Learn more in the [documentation](https://docs.gpt4all.io).
 
-The goal is simple - be the best instruction tuned assistant-style language model that any person or enterprise can freely use, distribute and build on.
-
-A GPT4All model is a 3GB - 8GB file that you can download and plug into the GPT4All open-source ecosystem software. **Nomic AI** supports and maintains this software ecosystem to enforce quality and security alongside spearheading the effort to allow any person or enterprise to easily train and deploy their own on-edge large language models. 
+A GPT4All model is a 3GB - 8GB file that you can download and plug into the GPT4All software. **Nomic AI** supports and maintains this software ecosystem to enforce quality and security alongside spearheading the effort to allow any person or enterprise to easily deploy their own on-edge large language models.
 
 
-### Chat Client
-Run any GPT4All model natively on your home desktop with the auto-updating desktop chat client. See <a href="https://gpt4all.io">GPT4All Website</a> for a full list of open-source models you can run with this powerful desktop application.
+### Installation
 
-Direct Installer Links:
+The recommended way to install GPT4All is to use one of the online installers linked above in this README, which are also available at the [GPT4All website](https://gpt4all.io/). These require an internet connection at install time, are slightly easier to use on macOS due to code signing, and provide a version of GPT4All that can check for updates.
 
-* [macOS](https://gpt4all.io/installers/gpt4all-installer-darwin.dmg)
+An alternative way to install GPT4All is to use one of the offline installers available on the [Releases page](https://github.com/nomic-ai/gpt4all/releases). These do not require an internet connection at install time, and can be used to install an older version of GPT4All if so desired. But using these requires acknowledging a security warning on macOS, and they provide a version of GPT4All that is unable to notify you of updates, so you should enable notifications for Releases on this repository (Watch > Custom > Releases) or sign up for announcements in our [Discord server](https://discord.gg/mGZE39AS3e).
 
-* [Windows](https://gpt4all.io/installers/gpt4all-installer-win64.exe)
 
-* [Ubuntu](https://gpt4all.io/installers/gpt4all-installer-linux.run)
+### What's New
+- **October 19th, 2023**: GGUF Support Launches with Support for:
+    - Mistral 7b base model, an updated model gallery on [gpt4all.io](https://gpt4all.io), several new local code models including Rift Coder v1.5
+    - [Nomic Vulkan](https://blog.nomic.ai/posts/gpt4all-gpu-inference-with-vulkan) support for Q4\_0 and Q4\_1 quantizations in GGUF.
+    - Offline build support for running old versions of the GPT4All Local LLM Chat Client.
+- **September 18th, 2023**: [Nomic Vulkan](https://blog.nomic.ai/posts/gpt4all-gpu-inference-with-vulkan) launches supporting local LLM inference on NVIDIA and AMD GPUs.
+- **July 2023**: Stable support for LocalDocs, a feature that allows you to privately and locally chat with your data.
+- **June 28th, 2023**: [Docker-based API server] launches allowing inference of local LLMs from an OpenAI-compatible HTTP endpoint.
 
-Find the most up-to-date information on the [GPT4All Website](https://gpt4all.io/)
+[Docker-based API server]: https://github.com/nomic-ai/gpt4all/tree/cef74c2be20f5b697055d5b8b506861c7b997fab/gpt4all-api
 
-### Chat Client building and running
 
-* Follow the visual instructions on the chat client [build_and_run](gpt4all-chat/build_and_run.md) page
+### Building From Source
+
+* Follow the instructions [here](gpt4all-chat/build_and_run.md) to build the GPT4All Chat UI from source.
+
 
 ### Bindings
 
-* <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/python/README.md">:snake: Official Python Bindings</a> [![Downloads](https://static.pepy.tech/badge/gpt4all/week)](https://pepy.tech/project/gpt4all)
-* <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/typescript">:computer: Official Typescript Bindings</a>
-* <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/golang">:computer: Official GoLang Bindings</a>
-* <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/csharp">:computer: Official C# Bindings</a>
-* <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/java">:computer: Official Java Bindings</a>
+* :snake: <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/python">Official Python Bindings</a> [![Downloads](https://static.pepy.tech/badge/gpt4all/week)](https://pepy.tech/project/gpt4all)
+* :computer: <a href="https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/typescript">Typescript Bindings</a>
+
 
 ### Integrations
 
-* 🗃️ [Weaviate Vector Database](https://github.com/weaviate/weaviate) - [module docs](https://weaviate.io/developers/weaviate/modules/retriever-vectorizer-modules/text2vec-gpt4all)
+* :parrot::link: [Langchain](https://python.langchain.com/en/latest/modules/models/llms/integrations/gpt4all.html)
+* :card_file_box: [Weaviate Vector Database](https://github.com/weaviate/weaviate) - [module docs](https://weaviate.io/developers/weaviate/modules/retriever-vectorizer-modules/text2vec-gpt4all)
+
 
 ## Contributing
 GPT4All welcomes contributions, involvement, and discussion from the open source community!
@@ -76,6 +77,59 @@ Check project discord, with project owners, or through existing issues/PRs to av
 Please make sure to tag all of the above with relevant project identifiers or your contribution could potentially get lost.
 Example tags: `backend`, `bindings`, `python-bindings`, `documentation`, etc.
 
+
+## GPT4All 2024 Roadmap
+To contribute to the development of any of the below roadmap items, make or find the corresponding issue and cross-reference the [in-progress task](https://github.com/orgs/nomic-ai/projects/2/views/1).
+
+Each item should have an issue link below.
+
+- Chat UI Language Localization (localize UI into the native languages of users)
+    - [ ] Chinese
+    - [ ] German
+    - [ ] French
+    - [ ] Portuguese
+    - [ ] Your native language here. 
+- UI Redesign: an internal effort at Nomic to improve the UI/UX of gpt4all for all users.
+    - [ ] Design new user interface and gather community feedback
+    - [ ] Implement the new user interface and experience.
+- Installer and Update Improvements
+    - [ ] Seamless native installation and update process on OSX
+    - [ ] Seamless native installation and update process on Windows
+    - [ ] Seamless native installation and update process on Linux
+- Model discoverability improvements:
+    - [x] Support huggingface model discoverability
+    - [ ] Support Nomic hosted model discoverability
+- LocalDocs (towards a local perplexity)
+    - Multilingual LocalDocs Support
+        - [ ] Create a multilingual experience
+        - [ ] Incorporate a multilingual embedding model
+        - [ ] Specify a preferred multilingual LLM for localdocs
+    - Improved RAG techniques
+        - [ ] Query augmentation and re-writing
+        - [ ] Improved chunking and text extraction from arbitrary modalities
+            - [ ] Custom PDF extractor past the QT default (charts, tables, text)
+        - [ ] Faster indexing and local exact search with v1.5 hamming embeddings and reranking (skip ANN index construction!)
+    - Support queries like 'summarize X document'
+    - Multimodal LocalDocs support with Nomic Embed
+    - Nomic Dataset Integration with real-time LocalDocs
+        - [ ] Include an option to allow the export of private LocalDocs collections to Nomic Atlas for debugging data/chat quality
+        - [ ] Allow optional sharing of LocalDocs collections between users.
+        - [ ] Allow the import of a LocalDocs collection from an Atlas Datasets
+            - Chat with live version of Wikipedia, Chat with Pubmed, chat with the latest snapshot of world news.
+- First class Multilingual LLM Support
+    - [ ] Recommend and set a default LLM for German
+    - [ ] Recommend and set a default LLM for English
+    - [ ] Recommend and set a default LLM for Chinese
+    - [ ] Recommend and set a default LLM for Spanish
+
+- Server Mode improvements
+    - Improved UI and new requested features:
+        - [ ] Fix outstanding bugs and feature requests around networking configurations.
+        - [ ] Support Nomic Embed inferencing
+        - [ ] First class documentation
+        - [ ] Improving developer use and quality of server mode (e.g. support larger batches)
+
+
 ## Technical Reports
 
 <p align="center">
@@ -90,6 +144,7 @@ Example tags: `backend`, `bindings`, `python-bindings`, `documentation`, etc.
 <a href="https://s3.amazonaws.com/static.nomic.ai/gpt4all/2023_GPT4All_Technical_Report.pdf">:green_book: Technical Report 1: GPT4All</a>
 </p>
 
+
 ## Citation
 
 If you utilize this repository, models or data in a downstream project, please consider citing it with:

gpt4all-api/.gitignore

@@ -1,112 +0,0 @@
-# Byte-compiled / optimized / DLL files
-__pycache__/
-app/__pycache__/
-gpt4all_api/__pycache__/
-gpt4all_api/app/api_v1/__pycache__/
-*.py[cod]
-*$py.class
-
-# C extensions
-*.so
-
-# VS Code
-.vscode/
-
-# Distribution / packaging
-.Python
-build/
-develop-eggs/
-dist/
-downloads/
-eggs/
-.eggs/
-lib64/
-parts/
-sdist/
-var/
-wheels/
-*.egg-info/
-.installed.cfg
-*.egg
-MANIFEST
-
-# PyInstaller
-#  Usually these files are written by a python script from a template
-#  before PyInstaller builds the exe, so as to inject date/other infos into it.
-*.manifest
-*.spec
-
-# Installer logs
-pip-log.txt
-pip-delete-this-directory.txt
-
-# Unit test / coverage reports
-htmlcov/
-.tox/
-.coverage
-.coverage.*
-.cache
-nosetests.xml
-coverage.xml
-*.cover
-.hypothesis/
-.pytest_cache/
-
-# Translations
-*.mo
-*.pot
-
-# Django stuff:
-*.log
-local_settings.py
-db.sqlite3
-
-# Flask stuff:
-instance/
-.webassets-cache
-
-# Scrapy stuff:
-.scrapy
-
-# Sphinx documentation
-docs/_build/
-
-# PyBuilder
-target/
-
-# Jupyter Notebook
-.ipynb_checkpoints
-
-# pyenv
-.python-version
-
-# celery beat schedule file
-celerybeat-schedule
-
-# SageMath parsed files
-*.sage.py
-
-# Environments
-.env
-.venv
-env/
-venv/
-ENV/
-env.bak/
-venv.bak/
-
-# Spyder project settings
-.spyderproject
-.spyproject
-
-# Rope project settings
-.ropeproject
-
-# mkdocs documentation
-/site
-
-# mypy
-.mypy_cache/
-
-*.lock
-*.cache

gpt4all-api/.isort.cfg

@@ -1,7 +0,0 @@
-[settings]
-known_third_party=geopy,nltk,np,numpy,pandas,pysbd,fire,torch
-
-line_length=120
-include_trailing_comma=True
-multi_line_output=3
-use_parentheses=True

gpt4all-api/LICENSE

@@ -1,13 +0,0 @@
-Copyright 2023 Nomic, Inc.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.

gpt4all-api/README.md

@@ -1,87 +0,0 @@
-# GPT4All REST API
-This directory contains the source code to run and build docker images that run a FastAPI app
-for serving inference from GPT4All models. The API matches the OpenAI API spec.
-
-## Tutorial
-
-The following tutorial assumes that you have checked out this repo and cd'd into it.
-
-### Starting the app
-
-First change your working directory to `gpt4all/gpt4all-api`.
-
-Now you can build the FastAPI docker image. You only have to do this on initial build or when you add new dependencies to the requirements.txt file:
-```bash
-DOCKER_BUILDKIT=1 docker build -t gpt4all_api --progress plain -f gpt4all_api/Dockerfile.buildkit .
-```
-
-Then, start the backend with:
-
-```bash
-docker compose up --build
-```
-
-This will run both the API and locally hosted GPU inference server. If you want to run the API without the GPU inference server, you can run:
-
-```bash
-docker compose up --build gpt4all_api
-```
-
-To run the API with the GPU inference server, you will need to include environment variables (like the `MODEL_ID`). Edit the `.env` file and run
-```bash
-docker compose --env-file .env up --build
-```
-
-
-#### Spinning up your app
-Run `docker compose up` to spin up the backend. Monitor the logs for errors in-case you forgot to set an environment variable above.
-
-
-#### Development
-Run
-
-```bash
-docker compose up --build
-```
-and edit files in the `api` directory. The api will hot-reload on changes.
-
-You can run the unit tests with
-
-```bash
-make test
-```
-
-#### Viewing API documentation
-
-Once the FastAPI ap is started you can access its documentation and test the search endpoint by going to:
-```
-localhost:80/docs
-```
-
-This documentation should match the OpenAI OpenAPI spec located at https://github.com/openai/openai-openapi/blob/master/openapi.yaml
-
-
-#### Running inference
-```python
-import openai
-openai.api_base = "http://localhost:4891/v1"
-
-openai.api_key = "not needed for a local LLM"
-
-
-def test_completion():
-    model = "gpt4all-j-v1.3-groovy"
-    prompt = "Who is Michael Jordan?"
-    response = openai.Completion.create(
-        model=model,
-        prompt=prompt,
-        max_tokens=50,
-        temperature=0.28,
-        top_p=0.95,
-        n=1,
-        echo=True,
-        stream=False
-    )
-    assert len(response['choices'][0]['text']) > len(prompt)
-    print(response)
-```

gpt4all-api/docker-compose.gpu.yaml

@@ -1,24 +0,0 @@
-version: "3.8"
-
-services:
-  gpt4all_gpu:
-    image: ghcr.io/huggingface/text-generation-inference:0.9.3
-    container_name: gpt4all_gpu
-    restart: always #restart on error (usually code compilation from save during bad state)
-    environment:
-      - HUGGING_FACE_HUB_TOKEN=token
-      - USE_FLASH_ATTENTION=false
-      - MODEL_ID=''
-      - NUM_SHARD=1
-    command: --model-id $MODEL_ID --num-shard $NUM_SHARD
-    volumes:
-      - ./:/data
-    ports:
-      - "8080:80"
-    shm_size: 1g
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              capabilities: [gpu]

gpt4all-api/docker-compose.yaml

@@ -1,19 +0,0 @@
-version: "3.8"
-
-services:
-  gpt4all_api:
-    image: gpt4all_api
-    container_name: gpt4all_api
-    restart: always #restart on error (usually code compilation from save during bad state)
-    ports:
-      - "4891:4891"
-    environment:
-      - APP_ENVIRONMENT=dev
-      - WEB_CONCURRENCY=2
-      - LOGLEVEL=debug
-      - PORT=4891
-      - model=ggml-mpt-7b-chat.bin
-      - inference_mode=cpu
-    volumes:
-      - './gpt4all_api/app:/app'
-    command: ["/start-reload.sh"]

gpt4all-api/gpt4all_api/Dockerfile.buildkit

@@ -1,23 +0,0 @@
-# syntax=docker/dockerfile:1.0.0-experimental
-FROM tiangolo/uvicorn-gunicorn:python3.11
-
-ARG MODEL_BIN=ggml-mpt-7b-chat.bin
-
-# Put first so anytime this file changes other cached layers are invalidated.
-COPY gpt4all_api/requirements.txt /requirements.txt
-
-RUN pip install --upgrade pip
-
-# Run various pip install commands with ssh keys from host machine.
-RUN --mount=type=ssh pip install -r /requirements.txt && \
-  rm -Rf /root/.cache && rm -Rf /tmp/pip-install*
-
-# Finally, copy app and client.
-COPY gpt4all_api/app /app
-
-RUN mkdir -p /models
-
-# Include the following line to bake a model into the image and not have to download it on API start.
-RUN wget -q --show-progress=off https://gpt4all.io/models/${MODEL_BIN} -P /models \
-  && md5sum /models/${MODEL_BIN}
-

gpt4all-api/gpt4all_api/README.md

@@ -1 +0,0 @@
-# FastAPI app for serving GPT4All models

gpt4all-api/gpt4all_api/app/api_v1/api.py

@@ -1,9 +0,0 @@
-from api_v1.routes import chat, completions, engines, health
-from fastapi import APIRouter
-
-router = APIRouter()
-
-router.include_router(chat.router)
-router.include_router(completions.router)
-router.include_router(engines.router)
-router.include_router(health.router)

gpt4all-api/gpt4all_api/app/api_v1/events.py

@@ -1,29 +0,0 @@
-import logging
-
-from api_v1.settings import settings
-from fastapi import HTTPException
-from fastapi.responses import JSONResponse
-from starlette.requests import Request
-
-log = logging.getLogger(__name__)
-
-
-startup_msg_fmt = """
- Starting up GPT4All API
-"""
-
-
-async def on_http_error(request: Request, exc: HTTPException):
-    return JSONResponse({'detail': exc.detail}, status_code=exc.status_code)
-
-
-async def on_startup(app):
-    startup_msg = startup_msg_fmt.format(settings=settings)
-    log.info(startup_msg)
-
-
-def startup_event_handler(app):
-    async def start_app() -> None:
-        await on_startup(app)
-
-    return start_app

gpt4all-api/gpt4all_api/app/api_v1/routes/chat.py

@@ -1,61 +0,0 @@
-import logging
-import time
-from typing import Dict, List
-
-from api_v1.settings import settings
-from fastapi import APIRouter, Depends, Response, Security, status
-from pydantic import BaseModel, Field
-
-logger = logging.getLogger(__name__)
-logger.setLevel(logging.DEBUG)
-
-### This should follow https://github.com/openai/openai-openapi/blob/master/openapi.yaml
-
-
-class ChatCompletionMessage(BaseModel):
-    role: str
-    content: str
-
-
-class ChatCompletionRequest(BaseModel):
-    model: str = Field(..., description='The model to generate a completion from.')
-    messages: List[ChatCompletionMessage] = Field(..., description='The model to generate a completion from.')
-
-
-class ChatCompletionChoice(BaseModel):
-    message: ChatCompletionMessage
-    index: int
-    finish_reason: str
-
-
-class ChatCompletionUsage(BaseModel):
-    prompt_tokens: int
-    completion_tokens: int
-    total_tokens: int
-
-
-class ChatCompletionResponse(BaseModel):
-    id: str
-    object: str = 'text_completion'
-    created: int
-    model: str
-    choices: List[ChatCompletionChoice]
-    usage: ChatCompletionUsage
-
-
-router = APIRouter(prefix="/chat", tags=["Completions Endpoints"])
-
-
-@router.post("/completions", response_model=ChatCompletionResponse)
-async def chat_completion(request: ChatCompletionRequest):
-    '''
-    Completes a GPT4All model response.
-    '''
-
-    return ChatCompletionResponse(
-        id='asdf',
-        created=time.time(),
-        model=request.model,
-        choices=[{}],
-        usage={'prompt_tokens': 0, 'completion_tokens': 0, 'total_tokens': 0},
-    )

gpt4all-api/gpt4all_api/app/api_v1/routes/completions.py

@@ -1,215 +0,0 @@
-import json
-from typing import List, Dict, Iterable, AsyncIterable
-import logging
-import time
-from typing import Dict, List, Union, Optional
-from uuid import uuid4
-import aiohttp
-import asyncio
-from api_v1.settings import settings
-from fastapi import APIRouter, Depends, Response, Security, status, HTTPException
-from fastapi.responses import StreamingResponse
-from gpt4all import GPT4All
-from pydantic import BaseModel, Field
-
-logger = logging.getLogger(__name__)
-logger.setLevel(logging.DEBUG)
-
-
-### This should follow https://github.com/openai/openai-openapi/blob/master/openapi.yaml
-
-
-class CompletionRequest(BaseModel):
-    model: str = Field(settings.model, description='The model to generate a completion from.')
-    prompt: Union[List[str], str] = Field(..., description='The prompt to begin completing from.')
-    max_tokens: int = Field(None, description='Max tokens to generate')
-    temperature: float = Field(settings.temp, description='Model temperature')
-    top_p: Optional[float] = Field(settings.top_p, description='top_p')
-    top_k: Optional[int] = Field(settings.top_k, description='top_k')
-    n: int = Field(1, description='How many completions to generate for each prompt')
-    stream: bool = Field(False, description='Stream responses')
-    repeat_penalty: float = Field(settings.repeat_penalty, description='Repeat penalty')
-
-
-class CompletionChoice(BaseModel):
-    text: str
-    index: int
-    logprobs: float
-    finish_reason: str
-
-
-class CompletionUsage(BaseModel):
-    prompt_tokens: int
-    completion_tokens: int
-    total_tokens: int
-
-
-class CompletionResponse(BaseModel):
-    id: str
-    object: str = 'text_completion'
-    created: int
-    model: str
-    choices: List[CompletionChoice]
-    usage: CompletionUsage
-
-
-class CompletionStreamResponse(BaseModel):
-    id: str
-    object: str = 'text_completion'
-    created: int
-    model: str
-    choices: List[CompletionChoice]
-
-
-router = APIRouter(prefix="/completions", tags=["Completion Endpoints"])
-
-def stream_completion(output: Iterable, base_response: CompletionStreamResponse):
-    """
-    Streams a GPT4All output to the client.
-
-    Args:
-        output: The output of GPT4All.generate(), which is an iterable of tokens.
-        base_response: The base response object, which is cloned and modified for each token.
-
-    Returns:
-        A Generator of CompletionStreamResponse objects, which are serialized to JSON Event Stream format.
-    """
-    for token in output:
-        chunk = base_response.copy()
-        chunk.choices = [dict(CompletionChoice(
-            text=token,
-            index=0,
-            logprobs=-1,
-            finish_reason=''
-        ))]
-        yield f"data: {json.dumps(dict(chunk))}\n\n"
-
-async def gpu_infer(payload, header):
-    async with aiohttp.ClientSession() as session:
-        try:
-            async with session.post(
-                settings.hf_inference_server_host, headers=header, data=json.dumps(payload)
-            ) as response:
-                resp = await response.json()
-            return resp
-
-        except aiohttp.ClientError as e:
-            # Handle client-side errors (e.g., connection error, invalid URL)
-            logger.error(f"Client error: {e}")
-        except aiohttp.ServerError as e:
-            # Handle server-side errors (e.g., internal server error)
-            logger.error(f"Server error: {e}")
-        except json.JSONDecodeError as e:
-            # Handle JSON decoding errors
-            logger.error(f"JSON decoding error: {e}")
-        except Exception as e:
-            # Handle other unexpected exceptions
-            logger.error(f"Unexpected error: {e}")
-
-@router.post("/", response_model=CompletionResponse)
-async def completions(request: CompletionRequest):
-    '''
-    Completes a GPT4All model response.
-    '''
-    if settings.inference_mode == "gpu":
-        params = request.dict(exclude={'model', 'prompt', 'max_tokens', 'n'})
-        params["max_new_tokens"] = request.max_tokens
-        params["num_return_sequences"] = request.n
-
-        header = {"Content-Type": "application/json"}
-        if isinstance(request.prompt, list):
-            tasks = []
-            for prompt in request.prompt:
-                payload = {"parameters": params}
-                payload["inputs"] = prompt
-                task = gpu_infer(payload, header)
-                tasks.append(task)
-            results = await asyncio.gather(*tasks)
-
-            choices = []
-            for response in results:
-                scores = response["scores"] if "scores" in response else -1.0
-                choices.append(
-                    dict(
-                        CompletionChoice(
-                            text=response["generated_text"], index=0, logprobs=scores, finish_reason='stop'
-                        )
-                    )
-                )
-
-            return CompletionResponse(
-                id=str(uuid4()),
-                created=time.time(),
-                model=request.model,
-                choices=choices,
-                usage={'prompt_tokens': 0, 'completion_tokens': 0, 'total_tokens': 0},
-            )
-
-        else:
-            payload = {"parameters": params}
-            # If streaming, we need to return a StreamingResponse
-            payload["inputs"] = request.prompt
-
-            resp = await gpu_infer(payload, header)
-
-            output = resp["generated_text"]
-            # this returns all logprobs
-            scores = resp["scores"] if "scores" in resp else -1.0
-
-            return CompletionResponse(
-                id=str(uuid4()),
-                created=time.time(),
-                model=request.model,
-                choices=[dict(CompletionChoice(text=output, index=0, logprobs=scores, finish_reason='stop'))],
-                usage={'prompt_tokens': 0, 'completion_tokens': 0, 'total_tokens': 0},
-            )
-
-    else:
-
-        if request.model != settings.model:
-            raise HTTPException(status_code=400,
-                                detail=f"The GPT4All inference server is booted to only infer: `{settings.model}`")
-
-        if isinstance(request.prompt, list):
-            if len(request.prompt) > 1:
-                raise HTTPException(status_code=400, detail="Can only infer one inference per request in CPU mode.")
-            else:
-                request.prompt = request.prompt[0]
-
-        model = GPT4All(model_name=settings.model, model_path=settings.gpt4all_path)
-
-        output = model.generate(prompt=request.prompt,
-                                max_tokens=request.max_tokens,
-                                streaming=request.stream,
-                                top_k=request.top_k,
-                                top_p=request.top_p,
-                                temp=request.temperature,
-                                )
-
-        # If streaming, we need to return a StreamingResponse
-        if request.stream:
-            base_chunk = CompletionStreamResponse(
-                id=str(uuid4()),
-                created=time.time(),
-                model=request.model,
-                choices=[]
-            )
-            return StreamingResponse((response for response in stream_completion(output, base_chunk)),
-                                     media_type="text/event-stream")
-        else:
-            return CompletionResponse(
-                id=str(uuid4()),
-                created=time.time(),
-                model=request.model,
-                choices=[dict(CompletionChoice(
-                    text=output,
-                    index=0,
-                    logprobs=-1,
-                    finish_reason='stop'
-                ))],
-                usage={
-                    'prompt_tokens': 0,  # TODO how to compute this?
-                    'completion_tokens': 0,
-                    'total_tokens': 0
-                }
-            )

gpt4all-api/gpt4all_api/app/api_v1/routes/embeddings.py

@@ -1,65 +0,0 @@
-from typing import List, Union
-from fastapi import APIRouter
-from api_v1.settings import settings
-from gpt4all import Embed4All
-from pydantic import BaseModel, Field
-
-### This should follow https://github.com/openai/openai-openapi/blob/master/openapi.yaml
-
-
-class EmbeddingRequest(BaseModel):
-    model: str = Field(
-        settings.model, description="The model to generate an embedding from."
-    )
-    input: Union[str, List[str], List[int], List[List[int]]] = Field(
-        ..., description="Input text to embed, encoded as a string or array of tokens."
-    )
-
-
-class EmbeddingUsage(BaseModel):
-    prompt_tokens: int = 0
-    total_tokens: int = 0
-
-
-class Embedding(BaseModel):
-    index: int = 0
-    object: str = "embedding"
-    embedding: List[float]
-
-
-class EmbeddingResponse(BaseModel):
-    object: str = "list"
-    model: str
-    data: List[Embedding]
-    usage: EmbeddingUsage
-
-
-router = APIRouter(prefix="/embeddings", tags=["Embedding Endpoints"])
-
-embedder = Embed4All()
-
-
-def get_embedding(data: EmbeddingRequest) -> EmbeddingResponse:
-    """
-    Calculates the embedding for the given input using a specified model.
-
-    Args:
-        data (EmbeddingRequest): An EmbeddingRequest object containing the input data
-        and model name.
-
-    Returns:
-        EmbeddingResponse: An EmbeddingResponse object encapsulating the calculated embedding,
-        usage info, and the model name.
-    """
-    embedding = embedder.embed(data.input)
-    return EmbeddingResponse(
-        data=[Embedding(embedding=embedding)], usage=EmbeddingUsage(), model=data.model
-    )
-
-
-@router.post("/", response_model=EmbeddingResponse)
-def embeddings(data: EmbeddingRequest):
-    """
-    Creates a GPT4All embedding
-    """
-    return get_embedding(data)

gpt4all-api/gpt4all_api/app/api_v1/routes/engines.py

@@ -1,40 +0,0 @@
-import logging
-from typing import Dict, List
-
-from api_v1.settings import settings
-from fastapi import APIRouter, Depends, Response, Security, status
-from pydantic import BaseModel, Field
-
-logger = logging.getLogger(__name__)
-logger.setLevel(logging.DEBUG)
-
-### This should follow https://github.com/openai/openai-openapi/blob/master/openapi.yaml
-
-
-class ListEnginesResponse(BaseModel):
-    data: List[Dict] = Field(..., description="All available models.")
-
-
-class EngineResponse(BaseModel):
-    data: List[Dict] = Field(..., description="All available models.")
-
-
-router = APIRouter(prefix="/engines", tags=["Search Endpoints"])
-
-
-@router.get("/", response_model=ListEnginesResponse)
-async def list_engines():
-    '''
-    List all available GPT4All models from
-    https://raw.githubusercontent.com/nomic-ai/gpt4all/main/gpt4all-chat/metadata/models.json
-    '''
-    raise NotImplementedError()
-    return ListEnginesResponse(data=[])
-
-
-@router.get("/{engine_id}", response_model=EngineResponse)
-async def retrieve_engine(engine_id: str):
-    ''' '''
-
-    raise NotImplementedError()
-    return EngineResponse()

gpt4all-api/gpt4all_api/app/api_v1/routes/health.py

@@ -1,13 +0,0 @@
-import logging
-from fastapi import APIRouter
-from fastapi.responses import JSONResponse
-
-log = logging.getLogger(__name__)
-
-router = APIRouter(prefix="/health", tags=["Health"])
-
-
-@router.get('/', response_class=JSONResponse)
-async def health_check():
-    """Runs a health check on this instance of the API."""
-    return JSONResponse({'status': 'ok'}, headers={'Access-Control-Allow-Origin': '*'})

gpt4all-api/gpt4all_api/app/api_v1/settings.py

@@ -1,19 +0,0 @@
-from pydantic import BaseSettings
-
-
-class Settings(BaseSettings):
-    app_environment = 'dev'
-    model: str = 'ggml-mpt-7b-chat.bin'
-    gpt4all_path: str = '/models'
-    inference_mode: str = "cpu"
-    hf_inference_server_host: str = "http://gpt4all_gpu:80/generate"
-    sentry_dns: str = None
-
-    temp: float = 0.18
-    top_p: float = 1.0
-    top_k: int = 50
-    repeat_penalty: float = 1.18
-
-
-
-settings = Settings()

gpt4all-api/gpt4all_api/app/docs.py

@@ -1,3 +0,0 @@
-desc = 'GPT4All API'
-
-endpoint_paths = {'health': '/health'}

gpt4all-api/gpt4all_api/app/main.py

@@ -1,84 +0,0 @@
-import logging
-import os
-
-import docs
-from api_v1 import events
-from api_v1.api import router as v1_router
-from api_v1.settings import settings
-from fastapi import FastAPI, HTTPException, Request
-from fastapi.logger import logger as fastapi_logger
-from starlette.middleware.cors import CORSMiddleware
-
-logger = logging.getLogger(__name__)
-
-app = FastAPI(title='GPT4All API', description=docs.desc)
-
-# CORS Configuration (in-case you want to deploy)
-app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_credentials=True,
-    allow_methods=["GET", "POST", "OPTIONS"],
-    allow_headers=["*"],
-)
-
-logger.info('Adding v1 endpoints..')
-
-# add v1
-app.include_router(v1_router, prefix='/v1')
-app.add_event_handler('startup', events.startup_event_handler(app))
-app.add_exception_handler(HTTPException, events.on_http_error)
-
-
-@app.on_event("startup")
-async def startup():
-    global model
-    if settings.inference_mode == "cpu":
-        logger.info(f"Downloading/fetching model: {os.path.join(settings.gpt4all_path, settings.model)}")
-        from gpt4all import GPT4All
-
-        model = GPT4All(model_name=settings.model, model_path=settings.gpt4all_path)
-
-        logger.info(f"GPT4All API is ready to infer from {settings.model} on CPU.")
-
-    else:
-        # is it possible to do this once the server is up?
-        ## TODO block until HF inference server is up.
-        logger.info(f"GPT4All API is ready to infer from {settings.model} on CPU.")
-
-
-
-@app.on_event("shutdown")
-async def shutdown():
-    logger.info("Shutting down API")
-
-
-if settings.sentry_dns is not None:
-    import sentry_sdk
-
-    def traces_sampler(sampling_context):
-        if 'health' in sampling_context['transaction_context']['name']:
-            return False
-
-    sentry_sdk.init(
-        dsn=settings.sentry_dns, traces_sample_rate=0.1, traces_sampler=traces_sampler, send_default_pii=False
-    )
-
-# This is needed to get logs to show up in the app
-if "gunicorn" in os.environ.get("SERVER_SOFTWARE", ""):
-    gunicorn_error_logger = logging.getLogger("gunicorn.error")
-    gunicorn_logger = logging.getLogger("gunicorn")
-
-    root_logger = logging.getLogger()
-    fastapi_logger.setLevel(gunicorn_logger.level)
-    fastapi_logger.handlers = gunicorn_error_logger.handlers
-    root_logger.setLevel(gunicorn_logger.level)
-
-    uvicorn_logger = logging.getLogger("uvicorn.access")
-    uvicorn_logger.handlers = gunicorn_error_logger.handlers
-else:
-    # https://github.com/tiangolo/fastapi/issues/2019
-    LOG_FORMAT2 = (
-        "[%(asctime)s %(process)d:%(threadName)s] %(name)s - %(levelname)s - %(message)s | %(filename)s:%(lineno)d"
-    )
-    logging.basicConfig(level=logging.INFO, format=LOG_FORMAT2)

gpt4all-api/gpt4all_api/app/tests/test_endpoints.py

@@ -1,59 +0,0 @@
-"""
-Use the OpenAI python API to test gpt4all models.
-"""
-from typing import List, get_args
-
-import openai
-
-openai.api_base = "http://localhost:4891/v1"
-
-openai.api_key = "not needed for a local LLM"
-
-
-def test_completion():
-    model = "ggml-mpt-7b-chat.bin"
-    prompt = "Who is Michael Jordan?"
-    response = openai.Completion.create(
-        model=model, prompt=prompt, max_tokens=50, temperature=0.28, top_p=0.95, n=1, echo=True, stream=False
-    )
-    assert len(response['choices'][0]['text']) > len(prompt)
-
-def test_streaming_completion():
-    model = "ggml-mpt-7b-chat.bin"
-    prompt = "Who is Michael Jordan?"
-    tokens = []
-    for resp in openai.Completion.create(
-            model=model,
-            prompt=prompt,
-            max_tokens=50,
-            temperature=0.28,
-            top_p=0.95,
-            n=1,
-            echo=True,
-            stream=True):
-        tokens.append(resp.choices[0].text)
-
-    assert (len(tokens) > 0)
-    assert (len("".join(tokens)) > len(prompt))
-
-
-def test_batched_completion():
-    model = "ggml-mpt-7b-chat.bin"
-    prompt = "Who is Michael Jordan?"
-    response = openai.Completion.create(
-        model=model, prompt=[prompt] * 3, max_tokens=50, temperature=0.28, top_p=0.95, n=1, echo=True, stream=False
-    )
-    assert len(response['choices'][0]['text']) > len(prompt)
-    assert len(response['choices']) == 3
-
-
-def test_embedding():
-    model = "ggml-all-MiniLM-L6-v2-f16.bin"
-    prompt = "Who is Michael Jordan?"
-    response = openai.Embedding.create(model=model, input=prompt)
-    output = response["data"][0]["embedding"]
-    args = get_args(List[float])
-
-    assert response["model"] == model
-    assert isinstance(output, list)
-    assert all(isinstance(x, args) for x in output)

gpt4all-api/gpt4all_api/requirements.txt

@@ -1,12 +0,0 @@
-aiohttp>=3.6.2
-aiofiles
-pydantic>=1.4.0,<2.0.0
-requests>=2.24.0
-ujson>=2.0.2
-fastapi>=0.95.0
-Jinja2>=3.0
-gpt4all>=1.0.0
-pytest
-openai
-black
-isort

gpt4all-api/makefile

@@ -1,46 +0,0 @@
-ROOT_DIR:=$(shell dirname $(realpath $(lastword $(MAKEFILE_LIST))))
-APP_NAME:=gpt4all_api
-PYTHON:=python3.8
-SHELL := /bin/bash
-
-all: dependencies
-
-fresh: clean dependencies
-
-testenv: clean_testenv test_build
-	docker compose -f docker-compose.yaml up --build
-
-testenv_gpu: clean_testenv test_build
-	docker compose -f docker-compose.yaml -f docker-compose.gpu.yaml up --build
-
-testenv_d: clean_testenv test_build
-	docker compose up --build -d
-
-test:
-	docker compose exec $(APP_NAME) pytest -svv --disable-warnings -p no:cacheprovider /app/tests
-
-test_build:
-    DOCKER_BUILDKIT=1 docker build -t $(APP_NAME) --progress plain -f $(APP_NAME)/Dockerfile.buildkit .
-
-clean_testenv:
-	docker compose down -v
-
-fresh_testenv: clean_testenv testenv
-
-venv:
-	if [ ! -d $(ROOT_DIR)/env ]; then $(PYTHON) -m venv $(ROOT_DIR)/env; fi
-
-dependencies: venv
-	source $(ROOT_DIR)/env/bin/activate; $(PYTHON) -m pip install -r $(ROOT_DIR)/$(APP_NAME)/requirements.txt
-
-clean: clean_testenv
-	# Remove existing environment
-	rm -rf $(ROOT_DIR)/env;
-	rm -rf $(ROOT_DIR)/$(APP_NAME)/*.pyc;
-
-
-black:
-	source $(ROOT_DIR)/env/bin/activate; black -l 120 -S --target-version py38 $(APP_NAME)
-
-isort:
-	source $(ROOT_DIR)/env/bin/activate; isort  --ignore-whitespace --atomic -w 120 $(APP_NAME)

gpt4all-backend/CMakeLists.txt

@@ -2,15 +2,23 @@ cmake_minimum_required(VERSION 3.16)
 set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
-if(APPLE)
-  option(BUILD_UNIVERSAL "Build a Universal binary on macOS" ON)
-  if(BUILD_UNIVERSAL)
+if (APPLE)
+    option(BUILD_UNIVERSAL "Build a Universal binary on macOS" ON)
+else()
+    option(LLMODEL_KOMPUTE "llmodel: use Kompute"              ON)
+    option(LLMODEL_VULKAN  "llmodel: use Vulkan"               OFF)
+    option(LLMODEL_CUDA    "llmodel: use CUDA"                 ON)
+    option(LLMODEL_ROCM    "llmodel: use ROCm"                 OFF)
+endif()
+
+if (APPLE)
+  if (BUILD_UNIVERSAL)
     # Build a Universal binary on macOS
     # This requires that the found Qt library is compiled as Universal binaries.
     set(CMAKE_OSX_ARCHITECTURES "arm64;x86_64" CACHE STRING "" FORCE)
   else()
     # Build for the host architecture on macOS
-    if(NOT CMAKE_OSX_ARCHITECTURES)
+    if (NOT CMAKE_OSX_ARCHITECTURES)
       set(CMAKE_OSX_ARCHITECTURES "${CMAKE_HOST_SYSTEM_PROCESSOR}" CACHE STRING "" FORCE)
     endif()
   endif()
@@ -20,7 +28,7 @@ endif()
 include_directories("${CMAKE_CURRENT_BINARY_DIR}")
 
 set(LLMODEL_VERSION_MAJOR 0)
-set(LLMODEL_VERSION_MINOR 4)
+set(LLMODEL_VERSION_MINOR 5)
 set(LLMODEL_VERSION_PATCH 0)
 set(LLMODEL_VERSION "${LLMODEL_VERSION_MAJOR}.${LLMODEL_VERSION_MINOR}.${LLMODEL_VERSION_PATCH}")
 project(llmodel VERSION ${LLMODEL_VERSION} LANGUAGES CXX C)
@@ -39,15 +47,35 @@ else()
     message(STATUS "Interprocedural optimization support detected")
 endif()
 
-if(NOT APPLE)
-  set(LLAMA_KOMPUTE YES)
-endif()
-
+set(DIRECTORY llama.cpp-mainline)
 include(llama.cpp.cmake)
 
-set(BUILD_VARIANTS default avxonly)
-if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
-    set(BUILD_VARIANTS ${BUILD_VARIANTS} metal)
+set(BUILD_VARIANTS)
+set(GPTJ_BUILD_VARIANT cpu)
+if (APPLE)
+    list(APPEND BUILD_VARIANTS metal)
+endif()
+if (LLMODEL_KOMPUTE)
+    list(APPEND BUILD_VARIANTS kompute kompute-avxonly)
+    set(GPTJ_BUILD_VARIANT kompute)
+else()
+    list(PREPEND BUILD_VARIANTS cpu cpu-avxonly)
+endif()
+if (LLMODEL_VULKAN)
+    list(APPEND BUILD_VARIANTS vulkan vulkan-avxonly)
+endif()
+if (LLMODEL_CUDA)
+    include(CheckLanguage)
+    check_language(CUDA)
+    if (NOT CMAKE_CUDA_COMPILER)
+        message(WARNING "CUDA Toolkit not found. To build without CUDA, use -DLLMODEL_CUDA=OFF.")
+    endif()
+    enable_language(CUDA)
+    list(APPEND BUILD_VARIANTS cuda cuda-avxonly)
+endif()
+if (LLMODEL_ROCM)
+    enable_language(HIP)
+    list(APPEND BUILD_VARIANTS rocm rocm-avxonly)
 endif()
 
 set(CMAKE_VERBOSE_MAKEFILE ON)
@@ -55,24 +83,34 @@ set(CMAKE_VERBOSE_MAKEFILE ON)
 # Go through each build variant
 foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
     # Determine flags
-    if (BUILD_VARIANT STREQUAL avxonly)
-        set(GPT4ALL_ALLOW_NON_AVX NO)
+    if (BUILD_VARIANT MATCHES avxonly)
+        set(GPT4ALL_ALLOW_NON_AVX OFF)
     else()
-        set(GPT4ALL_ALLOW_NON_AVX YES)
+        set(GPT4ALL_ALLOW_NON_AVX ON)
     endif()
     set(LLAMA_AVX2 ${GPT4ALL_ALLOW_NON_AVX})
     set(LLAMA_F16C ${GPT4ALL_ALLOW_NON_AVX})
     set(LLAMA_FMA  ${GPT4ALL_ALLOW_NON_AVX})
 
-    if (BUILD_VARIANT STREQUAL metal)
-        set(LLAMA_METAL YES)
-    else()
-        set(LLAMA_METAL NO)
+    set(LLAMA_METAL   OFF)
+    set(LLAMA_KOMPUTE OFF)
+    set(LLAMA_VULKAN  OFF)
+    set(LLAMA_CUDA    OFF)
+    set(LLAMA_ROCM    OFF)
+    if (BUILD_VARIANT MATCHES metal)
+        set(LLAMA_METAL   ON)
+    elseif (BUILD_VARIANT MATCHES kompute)
+        set(LLAMA_KOMPUTE ON)
+    elseif (BUILD_VARIANT MATCHES vulkan)
+        set(LLAMA_VULKAN  ON)
+    elseif (BUILD_VARIANT MATCHES cuda)
+        set(LLAMA_CUDA    ON)
+    elseif (BUILD_VARIANT MATCHES rocm)
+        set(LLAMA_HIPBLAS ON)
     endif()
 
     # Include GGML
-    set(LLAMA_K_QUANTS YES)
-    include_ggml(llama.cpp-mainline -mainline-${BUILD_VARIANT} ON)
+    include_ggml(-mainline-${BUILD_VARIANT})
 
     # Function for preparing individual implementations
     function(prepare_target TARGET_NAME BASE_LIB)
@@ -97,35 +135,14 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
         LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
     prepare_target(llamamodel-mainline llama-mainline)
 
-    add_library(replit-mainline-${BUILD_VARIANT} SHARED
-    replit.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-    target_compile_definitions(replit-mainline-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-    prepare_target(replit-mainline llama-mainline)
+    if (BUILD_VARIANT MATCHES ${GPTJ_BUILD_VARIANT})
+        add_library(gptj-${BUILD_VARIANT} SHARED
+            gptj.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
+        prepare_target(gptj llama-mainline)
+    endif()
 
-    if (NOT LLAMA_METAL)
-# FIXME: These need to be forward ported to latest ggml
-#        add_library(gptj-${BUILD_VARIANT} SHARED
-#            gptj.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-#        prepare_target(gptj ggml-230511)
-
-        add_library(falcon-${BUILD_VARIANT} SHARED
-            falcon.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-        target_compile_definitions(falcon-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-        prepare_target(falcon llama-mainline)
-# FIXME: These need to be forward ported to latest ggml
-#        add_library(mpt-${BUILD_VARIANT} SHARED
-#            mpt.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-#        prepare_target(mpt ggml-230511)
-
-        add_library(bert-${BUILD_VARIANT} SHARED
-            bert.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-        target_compile_definitions(bert-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-        prepare_target(bert llama-mainline)
-
-        add_library(starcoder-${BUILD_VARIANT} SHARED
-            starcoder.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-        target_compile_definitions(starcoder-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-        prepare_target(starcoder llama-mainline)
+    if (BUILD_VARIANT STREQUAL cuda)
+        set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
     endif()
 endforeach()
 
@@ -134,8 +151,6 @@ add_library(llmodel
     llmodel_c.h llmodel_c.cpp
     dlhandle.h
 )
-target_link_libraries(llmodel PRIVATE ggml-mainline-default)
-target_compile_definitions(llmodel PRIVATE GGML_BUILD_VARIANT="default")
 target_compile_definitions(llmodel PRIVATE LIB_FILE_EXT="${CMAKE_SHARED_LIBRARY_SUFFIX}")
 
 set_target_properties(llmodel PROPERTIES

gpt4all-backend/bert_impl.h

@@ -1,44 +0,0 @@
-#ifndef BERT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#error This file is NOT meant to be included outside of bert.cpp. Doing so is DANGEROUS. Be sure to know what you are doing before proceeding to #define BERT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#endif
-#ifndef BERT_H
-#define BERT_H
-
-#include <string>
-#include <functional>
-#include <vector>
-#include <memory>
-#include "llmodel.h"
-
-struct BertPrivate;
-class Bert : public LLModel {
-public:
-    Bert();
-    ~Bert();
-
-    bool supportsEmbedding() const override { return true; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
-    bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
-    size_t stateSize() const override;
-    size_t saveState(uint8_t *dest) const override;
-    size_t restoreState(const uint8_t *src) override;
-    void setThreadCount(int32_t n_threads) override;
-    int32_t threadCount() const override;
-
-    std::vector<float> embedding(const std::string &text) override;
-
-private:
-    std::unique_ptr<BertPrivate> d_ptr;
-
-protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    Token sampleToken(PromptContext &ctx) const override;
-    std::string tokenToString(Token) const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
-};
-
-#endif // BERT_H

gpt4all-backend/dlhandle.h

@@ -53,17 +53,20 @@ public:
     }
 };
 #else
+#include <algorithm>
+#include <filesystem>
 #include <string>
 #include <exception>
 #include <stdexcept>
+
+#define WIN32_LEAN_AND_MEAN
 #ifndef NOMINMAX
-    #define NOMINMAX
+#   define NOMINMAX
 #endif
 #include <windows.h>
 #include <libloaderapi.h>
 
 
-
 class Dlhandle {
     HMODULE chandle;
 
@@ -75,7 +78,9 @@ public:
 
     Dlhandle() : chandle(nullptr) {}
     Dlhandle(const std::string& fpath) {
-        chandle = LoadLibraryExA(fpath.c_str(), NULL, LOAD_LIBRARY_SEARCH_DEFAULT_DIRS | LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR);
+        std::string afpath = std::filesystem::absolute(fpath).string();
+        std::replace(afpath.begin(), afpath.end(), '/', '\\');
+        chandle = LoadLibraryExA(afpath.c_str(), NULL, LOAD_LIBRARY_SEARCH_DEFAULT_DIRS | LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR);
         if (!chandle) {
             throw Exception("dlopen(\""+fpath+"\"): Error");
         }

gpt4all-backend/falcon.cpp

@@ -1,985 +0,0 @@
-#include "ggml.h"
-#define FALCON_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#include "falcon_impl.h"
-#include "llama.h"
-#include "llama-util.h"
-#include "utils.h"
-#include "llmodel_shared.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <iostream>
-#include <sstream>
-
-namespace {
-const char *modelType_ = "Falcon";
-}
-
-// commented out 40B support as it presently would require forking ggml/llama.cpp
-// can re-add once mainline ggml supports it
-
-#define FALCON_MAGIC 0x67676a74
-
-// default hparams (Falcon 7B)
-struct falcon_hparams {
-    int32_t n_vocab = 65024;
-    int32_t n_embd  = 4544;
-    int32_t n_head  = 71;
-    int32_t n_head_kv = 1;
-    int32_t n_layer = 32;
-    int32_t falcon_version = 7; // 7 for Falcon-7B, 40 for Falcon-40B
-    int32_t ftype   = 1;
-    int32_t n_ctx   = 2048;
-};
-
-struct falcon_layer {
-    // normalization
-    struct ggml_tensor* input_layernorm;
-    struct ggml_tensor* input_layernorm_b;
-    //struct ggml_tensor* attention_norm;    // Falcon-40B only
-    //struct ggml_tensor* attention_norm_b;  // Falcon-40B only
-
-    // attention
-    struct ggml_tensor* query_key_value;
-    struct ggml_tensor* wo;
-
-    // ff
-    struct ggml_tensor* ffn_up;
-    struct ggml_tensor* ffn_down;
-};
-
-struct falcon_model {
-    falcon_hparams hparams;
-
-    struct ggml_tensor* tok_embeddings;
-    struct ggml_tensor* output_norm;
-    struct ggml_tensor* output_norm_b;
-    struct ggml_tensor* lm_head;
-
-    std::vector<falcon_layer> layers;
-
-    // key + value memory
-    llm_kv_cache kv_self;
-
-    struct ggml_context* ctx;
-    std::map<std::string, struct ggml_tensor*> tensors;
-
-    llm_buffer eval_buf;
-    llm_buffer work_buf;
-    llm_buffer scr0_buf;
-    llm_buffer scr1_buf;
-};
-
-static bool kv_cache_init(
-        const struct falcon_hparams & hparams,
-              struct llm_kv_cache & cache,
-                         ggml_type   wtype,
-                               int   n_ctx) {
-    const int n_embd  = hparams.n_embd;
-    const int dim_head  = n_embd / hparams.n_head;
-    const int dim_kv = dim_head * hparams.n_head_kv;
-    const int n_layer = hparams.n_layer;
-
-    const int64_t n_mem      = (int64_t)n_layer*n_ctx;
-    const int64_t n_elements = dim_kv * n_mem;
-    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2_MiB);
-    struct ggml_init_params params;
-    params.mem_size   = cache.buf.size;
-    params.mem_buffer = cache.buf.addr;
-    params.no_alloc   = false;
-
-    cache.ctx = ggml_init(params);
-    if (!cache.ctx) {
-        fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
-        return false;
-    }
-
-    cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
-    cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
-    return true;
-}
-
-// load the model's weights from a file
-bool falcon_model_load(const std::string & fname, falcon_model & model, gpt_vocab & vocab, size_t *mem_req) {
-    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
-    if (mem_req) {
-        *mem_req = 0;
-    }
-
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
-    // verify magic
-    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        if (magic != FALCON_MAGIC) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
-        }
-    }
-
-    uint32_t format_version;
-    fin.read((char *) &format_version, sizeof(format_version));
-
-    // load hparams
-    {
-        auto & hparams = model.hparams;
-
-        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
-        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
-        fin.read((char *) &hparams.n_head_kv, sizeof(hparams.n_head_kv));
-        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *) &hparams.falcon_version, sizeof(hparams.falcon_version));
-        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
-
-        if (hparams.falcon_version != 7) { // && hparams.falcon_version != 40) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad Falcon version: %d)\n", __func__, fname.c_str(), hparams.falcon_version);
-            return false;
-        }
-
-        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
-
-        printf("%s: n_vocab   = %d\n", __func__, hparams.n_vocab);
-        printf("%s: n_embd    = %d\n", __func__, hparams.n_embd);
-        printf("%s: n_head    = %d\n", __func__, hparams.n_head);
-        printf("%s: n_head_kv = %d\n", __func__, hparams.n_head_kv);
-        printf("%s: n_layer   = %d\n", __func__, hparams.n_layer);
-        printf("%s: ftype     = %d\n", __func__, hparams.ftype);
-        printf("%s: qntvr     = %d\n", __func__, qntvr);
-
-        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
-    }
-
-    // load vocab
-    {
-        const int32_t n_vocab = model.hparams.n_vocab;
-
-        std::string word;
-        std::vector<char> buf(128);
-
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
-
-            buf.resize(len);
-            fin.read((char *) buf.data(), len);
-            word.assign(buf.data(), len);
-
-            uint32_t dummy;
-            fin.read((char *) &dummy, sizeof(dummy));
-
-            vocab.token_to_id[word] = i;
-            vocab.id_to_token[i] = word;
-        }
-    }
-
-    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
-    // in order to save memory and also to speed up the computation
-    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
-    if (wtype == GGML_TYPE_COUNT) {
-        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
-                __func__, fname.c_str(), model.hparams.ftype);
-        return false;
-    }
-
-    auto & ctx = model.ctx;
-
-    size_t ctx_size = 0;
-
-    {
-        const auto& hparams = model.hparams;
-
-        const int n_embd = hparams.n_embd;
-        const int n_head = hparams.n_head;
-        const int n_head_kv = hparams.n_head_kv;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx = hparams.n_ctx;
-        const int n_ff = 4 * model.hparams.n_embd;
-        const int n_vocab = hparams.n_vocab;
-        const int head_dim = hparams.n_embd / hparams.n_head;
-
-        ctx_size += ggml_tensor_overhead() + ggml_type_sizef(wtype) * n_embd * n_vocab;  // tok_embeddings
-        ctx_size += ggml_tensor_overhead() + ggml_type_sizef(GGML_TYPE_F32) * n_embd;  // output_norm
-        ctx_size += ggml_tensor_overhead() + ggml_type_sizef(GGML_TYPE_F32) * n_embd;  // output_norm_b
-        ctx_size += ggml_tensor_overhead() + ggml_type_sizef(wtype) * n_embd * n_vocab;  // lm_head
-
-        // if (hparams.version == 40) { // Falcon-40B
-        //     ctx_size += n_layer * ggml_sizeof_tensor_1d(GGML_TYPE_F32, n_embd);  // attention_norm
-        //     ctx_size += n_layer * ggml_sizeof_tensor_1d(GGML_TYPE_F32, n_embd);  // attention_norm_b
-        // }
-        ctx_size += n_layer * (ggml_tensor_overhead() + ggml_type_sizef(GGML_TYPE_F32) * n_embd);  // input_layernorm
-        ctx_size += n_layer * (ggml_tensor_overhead() + ggml_type_sizef(GGML_TYPE_F32) * n_embd);  // input_layernorm_b
-        ctx_size += n_layer * (ggml_tensor_overhead() + ggml_type_sizef(wtype) * n_embd * (n_head_kv * 2 + n_head) * head_dim);  // query_key_value
-        ctx_size += n_layer * (ggml_tensor_overhead() + ggml_type_sizef(wtype) * n_embd * n_embd);  // wo
-        ctx_size += n_layer * (ggml_tensor_overhead() + ggml_type_sizef(wtype) * n_embd * n_ff);  // ffn_up
-        ctx_size += n_layer * (ggml_tensor_overhead() + ggml_type_sizef(wtype) * n_ff * n_embd);  // ffn_down
- 
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
-
-    if (mem_req) {
-        const int n_embd  = model.hparams.n_embd;
-        const int dim_head  = n_embd / model.hparams.n_head;
-        const int dim_kv = dim_head * model.hparams.n_head_kv;
-        const int n_layer = model.hparams.n_layer;
-
-        const int64_t n_mem      = (int64_t)n_layer*model.hparams.n_ctx;
-        const int64_t n_elements = dim_kv * n_mem;
-        size_t kv_cache_size = 2u*n_elements*ggml_type_size(wtype) + 2_MiB;
-        *mem_req = ctx_size + kv_cache_size;
-        return false;
-    }
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-            .no_alloc   = false,
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
-    // prepare memory for the weights
-    {
-        const auto& hparams = model.hparams;
-
-        const int n_embd = hparams.n_embd;
-        const int n_head = hparams.n_head;
-        const int n_head_kv = hparams.n_head_kv;
-        const int n_layer = hparams.n_layer;
-        const int n_ff = 4 * model.hparams.n_embd;
-        const int n_vocab = hparams.n_vocab;
-        const int head_dim = hparams.n_embd / hparams.n_head;
-
-        model.layers.resize(n_layer);
-
-        model.tok_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
-
-        model.output_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-        model.output_norm_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-        model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
-
-        // map by name
-        model.tensors["transformer.word_embeddings.weight"] =
-            model.tok_embeddings;
-
-        model.tensors["transformer.ln_f.weight"] = model.output_norm;
-        model.tensors["transformer.ln_f.bias"] = model.output_norm_b;
-        model.tensors["lm_head.weight"] = model.lm_head;
-
-        for (int i = 0; i < n_layer; ++i) {
-            auto& layer = model.layers[i];
-
-            layer.input_layernorm =
-                ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.input_layernorm_b =
-                ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-            // if (hparams.version == 40) { // for Falcon-40B only
-            //     layer.attention_norm =
-            //         ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            //     layer.attention_norm_b =
-            //         ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            // }
-
-            // query_key_value shape for config.multi_query == True:
-            layer.query_key_value = ggml_new_tensor_2d(
-                ctx, wtype, n_embd, (n_head_kv * 2 + n_head) * head_dim);
-            layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
-
-            layer.ffn_up = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff);
-            layer.ffn_down = ggml_new_tensor_2d(ctx, wtype, n_ff, n_embd);
-
-            // map by name
-            // if (hparams.version == 40) {
-            //     // Falcon-40B:
-            //     model.tensors["transformer.h." + std::to_string(i) +
-            //                   ".ln_mlp.weight"] = layer.input_layernorm;
-            //     model.tensors["transformer.h." + std::to_string(i) +
-            //                   ".ln_mlp.bias"] = layer.input_layernorm_b;
-            //     model.tensors["transformer.h." + std::to_string(i) +
-            //                   ".ln_attn.weight"] = layer.attention_norm;
-            //     model.tensors["transformer.h." + std::to_string(i) +
-            //                   ".ln_attn.bias"] = layer.attention_norm_b;
-            // } else {
-            // Falcon-7B:
-            model.tensors["transformer.h." + std::to_string(i) +
-                          ".input_layernorm.weight"] = layer.input_layernorm;
-            model.tensors["transformer.h." + std::to_string(i) +
-                          ".input_layernorm.bias"] = layer.input_layernorm_b;
-            //}
-
-            model.tensors["transformer.h." + std::to_string(i) +
-                          ".self_attention.query_key_value.weight"] =
-                layer.query_key_value;
-            model.tensors["transformer.h." + std::to_string(i) +
-                          ".self_attention.dense.weight"] = layer.wo;
-
-            model.tensors["transformer.h." + std::to_string(i) +
-                          ".mlp.dense_h_to_4h.weight"] = layer.ffn_up;
-            model.tensors["transformer.h." + std::to_string(i) +
-                          ".mlp.dense_4h_to_h.weight"] = layer.ffn_down;
-        }
-    }
-
-    // key + value memory
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_head_kv = hparams.n_head_kv;
-        const int head_dim = hparams.n_embd / hparams.n_head;
-
-        const int64_t n_mem      = n_layer*n_ctx;
-        const int64_t n_elements = head_dim*n_mem;
-
-        if (!kv_cache_init(hparams, model.kv_self, GGML_TYPE_F32, model.hparams.n_ctx)) {
-            fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
-            ggml_free(ctx);
-            return false;
-        }
-        const size_t memory_size = ggml_nbytes(model.kv_self.k) + ggml_nbytes(model.kv_self.v);
-
-        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
-    }
-
-    // load weights
-    {
-        int n_tensors = 0;
-        size_t total_size = 0;
-
-        printf("%s: ", __func__);
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            int32_t ttype;
-
-            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-            fin.seekg(-static_cast<ptrdiff_t>(fin.tellg()) & 31, std::ios_base::cur);
-
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                return false;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                return false;
-            }
-
-            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%5d, %5d], expected [%5d, %5d]\n",
-                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            // for debugging
-            if (0) {
-                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
-            }
-
-            const size_t bpe = ggml_type_size(ggml_type(ttype));
-
-            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
-                return false;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-            total_size += ggml_nbytes(tensor);
-            if (++n_tensors % 8 == 0) {
-                printf(".");
-                fflush(stdout);
-            }
-        }
-
-        printf(" done\n");
-
-        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
-    }
-
-    fin.close();
-
-    model.eval_buf.resize(1280u * 1024 * 1024);
-    model.scr0_buf.resize(256u * 1024 * 1024);
-    model.scr1_buf.resize(256u * 1024 * 1024);
-    return true;
-}
-
-
-// evaluate the transformer
-//
-//   - model:     the model
-//   - n_threads: number of threads to use
-//   - n_past:    the context size so far
-//   - embd_inp:  the embeddings of the tokens in the context
-//   - embd_w:    the predicted logits for the next token
-//
-bool falcon_eval(
-        falcon_model & model,
-        const int n_threads,
-        const int n_past,
-        const std::vector<gpt_vocab::id> & embd_inp,
-              std::vector<float>         & embd_w,
-              size_t                     & mem_per_token) {
-    const int N = embd_inp.size();
-
-    const auto & hparams = model.hparams;
-
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_ctx   = hparams.n_ctx;
-    const int n_head  = hparams.n_head;
-    const int n_head_kv = hparams.n_head_kv;
-    const int n_vocab = hparams.n_vocab;
-    const int version = hparams.falcon_version;
-    const size_t head_dim = n_embd / n_head;
-
-   struct ggml_init_params eval_ctx_params = {
-        .mem_size = model.eval_buf.size,
-        .mem_buffer = model.eval_buf.addr,
-        .no_alloc = false,
-    };
-
-    struct ggml_context * ctx0 = ggml_init(eval_ctx_params);
-    struct ggml_cgraph gf = {};
-
-    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
-
-    // wte
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
-    struct ggml_tensor* repeat_dummy = ggml_new_tensor_3d(ctx0, inpL->type, head_dim, N + n_past, n_head);
-
-    ggml_type wtype = GGML_TYPE_F32;
-    const int sizeof_wtype = ggml_type_sizef(wtype);
-
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * cur;
-        struct ggml_tensor * layernorm_output;
-
-        ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
-
-        // self-attention
-        {
-            layernorm_output = ggml_norm(ctx0, inpL);
-
-            layernorm_output = ggml_add(ctx0,
-                    ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model.layers[il].input_layernorm, layernorm_output),
-                        layernorm_output),
-                    ggml_repeat(ctx0, model.layers[il].input_layernorm_b, layernorm_output));
-
-            // if (version == 40) { // Falcon-40B only
-            //     cur = ggml_norm(ctx0, inpL);
-
-            //     cur = ggml_add(ctx0,
-            //             ggml_mul(ctx0,
-            //                 ggml_repeat(ctx0, model.layers[il].attention_norm, cur),
-            //                 cur),
-            //             ggml_repeat(ctx0, model.layers[il].attention_norm_b, cur));
-            // }
-            // else {
-            cur = layernorm_output;
-            // }
-
-            // compute QKV
-
-            cur = ggml_mul_mat(ctx0, model.layers[il].query_key_value, cur);
-
-            // Note that the strides for Kcur, Vcur are set up so that the
-            // resulting views are misaligned with the tensor's storage
-            // (by applying the K/V offset we shift the tensor's original
-            // view to stick out behind the viewed QKV tensor's allocated
-            // memory, so to say). This is ok because no actual accesses
-            // happen to that out-of-range memory, but it can require some
-            // trickery when trying to accurately dump these views for
-            // debugging.
-
-            struct ggml_tensor * Qcur = ggml_view_3d(
-                ctx0, cur, head_dim, n_head, N,
-                head_dim * sizeof_wtype,
-                head_dim * (n_head + 2 * n_head_kv) * sizeof_wtype,
-                0);
-
-            struct ggml_tensor * Kcur = ggml_view_3d(
-                ctx0, cur, head_dim, n_head_kv, N,
-                head_dim * sizeof_wtype,
-                head_dim * (n_head + 2 * n_head_kv) * sizeof_wtype,
-                head_dim * n_head * sizeof_wtype);
-
-            struct ggml_tensor * Vcur = ggml_view_3d(
-                ctx0, cur, head_dim, n_head_kv, N,
-                head_dim * sizeof_wtype,
-                head_dim * (n_head + 2 * n_head_kv) * sizeof_wtype,
-                head_dim * (n_head + n_head_kv) * sizeof_wtype);
-
-            // using mode = 2 for neox mode
-            Qcur = ggml_rope_inplace(ctx0, Qcur, n_past, head_dim, 2, n_ctx);
-            Kcur = ggml_rope_inplace(ctx0, Kcur, n_past, head_dim, 2, n_ctx);
-
-            // store key and value to memory
-            {
-                struct ggml_tensor* k = ggml_view_1d(
-                    ctx0, model.kv_self.k, N * n_head_kv * head_dim,
-                    (ggml_element_size(model.kv_self.k) * n_head_kv * head_dim) *
-                        (il * n_ctx + n_past));
-                struct ggml_tensor* v = ggml_view_1d(
-                    ctx0, model.kv_self.v, N * n_head_kv * head_dim,
-                    (ggml_element_size(model.kv_self.v) * n_head_kv * head_dim) *
-                        (il * n_ctx + n_past));
-
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
-            }
-
-            struct ggml_tensor * K = ggml_permute(
-                ctx0,
-                ggml_view_3d(
-                    ctx0,
-                    model.kv_self.k,
-                    head_dim, n_head_kv, n_past + N,
-                    head_dim * sizeof_wtype,
-                    head_dim * n_head_kv * sizeof_wtype,
-                    il * n_ctx * ggml_element_size(model.kv_self.k) * n_head_kv * head_dim),
-                0, 2, 1, 3);
-
-            // K * Q
-
-            // changed from repeat2 back to repeat, will not support 40B!
-            K = ggml_cont(ctx0, ggml_repeat(ctx0, K, repeat_dummy));
-
-            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale_inplace(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(head_dim)))
-                        );
-
-            // KQ_masked = mask_past(KQ_scaled)
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
-
-            // KQ = soft_max(KQ_masked)
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
-
-            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
-            struct ggml_tensor* V = ggml_permute(
-                ctx0,
-                ggml_view_3d(
-                    ctx0,
-                    model.kv_self.v,
-                    head_dim, n_head_kv, n_past + N,
-                    head_dim * sizeof_wtype,
-                    head_dim * n_head_kv * sizeof_wtype,
-                    il * n_ctx * ggml_element_size(model.kv_self.v) * n_head_kv * head_dim),
-                0, 2, 1, 3);
-
-            // changed from repeat2 back to repeat, will not support 40B!
-            V = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_repeat(ctx0, V, repeat_dummy)));
-
-            // KQV = transpose(V) * KQ_soft_max
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-
-            // projection
-            {
-                cur = ggml_mul_mat(ctx0,
-                        model.layers[il].wo,
-                        cur);
-            }
-        }
-
-        ggml_set_scratch(ctx0, {0, model.scr1_buf.size, model.scr1_buf.addr, });
-
-        struct ggml_tensor* inpFF = layernorm_output;
-        struct ggml_tensor* attn_out = ggml_cpy(
-            ctx0, cur, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-
-        {
-            cur = ggml_mul_mat(ctx0, model.layers[il].ffn_up, inpFF);
-            cur = ggml_gelu(ctx0, cur);
-            cur = ggml_mul_mat(ctx0, model.layers[il].ffn_down, cur);
-        }
-
-        cur = ggml_add(ctx0, cur, attn_out);
-        cur = ggml_add(ctx0, cur, inpL);
-        // input for next layer
-        inpL = cur;
-    }
-
-    ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
-
-    // norm
-    {
-        inpL = ggml_norm(ctx0, inpL);
-
-        // inpL = ln_f_g*inpL + ln_f_b
-        inpL = ggml_add(ctx0,
-                ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.output_norm, inpL),
-                    inpL),
-                ggml_repeat(ctx0, model.output_norm_b, inpL));
-    }
-
-    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
-
-    // lm_head
-    {
-        inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
-
-        //inpL = ggml_add(ctx0,
-        //        ggml_repeat(ctx0, model.lmh_b, inpL),
-        //        inpL);
-    }
-
-    // logits -> probs
-    //inpL = ggml_soft_max_inplace(ctx0, inpL);
-
-    // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute_g4a(model.work_buf, &gf, n_threads);
-  
-
-    //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
-    //}
-
-    //embd_w.resize(n_vocab*N);
-    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
-
-    // return result for just the last token
-    embd_w.resize(n_vocab);
-    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
-
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
-
-    ggml_free(ctx0);
-
-    return true;
-}
-
-
-#define MAX_RNG_STATE 64*1024
-size_t falcon_get_state_size(const falcon_model &model) {
-    const size_t s_rng_size        = sizeof(size_t);
-    const size_t s_rng             = MAX_RNG_STATE;
-    const size_t s_kv_size         = sizeof(size_t);
-    const size_t s_kv_ntok         = sizeof(int);
-    const size_t s_kv              = model.kv_self.buf.size;
-    const size_t s_total = (
-        + s_rng_size
-        + s_rng
-        + s_kv_size
-        + s_kv_ntok
-        + s_kv
-    );
-    return s_total;
-}
-
-size_t falcon_copy_state_data(const falcon_model &model, const std::mt19937 &rng, uint8_t *dest)
-{
-    uint8_t * out = dest;
-    // copy rng
-    {
-        std::stringstream rng_ss;
-        rng_ss << rng;
-
-        const size_t rng_size = rng_ss.str().size();
-        char rng_buf[MAX_RNG_STATE];
-
-        memset(&rng_buf[0], 0, MAX_RNG_STATE);
-        memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
-
-        memcpy(out, &rng_size,   sizeof(rng_size));   out += sizeof(rng_size);
-        memcpy(out, &rng_buf[0], MAX_RNG_STATE); out += MAX_RNG_STATE;
-    }
-
-    // copy kv cache
-    {
-        const size_t kv_size = model.kv_self.buf.size;
-        const int    kv_ntok = model.kv_self.n;
-
-        memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size);
-        memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok);
-
-        if (kv_size) {
-            memcpy(out, model.kv_self.buf.addr, kv_size); out += kv_size;
-        }
-    }
-
-    const size_t written  = out - dest;
-    assert(written == falcon_get_state_size(model));
-    fflush(stdout);
-    return written;
-}
-
-size_t falcon_set_state_data(falcon_model *model, std::mt19937 *rng, const uint8_t *src)
-{
-    const uint8_t * in = src;
-
-    // set rng
-    {
-        size_t rng_size;
-        char   rng_buf[MAX_RNG_STATE];
-
-        memcpy(&rng_size,   in, sizeof(rng_size));    in += sizeof(rng_size);
-        memcpy(&rng_buf[0], in, MAX_RNG_STATE); in += MAX_RNG_STATE;
-
-        std::stringstream rng_ss;
-        rng_ss.str(std::string(&rng_buf[0], rng_size));
-        rng_ss >> *rng;
-
-        assert(rng_ss.fail() == false);
-    }
-
-    // set kv cache
-    {
-        size_t kv_size;
-        int kv_ntok;
-
-        memcpy(&kv_size, in, sizeof(kv_size)); in += sizeof(kv_size);
-        memcpy(&kv_ntok, in, sizeof(kv_ntok)); in += sizeof(kv_ntok);
-
-        if (kv_size) {
-            assert(model->kv_self.buf.size == kv_size);
-
-            void * k_data = model->kv_self.k->data; // remember data pointers
-            void * v_data = model->kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
-
-            memcpy(model->kv_self.buf.addr, in, kv_size); in += kv_size;
-
-            model->kv_self.k->data = k_data; // restore correct data pointers
-            model->kv_self.v->data = v_data;
-
-        }
-
-        model->kv_self.n = kv_ntok;
-    }
-
-    const size_t nread    = in - src;
-    assert(nread == falcon_get_state_size(*model));
-    fflush(stdout);
-    return nread;
-}
-
-struct FalconPrivate {
-    const std::string modelPath;
-    bool modelLoaded;
-    gpt_vocab vocab;
-    falcon_model *model = nullptr;
-    int64_t n_threads = 0;
-    size_t mem_per_token = 0;
-    std::mt19937 rng;
-};
-
-Falcon::Falcon() : d_ptr(new FalconPrivate) {
-    d_ptr->model = new falcon_model;
-    d_ptr->model->ctx = nullptr;
-    d_ptr->modelLoaded = false;
-}
-
-Falcon::~Falcon() {
-    if(d_ptr->model->ctx) {
-        ggml_free(d_ptr->model->ctx);
-        d_ptr->model->ctx = nullptr;
-    }
-    delete d_ptr->model;
-}
-
-bool Falcon::loadModel(const std::string &modelPath)
-{
-    std::mt19937 rng(time(NULL));
-    d_ptr->rng = rng;
-
-    // load the model
-    if (!falcon_model_load(modelPath, *d_ptr->model, d_ptr->vocab, nullptr)) {
-        std::cerr << "FALCON ERROR: failed to load model from " <<  modelPath;
-        return false;
-    }
-
-    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    d_ptr->modelLoaded = true;
-    fflush(stdout);
-    return true;
-}
-
-bool Falcon::isModelLoaded() const
-{
-    return d_ptr -> modelLoaded;
-}
-
-size_t Falcon::requiredMem(const std::string &modelPath)
-{
-    falcon_model dummy_model;
-    gpt_vocab dummy_vocab;
-    size_t mem_req;
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-    falcon_model_load(modelPath, dummy_model, dummy_vocab, &mem_req);
-    return mem_req;
-}
-
-size_t Falcon::stateSize() const
-{
-    return falcon_get_state_size(*d_ptr->model);
-}
-
-size_t Falcon::saveState(uint8_t *dest) const
-{
-    return falcon_copy_state_data(*d_ptr->model, d_ptr->rng, dest);
-}
-
-size_t Falcon::restoreState(const uint8_t *src)
-{
-    return falcon_set_state_data(d_ptr->model, &d_ptr->rng, src);
-}
-
-void Falcon::setThreadCount(int32_t n_threads)
-{
-    d_ptr->n_threads = n_threads;
-}
-
-int32_t Falcon::threadCount() const
-{
-    return d_ptr->n_threads;
-}
-
-std::vector<LLModel::Token> Falcon::tokenize(PromptContext &, const std::string &str) const
-{
-    return ::gpt_tokenize(d_ptr->vocab, str);
-}
-
-LLModel::Token Falcon::sampleToken(PromptContext &promptCtx) const
-{
-    const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
-    return gpt_sample_top_k_top_p(d_ptr->model->hparams.n_vocab,
-        promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
-        n_prev_toks,
-        promptCtx.logits,
-        promptCtx.top_k, promptCtx.top_p, promptCtx.temp,
-        promptCtx.repeat_penalty,
-        d_ptr->rng);
-}
-
-std::string Falcon::tokenToString(Token id) const
-{
-    return d_ptr->vocab.id_to_token[id];
-}
-
-bool Falcon::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
-{
-    // determine the required inference memory per token:
-    static bool initialized = false;
-    if (!initialized) {
-        falcon_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, ctx.logits,
-            d_ptr->mem_per_token);
-        initialized = true;
-    }
-
-    return falcon_eval(*d_ptr->model, d_ptr->n_threads, ctx.n_past, tokens, ctx.logits, d_ptr->mem_per_token);
-}
-
-int32_t Falcon::contextLength() const
-{
-    return d_ptr->model->hparams.n_ctx;
-}
-
-const std::vector<LLModel::Token> &Falcon::endTokens() const
-{
-    static const std::vector<LLModel::Token> out = { 11 };
-    return out;
-}
-
-#if defined(_WIN32)
-#define DLL_EXPORT __declspec(dllexport)
-#else
-#define DLL_EXPORT __attribute__ ((visibility ("default")))
-#endif
-
-extern "C" {
-DLL_EXPORT bool is_g4a_backend_model_implementation() {
-    return true;
-}
-
-DLL_EXPORT const char *get_model_type() {
-    return modelType_;
-}
-
-DLL_EXPORT const char *get_build_variant() {
-    return GGML_BUILD_VARIANT;
-}
-
-DLL_EXPORT bool magic_match(std::istream& f) {
-    uint32_t magic = 0;
-    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
-    uint32_t version = 0;
-    f.read(reinterpret_cast<char*>(&version), sizeof(version));
-    if (magic != FALCON_MAGIC) {
-         return false; 
-    }
-    falcon_hparams hparams;
-    f.read(reinterpret_cast<char*>(&hparams), sizeof(hparams));
-    // we're matching the file format of existing pre-converted models
-    // compatible with ctransformers llama.cpp based format, which also
-    // unfortunately shares its magic number what llama uses, so we now
-    // differentiate by n_vocab
-    // give some wiggle room over the max to allow for finetunes that expand the
-    // vocabulary
-    if (!(hparams.n_vocab >= 65024 && hparams.n_vocab <= 65100)) {
-        return false;
-    }
-    if (hparams.falcon_version != 7) {
-        return false;
-    }
-    return true;
-}
-
-DLL_EXPORT LLModel *construct() {
-    return new Falcon;
-}
-}

gpt4all-backend/falcon_impl.h

@@ -1,42 +0,0 @@
-#ifndef FALCON_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#error This file is NOT meant to be included outside of falcon.cpp. Doing so is DANGEROUS. Be sure to know what you are doing before proceeding to #define FALCON_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#endif
-#ifndef FALCON_H
-#define FALCON_H
-
-#include <string>
-#include <functional>
-#include <vector>
-#include <memory>
-#include "llmodel.h"
-
-struct FalconPrivate;
-class Falcon : public LLModel {
-public:
-    Falcon();
-    ~Falcon();
-
-    bool supportsEmbedding() const override { return false; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
-    bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
-    size_t stateSize() const override;
-    size_t saveState(uint8_t *dest) const override;
-    size_t restoreState(const uint8_t *src) override;
-    void setThreadCount(int32_t n_threads) override;
-    int32_t threadCount() const override;
-
-private:
-    std::unique_ptr<FalconPrivate> d_ptr;
-
-protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    Token sampleToken(PromptContext &ctx) const override;
-    std::string tokenToString(Token) const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
-};
-
-#endif // Falcon_H

gpt4all-backend/gptj.cpp

@@ -9,7 +9,6 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
-#include <fstream>
 #include <map>
 #include <string>
 #include <vector>
@@ -42,7 +41,7 @@ struct gptj_hparams {
     int32_t n_head  = 16;
     int32_t n_layer = 28;
     int32_t n_rot   = 64;
-    int32_t f16     = 1;
+    float norm_eps  = 1e-5;
 };
 
 struct gptj_layer {
@@ -128,216 +127,149 @@ static bool kv_cache_init(
     return true;
 }
 
-// load the model's weights from a stream
-bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & model, gpt_vocab & vocab, size_t * mem_req = nullptr) {
+// load the model's weights from a file path
+bool gptj_model_load(const std::string &fname, gptj_model & model, gpt_vocab & vocab, size_t * mem_req = nullptr) {
     printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
     if(mem_req != nullptr) {
         *mem_req = 0;
     }
 
-    // verify magic
-    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        if (magic != 0x67676d6c) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
-        }
+    // create the ggml context
+    struct gguf_init_params params = {
+        /*.no_alloc = */ false,
+        /*.ctx      = */ &model.ctx,
+    };
+
+    gguf_context *ggufctx = gguf_init_from_file(fname.c_str(), params);
+    if (!ggufctx) {
+        fprintf(stderr, "%s: gguf_init_from_file() failed\n", __func__);
+        return false;
     }
 
     // load hparams
     {
         auto & hparams = model.hparams;
 
-        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
-        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
-        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
-        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
-        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+        bool ok = false;
+        int keyidx;
+
+        do {
+            keyidx = gguf_find_key(ggufctx, "gptj.context_length");
+            if (keyidx == -1) { break; }
+            hparams.n_ctx = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.embedding_length");
+            if (keyidx == -1) { break; }
+            hparams.n_embd = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.attention.head_count");
+            if (keyidx == -1) { break; }
+            hparams.n_head = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.block_count");
+            if (keyidx == -1) { break; }
+            hparams.n_layer = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.rope.dimension_count");
+            if (keyidx == -1) { break; }
+            hparams.n_rot = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.attention.layer_norm_epsilon");
+            if (keyidx == -1) { break; }
+            hparams.norm_eps = gguf_get_val_f32(ggufctx, keyidx);
+
+            ok = true;
+        } while (false);
+
+        if (!ok) {
+            fprintf(stderr, "%s: required hparam missing!\n", __func__);
+            return false;
+        }
 
-        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
         printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
         printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
         printf("%s: n_head  = %d\n", __func__, hparams.n_head);
         printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
         printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
-        printf("%s: f16     = %d\n", __func__, hparams.f16);
     }
 
     // load vocab
     {
-        int32_t n_vocab = 0;
-        fin.read((char *) &n_vocab, sizeof(n_vocab));
+        auto & hparams = model.hparams;
 
-        if (n_vocab != model.hparams.n_vocab) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
-                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+        int keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.model");
+        if (keyidx == -1) {
+            fprintf(stderr, "%s: tokenizer model not found!\n", __func__);
+            return false;
+        }
+        if (strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
+            fprintf(stderr, "%s: tokenizer model not supported!\n", __func__);
             return false;
         }
 
-        std::string word;
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
+        int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
+        if (tokens_keyidx == -1) {
+            fprintf(stderr, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            return false;
+        }
 
-            word.resize(len);
-            fin.read((char *) word.data(), len);
+        hparams.n_vocab = gguf_get_arr_n(ggufctx, tokens_keyidx);
+        printf("%s: gpt2 tokenizer vocab = %d\n", __func__, int(hparams.n_vocab));
 
+        for (int i = 0; i < hparams.n_vocab; i++) {
+            std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
             vocab.token_to_id[word] = i;
             vocab.id_to_token[i] = word;
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
-    // in order to save memory and also to speed up the computation
-    ggml_type wtype = GGML_TYPE_COUNT;
-    switch (model.hparams.f16) {
-        case 0: wtype = GGML_TYPE_F32;  break;
-        case 1: wtype = GGML_TYPE_F16;  break;
-        case 2: wtype = GGML_TYPE_Q4_0; break;
-        case 3: wtype = GGML_TYPE_Q4_1; break;
-        case 5: wtype = GGML_TYPE_Q4_2; break;
-        default:
-                {
-                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
-                            __func__, fname.c_str(), model.hparams.f16);
-                    return false;
-                }
-    }
-
     auto & ctx = model.ctx;
 
-    size_t ctx_size = 0;
-
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);         // lmh_g
-        ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
-
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
-
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
-
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
-
-        ctx_size += (5 + 10*n_layer)*256; // object overhead
-
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
+    size_t ctx_size = ggml_get_mem_size(ctx);
+    printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size / (1024.0 * 1024.0));
 
     if (mem_req != nullptr) {
-        *mem_req += ctx_size;
-        const int n_embd  = model.hparams.n_embd;
-        const int n_layer = model.hparams.n_layer;
-
-        const int64_t n_mem      = (int64_t)n_layer*model.hparams.n_ctx;
-        const int64_t n_elements = n_embd*n_mem;
-
-        *mem_req += (2u*n_elements*ggml_type_size(wtype) + 2_MiB);
+        *mem_req = ctx_size;
+        gguf_free(ggufctx);
         return false;
     }
 
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-            .no_alloc = false
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
     // prepare memory for the weights
     {
         const auto & hparams = model.hparams;
+        model.layers.resize(hparams.n_layer);
 
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_vocab = hparams.n_vocab;
+        model.wte    = ggml_get_tensor(ctx, "token_embd.weight");
 
-        model.layers.resize(n_layer);
+        model.ln_f_g = ggml_get_tensor(ctx, "output_norm.weight");
+        model.ln_f_b = ggml_get_tensor(ctx, "output_norm.bias");
 
-        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.lmh_g  = ggml_get_tensor(ctx, "output.weight");
+        model.lmh_b  = ggml_get_tensor(ctx, "output.bias");
 
-        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        auto name = [](int i, std::string n) {
+            static std::string key;
+            key = "blk." + std::to_string(i) + "." + n;
+            return key.c_str();
+        };
 
-        model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
-        model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
-
-        // map by name
-        model.tensors["transformer.wte.weight"] = model.wte;
-
-        model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
-        model.tensors["transformer.ln_f.bias"]   = model.ln_f_b;
-
-        model.tensors["lm_head.weight"] = model.lmh_g;
-        model.tensors["lm_head.bias"]   = model.lmh_b;
-
-        for (int i = 0; i < n_layer; ++i) {
+        for (int i = 0; i < hparams.n_layer; ++i) {
             auto & layer = model.layers[i];
 
-            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_g          = ggml_get_tensor(ctx, name(i, "attn_norm.weight"));
+            layer.ln_1_b          = ggml_get_tensor(ctx, name(i, "attn_norm.bias"));
 
-            layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
-            layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
-            layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_q_proj_w = ggml_get_tensor(ctx, name(i, "attn_q.weight"));
+            layer.c_attn_k_proj_w = ggml_get_tensor(ctx, name(i, "attn_k.weight"));
+            layer.c_attn_v_proj_w = ggml_get_tensor(ctx, name(i, "attn_v.weight"));
 
-            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_proj_w   = ggml_get_tensor(ctx, name(i, "attn_output.weight"));
 
-            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
-            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+            layer.c_mlp_fc_w      = ggml_get_tensor(ctx, name(i, "ffn_up.weight"));
+            layer.c_mlp_fc_b      = ggml_get_tensor(ctx, name(i, "ffn_up.bias"));
 
-            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
-            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-
-            // map by name
-            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"]          = layer.ln_1_g;
-            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"]            = layer.ln_1_b;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"]   = layer.c_attn_q_proj_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"]   = layer.c_attn_k_proj_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"]   = layer.c_attn_v_proj_w;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"]     = layer.c_mlp_fc_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"]       = layer.c_mlp_fc_b;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"]    = layer.c_mlp_proj_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"]      = layer.c_mlp_proj_b;
+            layer.c_mlp_proj_w    = ggml_get_tensor(ctx, name(i, "ffn_down.weight"));
+            layer.c_mlp_proj_b    = ggml_get_tensor(ctx, name(i, "ffn_down.bias"));
         }
     }
 
@@ -354,113 +286,12 @@ bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & m
         printf("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
     }
 
-    // load weights
-    {
-        int n_tensors = 0;
-        size_t total_size = 0;
-
-        printf("%s: ", __func__);
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            int32_t ftype;
-
-            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                return false;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                return false;
-            }
-
-            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
-                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            if (0) {
-                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
-                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
-            }
-
-            size_t bpe = 0;
-
-            switch (ftype) {
-                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
-                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
-                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
-                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
-                default:
-                        {
-                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
-                            return false;
-                        }
-            };
-
-            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
-                return false;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
-            total_size += ggml_nbytes(tensor);
-            if (++n_tensors % 8 == 0) {
-                printf(".");
-                fflush(stdout);
-            }
-        }
-
-        printf(" done\n");
-
-        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
-    }
-
     model.scr0_buf.resize(256u * 1024 * 1024);
     model.scr1_buf.resize(256u * 1024 * 1024);
 
     return true;
 }
 
-// load the model's weights from a file path
-bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
-
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
-    bool loaded = gptj_model_load(fname, fin, model, vocab);
-    fin.close();
-    return loaded;
-}
-
 // evaluate the transformer
 //
 //   - model:     the model
@@ -512,8 +343,14 @@ bool gptj_eval(
     };
 
     struct ggml_context * ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    int * data = (int *) KQ_pos->data;
+    for (int i = 0; i < N; ++i) {
+        data[i] = n_past + i;
+    }
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -526,7 +363,7 @@ bool gptj_eval(
         ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
         // norm
         {
-            cur = ggml_norm(ctx0, inpL);
+            cur = ggml_norm(ctx0, inpL, model.hparams.norm_eps);
 
             // cur = ln_1_g*cur + ln_1_b
             cur = ggml_add(ctx0,
@@ -540,48 +377,44 @@ bool gptj_eval(
 
         // self-attention
         {
-            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
-            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
-            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur);
+            struct ggml_tensor * Qcur = ggml_rope(
+                ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N),
+                KQ_pos, n_rot, 0, 0
+            );
+            struct ggml_tensor * Kcur = ggml_rope(
+                ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N),
+                KQ_pos, n_rot, 0, 0
+            );
 
             // store key and value to memory
             {
-                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, model.kv_self.v, N*n_embd, (ggml_element_size(model.kv_self.v)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur));
 
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.kv_self.v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.kv_self.v),
+                        (il*n_ctx)*ggml_element_size(model.kv_self.v)*n_embd + n_past*ggml_element_size(model.kv_self.v));
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
             }
 
             // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_rope(ctx0,
-                            ggml_cpy(ctx0,
-                                Qcur,
-                                ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
-                            n_past, n_rot, 0),
-                        0, 2, 1, 3);
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
 
             // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
-                        ggml_rope(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
-                                n_embd/n_head, n_head, n_past + N),
-                            n_past, n_rot, 1),
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrt(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
@@ -590,17 +423,15 @@ bool gptj_eval(
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
 
             // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
-            struct ggml_tensor * V_trans =
-                ggml_cpy(ctx0,
-                        ggml_permute(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, model.kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd),
-                                n_embd/n_head, n_head, n_past + N),
-                            1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, model.kv_self.v->type, n_past + N, n_embd/n_head, n_head));
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.kv_self.v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.kv_self.v),
+                        n_ctx*ggml_element_size(model.kv_self.v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd);
 
             // KQV = transpose(V) * KQ_soft_max
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
@@ -656,7 +487,7 @@ bool gptj_eval(
 
     // norm
     {
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, model.hparams.norm_eps);
 
         // inpL = ln_f_g*inpL + ln_f_b
         inpL = ggml_add(ctx0,
@@ -680,13 +511,22 @@ bool gptj_eval(
     // logits -> probs
     //inpL = ggml_soft_max(ctx0, inpL);
 
+    ggml_build_forward_expand(gf, inpL);
+
     // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    {
+        std::unique_ptr<uint8_t []> data;
+        auto plan = ggml_graph_plan(gf, n_threads);
+        if (plan.work_size > 0) {
+            data.reset(new uint8_t[plan.work_size]);
+            plan.work_data = data.get();
+        }
+        ggml_graph_compute(gf, &plan);
+    }
 
     //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //    ggml_graph_print   (gf);
+    //    ggml_graph_dump_dot(gf, NULL, "gpt-2.dot");
     //}
 
     //embd_w.resize(n_vocab*N);
@@ -832,30 +672,34 @@ GPTJ::GPTJ()
     d_ptr->modelLoaded = false;
 }
 
-size_t GPTJ::requiredMem(const std::string &modelPath) {
+size_t GPTJ::requiredMem(const std::string &modelPath, int n_ctx, int ngl) {
+    (void)n_ctx;
+    (void)ngl;
     gptj_model dummy_model;
     gpt_vocab dummy_vocab;
     size_t mem_req;
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-    gptj_model_load(modelPath, fin, dummy_model, dummy_vocab, &mem_req);
+    gptj_model_load(modelPath, dummy_model, dummy_vocab, &mem_req);
     return mem_req;
 }
 
-bool GPTJ::loadModel(const std::string &modelPath) {
+bool GPTJ::loadModel(const std::string &modelPath, int n_ctx, int ngl) {
+    (void)n_ctx;
+    (void)ngl;
+    d_ptr->modelLoaded = false;
+
     std::mt19937 rng(time(NULL));
     d_ptr->rng = rng;
 
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-
     // load the model
-    if (!gptj_model_load(modelPath, fin, *d_ptr->model, d_ptr->vocab)) {
+    bool ok = gptj_model_load(modelPath, *d_ptr->model, d_ptr->vocab);
+    fflush(stdout);
+    if (!ok) {
         std::cerr << "GPT-J ERROR: failed to load model from " <<  modelPath;
         return false;
     }
 
     d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
     d_ptr->modelLoaded = true;
-    fflush(stdout);
     return true;
 }
 
@@ -893,8 +737,10 @@ size_t GPTJ::restoreState(const uint8_t *src)
     return gptj_set_state_data(d_ptr->model, &d_ptr->rng, src);
 }
 
-std::vector<LLModel::Token> GPTJ::tokenize(PromptContext &, const std::string &str) const
+std::vector<LLModel::Token> GPTJ::tokenize(PromptContext &ctx, const std::string &str, bool special) const
 {
+    (void)ctx;
+    (void)special;
     return ::gpt_tokenize(d_ptr->vocab, str);
 }
 
@@ -939,6 +785,18 @@ const std::vector<LLModel::Token> &GPTJ::endTokens() const
     return fres;
 }
 
+const char *get_arch_name(gguf_context *ctx_gguf) {
+    const int kid = gguf_find_key(ctx_gguf, "general.architecture");
+    if (kid == -1)
+        throw std::runtime_error("key not found in model: general.architecture");
+
+    enum gguf_type ktype = gguf_get_kv_type(ctx_gguf, kid);
+    if (ktype != GGUF_TYPE_STRING)
+        throw std::runtime_error("key general.architecture has wrong type");
+
+    return gguf_get_val_str(ctx_gguf, kid);
+}
+
 #if defined(_WIN32)
 #define DLL_EXPORT __declspec(dllexport)
 #else
@@ -958,15 +816,29 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(std::istream& f) {
-    uint32_t magic = 0;
-    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
-    gptj_hparams hparams;
-    f.read(reinterpret_cast<char*>(&hparams), sizeof(hparams));
-    if (!(hparams.n_vocab >= 50300 && hparams.n_vocab <= 50400)) {
-        return false; // not a gptj.
+DLL_EXPORT char *get_file_arch(const char *fname) {
+    struct ggml_context * ctx_meta = NULL;
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ &ctx_meta,
+    };
+    gguf_context *ctx_gguf = gguf_init_from_file(fname, params);
+
+    char *arch = nullptr;
+    if (ctx_gguf && gguf_get_version(ctx_gguf) <= 3) {
+        try {
+            arch = strdup(get_arch_name(ctx_gguf));
+        } catch (const std::runtime_error &) {
+            // cannot read key -> return null
+        }
     }
-    return magic == 0x67676d6c;
+
+    gguf_free(ctx_gguf);
+    return arch;
+}
+
+DLL_EXPORT bool is_arch_supported(const char *arch) {
+    return !strcmp(arch, "gptj");
 }
 
 DLL_EXPORT LLModel *construct() {

gpt4all-backend/gptj_impl.h

@@ -17,9 +17,9 @@ public:
 
     bool supportsEmbedding() const override { return false; }
     bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
+    bool loadModel(const std::string &modelPath, int n_ctx, int ngl) override;
     bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
+    size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) override;
     size_t stateSize() const override;
     size_t saveState(uint8_t *dest) const override;
     size_t restoreState(const uint8_t *src) override;
@@ -30,12 +30,13 @@ private:
     GPTJPrivate *d_ptr;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
+    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) const override;
     Token sampleToken(PromptContext &ctx) const override;
-    std::string tokenToString(Token) const override;
+    std::string tokenToString(Token id) const override;
     bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
     int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
+    const std::vector<Token> &endTokens() const override;
+    bool shouldAddBOS() const override { return false; }
 };
 
 #endif // GPTJ_H

gpt4all-backend/llamamodel_impl.h

@@ -4,44 +4,70 @@
 #ifndef LLAMAMODEL_H
 #define LLAMAMODEL_H
 
-#include <string>
 #include <functional>
+#include <memory>
+#include <string>
 #include <vector>
 #include "llmodel.h"
 
 struct LLamaPrivate;
+struct EmbModelSpec;
+
 class LLamaModel : public LLModel {
 public:
     LLamaModel();
     ~LLamaModel();
 
-    bool supportsEmbedding() const override { return false; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
+    bool supportsEmbedding() const override { return m_supportsEmbedding; }
+    bool supportsCompletion() const override { return m_supportsCompletion; }
+    bool loadModel(const std::string &modelPath, int n_ctx, int ngl) override;
+    bool isModelBlacklisted(const std::string &modelPath) const override;
+    bool isEmbeddingModel(const std::string &modelPath) const override;
     bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
+    size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) override;
     size_t stateSize() const override;
     size_t saveState(uint8_t *dest) const override;
     size_t restoreState(const uint8_t *src) override;
     void setThreadCount(int32_t n_threads) override;
     int32_t threadCount() const override;
-    std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired) override;
-    bool initializeGPUDevice(size_t memoryRequired, const std::string& device) override;
-    bool initializeGPUDevice(const GPUDevice &device) override;
-    bool initializeGPUDevice(int device) override;
-    bool hasGPUDevice() override;
-    bool usingGPUDevice() override;
+    std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired = 0) const override;
+    bool initializeGPUDevice(size_t memoryRequired, const std::string &name) const override;
+    bool initializeGPUDevice(int device, std::string *unavail_reason = nullptr) const override;
+    bool hasGPUDevice() const override;
+    bool usingGPUDevice() const override;
+    const char *backendName() const override;
+    const char *gpuDeviceName() const override;
+
+    size_t embeddingSize() const override;
+    // user-specified prefix
+    void embed(const std::vector<std::string> &texts, float *embeddings, std::optional<std::string> prefix,
+               int dimensionality = -1, size_t *tokenCount = nullptr, bool doMean = true, bool atlas = false,
+               EmbedCancelCallback *cancelCb = nullptr) override;
+    // automatic prefix
+    void embed(const std::vector<std::string> &texts, float *embeddings, bool isRetrieval, int dimensionality = -1,
+               size_t *tokenCount = nullptr, bool doMean = true, bool atlas = false) override;
 
 private:
-    LLamaPrivate *d_ptr;
+    void testModel(); // used for CUDA to eagerly allocate memory
+
+    std::unique_ptr<LLamaPrivate> d_ptr;
+    bool m_supportsEmbedding = false;
+    bool m_supportsCompletion = false;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    std::string tokenToString(Token) const override;
-    Token sampleToken(PromptContext& ctx) const override;
-    bool evalTokens(PromptContext& ctx, const std::vector<int32_t> &tokens) const override;
+    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) const override;
+    std::string tokenToString(Token id) const override;
+    Token sampleToken(PromptContext &ctx) const override;
+    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
     int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
+    const std::vector<Token> &endTokens() const override;
+    bool shouldAddBOS() const override;
+    int32_t maxContextLength(std::string const &modelPath) const override;
+    int32_t layerCount(std::string const &modelPath) const override;
+
+    void embedInternal(const std::vector<std::string> &texts, float *embeddings, std::string prefix, int dimensionality,
+                       size_t *tokenCount, bool doMean, bool atlas, EmbedCancelCallback *cancelCb,
+                       const EmbModelSpec *spec);
 };
 
 #endif // LLAMAMODEL_H

gpt4all-backend/llmodel.cpp

@@ -2,47 +2,62 @@
 #include "dlhandle.h"
 #include "sysinfo.h"
 
-#include <iostream>
-#include <string>
-#include <vector>
-#include <fstream>
-#include <filesystem>
 #include <cassert>
 #include <cstdlib>
+#include <filesystem>
+#include <fstream>
+#include <iostream>
+#include <memory>
+#include <optional>
+#include <regex>
 #include <sstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#ifdef _WIN32
+#   define WIN32_LEAN_AND_MEAN
+#   ifndef NOMINMAX
+#       define NOMINMAX
+#   endif
+#   include <windows.h>
+#endif
+
 #ifdef _MSC_VER
-#include <intrin.h>
+#   include <intrin.h>
+#endif
+
+#ifndef __APPLE__
+static const std::string DEFAULT_BACKENDS[] = {"kompute", "cpu"};
+#elif defined(__aarch64__)
+static const std::string DEFAULT_BACKENDS[] = {"metal", "cpu"};
+#else
+static const std::string DEFAULT_BACKENDS[] = {"cpu"};
 #endif
 
 std::string s_implementations_search_path = ".";
 
-static bool has_at_least_minimal_hardware() {
-#if defined(__x86_64__) || defined(_M_X64)
-    #ifndef _MSC_VER
-        return __builtin_cpu_supports("avx");
-    #else
-        int cpuInfo[4];
-        __cpuid(cpuInfo, 1);
-        return cpuInfo[2] & (1 << 28);
-    #endif
-#else
-    return true; // Don't know how to handle non-x86_64
-#endif
-}
+#if !(defined(__x86_64__) || defined(_M_X64))
+    // irrelevant on non-x86_64
+    #define cpu_supports_avx()  -1
+    #define cpu_supports_avx2() -1
+#elif defined(_MSC_VER)
+    // MSVC
+    static int get_cpu_info(int func_id, int reg_id) {
+        int info[4];
+        __cpuid(info, func_id);
+        return info[reg_id];
+    }
 
-static bool requires_avxonly() {
-#if defined(__x86_64__) || defined(_M_X64)
-    #ifndef _MSC_VER
-        return !__builtin_cpu_supports("avx2");
-    #else
-        int cpuInfo[4];
-        __cpuidex(cpuInfo, 7, 0);
-        return !(cpuInfo[1] & (1 << 5));
-    #endif
+    // AVX via EAX=1: Processor Info and Feature Bits, bit 28 of ECX
+    #define cpu_supports_avx()  !!(get_cpu_info(1, 2) & (1 << 28))
+    // AVX2 via EAX=7, ECX=0: Extended Features, bit 5 of EBX
+    #define cpu_supports_avx2() !!(get_cpu_info(7, 1) & (1 <<  5))
 #else
-    return false; // Don't know how to handle non-x86_64
+    // gcc/clang
+    #define cpu_supports_avx()  !!__builtin_cpu_supports("avx")
+    #define cpu_supports_avx2() !!__builtin_cpu_supports("avx2")
 #endif
-}
 
 LLModel::Implementation::Implementation(Dlhandle &&dlhandle_)
     : m_dlhandle(new Dlhandle(std::move(dlhandle_))) {
@@ -52,14 +67,17 @@ LLModel::Implementation::Implementation(Dlhandle &&dlhandle_)
     auto get_build_variant = m_dlhandle->get<const char *()>("get_build_variant");
     assert(get_build_variant);
     m_buildVariant = get_build_variant();
-    m_magicMatch = m_dlhandle->get<bool(std::ifstream&)>("magic_match");
-    assert(m_magicMatch);
+    m_getFileArch = m_dlhandle->get<char *(const char *)>("get_file_arch");
+    assert(m_getFileArch);
+    m_isArchSupported = m_dlhandle->get<bool(const char *)>("is_arch_supported");
+    assert(m_isArchSupported);
     m_construct = m_dlhandle->get<LLModel *()>("construct");
     assert(m_construct);
 }
 
 LLModel::Implementation::Implementation(Implementation &&o)
-    : m_magicMatch(o.m_magicMatch)
+    : m_getFileArch(o.m_getFileArch)
+    , m_isArchSupported(o.m_isArchSupported)
     , m_construct(o.m_construct)
     , m_modelType(o.m_modelType)
     , m_buildVariant(o.m_buildVariant)
@@ -68,19 +86,44 @@ LLModel::Implementation::Implementation(Implementation &&o)
 }
 
 LLModel::Implementation::~Implementation() {
-    if (m_dlhandle) delete m_dlhandle;
+    delete m_dlhandle;
 }
 
-bool LLModel::Implementation::isImplementation(const Dlhandle &dl) {
+static bool isImplementation(const Dlhandle &dl) {
     return dl.get<bool(uint32_t)>("is_g4a_backend_model_implementation");
 }
 
+// Add the CUDA Toolkit to the DLL search path on Windows.
+// This is necessary for chat.exe to find CUDA when started from Qt Creator.
+static void addCudaSearchPath() {
+#ifdef _WIN32
+    if (const auto *cudaPath = _wgetenv(L"CUDA_PATH")) {
+        auto libDir = std::wstring(cudaPath) + L"\\bin";
+        if (!AddDllDirectory(libDir.c_str())) {
+            auto err = GetLastError();
+            std::wcerr << L"AddDllDirectory(\"" << libDir << L"\") failed with error 0x" << std::hex << err << L"\n";
+        }
+    }
+#endif
+}
+
 const std::vector<LLModel::Implementation> &LLModel::Implementation::implementationList() {
+    if (cpu_supports_avx() == 0) {
+        throw std::runtime_error("CPU does not support AVX");
+    }
+
     // NOTE: allocated on heap so we leak intentionally on exit so we have a chance to clean up the
     // individual models without the cleanup of the static list interfering
     static auto* libs = new std::vector<Implementation>([] () {
         std::vector<Implementation> fres;
 
+        addCudaSearchPath();
+
+        std::string impl_name_re = "(gptj|llamamodel-mainline)-(cpu|metal|kompute|vulkan|cuda)";
+        if (cpu_supports_avx2() == 0) {
+            impl_name_re += "-avxonly";
+        }
+        std::regex re(impl_name_re);
         auto search_in_directory = [&](const std::string& paths) {
             std::stringstream ss(paths);
             std::string path;
@@ -90,13 +133,15 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
                 // Iterate over all libraries
                 for (const auto& f : std::filesystem::directory_iterator(fs_path)) {
                     const std::filesystem::path& p = f.path();
+
                     if (p.extension() != LIB_FILE_EXT) continue;
+                    if (!std::regex_search(p.stem().string(), re)) continue;
+
                     // Add to list if model implementation
                     try {
                         Dlhandle dl(p.string());
-                        if (!Implementation::isImplementation(dl)) {
+                        if (!isImplementation(dl))
                             continue;
-                        }
                         fres.emplace_back(Implementation(std::move(dl)));
                     } catch (...) {}
                 }
@@ -111,65 +156,149 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
     return *libs;
 }
 
-const LLModel::Implementation* LLModel::Implementation::implementation(std::ifstream& f, const std::string& buildVariant) {
-    for (const auto& i : implementationList()) {
-        f.seekg(0);
-        if (!i.m_magicMatch(f)) continue;
-        if (buildVariant != i.m_buildVariant) continue;
-        return &i;
+static std::string applyCPUVariant(const std::string &buildVariant) {
+    if (buildVariant != "metal" && cpu_supports_avx2() == 0) {
+        return buildVariant + "-avxonly";
     }
+    return buildVariant;
+}
+
+const LLModel::Implementation* LLModel::Implementation::implementation(const char *fname, const std::string& buildVariant) {
+    bool buildVariantMatched = false;
+    std::optional<std::string> archName;
+    for (const auto& i : implementationList()) {
+        if (buildVariant != i.m_buildVariant) continue;
+        buildVariantMatched = true;
+
+        char *arch = i.m_getFileArch(fname);
+        if (!arch) continue;
+        archName = arch;
+
+        bool archSupported = i.m_isArchSupported(arch);
+        free(arch);
+        if (archSupported) return &i;
+    }
+
+    if (!buildVariantMatched)
+        return nullptr;
+    if (!archName)
+        throw UnsupportedModelError("Unsupported file format");
+
+    throw BadArchError(std::move(*archName));
+}
+
+LLModel *LLModel::Implementation::construct(const std::string &modelPath, const std::string &backend, int n_ctx) {
+    std::vector<std::string> desiredBackends;
+    if (backend != "auto") {
+        desiredBackends.push_back(backend);
+    } else {
+        desiredBackends.insert(desiredBackends.end(), DEFAULT_BACKENDS, std::end(DEFAULT_BACKENDS));
+    }
+
+    for (const auto &desiredBackend: desiredBackends) {
+        const auto *impl = implementation(modelPath.c_str(), applyCPUVariant(desiredBackend));
+
+        if (impl) {
+            // Construct llmodel implementation
+            auto *fres = impl->m_construct();
+            fres->m_implementation = impl;
+
+#if defined(__APPLE__) && defined(__aarch64__) // FIXME: See if metal works for intel macs
+            /* TODO(cebtenzzre): after we fix requiredMem, we should change this to happen at
+             * load time, not construct time. right now n_ctx is incorrectly hardcoded 2048 in
+             * most (all?) places where this is called, causing underestimation of required
+             * memory. */
+            if (backend == "auto" && desiredBackend == "metal") {
+                // on a 16GB M2 Mac a 13B q4_0 (0.52) works for me but a 13B q4_K_M (0.55) does not
+                size_t req_mem = fres->requiredMem(modelPath, n_ctx, 100);
+                if (req_mem >= size_t(0.53f * getSystemTotalRAMInBytes())) {
+                    delete fres;
+                    continue;
+                }
+            }
+#else
+            (void)n_ctx;
+#endif
+
+            return fres;
+        }
+    }
+
+    throw MissingImplementationError("Could not find any implementations for backend: " + backend);
+}
+
+LLModel *LLModel::Implementation::constructGlobalLlama(const std::optional<std::string> &backend) {
+    static std::unordered_map<std::string, std::unique_ptr<LLModel>> implCache;
+
+    const std::vector<Implementation> *impls;
+    try {
+        impls = &implementationList();
+    } catch (const std::runtime_error &e) {
+        std::cerr << __func__ << ": implementationList failed: " << e.what() << "\n";
+        return nullptr;
+    }
+
+    std::vector<std::string> desiredBackends;
+    if (backend) {
+        desiredBackends.push_back(backend.value());
+    } else {
+        desiredBackends.insert(desiredBackends.end(), DEFAULT_BACKENDS, std::end(DEFAULT_BACKENDS));
+    }
+
+    const Implementation *impl = nullptr;
+
+    for (const auto &desiredBackend: desiredBackends) {
+        auto cacheIt = implCache.find(desiredBackend);
+        if (cacheIt != implCache.end())
+            return cacheIt->second.get(); // cached
+
+        for (const auto &i: *impls) {
+            if (i.m_modelType == "LLaMA" && i.m_buildVariant == applyCPUVariant(desiredBackend)) {
+                impl = &i;
+                break;
+            }
+        }
+
+        if (impl) {
+            auto *fres = impl->m_construct();
+            fres->m_implementation = impl;
+            implCache[desiredBackend] = std::unique_ptr<LLModel>(fres);
+            return fres;
+        }
+    }
+
+    std::cerr << __func__ << ": could not find Llama implementation for backend: " << backend.value_or("default") << "\n";
     return nullptr;
 }
 
-LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::string buildVariant) {
-
-    if (!has_at_least_minimal_hardware())
-        return nullptr;
-
-    // Read magic
-    std::ifstream f(modelPath, std::ios::binary);
-    if (!f) return nullptr;
-    // Get correct implementation
-    const Implementation* impl = nullptr;
-
-    #if defined(__APPLE__) && defined(__arm64__) // FIXME: See if metal works for intel macs
-        if (buildVariant == "auto") {
-            size_t total_mem = getSystemTotalRAMInBytes();
-            impl = implementation(f, "metal");
-            if(impl) {
-                LLModel* metalimpl = impl->m_construct();
-                metalimpl->m_implementation = impl;
-                size_t req_mem = metalimpl->requiredMem(modelPath);
-                float req_to_total = (float) req_mem / (float) total_mem;
-                // on a 16GB M2 Mac a 13B q4_0 (0.52) works for me but a 13B q4_K_M (0.55) does not
-                if (req_to_total >= 0.53) {
-                    delete metalimpl;
-                    impl = nullptr;
-                } else {
-                    return metalimpl;
-                }
-            }
+std::vector<LLModel::GPUDevice> LLModel::Implementation::availableGPUDevices(size_t memoryRequired) {
+    std::vector<LLModel::GPUDevice> devices;
+#ifndef __APPLE__
+    static const std::string backends[] = {"kompute", "cuda"};
+    for (const auto &backend: backends) {
+        auto *llama = constructGlobalLlama(backend);
+        if (llama) {
+            auto backendDevs = llama->availableGPUDevices(memoryRequired);
+            devices.insert(devices.end(), backendDevs.begin(), backendDevs.end());
         }
-    #endif
-
-    if (!impl) {
-        //TODO: Auto-detect CUDA/OpenCL
-        if (buildVariant == "auto") {
-            if (requires_avxonly()) {
-                buildVariant = "avxonly";
-            } else {
-                buildVariant = "default";
-            }
-        }
-        impl = implementation(f, buildVariant);
-        if (!impl) return nullptr;
     }
-    f.close();
+#endif
+    return devices;
+}
 
-    // Construct and return llmodel implementation
-    auto fres = impl->m_construct();
-    fres->m_implementation = impl;
-    return fres;
+int32_t LLModel::Implementation::maxContextLength(const std::string &modelPath) {
+    auto *llama = constructGlobalLlama();
+    return llama ? llama->maxContextLength(modelPath) : -1;
+}
+
+int32_t LLModel::Implementation::layerCount(const std::string &modelPath) {
+    auto *llama = constructGlobalLlama();
+    return llama ? llama->layerCount(modelPath) : -1;
+}
+
+bool LLModel::Implementation::isEmbeddingModel(const std::string &modelPath) {
+    auto *llama = constructGlobalLlama();
+    return llama && llama->isEmbeddingModel(modelPath);
 }
 
 void LLModel::Implementation::setImplementationsSearchPath(const std::string& path) {
@@ -179,3 +308,11 @@ void LLModel::Implementation::setImplementationsSearchPath(const std::string& pa
 const std::string& LLModel::Implementation::implementationsSearchPath() {
     return s_implementations_search_path;
 }
+
+bool LLModel::Implementation::hasSupportedCPU() {
+    return cpu_supports_avx() != 0;
+}
+
+int LLModel::Implementation::cpuSupportsAVX2() {
+    return cpu_supports_avx2();
+}

gpt4all-backend/llmodel.h

@@ -1,13 +1,18 @@
 #ifndef LLMODEL_H
 #define LLMODEL_H
 
-#include <string>
-#include <functional>
-#include <vector>
-#include <string_view>
-#include <fstream>
+#include <algorithm>
 #include <cstdint>
+#include <fstream>
+#include <functional>
 #include <limits>
+#include <optional>
+#include <string>
+#include <string_view>
+#include <unordered_map>
+#include <vector>
+
+using namespace std::string_literals;
 
 #define LLMODEL_MAX_PROMPT_BATCH 128
 
@@ -15,28 +20,93 @@ class Dlhandle;
 class LLModel {
 public:
     using Token = int32_t;
+
+    class BadArchError: public std::runtime_error {
+    public:
+        BadArchError(std::string arch)
+            : runtime_error("Unsupported model architecture: " + arch)
+            , m_arch(std::move(arch))
+            {}
+
+        const std::string &arch() const noexcept { return m_arch; }
+
+    private:
+        std::string m_arch;
+    };
+
+    class MissingImplementationError: public std::runtime_error {
+    public:
+        using std::runtime_error::runtime_error;
+    };
+
+    class UnsupportedModelError: public std::runtime_error {
+    public:
+        using std::runtime_error::runtime_error;
+    };
+
+    struct GPUDevice {
+        const char *backend;
+        int index;
+        int type;
+        size_t heapSize;
+        std::string name;
+        std::string vendor;
+
+        GPUDevice(const char *backend, int index, int type, size_t heapSize, std::string name, std::string vendor):
+            backend(backend), index(index), type(type), heapSize(heapSize), name(std::move(name)),
+            vendor(std::move(vendor)) {}
+
+        std::string selectionName() const { return m_backendNames.at(backend) + ": " + name; }
+        std::string reportedName()  const { return name + " (" + m_backendNames.at(backend) + ")"; }
+
+        static std::string updateSelectionName(const std::string &name) {
+            if (name == "Auto" || name == "CPU" || name == "Metal")
+                return name;
+            auto it = std::find_if(m_backendNames.begin(), m_backendNames.end(), [&name](const auto &entry) {
+                return name.starts_with(entry.second + ": ");
+            });
+            if (it != m_backendNames.end())
+                return name;
+            return "Vulkan: " + name; // previously, there were only Vulkan devices
+        }
+
+    private:
+        static inline const std::unordered_map<std::string, std::string> m_backendNames {
+            {"cuda", "CUDA"}, {"kompute", "Vulkan"},
+        };
+    };
+
     class Implementation {
     public:
-        Implementation(Dlhandle&&);
-        Implementation(const Implementation&) = delete;
-        Implementation(Implementation&&);
+        Implementation(const Implementation &) = delete;
+        Implementation(Implementation &&);
         ~Implementation();
 
         std::string_view modelType() const { return m_modelType; }
         std::string_view buildVariant() const { return m_buildVariant; }
 
-        static bool isImplementation(const Dlhandle&);
-        static const std::vector<Implementation>& implementationList();
-        static const Implementation *implementation(std::ifstream& f, const std::string& buildVariant);
-        static LLModel *construct(const std::string &modelPath, std::string buildVariant = "auto");
-        static void setImplementationsSearchPath(const std::string& path);
-        static const std::string& implementationsSearchPath();
+        static LLModel *construct(const std::string &modelPath, const std::string &backend = "auto", int n_ctx = 2048);
+        static std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired = 0);
+        static int32_t maxContextLength(const std::string &modelPath);
+        static int32_t layerCount(const std::string &modelPath);
+        static bool isEmbeddingModel(const std::string &modelPath);
+        static void setImplementationsSearchPath(const std::string &path);
+        static const std::string &implementationsSearchPath();
+        static bool hasSupportedCPU();
+        // 0 for no, 1 for yes, -1 for non-x86_64
+        static int cpuSupportsAVX2();
 
     private:
-        bool (*m_magicMatch)(std::ifstream& f);
+        Implementation(Dlhandle &&);
+
+        static const std::vector<Implementation> &implementationList();
+        static const Implementation *implementation(const char *fname, const std::string &buildVariant);
+        static LLModel *constructGlobalLlama(const std::optional<std::string> &backend = std::nullopt);
+
+        char *(*m_getFileArch)(const char *fname);
+        bool (*m_isArchSupported)(const char *arch);
         LLModel *(*m_construct)();
 
-    private:
         std::string_view m_modelType;
         std::string_view m_buildVariant;
         Dlhandle *m_dlhandle;
@@ -50,68 +120,110 @@ public:
         int32_t n_predict = 200;
         int32_t top_k = 40;
         float   top_p = 0.9f;
+        float   min_p = 0.0f;
         float   temp = 0.9f;
-        int32_t n_batch = 9;
+        int32_t n_batch = 128;
         float   repeat_penalty = 1.10f;
         int32_t repeat_last_n = 64;     // last n tokens to penalize
-        float   contextErase = 0.75f;   // percent of context to erase if we exceed the context
-            // window
+        float   contextErase = 0.75f;   // percent of context to erase if we exceed the context window
+        int32_t n_last_batch_tokens = 0;
     };
 
-    struct GPUDevice {
-        int index = 0;
-        int type = 0;
-        size_t heapSize = 0;
-        std::string name;
-        std::string vendor;
-    };
+    using ProgressCallback = std::function<bool(float progress)>;
 
     explicit LLModel() {}
     virtual ~LLModel() {}
 
     virtual bool supportsEmbedding() const = 0;
     virtual bool supportsCompletion() const = 0;
-    virtual bool loadModel(const std::string &modelPath) = 0;
+    virtual bool loadModel(const std::string &modelPath, int n_ctx, int ngl) = 0;
+    virtual bool isModelBlacklisted(const std::string &modelPath) const { (void)modelPath; return false; };
+    virtual bool isEmbeddingModel(const std::string &modelPath) const { (void)modelPath; return false; }
     virtual bool isModelLoaded() const = 0;
-    virtual size_t requiredMem(const std::string &modelPath) = 0;
+    virtual size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) = 0;
     virtual size_t stateSize() const { return 0; }
-    virtual size_t saveState(uint8_t */*dest*/) const { return 0; }
-    virtual size_t restoreState(const uint8_t */*src*/) { return 0; }
+    virtual size_t saveState(uint8_t *dest) const { (void)dest; return 0; }
+    virtual size_t restoreState(const uint8_t *src) { (void)src; return 0; }
 
     // This method requires the model to return true from supportsCompletion otherwise it will throw
     // an error
     virtual void prompt(const std::string &prompt,
+                        const std::string &promptTemplate,
                         std::function<bool(int32_t)> promptCallback,
                         std::function<bool(int32_t, const std::string&)> responseCallback,
                         std::function<bool(bool)> recalculateCallback,
-                        PromptContext &ctx);
+                        PromptContext &ctx,
+                        bool special = false,
+                        std::string *fakeReply = nullptr);
 
-    virtual std::vector<float> embedding(const std::string &text);
+    using EmbedCancelCallback = bool(unsigned *batchSizes, unsigned nBatch, const char *backend);
 
-    virtual void setThreadCount(int32_t /*n_threads*/) {}
+    virtual size_t embeddingSize() const {
+        throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
+    }
+    // user-specified prefix
+    virtual void embed(const std::vector<std::string> &texts, float *embeddings, std::optional<std::string> prefix,
+                       int dimensionality = -1, size_t *tokenCount = nullptr, bool doMean = true, bool atlas = false,
+                       EmbedCancelCallback *cancelCb = nullptr);
+    // automatic prefix
+    virtual void embed(const std::vector<std::string> &texts, float *embeddings, bool isRetrieval,
+                       int dimensionality = -1, size_t *tokenCount = nullptr, bool doMean = true, bool atlas = false);
+
+    virtual void setThreadCount(int32_t n_threads) { (void)n_threads; }
     virtual int32_t threadCount() const { return 1; }
 
-    const Implementation& implementation() const {
+    const Implementation &implementation() const {
         return *m_implementation;
     }
 
-    virtual std::vector<GPUDevice> availableGPUDevices(size_t /*memoryRequired*/) { return std::vector<GPUDevice>(); }
-    virtual bool initializeGPUDevice(size_t /*memoryRequired*/, const std::string& /*device*/) { return false; }
-    virtual bool initializeGPUDevice(const GPUDevice &/*device*/) { return false; }
-    virtual bool initializeGPUDevice(int /*device*/) { return false; }
-    virtual bool hasGPUDevice() { return false; }
-    virtual bool usingGPUDevice() { return false; }
-    static std::vector<GPUDevice> availableGPUDevices();
+    virtual std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired) const {
+        (void)memoryRequired;
+        return {};
+    }
+
+    virtual bool initializeGPUDevice(size_t memoryRequired, const std::string &name) const {
+        (void)memoryRequired;
+        (void)name;
+        return false;
+    }
+
+    virtual bool initializeGPUDevice(int device, std::string *unavail_reason = nullptr) const {
+        (void)device;
+        if (unavail_reason) {
+            *unavail_reason = "model has no GPU support";
+        }
+        return false;
+    }
+
+    virtual bool hasGPUDevice() const { return false; }
+    virtual bool usingGPUDevice() const { return false; }
+    virtual const char *backendName() const { return "cpu"; }
+    virtual const char *gpuDeviceName() const { return nullptr; }
+
+    void setProgressCallback(ProgressCallback callback) { m_progressCallback = callback; }
 
 protected:
     // These are pure virtual because subclasses need to implement as the default implementation of
     // 'prompt' above calls these functions
-    virtual std::vector<Token> tokenize(PromptContext &, const std::string&) const = 0;
-    virtual std::string tokenToString(Token) const = 0;
+    virtual std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special = false) const = 0;
+    virtual std::string tokenToString(Token id) const = 0;
     virtual Token sampleToken(PromptContext &ctx) const = 0;
-    virtual bool evalTokens(PromptContext &/*ctx*/, const std::vector<int32_t>& /*tokens*/) const = 0;
+    virtual bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const = 0;
     virtual int32_t contextLength() const = 0;
-    virtual const std::vector<Token>& endTokens() const = 0;
+    virtual const std::vector<Token> &endTokens() const = 0;
+    virtual bool shouldAddBOS() const = 0;
+
+    virtual int32_t maxContextLength(std::string const &modelPath) const
+    {
+        (void)modelPath;
+        return -1;
+    }
+
+    virtual int32_t layerCount(std::string const &modelPath) const
+    {
+        (void)modelPath;
+        return -1;
+    }
 
     // This is a helper function called from the default implementation of 'prompt' but it can be
     // shared by all base classes so it isn't virtual
@@ -119,6 +231,24 @@ protected:
 
     const Implementation *m_implementation = nullptr;
 
+    ProgressCallback m_progressCallback;
+    static bool staticProgressCallback(float progress, void* ctx)
+    {
+        LLModel* model = static_cast<LLModel*>(ctx);
+        if (model && model->m_progressCallback)
+            return model->m_progressCallback(progress);
+        return true;
+    }
+
+    void decodePrompt(std::function<bool(int32_t)> promptCallback,
+                      std::function<bool(int32_t, const std::string&)> responseCallback,
+                      std::function<bool(bool)> recalculateCallback,
+                      PromptContext &promptCtx,
+                      std::vector<Token> embd_inp);
+    void generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
+                          std::function<bool(bool)> recalculateCallback,
+                          PromptContext &promptCtx);
+
 private:
     friend class LLMImplementation;
 };

gpt4all-backend/llmodel_c.cpp

@@ -1,8 +1,11 @@
 #include "llmodel_c.h"
 #include "llmodel.h"
 
-#include <cstring>
 #include <cerrno>
+#include <cstring>
+#include <iostream>
+#include <memory>
+#include <optional>
 #include <utility>
 
 struct LLModelWrapper {
@@ -11,121 +14,98 @@ struct LLModelWrapper {
     ~LLModelWrapper() { delete llModel; }
 };
 
-
-thread_local static std::string last_error_message;
-
-
 llmodel_model llmodel_model_create(const char *model_path) {
-    auto fres = llmodel_model_create2(model_path, "auto", nullptr);
+    const char *error;
+    auto fres = llmodel_model_create2(model_path, "auto", &error);
     if (!fres) {
-        fprintf(stderr, "Invalid model file\n");
+        fprintf(stderr, "Unable to instantiate model: %s\n", error);
     }
     return fres;
 }
 
-llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, llmodel_error *error) {
-    auto wrapper = new LLModelWrapper;
-    int error_code = 0;
+static void llmodel_set_error(const char **errptr, const char *message) {
+    thread_local static std::string last_error_message;
+    if (errptr) {
+        last_error_message = message;
+        *errptr = last_error_message.c_str();
+    }
+}
 
+llmodel_model llmodel_model_create2(const char *model_path, const char *backend, const char **error) {
+    LLModel *llModel;
     try {
-        wrapper->llModel = LLModel::Implementation::construct(model_path, build_variant);
+        llModel = LLModel::Implementation::construct(model_path, backend);
     } catch (const std::exception& e) {
-        error_code = EINVAL;
-        last_error_message = e.what();
+        llmodel_set_error(error, e.what());
+        return nullptr;
     }
 
-    if (!wrapper->llModel) {
-        delete std::exchange(wrapper, nullptr);
-        // Get errno and error message if none
-        if (error_code == 0) {
-            if (errno != 0) {
-                error_code = errno;
-                last_error_message = std::strerror(error_code);
-            } else {
-                error_code = ENOTSUP;
-                last_error_message = "Model format not supported (no matching implementation found)";
-            }
-        }
-        // Set error argument
-        if (error) {
-            error->message = last_error_message.c_str();
-            error->code = error_code;
-        }
-    }
-    return reinterpret_cast<llmodel_model*>(wrapper);
+    auto wrapper = new LLModelWrapper;
+    wrapper->llModel = llModel;
+    return wrapper;
 }
 
 void llmodel_model_destroy(llmodel_model model) {
-    delete reinterpret_cast<LLModelWrapper*>(model);
+    delete static_cast<LLModelWrapper *>(model);
 }
 
-size_t llmodel_required_mem(llmodel_model model, const char *model_path)
+size_t llmodel_required_mem(llmodel_model model, const char *model_path, int n_ctx, int ngl)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
-    return wrapper->llModel->requiredMem(model_path);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
+    return wrapper->llModel->requiredMem(model_path, n_ctx, ngl);
 }
 
-bool llmodel_loadModel(llmodel_model model, const char *model_path)
+bool llmodel_loadModel(llmodel_model model, const char *model_path, int n_ctx, int ngl)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
-    return wrapper->llModel->loadModel(model_path);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
+
+    std::string modelPath(model_path);
+    if (wrapper->llModel->isModelBlacklisted(modelPath)) {
+        size_t slash = modelPath.find_last_of("/\\");
+        auto basename = slash == std::string::npos ? modelPath : modelPath.substr(slash + 1);
+        std::cerr << "warning: model '" << basename << "' is out-of-date, please check for an updated version\n";
+    }
+    return wrapper->llModel->loadModel(modelPath, n_ctx, ngl);
 }
 
 bool llmodel_isModelLoaded(llmodel_model model)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->isModelLoaded();
 }
 
 uint64_t llmodel_get_state_size(llmodel_model model)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->stateSize();
 }
 
 uint64_t llmodel_save_state_data(llmodel_model model, uint8_t *dest)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->saveState(dest);
 }
 
 uint64_t llmodel_restore_state_data(llmodel_model model, const uint8_t *src)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->restoreState(src);
 }
 
-// Wrapper functions for the C callbacks
-bool prompt_wrapper(int32_t token_id, void *user_data) {
-    llmodel_prompt_callback callback = reinterpret_cast<llmodel_prompt_callback>(user_data);
-    return callback(token_id);
-}
-
-bool response_wrapper(int32_t token_id, const std::string &response, void *user_data) {
-    llmodel_response_callback callback = reinterpret_cast<llmodel_response_callback>(user_data);
-    return callback(token_id, response.c_str());
-}
-
-bool recalculate_wrapper(bool is_recalculating, void *user_data) {
-    llmodel_recalculate_callback callback = reinterpret_cast<llmodel_recalculate_callback>(user_data);
-    return callback(is_recalculating);
-}
-
 void llmodel_prompt(llmodel_model model, const char *prompt,
+                    const char *prompt_template,
                     llmodel_prompt_callback prompt_callback,
                     llmodel_response_callback response_callback,
                     llmodel_recalculate_callback recalculate_callback,
-                    llmodel_prompt_context *ctx)
+                    llmodel_prompt_context *ctx,
+                    bool special,
+                    const char *fake_reply)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
 
-    // Create std::function wrappers that call the C function pointers
-    std::function<bool(int32_t)> prompt_func =
-        std::bind(&prompt_wrapper, std::placeholders::_1, reinterpret_cast<void*>(prompt_callback));
-    std::function<bool(int32_t, const std::string&)> response_func =
-        std::bind(&response_wrapper, std::placeholders::_1, std::placeholders::_2, reinterpret_cast<void*>(response_callback));
-    std::function<bool(bool)> recalc_func =
-        std::bind(&recalculate_wrapper, std::placeholders::_1, reinterpret_cast<void*>(recalculate_callback));
+    auto response_func = [response_callback](int32_t token_id, const std::string &response) {
+        return response_callback(token_id, response.c_str());
+    };
 
     if (size_t(ctx->n_past) < wrapper->promptContext.tokens.size())
         wrapper->promptContext.tokens.resize(ctx->n_past);
@@ -136,14 +116,20 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     wrapper->promptContext.n_predict = ctx->n_predict;
     wrapper->promptContext.top_k = ctx->top_k;
     wrapper->promptContext.top_p = ctx->top_p;
+    wrapper->promptContext.min_p = ctx->min_p;
     wrapper->promptContext.temp = ctx->temp;
     wrapper->promptContext.n_batch = ctx->n_batch;
     wrapper->promptContext.repeat_penalty = ctx->repeat_penalty;
     wrapper->promptContext.repeat_last_n = ctx->repeat_last_n;
     wrapper->promptContext.contextErase = ctx->context_erase;
 
+    std::string fake_reply_str;
+    if (fake_reply) { fake_reply_str = fake_reply; }
+    auto *fake_reply_p = fake_reply ? &fake_reply_str : nullptr;
+
     // Call the C++ prompt method
-    wrapper->llModel->prompt(prompt, prompt_func, response_func, recalc_func, wrapper->promptContext);
+    wrapper->llModel->prompt(prompt, prompt_template, prompt_callback, response_func, recalculate_callback,
+                             wrapper->promptContext, special, fake_reply_p);
 
     // Update the C context by giving access to the wrappers raw pointers to std::vector data
     // which involves no copies
@@ -158,6 +144,7 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     ctx->n_predict = wrapper->promptContext.n_predict;
     ctx->top_k = wrapper->promptContext.top_k;
     ctx->top_p = wrapper->promptContext.top_p;
+    ctx->min_p = wrapper->promptContext.min_p;
     ctx->temp = wrapper->promptContext.temp;
     ctx->n_batch = wrapper->promptContext.n_batch;
     ctx->repeat_penalty = wrapper->promptContext.repeat_penalty;
@@ -165,38 +152,58 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     ctx->context_erase = wrapper->promptContext.contextErase;
 }
 
-float *llmodel_embedding(llmodel_model model, const char *text, size_t *embedding_size)
-{
-    if (model == nullptr || text == nullptr || !strlen(text)) {
-        *embedding_size = 0;
+float *llmodel_embed(
+    llmodel_model model, const char **texts, size_t *embedding_size, const char *prefix, int dimensionality,
+    size_t *token_count, bool do_mean, bool atlas, llmodel_emb_cancel_callback cancel_cb, const char **error
+) {
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
+
+    if (!texts || !*texts) {
+        llmodel_set_error(error, "'texts' is NULL or empty");
         return nullptr;
     }
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
-    std::vector<float> embeddingVector = wrapper->llModel->embedding(text);
-    float *embedding = (float *)malloc(embeddingVector.size() * sizeof(float));
-    if (embedding == nullptr) {
-        *embedding_size = 0;
+
+    std::vector<std::string> textsVec;
+    while (*texts) { textsVec.emplace_back(*texts++); }
+
+    size_t embd_size;
+    float *embedding;
+
+    try {
+        embd_size = wrapper->llModel->embeddingSize();
+        if (dimensionality > 0 && dimensionality < int(embd_size))
+            embd_size = dimensionality;
+
+        embd_size *= textsVec.size();
+
+        std::optional<std::string> prefixStr;
+        if (prefix) { prefixStr = prefix; }
+
+        embedding = new float[embd_size];
+        wrapper->llModel->embed(textsVec, embedding, prefixStr, dimensionality, token_count, do_mean, atlas, cancel_cb);
+    } catch (std::exception const &e) {
+        llmodel_set_error(error, e.what());
         return nullptr;
     }
-    std::copy(embeddingVector.begin(), embeddingVector.end(), embedding);
-    *embedding_size = embeddingVector.size();
+
+    *embedding_size = embd_size;
     return embedding;
 }
 
 void llmodel_free_embedding(float *ptr)
 {
-    free(ptr);
+    delete[] ptr;
 }
 
 void llmodel_setThreadCount(llmodel_model model, int32_t n_threads)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     wrapper->llModel->setThreadCount(n_threads);
 }
 
 int32_t llmodel_threadCount(llmodel_model model)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->threadCount();
 }
 
@@ -210,56 +217,80 @@ const char *llmodel_get_implementation_search_path()
     return LLModel::Implementation::implementationsSearchPath().c_str();
 }
 
-struct llmodel_gpu_device* llmodel_available_gpu_devices(llmodel_model model, size_t memoryRequired, int* num_devices)
-{
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
-    std::vector<LLModel::GPUDevice> devices = wrapper->llModel->availableGPUDevices(memoryRequired);
+// RAII wrapper around a C-style struct
+struct llmodel_gpu_device_cpp: llmodel_gpu_device {
+    llmodel_gpu_device_cpp() = default;
 
-    // Set the num_devices
+    llmodel_gpu_device_cpp(const llmodel_gpu_device_cpp  &) = delete;
+    llmodel_gpu_device_cpp(      llmodel_gpu_device_cpp &&) = delete;
+
+    const llmodel_gpu_device_cpp &operator=(const llmodel_gpu_device_cpp  &) = delete;
+          llmodel_gpu_device_cpp &operator=(      llmodel_gpu_device_cpp &&) = delete;
+
+    ~llmodel_gpu_device_cpp() {
+        free(const_cast<char *>(name));
+        free(const_cast<char *>(vendor));
+    }
+};
+
+static_assert(sizeof(llmodel_gpu_device_cpp) == sizeof(llmodel_gpu_device));
+
+struct llmodel_gpu_device *llmodel_available_gpu_devices(size_t memoryRequired, int *num_devices)
+{
+    static thread_local std::unique_ptr<llmodel_gpu_device_cpp[]> c_devices;
+
+    auto devices = LLModel::Implementation::availableGPUDevices(memoryRequired);
     *num_devices = devices.size();
 
-    if (*num_devices == 0) return nullptr;  // Return nullptr if no devices are found
+    if (devices.empty()) { return nullptr; /* no devices */ }
 
-    // Allocate memory for the output array
-    struct llmodel_gpu_device* output = (struct llmodel_gpu_device*) malloc(*num_devices * sizeof(struct llmodel_gpu_device));
-
-    for (int i = 0; i < *num_devices; i++) {
-        output[i].index = devices[i].index;
-        output[i].type = devices[i].type;
-        output[i].heapSize = devices[i].heapSize;
-        output[i].name = strdup(devices[i].name.c_str());  // Convert std::string to char* and allocate memory
-        output[i].vendor = strdup(devices[i].vendor.c_str());  // Convert std::string to char* and allocate memory
+    c_devices = std::make_unique<llmodel_gpu_device_cpp[]>(devices.size());
+    for (unsigned i = 0; i < devices.size(); i++) {
+        const auto &dev  =   devices[i];
+              auto &cdev = c_devices[i];
+        cdev.backend  = dev.backend;
+        cdev.index    = dev.index;
+        cdev.type     = dev.type;
+        cdev.heapSize = dev.heapSize;
+        cdev.name     = strdup(dev.name.c_str());
+        cdev.vendor   = strdup(dev.vendor.c_str());
     }
 
-    return output;
+    return c_devices.get();
 }
 
 bool llmodel_gpu_init_gpu_device_by_string(llmodel_model model, size_t memoryRequired, const char *device)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->initializeGPUDevice(memoryRequired, std::string(device));
 }
 
 bool llmodel_gpu_init_gpu_device_by_struct(llmodel_model model, const llmodel_gpu_device *device)
 {
-    LLModel::GPUDevice d;
-    d.index = device->index;
-    d.type = device->type;
-    d.heapSize = device->heapSize;
-    d.name = device->name;
-    d.vendor = device->vendor;
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
-    return wrapper->llModel->initializeGPUDevice(d);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
+    return wrapper->llModel->initializeGPUDevice(device->index);
 }
 
 bool llmodel_gpu_init_gpu_device_by_int(llmodel_model model, int device)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->initializeGPUDevice(device);
 }
 
 bool llmodel_has_gpu_device(llmodel_model model)
 {
-    LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
+    const auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->hasGPUDevice();
 }
+
+const char *llmodel_model_backend_name(llmodel_model model)
+{
+    const auto *wrapper = static_cast<LLModelWrapper *>(model);
+    return wrapper->llModel->backendName();
+}
+
+const char *llmodel_model_gpu_device_name(llmodel_model model)
+{
+    const auto *wrapper = static_cast<LLModelWrapper *>(model);
+    return wrapper->llModel->gpuDeviceName();
+}

gpt4all-backend/llmodel_c.h

@@ -23,17 +23,6 @@ extern "C" {
  */
 typedef void *llmodel_model;
 
-/**
- * Structure containing any errors that may eventually occur
- */
-struct llmodel_error {
-    const char *message;  // Human readable error description; Thread-local; guaranteed to survive until next llmodel C API call
-    int code;             // errno; 0 if none
-};
-#ifndef __cplusplus
-typedef struct llmodel_error llmodel_error;
-#endif
-
 /**
  * llmodel_prompt_context structure for holding the prompt context.
  * NOTE: The implementation takes care of all the memory handling of the raw logits pointer and the
@@ -50,6 +39,7 @@ struct llmodel_prompt_context {
     int32_t n_predict;      // number of tokens to predict
     int32_t top_k;          // top k logits to sample from
     float top_p;            // nucleus sampling probability threshold
+    float min_p;            // Min P sampling
     float temp;             // temperature to adjust model's output distribution
     int32_t n_batch;        // number of predictions to generate in parallel
     float repeat_penalty;   // penalty factor for repeated tokens
@@ -58,9 +48,10 @@ struct llmodel_prompt_context {
 };
 
 struct llmodel_gpu_device {
-    int index = 0;
-    int type = 0;           // same as VkPhysicalDeviceType
-    size_t heapSize = 0;
+    const char * backend;
+    int index;
+    int type; // same as VkPhysicalDeviceType
+    size_t heapSize;
     const char * name;
     const char * vendor;
 };
@@ -92,6 +83,15 @@ typedef bool (*llmodel_response_callback)(int32_t token_id, const char *response
  */
 typedef bool (*llmodel_recalculate_callback)(bool is_recalculating);
 
+/**
+ * Embedding cancellation callback for use with llmodel_embed.
+ * @param batch_sizes The number of tokens in each batch that will be embedded.
+ * @param n_batch The number of batches that will be embedded.
+ * @param backend The backend that will be used for embedding. One of "cpu", "kompute", "cuda", or "metal".
+ * @return True to cancel llmodel_embed, false to continue.
+ */
+typedef bool (*llmodel_emb_cancel_callback)(unsigned *batch_sizes, unsigned n_batch, const char *backend);
+
 /**
  * Create a llmodel instance.
  * Recognises correct model type from file at model_path
@@ -104,11 +104,11 @@ DEPRECATED llmodel_model llmodel_model_create(const char *model_path);
  * Create a llmodel instance.
  * Recognises correct model type from file at model_path
  * @param model_path A string representing the path to the model file; will only be used to detect model type.
- * @param build_variant A string representing the implementation to use (auto, default, avxonly, ...),
- * @param error A pointer to a llmodel_error; will only be set on error.
+ * @param backend A string representing the implementation to use. One of 'auto', 'cpu', 'metal', 'kompute', or 'cuda'.
+ * @param error A pointer to a string; will only be set on error.
  * @return A pointer to the llmodel_model instance; NULL on error.
  */
-llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, llmodel_error *error);
+llmodel_model llmodel_model_create2(const char *model_path, const char *backend, const char **error);
 
 /**
  * Destroy a llmodel instance.
@@ -121,17 +121,21 @@ void llmodel_model_destroy(llmodel_model model);
  * Estimate RAM requirement for a model file
  * @param model A pointer to the llmodel_model instance.
  * @param model_path A string representing the path to the model file.
+ * @param n_ctx Maximum size of context window
+ * @param ngl Number of GPU layers to use (Vulkan)
  * @return size greater than 0 if the model was parsed successfully, 0 if file could not be parsed.
  */
-size_t llmodel_required_mem(llmodel_model model, const char *model_path);
+size_t llmodel_required_mem(llmodel_model model, const char *model_path, int n_ctx, int ngl);
 
 /**
  * Load a model from a file.
  * @param model A pointer to the llmodel_model instance.
  * @param model_path A string representing the path to the model file.
+ * @param n_ctx Maximum size of context window
+ * @param ngl Number of GPU layers to use (Vulkan)
  * @return true if the model was loaded successfully, false otherwise.
  */
-bool llmodel_loadModel(llmodel_model model, const char *model_path);
+bool llmodel_loadModel(llmodel_model model, const char *model_path, int n_ctx, int ngl);
 
 /**
  * Check if a model is loaded.
@@ -170,29 +174,48 @@ uint64_t llmodel_restore_state_data(llmodel_model model, const uint8_t *src);
  * Generate a response using the model.
  * @param model A pointer to the llmodel_model instance.
  * @param prompt A string representing the input prompt.
+ * @param prompt_template A string representing the input prompt template.
  * @param prompt_callback A callback function for handling the processing of prompt.
  * @param response_callback A callback function for handling the generated response.
  * @param recalculate_callback A callback function for handling recalculation requests.
+ * @param special True if special tokens in the prompt should be processed, false otherwise.
+ * @param fake_reply A string to insert into context as the model's reply, or NULL to generate one.
  * @param ctx A pointer to the llmodel_prompt_context structure.
  */
 void llmodel_prompt(llmodel_model model, const char *prompt,
+                    const char *prompt_template,
                     llmodel_prompt_callback prompt_callback,
                     llmodel_response_callback response_callback,
                     llmodel_recalculate_callback recalculate_callback,
-                    llmodel_prompt_context *ctx);
+                    llmodel_prompt_context *ctx,
+                    bool special,
+                    const char *fake_reply);
 
 /**
  * Generate an embedding using the model.
  * NOTE: If given NULL pointers for the model or text, or an empty text, a NULL pointer will be
  * returned. Bindings should signal an error when NULL is the return value.
  * @param model A pointer to the llmodel_model instance.
- * @param text A string representing the text to generate an embedding for.
+ * @param texts A pointer to a NULL-terminated array of strings representing the texts to generate an
+ * embedding for.
  * @param embedding_size A pointer to a size_t type that will be set by the call indicating the length
  * of the returned floating point array.
+ * @param prefix The model-specific prefix representing the embedding task, without the trailing colon. NULL for no
+ * prefix.
+ * @param dimensionality The embedding dimension, for use with Matryoshka-capable models. Set to -1 to for full-size.
+ * @param token_count Return location for the number of prompt tokens processed, or NULL.
+ * @param do_mean True to average multiple embeddings if the text is longer than the model can accept, False to
+ * truncate.
+ * @param atlas Try to be fully compatible with the Atlas API. Currently, this means texts longer than 8192 tokens with
+ * long_text_mode="mean" will raise an error. Disabled by default.
+ * @param cancel_cb Cancellation callback, or NULL. See the documentation of llmodel_emb_cancel_callback.
+ * @param error Return location for a malloc()ed string that will be set on error, or NULL.
  * @return A pointer to an array of floating point values passed to the calling method which then will
- * be responsible for lifetime of this memory.
+ * be responsible for lifetime of this memory. NULL if an error occurred.
  */
-float *llmodel_embedding(llmodel_model model, const char *text, size_t *embedding_size);
+float *llmodel_embed(llmodel_model model, const char **texts, size_t *embedding_size, const char *prefix,
+                     int dimensionality, size_t *token_count, bool do_mean, bool atlas,
+                     llmodel_emb_cancel_callback cancel_cb, const char **error);
 
 /**
  * Frees the memory allocated by the llmodel_embedding function.
@@ -230,9 +253,10 @@ const char *llmodel_get_implementation_search_path();
 
 /**
  * Get a list of available GPU devices given the memory required.
+ * @param memoryRequired The minimum amount of VRAM, in bytes
  * @return A pointer to an array of llmodel_gpu_device's whose number is given by num_devices.
  */
-struct llmodel_gpu_device* llmodel_available_gpu_devices(llmodel_model model, size_t memoryRequired, int* num_devices);
+struct llmodel_gpu_device* llmodel_available_gpu_devices(size_t memoryRequired, int* num_devices);
 
 /**
  * Initializes a GPU device based on a specified string criterion.
@@ -272,6 +296,16 @@ bool llmodel_gpu_init_gpu_device_by_int(llmodel_model model, int device);
  */
 bool llmodel_has_gpu_device(llmodel_model model);
 
+/**
+ * @return The name of the llama.cpp backend currently in use. One of "cpu", "kompute", or "metal".
+ */
+const char *llmodel_model_backend_name(llmodel_model model);
+
+/**
+ * @return The name of the GPU device currently in use, or NULL for backends other than Kompute.
+ */
+const char *llmodel_model_gpu_device_name(llmodel_model model);
+
 #ifdef __cplusplus
 }
 #endif

gpt4all-backend/llmodel_shared.cpp

@@ -2,15 +2,21 @@
 
 #include <cassert>
 #include <iostream>
+#include <regex>
+#include <string>
 #include <unordered_set>
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-vulkan.h"
-#endif
-
+// TODO(cebtenzzre): replace this with llama_kv_cache_seq_shift for llamamodel (GPT-J needs this as-is)
 void LLModel::recalculateContext(PromptContext &promptCtx, std::function<bool(bool)> recalculate) {
-    size_t i = 0;
-    promptCtx.n_past = 0;
+    int n_keep = shouldAddBOS();
+    const int32_t n_discard = (promptCtx.n_ctx - n_keep) * promptCtx.contextErase;
+
+    // Erase the first percentage of context from the tokens
+    std::cerr << implementation().modelType() << ": reached the end of the context window so resizing\n";
+    promptCtx.tokens.erase(promptCtx.tokens.begin() + n_keep, promptCtx.tokens.begin() + n_keep + n_discard);
+
+    size_t i = n_keep;
+    promptCtx.n_past = n_keep;
     while (i < promptCtx.tokens.size()) {
         size_t batch_end = std::min(i + promptCtx.n_batch, promptCtx.tokens.size());
         std::vector<int32_t> batch(promptCtx.tokens.begin() + i, promptCtx.tokens.begin() + batch_end);
@@ -30,11 +36,36 @@ stop_generating:
     recalculate(false);
 }
 
+static bool parsePromptTemplate(const std::string &tmpl, std::vector<std::smatch> &placeholders, std::string &err) {
+    static const std::regex placeholderRegex(R"(%[1-2](?![0-9]))");
+
+    auto it = std::sregex_iterator(tmpl.begin(), tmpl.end(), placeholderRegex);
+    placeholders.clear();
+    placeholders.insert(placeholders.end(), it, std::sregex_iterator());
+
+    if (placeholders.size() > 2) {
+        err = "ERROR: expected at most two placeholders, got " + std::to_string(placeholders.size());
+        return false;
+    }
+    if (placeholders.size() >= 1 && placeholders[0].str() != "%1") {
+        err = "ERROR: first placeholder must be %1, got " + placeholders[0].str();
+        return false;
+    }
+    if (placeholders.size() >= 2 && placeholders[1].str() != "%2") {
+        err = "ERROR: second placeholder must be %2, got " + placeholders[1].str();
+        return false;
+    }
+    return true;
+}
+
 void LLModel::prompt(const std::string &prompt,
+                     const std::string &promptTemplate,
                      std::function<bool(int32_t)> promptCallback,
                      std::function<bool(int32_t, const std::string&)> responseCallback,
                      std::function<bool(bool)> recalculateCallback,
-                     PromptContext &promptCtx)
+                     PromptContext &promptCtx,
+                     bool special,
+                     std::string *fakeReply)
 {
     if (!isModelLoaded()) {
         std::cerr << implementation().modelType() << " ERROR: prompt won't work with an unloaded model!\n";
@@ -42,15 +73,89 @@ void LLModel::prompt(const std::string &prompt,
     }
 
     if (!supportsCompletion()) {
-        std::string errorMessage = "ERROR: this model does not support text completion or chat!\n";
+        std::string errorMessage = "ERROR: this model does not support text completion or chat!";
         responseCallback(-1, errorMessage);
-        std::cerr << implementation().modelType() << errorMessage;
+        std::cerr << implementation().modelType() << " " << errorMessage << "\n";
         return;
     }
 
-    // tokenize the prompt
-    std::vector<Token> embd_inp = tokenize(promptCtx, prompt);
+    // parse the prompt template
+    std::vector<std::smatch> placeholders;
+    {
+        std::string err;
+        if (!parsePromptTemplate(promptTemplate, placeholders, err)) {
+            responseCallback(-1, err);
+            std::cerr << err << "\n";
+            return;
+        }
+    }
 
+    auto old_n_past = promptCtx.n_past; // prepare to fake n_past for tokenize
+
+    // tokenize the user prompt
+    std::vector<Token> embd_inp;
+    if (placeholders.empty()) {
+        // this is unusual, but well-defined
+        std::cerr << __func__ << ": prompt template has no placeholder\n";
+        embd_inp = tokenize(promptCtx, promptTemplate, true);
+    } else {
+        // template: beginning of user prompt
+        const auto &phUser = placeholders[0];
+        std::string userPrefix(phUser.prefix());
+        if (!userPrefix.empty()) {
+            embd_inp = tokenize(promptCtx, userPrefix, true);
+            promptCtx.n_past += embd_inp.size();
+        }
+
+        // user input (shouldn't have special token processing)
+        auto tokens = tokenize(promptCtx, prompt, special);
+        embd_inp.insert(embd_inp.end(), tokens.begin(), tokens.end());
+        promptCtx.n_past += tokens.size();
+
+        // template: end of user prompt + start of assistant prompt
+        size_t start = phUser.position() + phUser.length();
+        size_t end = placeholders.size() >= 2 ? placeholders[1].position() : promptTemplate.length();
+        auto userToAsst = promptTemplate.substr(start, end - start);
+        if (!userToAsst.empty()) {
+            tokens = tokenize(promptCtx, userToAsst, true);
+            embd_inp.insert(embd_inp.end(), tokens.begin(), tokens.end());
+            promptCtx.n_past += tokens.size();
+        }
+    }
+
+    promptCtx.n_past = old_n_past; // restore n_past so decodePrompt can increment it
+
+    // decode the user prompt
+    decodePrompt(promptCallback, responseCallback, recalculateCallback, promptCtx, embd_inp);
+
+    // decode the assistant's reply, either generated or spoofed
+    if (fakeReply == nullptr) {
+        generateResponse(responseCallback, recalculateCallback, promptCtx);
+    } else {
+        embd_inp = tokenize(promptCtx, *fakeReply, false);
+        decodePrompt(promptCallback, responseCallback, recalculateCallback, promptCtx, embd_inp);
+    }
+
+    // decode the rest of the prompt template
+    // template: end of assistant prompt
+    std::string asstSuffix;
+    if (placeholders.size() >= 2) {
+        size_t start = placeholders[1].position() + placeholders[1].length();
+        asstSuffix = promptTemplate.substr(start);
+    } else {
+        asstSuffix = "\n\n"; // default to a blank link, good for e.g. Alpaca
+    }
+    if (!asstSuffix.empty()) {
+        embd_inp = tokenize(promptCtx, asstSuffix, true);
+        decodePrompt(promptCallback, responseCallback, recalculateCallback, promptCtx, embd_inp);
+    }
+}
+
+void LLModel::decodePrompt(std::function<bool(int32_t)> promptCallback,
+                           std::function<bool(int32_t, const std::string&)> responseCallback,
+                           std::function<bool(bool)> recalculateCallback,
+                           PromptContext &promptCtx,
+                           std::vector<Token> embd_inp) {
     // save the context size
     promptCtx.n_ctx = contextLength();
 
@@ -73,11 +178,6 @@ void LLModel::prompt(const std::string &prompt,
 
         // Check if the context has run out...
         if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
-            const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
-            // Erase the first percentage of context from the tokens...
-            std::cerr << implementation().modelType() << ": reached the end of the context window so resizing\n";
-            promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
-            promptCtx.n_past = promptCtx.tokens.size();
             recalculateContext(promptCtx, recalculateCallback);
             assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
         }
@@ -92,13 +192,17 @@ void LLModel::prompt(const std::string &prompt,
             if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
                 promptCtx.tokens.erase(promptCtx.tokens.begin());
             promptCtx.tokens.push_back(batch.at(t));
+            promptCtx.n_past += 1;
             if (!promptCallback(batch.at(t)))
                 return;
         }
-        promptCtx.n_past += batch.size();
         i = batch_end;
     }
+}
 
+void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
+                               std::function<bool(bool)> recalculateCallback,
+                               PromptContext &promptCtx) {
     std::string cachedResponse;
     std::vector<Token> cachedTokens;
     std::unordered_set<std::string> reversePrompts
@@ -112,11 +216,6 @@ void LLModel::prompt(const std::string &prompt,
 
         // Check if the context has run out...
         if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
-            const int32_t erasePoint = promptCtx.n_ctx * promptCtx.contextErase;
-            // Erase the first percentage of context from the tokens...
-            std::cerr << implementation().modelType() << ": reached the end of the context window so resizing\n";
-            promptCtx.tokens.erase(promptCtx.tokens.begin(), promptCtx.tokens.begin() + erasePoint);
-            promptCtx.n_past = promptCtx.tokens.size();
             recalculateContext(promptCtx, recalculateCallback);
             assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
         }
@@ -126,8 +225,6 @@ void LLModel::prompt(const std::string &prompt,
             return;
         }
 
-        promptCtx.n_past += 1;
-
         // display text
         for (const auto token : endTokens()) {
             if (id == token) return;
@@ -162,6 +259,7 @@ void LLModel::prompt(const std::string &prompt,
             if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
                 promptCtx.tokens.erase(promptCtx.tokens.begin());
             promptCtx.tokens.push_back(t);
+            promptCtx.n_past += 1;
             //TODO: Conversion to std::string can be avoided here...
             if (!responseCallback(t, std::string(tokenToString(t))))
                 return;
@@ -170,34 +268,31 @@ void LLModel::prompt(const std::string &prompt,
     }
 }
 
-std::vector<float> LLModel::embedding(const std::string &/*text*/)
-{
-    if (!supportsCompletion()) {
-        std::string errorMessage = "ERROR: this model does not support generating embeddings!\n";
-        std::cerr << implementation().modelType() << errorMessage;
-    }
-    return std::vector<float>();
+void LLModel::embed(
+    const std::vector<std::string> &texts, float *embeddings, std::optional<std::string> prefix, int dimensionality,
+    size_t *tokenCount, bool doMean, bool atlas, EmbedCancelCallback *cancelCb
+) {
+    (void)texts;
+    (void)embeddings;
+    (void)prefix;
+    (void)dimensionality;
+    (void)tokenCount;
+    (void)doMean;
+    (void)atlas;
+    (void)cancelCb;
+    throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
 }
 
-std::vector<LLModel::GPUDevice> LLModel::availableGPUDevices()
-{
-#if defined(GGML_USE_KOMPUTE)
-    std::vector<ggml_vk_device> vkDevices = ggml_vk_available_devices(0);
-
-    std::vector<LLModel::GPUDevice> devices;
-    for(const auto& vkDevice : vkDevices) {
-        LLModel::GPUDevice device;
-        device.index = vkDevice.index;
-        device.type = vkDevice.type;
-        device.heapSize = vkDevice.heapSize;
-        device.name = vkDevice.name;
-        device.vendor = vkDevice.vendor;
-
-        devices.push_back(device);
-    }
-
-    return devices;
-#else
-    return std::vector<LLModel::GPUDevice>();
-#endif
+void LLModel::embed(
+    const std::vector<std::string> &texts, float *embeddings, bool isRetrieval, int dimensionality, size_t *tokenCount,
+    bool doMean, bool atlas
+) {
+    (void)texts;
+    (void)embeddings;
+    (void)isRetrieval;
+    (void)dimensionality;
+    (void)tokenCount;
+    (void)doMean;
+    (void)atlas;
+    throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
 }

gpt4all-backend/llmodel_shared.h

@@ -4,49 +4,6 @@
 #include <vector>
 #include <ggml.h>
 
-#if defined(GGML_USE_KOMPUTE)
-#include "ggml-vulkan.h"
-struct llm_buffer {
-    uint8_t * addr = NULL;
-    size_t size = 0;
-    ggml_vk_memory memory;
-
-    llm_buffer() = default;
-
-    void resize(size_t size) {
-        free();
-
-        if (!ggml_vk_has_device()) {
-            this->addr = new uint8_t[size];
-            this->size = size;
-        } else {
-            this->memory = ggml_vk_allocate(size);
-            this->addr = (uint8_t*)memory.data;
-            this->size = size;
-        }
-    }
-
-    void free() {
-        if (!memory.primaryMemory) {
-            delete[] addr;
-        } else if (memory.data) {
-            ggml_vk_free_memory(memory);
-        }
-        this->addr = NULL;
-        this->size = 0;
-    }
-
-    ~llm_buffer() {
-        free();
-    }
-
-    // disable copy and move
-    llm_buffer(const llm_buffer&) = delete;
-    llm_buffer(llm_buffer&&) = delete;
-    llm_buffer& operator=(const llm_buffer&) = delete;
-    llm_buffer& operator=(llm_buffer&&) = delete;
-};
-#else
 struct llm_buffer {
     uint8_t * addr = NULL;
     size_t size = 0;
@@ -61,7 +18,6 @@ struct llm_buffer {
         delete[] addr;
     }
 };
-#endif
 
 struct llm_kv_cache {
     struct ggml_tensor * k;
@@ -80,7 +36,6 @@ struct llm_kv_cache {
     }
 };
 
-#if LLAMA_DATE >= 230519
 inline void ggml_graph_compute_g4a(llm_buffer& buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
     if (plan.work_size > 0) {
@@ -89,4 +44,3 @@ inline void ggml_graph_compute_g4a(llm_buffer& buf, ggml_cgraph * graph, int n_t
     }
     ggml_graph_compute(graph, &plan);
 }
-#endif

gpt4all-backend/mpt.cpp

@@ -1,893 +0,0 @@
-#define MPT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#include "mpt_impl.h"
-
-#include "utils.h"
-#include "llmodel_shared.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <map>
-#include <random>
-#include <string>
-#include <vector>
-#include <iostream>
-#if defined(_WIN32) && defined(_MSC_VER)
-    #define WIN32_LEAN_AND_MEAN
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-    #include <windows.h>
-    #include <io.h>
-    #include <stdio.h>
-#else
-    #include <unistd.h>
-#endif
-#include <sstream>
-#include <thread>
-#include <unordered_set>
-#include <regex>
-#include <ggml.h>
-
-
-namespace {
-const char *modelType_ = "MPT";
-}
-
-// default hparams (MPT 7B)
-struct mpt_hparams {
-    int32_t n_vocab      = 50432;
-    int32_t n_ctx        = 2048;
-    int32_t n_embd       = 4096;
-    int32_t n_head       = 32;
-    int32_t n_layer      = 32;
-    float alibi_bias_max = 8;
-    float clip_qkv       = 0;
-    int32_t expand       = 4;
-    int32_t f16          = 1;
-};
-
-struct mpt_layer {
-    // normalization
-    struct ggml_tensor * norm_1_w;
-    struct ggml_tensor * norm_2_w;
-
-    // attention
-    struct ggml_tensor * attn_Wqkv_w;
-    struct ggml_tensor * attn_out_proj_w;
-
-    // ff
-    struct ggml_tensor * ffn_up_proj_w;
-    struct ggml_tensor * ffn_down_proj_w;
-};
-
-struct mpt_model {
-    mpt_hparams hparams;
-
-    // normalization
-    struct ggml_tensor * norm_f_w;
-
-    struct ggml_tensor * wte; // position embedding
-
-    // mpt does weight tying
-
-    std::vector<mpt_layer> layers;
-
-    struct llm_kv_cache kv_self;
-    struct ggml_context * ctx;
-    std::map<std::string, struct ggml_tensor *> tensors;
-
-
-    llm_buffer eval_buf;
-    llm_buffer scr0_buf;
-    llm_buffer scr1_buf;
-
-    ~mpt_model() {
-        if (ctx) {
-            ggml_free(ctx);
-        }
-    }
-};
-
-static bool kv_cache_init(
-        const struct mpt_hparams & hparams,
-             struct llm_kv_cache & cache,
-                         ggml_type   wtype,
-                               int   n_ctx) {
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-
-    const int64_t n_mem      = (int64_t)n_layer*n_ctx;
-    const int64_t n_elements = n_embd*n_mem;
-
-    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2_MiB);
-
-    struct ggml_init_params params;
-    params.mem_size   = cache.buf.size;
-    params.mem_buffer = cache.buf.addr;
-    params.no_alloc   = false;
-
-    cache.ctx = ggml_init(params);
-
-    if (!cache.ctx) {
-        fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
-        return false;
-    }
-
-    cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
-    cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
-
-    return true;
-}
-
-// load the model's weights from a stream. if mem_req ptr is passed the model is
-// only partially parsed to estimate required memory
-bool mpt_model_load(const std::string &fname, std::istream &fin, mpt_model & model, gpt_vocab & vocab, size_t * mem_req) {
-    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
-    if (mem_req != nullptr) {
-        *mem_req = 0;
-    }
-
-    // verify magic
-    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        if (magic != 0x67676d6d) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
-        }
-    }
-
-    // load hparams
-    {
-        auto & hparams = model.hparams;
-
-        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
-        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
-        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
-        fin.read((char *) &hparams.alibi_bias_max,  sizeof(hparams.alibi_bias_max));
-        fin.read((char *) &hparams.clip_qkv,  sizeof(hparams.clip_qkv));
-        fin.read((char *) &hparams.f16,   sizeof(hparams.f16));
-
-        printf("%s: n_vocab        = %d\n", __func__, hparams.n_vocab);
-        printf("%s: n_ctx          = %d\n", __func__, hparams.n_ctx);
-        printf("%s: n_embd         = %d\n", __func__, hparams.n_embd);
-        printf("%s: n_head         = %d\n", __func__, hparams.n_head);
-        printf("%s: n_layer        = %d\n", __func__, hparams.n_layer);
-        printf("%s: alibi_bias_max = %f\n", __func__, hparams.alibi_bias_max);
-        printf("%s: clip_qkv       = %f\n", __func__, hparams.clip_qkv);
-        printf("%s: ftype          = %d\n", __func__, hparams.f16);
-    }
-
-    // load vocab
-    {
-        int32_t n_vocab = model.hparams.n_vocab;
-        fin.read((char *) &n_vocab, sizeof(n_vocab));
-
-        if (n_vocab != model.hparams.n_vocab) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
-                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
-            return false;
-        }
-
-        std::string word;
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
-            bool special = false;
-            if (len & (1<<31)) {
-                len = len &~ (1<<31);
-                special = true;
-            }
-
-            if (len > 0) {
-                word.resize(len);
-                fin.read((char *) word.data(), len);
-                vocab.token_to_id[word] = i;
-                vocab.id_to_token[i] = word;
-            }
-
-            if(special) {
-                vocab.add_special_token(word);
-            }
-        }
-    }
-
-    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
-    // in order to save memory and also to speed up the computation
-    ggml_type wtype = GGML_TYPE_COUNT;
-    switch (model.hparams.f16) {
-        case 0: wtype = GGML_TYPE_F32;  break;
-        case 1: wtype = GGML_TYPE_F16;  break;
-        case 2: wtype = GGML_TYPE_Q4_0; break;
-        case 3: wtype = GGML_TYPE_Q4_1; break;
-        case 5: wtype = GGML_TYPE_Q4_2; break;
-        default:
-                {
-                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
-                            __func__, fname.c_str(), model.hparams.f16);
-                    return false;
-                }
-    }
-
-    auto & ctx = model.ctx;
-
-    size_t ctx_size = 0;
-
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-        const int expand  = hparams.expand;
-
-
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_w
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(GGML_TYPE_F32); // wte
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // norm_1_w
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // norm_2_w
-
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype)); // attn_Wqkv_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // attn_out_proj_w
-
-        ctx_size += n_layer*(expand*n_embd*n_embd*ggml_type_sizef(wtype));  // ffn_up_proj_w
-        ctx_size += n_layer*(expand*n_embd*n_embd*ggml_type_sizef(wtype)); // ffn_down_proj_w
-
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_v
-
-        // TODO probably less now?
-        ctx_size += (5 + 10*n_layer)*256; // object overhead
-
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
-
-    if (mem_req != nullptr) {
-        *mem_req += ctx_size;
-        const int n_embd  = model.hparams.n_embd;
-        const int n_layer = model.hparams.n_layer;
-
-        const int64_t n_mem      = (int64_t)n_layer*model.hparams.n_ctx;
-        const int64_t n_elements = n_embd*n_mem;
-
-        *mem_req += (2u*n_elements*ggml_type_size(wtype) + 2_MiB);
-        return false;
-    }
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-            .no_alloc   = false,
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
-    // prepare memory for the weights
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_vocab = hparams.n_vocab;
-        const int expand  = hparams.expand;
-
-        model.layers.resize(n_layer);
-
-        model.wte    = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-        model.norm_f_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-        // map by name
-        model.tensors["transformer.wte.weight"] = model.wte;
-        model.tensors["transformer.norm_f.weight"] = model.norm_f_w;
-
-        for (int i = 0; i < n_layer; ++i) {
-            auto & layer = model.layers[i];
-
-            layer.norm_1_w        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.norm_2_w        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-            layer.attn_Wqkv_w     = ggml_new_tensor_2d(ctx, wtype,        n_embd, n_embd * 3);
-            layer.attn_out_proj_w = ggml_new_tensor_2d(ctx, wtype,        n_embd, n_embd);
-            layer.ffn_up_proj_w   = ggml_new_tensor_2d(ctx, wtype,        n_embd, expand*n_embd);
-            layer.ffn_down_proj_w = ggml_new_tensor_2d(ctx, wtype, expand*n_embd, n_embd);
-
-            // map by name
-            model.tensors["transformer.blocks." + std::to_string(i) + ".norm_1.weight"]        = layer.norm_1_w;
-            model.tensors["transformer.blocks." + std::to_string(i) + ".norm_2.weight"]        = layer.norm_2_w;
-            model.tensors["transformer.blocks." + std::to_string(i) + ".attn.Wqkv.weight"]     = layer.attn_Wqkv_w;
-            model.tensors["transformer.blocks." + std::to_string(i) + ".attn.out_proj.weight"] = layer.attn_out_proj_w;
-
-            model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.up_proj.weight"]   = layer.ffn_up_proj_w;
-            model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.down_proj.weight"] = layer.ffn_down_proj_w;
-        }
-    }
-
-    // key + value memory
-    {
-        const auto & hparams = model.hparams;
-        if (!kv_cache_init(hparams, model.kv_self, GGML_TYPE_F16, model.hparams.n_ctx)) {
-            fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
-            ggml_free(ctx);
-            return false;
-        }
-
-        const size_t memory_size = ggml_nbytes(model.kv_self.k) + ggml_nbytes(model.kv_self.v);
-        printf("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
-    }
-
-    // load weights
-    {
-        int n_tensors = 0;
-        size_t total_size = 0;
-
-        printf("%s: ", __func__);
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            int32_t ttype;
-
-            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                return false;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                return false;
-            }
-
-            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
-                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            // for debugging
-            if (0) {
-                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
-            }
-
-            const size_t bpe = ggml_type_size(ggml_type(ttype));
-
-            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
-                return false;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ttype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
-            total_size += ggml_nbytes(tensor);
-            if (++n_tensors % 8 == 0) {
-                printf(".");
-                fflush(stdout);
-            }
-        }
-
-        printf(" done\n");
-
-        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
-    }
-
-    model.scr0_buf.resize(256u * 1024 * 1024);
-    model.scr1_buf.resize(256u * 1024 * 1024);
-
-    return true;
-}
-
-// load the model's weights from a file path
-bool mpt_model_load(const std::string & fname, mpt_model & model, gpt_vocab & vocab) {
-
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
-    bool loaded = mpt_model_load(fname, fin, model, vocab, nullptr);
-    fin.close();
-    return loaded;
-}
-
-bool mpt_eval(
-        mpt_model & model,
-        const int n_threads,
-        const int n_past,
-        const std::vector<int>           & embd_inp,
-              std::vector<float>         & embd_w,
-              size_t                     & mem_per_token) {
-    const int N = embd_inp.size();
-
-    const auto & hparams = model.hparams;
-
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_ctx   = hparams.n_ctx;
-    const int n_head  = hparams.n_head;
-    const int n_vocab = hparams.n_vocab;
-
-    const size_t init_buf_size = 1024_MiB;
-    if (!model.eval_buf.addr || model.eval_buf.size < init_buf_size)
-        model.eval_buf.resize(init_buf_size);
-
-    if (mem_per_token > 0 && mem_per_token*N > model.eval_buf.size) {
-        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
-        // printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, model.buf.size, buf_size_new);
-
-        // reallocate
-        model.eval_buf.resize(buf_size_new);
-        if (model.eval_buf.addr == nullptr) {
-            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, model.eval_buf.size);
-            return false;
-        }
-    }
-
-    struct ggml_init_params params = {
-        .mem_size   = model.eval_buf.size,
-        .mem_buffer = model.eval_buf.addr,
-        .no_alloc = false
-    };
-
-    struct ggml_context * ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
-
-    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
-
-    // wte
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
-
-    for (int il = 0; il < n_layer; ++il) {
-        ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
-
-        struct ggml_tensor * inpSA = inpL;
-        struct ggml_tensor * cur = inpSA;
-        // self-attention
-        {
-
-            // norm1
-            cur = ggml_norm(ctx0, cur);
-            cur = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].norm_1_w, cur),
-                    cur);
-            // compute QKV
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].attn_Wqkv_w,
-                    cur);
-
-            // TODO: clip_qkv
-            struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*ggml_element_size(cur)*n_embd));
-            struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*ggml_element_size(cur)*n_embd));
-            struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*ggml_element_size(cur)*n_embd));
-
-            // TODO: qk_ln? (seems to be False in MPT-7B configs)
-            {
-                Vcur = ggml_transpose(ctx0, Vcur);
-
-                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_2d(ctx0, model.kv_self.v, N, n_embd,
-                                        (   n_ctx)*ggml_element_size(model.kv_self.v),
-                                        (il*n_ctx)*ggml_element_size(model.kv_self.v)*n_embd + n_past*ggml_element_size(model.kv_self.v));
-
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
-            }
-            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, N),
-                        0, 2, 1, 3);
-
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        0, 2, 1, 3);
-
-            // K * Q
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
-
-
-            // Alibi
-            struct ggml_tensor * KQ_scaled_biased = ggml_alibi(ctx0, ggml_cont(ctx0, KQ_scaled), n_past, n_head);
-
-            // KQ_masked = mask_past(KQ_scaled)
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_biased, n_past);
-
-            // KQ = soft_max(KQ_masked)
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
-
-            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
-            struct ggml_tensor * V =
-                ggml_view_3d(ctx0, model.kv_self.v,
-                        n_past + N, n_embd/n_head, n_head,
-                        n_ctx*ggml_element_size(model.kv_self.v),
-                        n_ctx*ggml_element_size(model.kv_self.v)*n_embd/n_head,
-                        il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd);
-
-            // KQV = transpose(V) * KQ_soft_max
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-
-            // projection (no bias)
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].attn_out_proj_w,
-                    cur);
-        }
-
-        ggml_set_scratch(ctx0, {0, model.scr1_buf.size, model.scr1_buf.addr, });
-        // residual
-        struct ggml_tensor * resSA = ggml_add(ctx0, cur, inpSA);
-        // feed-forward network
-        {
-            cur = resSA;
-            // norm2
-            cur = ggml_norm(ctx0, cur);
-            cur = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].norm_2_w, cur),
-                    cur);
-            // ffn
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].ffn_up_proj_w,
-                    cur);
-            cur = ggml_gelu(ctx0, cur);
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].ffn_down_proj_w,
-                    cur);
-
-        }
-
-        // self-attention + FF
-        inpL = ggml_add(ctx0, cur, resSA);
-    }
-    ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
-
-    struct ggml_tensor * out = inpL;
-    // -> logits
-    {
-        out = ggml_norm(ctx0, out);
-        out = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.norm_f_w, out),
-                    out);
-        ggml_set_scratch(ctx0, { 0, 0, nullptr, });
-        out = ggml_mul_mat(ctx0, model.wte, out);
-    }
-
-
-    // run the computation
-    ggml_build_forward_expand(&gf, out);
-    ggml_graph_compute       (ctx0, &gf);
-
-
-    // return result for just the last token
-    embd_w.resize(n_vocab);
-    memcpy(embd_w.data(), (float *) ggml_get_data(out) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
-
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
-
-    ggml_free(ctx0);
-
-    return true;
-}
-
-
-#define MPT_MAX_RNG_STATE 64*1024
-
-size_t mpt_get_state_size(const mpt_model &model)
-{
-    // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
-    // for reference, std::mt19937(1337) serializes to 6701 bytes.
-    const size_t s_rng_size        = sizeof(size_t);
-    const size_t s_rng             = MPT_MAX_RNG_STATE;
-    const size_t s_kv_size         = sizeof(size_t);
-    const size_t s_kv_ntok         = sizeof(int);
-    const size_t s_kv              = model.kv_self.buf.size;
-    const size_t s_total = (
-        + s_rng_size
-        + s_rng
-        + s_kv_size
-        + s_kv_ntok
-        + s_kv
-    );
-    fflush(stdout);
-    return s_total;
-}
-
-size_t mpt_copy_state_data(const mpt_model &model, const std::mt19937 &rng, uint8_t *dest)
-{
-    uint8_t * out = dest;
-    fflush(stdout);
-    // copy rng
-    {
-        std::stringstream rng_ss;
-        rng_ss << rng;
-
-        const size_t rng_size = rng_ss.str().size();
-        char rng_buf[MPT_MAX_RNG_STATE];
-
-        memset(&rng_buf[0], 0, MPT_MAX_RNG_STATE);
-        memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
-
-        memcpy(out, &rng_size,   sizeof(rng_size));   out += sizeof(rng_size);
-        memcpy(out, &rng_buf[0], MPT_MAX_RNG_STATE); out += MPT_MAX_RNG_STATE;
-    }
-
-    // copy kv cache
-    {
-        const size_t kv_size = model.kv_self.buf.size;
-        const int    kv_ntok = model.kv_self.n;
-
-        memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size);
-        memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok);
-
-        if (kv_size) {
-            memcpy(out, model.kv_self.buf.addr, kv_size); out += kv_size;
-        }
-    }
-
-    const size_t written  = out - dest;
-    assert(written == mpt_get_state_size(model));
-    fflush(stdout);
-    return written;
-}
-
-size_t mpt_set_state_data(mpt_model *model, std::mt19937 *rng, const uint8_t *src)
-{
-    const uint8_t * in = src;
-
-    // set rng
-    {
-        size_t rng_size;
-        char   rng_buf[MPT_MAX_RNG_STATE];
-
-        memcpy(&rng_size,   in, sizeof(rng_size));    in += sizeof(rng_size);
-        memcpy(&rng_buf[0], in, MPT_MAX_RNG_STATE); in += MPT_MAX_RNG_STATE;
-
-        std::stringstream rng_ss;
-        rng_ss.str(std::string(&rng_buf[0], rng_size));
-        rng_ss >> *rng;
-
-        assert(rng_ss.fail() == false);
-    }
-
-    // set kv cache
-    {
-        size_t kv_size;
-        int kv_ntok;
-
-        memcpy(&kv_size, in, sizeof(kv_size)); in += sizeof(kv_size);
-        memcpy(&kv_ntok, in, sizeof(kv_ntok)); in += sizeof(kv_ntok);
-
-        if (kv_size) {
-            assert(model->kv_self.buf.size == kv_size);
-
-            void * k_data = model->kv_self.k->data; // remember data pointers
-            void * v_data = model->kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
-
-            memcpy(model->kv_self.buf.addr, in, kv_size); in += kv_size;
-
-            model->kv_self.k->data = k_data; // restore correct data pointers
-            model->kv_self.v->data = v_data;
-
-        }
-
-        model->kv_self.n = kv_ntok;
-    }
-
-    const size_t nread    = in - src;
-    assert(nread == mpt_get_state_size(*model));
-    fflush(stdout);
-    return nread;
-}
-
-struct MPTPrivate {
-    const std::string modelPath;
-    bool modelLoaded;
-    gpt_vocab vocab;
-    mpt_model *model = nullptr;
-    int64_t n_threads = 0;
-    size_t mem_per_token = 0;
-    std::mt19937 rng;
-    bool has_im_end = false;
-};
-
-MPT::MPT()
-    : d_ptr(new MPTPrivate) {
-    d_ptr->model = new mpt_model;
-    d_ptr->model->ctx = nullptr;
-    d_ptr->modelLoaded = false;
-}
-
-size_t MPT::requiredMem(const std::string &modelPath) {
-    mpt_model dummy_model;
-    gpt_vocab dummy_vocab;
-    size_t mem_req;
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-    mpt_model_load(modelPath, fin, dummy_model, dummy_vocab, &mem_req);
-    return mem_req;
-}
-
-bool MPT::loadModel(const std::string &modelPath) {
-    std::mt19937 rng(time(NULL));
-    d_ptr->rng = rng;
-
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-
-    // load the model
-    if (!mpt_model_load(modelPath, fin, *d_ptr->model, d_ptr->vocab, nullptr)) {
-        std::cerr << "MPT ERROR: failed to load model from " <<  modelPath;
-        return false;
-    }
-
-    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    d_ptr->modelLoaded = true;
-    d_ptr->has_im_end = d_ptr->vocab.token_to_id.find("<|im_end|>") != d_ptr->vocab.token_to_id.end();
-    fflush(stdout);
-    return true;
-}
-
-void MPT::setThreadCount(int32_t n_threads) {
-    d_ptr->n_threads = n_threads;
-}
-
-int32_t MPT::threadCount() const
-{
-    return d_ptr->n_threads;
-}
-
-MPT::~MPT()
-{
-    delete d_ptr->model;
-}
-
-bool MPT::isModelLoaded() const
-{
-    return d_ptr->modelLoaded;
-}
-
-size_t MPT::stateSize() const
-{
-    return mpt_get_state_size(*d_ptr->model);
-}
-
-size_t MPT::saveState(uint8_t *dest) const
-{
-    return mpt_copy_state_data(*d_ptr->model, d_ptr->rng, dest);
-}
-
-size_t MPT::restoreState(const uint8_t *src)
-{
-    return mpt_set_state_data(d_ptr->model, &d_ptr->rng, src);
-}
-
-std::vector<LLModel::Token> MPT::tokenize(PromptContext &, const std::string &str) const
-{
-    return ::gpt_tokenize(d_ptr->vocab, str);
-}
-
-std::string MPT::tokenToString(Token id) const
-{
-    return d_ptr->vocab.id_to_token[id];
-}
-
-LLModel::Token MPT::sampleToken(PromptContext &promptCtx) const
-{
-    const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
-    return gpt_sample_top_k_top_p(d_ptr->model->hparams.n_vocab,
-        promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
-        n_prev_toks,
-        promptCtx.logits,
-        promptCtx.top_k, promptCtx.top_p, promptCtx.temp,
-        promptCtx.repeat_penalty,
-        d_ptr->rng);
-}
-
-bool MPT::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
-{
-    // determine the required inference memory per token:
-    static bool initialized = false;
-    if (!initialized) {
-        mpt_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, ctx.logits,
-            d_ptr->mem_per_token);
-        initialized = true;
-    }
-
-    return mpt_eval(*d_ptr->model, d_ptr->n_threads, ctx.n_past, tokens, ctx.logits, d_ptr->mem_per_token);
-}
-
-int32_t MPT::contextLength() const
-{
-    return d_ptr->model->hparams.n_ctx;
-}
-
-const std::vector<LLModel::Token> &MPT::endTokens() const
-{
-    static const std::vector<LLModel::Token> fres = {0, d_ptr->vocab.token_to_id["<|im_end|>"]};
-    return fres;
-}
-
-#if defined(_WIN32)
-#define DLL_EXPORT __declspec(dllexport)
-#else
-#define DLL_EXPORT __attribute__ ((visibility ("default")))
-#endif
-
-extern "C" {
-DLL_EXPORT bool is_g4a_backend_model_implementation() {
-    return true;
-}
-
-DLL_EXPORT const char *get_model_type() {
-    return modelType_;
-}
-
-DLL_EXPORT const char *get_build_variant() {
-    return GGML_BUILD_VARIANT;
-}
-
-DLL_EXPORT bool magic_match(std::istream& f) {
-    uint32_t magic = 0;
-    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
-    return magic == 0x67676d6d;
-}
-
-DLL_EXPORT LLModel *construct() {
-    return new MPT;
-}
-}

gpt4all-backend/mpt_impl.h

@@ -1,41 +0,0 @@
-#ifndef MPT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#error This file is NOT meant to be included outside of mpt.cpp. Doing so is DANGEROUS. Be sure to know what you are doing before proceeding to #define MPT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#endif
-#ifndef MPT_H
-#define MPT_H
-
-#include <string>
-#include <functional>
-#include <vector>
-#include "llmodel.h"
-
-struct MPTPrivate;
-class MPT : public LLModel {
-public:
-    MPT();
-    ~MPT();
-
-    bool supportsEmbedding() const override { return false; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
-    bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
-    size_t stateSize() const override;
-    size_t saveState(uint8_t *dest) const override;
-    size_t restoreState(const uint8_t *src) override;
-    void setThreadCount(int32_t n_threads) override;
-    int32_t threadCount() const override;
-
-private:
-    MPTPrivate *d_ptr;
-
-protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    std::string tokenToString(Token) const override;
-    Token sampleToken(PromptContext &ctx) const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
-};
-
-#endif // MPT_H

gpt4all-backend/replit_impl.h

@@ -1,44 +0,0 @@
-#ifndef REPLIT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#error This file is NOT meant to be included outside of replit.cpp. Doing so is DANGEROUS. Be sure to know what you are doing before proceeding to #define REPLIT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#endif
-#ifndef REPLIT_H
-#define REPLIT_H
-
-#include <string>
-#include <functional>
-#include <vector>
-#include "llmodel.h"
-
-#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
-
-struct ReplitPrivate;
-class Replit : public LLModel {
-public:
-    Replit();
-    ~Replit();
-
-    bool supportsEmbedding() const override { return false; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
-    bool isModelLoaded() const override;
-    size_t requiredMem(const std::string & modelPath) override;
-    size_t stateSize() const override;
-    size_t saveState(uint8_t *dest) const override;
-    size_t restoreState(const uint8_t *src) override;
-    void setThreadCount(int32_t n_threads) override;
-    int32_t threadCount() const override;
-    bool usingGPUDevice() override;
-
-private:
-    ReplitPrivate *d_ptr;
-
-protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    std::string tokenToString(Token) const override;
-    Token sampleToken(PromptContext &ctx) const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
-};
-
-#endif // REPLIT_H

gpt4all-backend/scripts/convert_bert_hf_to_ggml.py

@@ -1,102 +0,0 @@
-import sys
-import struct
-import json
-import torch
-import numpy as np
-
-from transformers import AutoModel, AutoTokenizer
-
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-    sys.exit(1)
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-fname_out = sys.argv[1] + "/ggml-model.bin"
-
-with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-    encoder = json.load(f)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-with open(dir_model + "/vocab.txt", "r", encoding="utf-8") as f:
-    vocab = f.readlines()
-# possible data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-#
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-        sys.exit(1)
-    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
-
-
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-model = AutoModel.from_pretrained(dir_model, low_cpu_mem_usage=True)
-print (model)
-
-print(tokenizer.encode('I believe the meaning of life is'))
-
-list_vars = model.state_dict()
-for name in list_vars.keys():
-    print(name, list_vars[name].shape, list_vars[name].dtype)
-
-fout = open(fname_out, "wb")
-
-print(hparams)
-
-fout.write(struct.pack("i", 0x62657274)) # magic: ggml in hex
-fout.write(struct.pack("i", hparams["vocab_size"]))
-fout.write(struct.pack("i", hparams["max_position_embeddings"]))
-fout.write(struct.pack("i", hparams["hidden_size"]))
-fout.write(struct.pack("i", hparams["intermediate_size"]))
-fout.write(struct.pack("i", hparams["num_attention_heads"]))
-fout.write(struct.pack("i", hparams["num_hidden_layers"]))
-fout.write(struct.pack("i", ftype))
-
-for i in range(hparams["vocab_size"]):
-    text = vocab[i][:-1] # strips newline at the end
-    #print(f"{i}:{text}")
-    data = bytes(text, 'utf-8')
-    fout.write(struct.pack("i", len(data)))
-    fout.write(data)
-
-for name in list_vars.keys():
-    data = list_vars[name].squeeze().numpy()
-    if name in ['embeddings.position_ids', 'pooler.dense.weight', 'pooler.dense.bias']:
-        continue
-    print("Processing variable: " + name + " with shape: ", data.shape)
-
-    n_dims = len(data.shape);
-
-    # ftype == 0 -> float32, ftype == 1 -> float16
-    if ftype == 1 and name[-7:] == ".weight" and n_dims == 2:
-            print("  Converting to float16")
-            data = data.astype(np.float16)
-            l_type = 1
-    else:
-        l_type = 0
-
-    # header
-    str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), l_type))
-    for i in range(n_dims):
-        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str);
-
-    # data
-    data.tofile(fout)
-
-fout.close()
-
-print("Done. Output file: " + fname_out)
-print("")

gpt4all-backend/scripts/convert_bert_hf_to_gguf.py

@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import json
+import struct
+import sys
+from pathlib import Path
+
+import gguf
+import numpy as np
+from transformers import AutoConfig, AutoModel, AutoTokenizer
+
+
+if not 2 <= len(sys.argv) < 4:
+    print("Usage: {} dir-model [ftype]\n".format(Path(__file__).name))
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = Path(sys.argv[1])
+
+with open(dir_model / "vocab.txt", encoding="utf-8") as f:
+    vocab = f.readlines()
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+
+fname_out = dir_model / ("ggml-model-" + ftype_str[ftype] + ".gguf")
+
+
+ARCH = gguf.MODEL_ARCH.BERT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+config = AutoConfig.from_pretrained(dir_model)
+
+block_count = config.num_hidden_layers
+gguf_writer.add_name("BERT")
+gguf_writer.add_context_length(config.max_position_embeddings)
+gguf_writer.add_embedding_length(config.hidden_size)
+gguf_writer.add_feed_forward_length(config.intermediate_size)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(config.num_attention_heads)
+gguf_writer.add_file_type(ftype)
+
+print("gguf: get tokenizer metadata")
+
+try:
+    with open(dir_model / "tokenizer.json", encoding="utf-8") as f:
+        tokenizer_json = json.load(f)
+except FileNotFoundError as e:
+    print(f'Error: Missing {e.filename!r}', file=sys.stderr)
+    sys.exit(1)
+
+print("gguf: get wordpiece tokenizer vocab")
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+print(tokenizer.encode('I believe the meaning of life is'))
+
+tokens: list[bytearray] = []
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+for i in range(config.vocab_size):
+    try:
+        text = reverse_vocab[i]
+    except KeyError:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+
+gguf_writer.add_tokenizer_model("bert")  # wordpiece
+gguf_writer.add_token_list(tokens)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab.add_to_gguf(gguf_writer)
+
+print("gguf: get tensor metadata")
+
+model = AutoModel.from_pretrained(dir_model, config=config, low_cpu_mem_usage=True)
+print(model)
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    if name in ['embeddings.position_ids', 'pooler.dense.weight', 'pooler.dense.bias']:
+        continue
+    print("Processing variable:", name, "with shape:", data.shape)
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    if ftype == 1 and name[-7:] == ".weight" and n_dims == 2:
+        print("  Converting to float16")
+        data = data.astype(np.float16)
+        l_type = 1
+    else:
+        l_type = 0
+
+    # map tensor names
+    new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+    if new_name is None:
+        print("Can not map tensor '" + name + "'")
+        sys.exit()
+
+    gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+print("gguf: write tensors")
+gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print()

gpt4all-backend/scripts/convert_falcon_hf_to_ggml.py

@@ -1,143 +0,0 @@
-# Based on: https://github.com/KerfuffleV2/ggml-falcon/blob/feat-improve-falcon-convert-hf/examples/falcon/convert-hf-to-ggml.py
-# Convert Hugging Face fine-tuned bloom-like models to ggml format
-#
-# Usage:
-#
-#   python3 convert_falcon_hf_to_ggml.py model_directory output_directory [use-f32]
-#
-# This script is similar to "convert-pt-to-ggml.py"
-#
-
-import io
-import os
-import sys
-import struct
-import json
-import code
-import torch
-import numpy as np
-import gc
-
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
-
-# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-if len(sys.argv) < 3:
-    print("INFO: GGML V1 files produced are meant to be finalized through examples/falcon_quantize which will bring them to latest version and precision of choice");
-    print("Usage: python convert_falcon_hf_to_ggml.py model_directory output_directory [use-f32]")
-    print("  model_directory: name of the directory and model you convert (it should be a subdirectory)")
-    print("  output-directory: directory where the output file will be written")
-    print("  use-f32:    if present, use float32 instead of float16 (f32 is recommended)")
-    sys.exit(1)
-
-# num_parts = int(sys.argv[1])
-dir_model = sys.argv[1] # name and dir of model
-dir_out = sys.argv[2]   # output directory
-
-# make sure the output directory exists
-os.makedirs(dir_out, exist_ok=True)
-
-
-# possible data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-#
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-ftype = 1
-if len(sys.argv) > 3:
-    ftype = 0
-
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-# print(tokenizer)
-config = AutoConfig.from_pretrained(dir_model, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(dir_model, trust_remote_code=True, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True)
-hparams = config.to_dict()
-
-n_head = hparams["n_head"]
-n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
-head_dim = hparams["hidden_size"] // n_head
-print("* Loading model from: ", dir_model)
-
-fname_out = dir_out + f"/ggml-model-{dir_model.split('/')[-1]}-{ftype_str[ftype]}.bin"
-fout = open(fname_out, "wb")
-fout.write(struct.pack("i", 0x67676a74)) # magic: ggmf in hex (version 1) - possibly change to ggfc ?
-fout.write(struct.pack("i", 1)) # version
-fout.write(struct.pack("i", hparams["vocab_size"]))
-fout.write(struct.pack("i", hparams["hidden_size"]))
-fout.write(struct.pack("i", n_head))
-fout.write(struct.pack("i", n_head_kv))
-fout.write(struct.pack("i", hparams["n_layer"]))
-fout.write(struct.pack("i", 40 if "n_head_kv" in hparams else 7)) # obsolete field that breaks ggml compatibility - todo again remove one day
-fout.write(struct.pack("i", ftype))
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-byte_encoder = bytes_to_unicode()
-byte_decoder = {v:k for k, v in byte_encoder.items()}
-
-for i in range(hparams["vocab_size"]):
-    text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-    fout.write(struct.pack("i", len(text)))
-    fout.write(text)
-    fout.write(struct.pack("f", 0.0)) # falcon uses bpe on RefinedWeb - no probability scores used
-
-model = model.state_dict()
-for name in model.keys():
-    src = name
-    # The original query_key_value tensor contains n_head_kv "kv groups",
-    # each consisting of n_head/n_head_kv query weights followed by one key
-    # and one value weight (shared by all query heads in the kv group).
-    # This layout makes it a big pain to work with in GGML.
-    # So we rearrange them here,, so that we have n_head query weights
-    # followed by n_head_kv key weights followed by n_head_kv value weights,
-    # in contiguous fashion.
-
-    if "query_key_value" in src:
-        qkv = model[src].view(
-            n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
-
-        q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
-        k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
-        v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
-
-        model[src] = torch.cat((q,k,v)).reshape_as(model[src])
-    data = model[src].squeeze()
-    n_dims = len(data.shape)
-    # default type is fp32
-    ftype_cur = 1 if ftype == 1 and n_dims > 1 else 0
-    data = data.to(dtype = torch.float16 if ftype_cur == 1 else torch.float32).numpy()
-    print(f'  |', name, data.shape, '->', data.dtype)
-    # header
-    str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
-    for i in range(n_dims):
-        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str)
-
-    # data
-    data.tofile(fout)
-
-fout.close()
-
-print("Done. Output file: " + fname_out)
-print("")

gpt4all-backend/scripts/convert_gptj_to_gguf.py

@@ -0,0 +1,165 @@
+#!/usr/bin/env python3
+# Convert GPT-J-6B h5 transformer model to ggml format
+#
+# Load the model using GPTJForCausalLM.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "ftype" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+from __future__ import annotations
+
+import sys
+import struct
+import json
+from pathlib import Path
+
+import gguf
+import numpy as np
+from transformers import AutoConfig, AutoTokenizer, GPTJForCausalLM
+from transformers.models.gpt2 import tokenization_gpt2
+
+
+if not 2 <= len(sys.argv) < 4:
+    print("Usage: python {} dir-model [ftype]\n".format(Path(__file__).name))
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = Path(sys.argv[1])
+fname_out = dir_model / "ggml-model.gguf"
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+
+fname_out = dir_model / ("ggml-model-" + ftype_str[ftype] + ".gguf")
+
+
+ARCH = gguf.MODEL_ARCH.GPTJ
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+config = AutoConfig.from_pretrained(dir_model)
+
+block_count = config.n_layer
+gguf_writer.add_name("GPT-J")
+gguf_writer.add_context_length(config.n_positions)
+gguf_writer.add_embedding_length(config.n_embd)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(4 * config.n_embd)
+gguf_writer.add_head_count(config.n_head)
+gguf_writer.add_rope_dimension_count(config.rotary_dim)
+gguf_writer.add_layer_norm_eps(config.layer_norm_epsilon)
+gguf_writer.add_file_type(ftype)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = tokenization_gpt2.bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+tokens: list[bytearray] = []
+
+for i in range(config.vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[c])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+
+
+gguf_writer.add_tokenizer_model("gpt2")
+gguf_writer.add_token_list(tokens)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab.add_to_gguf(gguf_writer)
+
+print("gguf: get tensor metadata")
+
+model = GPTJForCausalLM.from_pretrained(dir_model, config=config, low_cpu_mem_usage=True)
+#print (model)
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable:", name, "with shape:", data.shape)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable:", name)
+        continue
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if ftype == 1 and name[-7:] == ".weight" and n_dims == 2:
+        print("  Converting to float16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1 or data.dtype != np.float32:
+        print("  Converting to float32")
+        data = data.astype(np.float32)
+        ftype_cur = 0
+
+    # map tensor names
+    new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+    if new_name is None:
+        print("Can not map tensor '" + name + "'")
+        sys.exit()
+
+    gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+print("gguf: write tensors")
+gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print()

gpt4all-backend/scripts/convert_mpt_hf_to_ggml.py

@@ -1,145 +0,0 @@
-# Convert Hugging Face fine-tuned bloom-like models to ggml format
-#
-# Usage:
-#
-#   python3 models/convert-h5-to-ggml.py 
-#
-# This script is similar to "convert-pt-to-ggml.py"
-#
-
-import io
-import os
-import sys
-import struct
-import json
-import code
-import torch
-import numpy as np
-
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BloomForCausalLM
-
-# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-if len(sys.argv) < 3:
-    print("Usage: python convert-hf-to-ggml.py model_name dir-output [use-f32]")
-    print("  model_name: name of the model to convert. Example: 'bigscience/bloomz-560m'")
-    print("  dir-output: directory where the output file will be written")
-    print("  use-f32:    if present, use float32 instead of float16")
-    sys.exit(1)
-
-model_name = sys.argv[1]
-dir_out = sys.argv[2]
-
-# make sure the output directory exists
-os.makedirs(dir_out, exist_ok=True)
-
-# possible data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-#
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-ftype = 1
-if len(sys.argv) > 3:
-    ftype = 0
-
-tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
-hparams = config.to_dict()
-print("Loading model: ", model_name)
-model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, config=config, torch_dtype=torch.float16 if ftype == 1 else torch.float32, low_cpu_mem_usage=True)
-print("Model loaded: ", model_name)
-
-
-fname_out = dir_out + f"/ggml-model-{model_name.split('/')[-1]}-{ftype_str[ftype]}.bin"
-fout = open(fname_out, "wb")
-vocab = tokenizer.vocab
-
-hparams["multiple_of"] = 1
-fout.write(struct.pack("I", 0x67676d6d)) # magic: ggml in hex
-fout.write(struct.pack("I", model.config.vocab_size))
-fout.write(struct.pack("I", model.config.max_seq_len))
-fout.write(struct.pack("I", model.config.n_layers))
-fout.write(struct.pack("I", model.config.n_heads))
-fout.write(struct.pack("I", model.config.d_model))
-fout.write(struct.pack("f", model.config.attn_config['alibi_bias_max']))
-clip_qkv = model.config.attn_config['clip_qkv']
-fout.write(struct.pack("f",  clip_qkv if clip_qkv is not None else 0))
-fout.write(struct.pack("I", ftype))
-
-# # Is this correct??
-# dot_token = tokenizer.encode(".")[0]
-# write tokens to ggml file 
-dot_token = tokenizer.encode('.')[0]
-fout.write(struct.pack("I", model.config.vocab_size))
-
-for i in range(model.config.vocab_size):
-    text = tokenizer.decode([dot_token, i]).encode('utf-8')
-    # remove the first byte (it's always '.')
-    text = text[1:]
-    enclen = len(text)
-    if i in tokenizer.all_special_ids:
-        print(f"special token: {text}")
-        enclen = enclen | 1<<31
-    fout.write(struct.pack("I", enclen))
-    fout.write(text)
-    
-list_vars = model.state_dict()
-for name in list_vars.keys():
-    data = list_vars[name].squeeze().numpy()
-    print("Processing variable: " + name + " with shape: ", data.shape)
-
-    n_dims = len(data.shape);
-
-    # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype_cur = 0;
-    if ftype != 0:
-        # Keep token embeddings in fp32
-        if name[-7:] == ".weight" and n_dims == 2 and ".wte" not in name:
-            print("  Converting to float16")
-            data = data.astype(np.float16)
-            ftype_cur = 1
-        else:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-    else:
-        if data.dtype != np.float32:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-
-    # header
-    str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
-    for i in range(n_dims):
-        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str);
-
-    # data
-    data.tofile(fout)
-
-fout.close()
-
-print("Done. Output file: " + fname_out)
-print("")

gpt4all-backend/scripts/convert_replit_hf_to_ggml.py

@@ -1,113 +0,0 @@
-from pathlib import Path
-import sys
-import struct
-import json
-import numpy as np
-from transformers import AutoModelForCausalLM, AutoTokenizer
-import sentencepiece.sentencepiece_model_pb2 as model
-
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-    sys.exit(1)
-
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-fname_out = sys.argv[1] + "/ggml-replit-code-v1-3b.bin"
-
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-sp_proto = model.ModelProto()
-sp_proto.ParseFromString(open(Path(sys.argv[1]) / "spiece.model", "rb").read())
-
-
-# possible data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-#
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-        sys.exit(1)
-    fname_out = sys.argv[1] + "/ggml-replit-code-v1-3b-" + ftype_str[ftype] + ".bin"
-
-
-tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(
-    dir_model, low_cpu_mem_usage=True, trust_remote_code=True
-)
-# print (model)
-
-# print(tokenizer.encode('I believe the meaning of life is'))
-
-list_vars = model.state_dict()
-for name in list_vars.keys():
-    print(name, list_vars[name].shape, list_vars[name].dtype)
-
-fout = open(fname_out, "wb")
-
-print(hparams)
-
-fout.write(struct.pack("i", 0x7265706c))  # magic: repl in hex
-fout.write(struct.pack("i", hparams["vocab_size"]))
-fout.write(struct.pack("i", hparams["max_seq_len"]))
-fout.write(struct.pack("i", hparams["d_model"]))
-fout.write(struct.pack("i", hparams["n_heads"]))
-fout.write(struct.pack("i", hparams["n_layers"]))
-fout.write(struct.pack("i", ftype))
-
-
-# TODO: temporary hack to not deal with implementing the tokenizer
-for piece in sp_proto.pieces:
-    encoded_piece = piece.piece.encode("utf-8")
-    fout.write(struct.pack("i", len(encoded_piece)))
-    fout.write(encoded_piece)
-    fout.write(struct.pack("f", piece.score))
-
-
-for name in list_vars.keys():
-    data = list_vars[name].squeeze().numpy()
-    print("Processing variable: " + name + " with shape: ", data.shape)
-
-    n_dims = len(data.shape)
-
-    # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype_cur = 0
-    if ftype != 0:
-        if name[-7:] == ".weight" and n_dims == 2:
-            print("  Converting to float16")
-            data = data.astype(np.float16)
-            ftype_cur = 1
-        else:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-    else:
-        if data.dtype != np.float32:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-
-    # header
-    str = name.encode("utf-8")
-    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
-    for i in range(n_dims):
-        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str)
-
-    # data
-    data.tofile(fout)
-
-fout.close()
-
-print("Done. Output file: " + fname_out)
-print("")

gpt4all-backend/starcoder_impl.h

@@ -1,42 +0,0 @@
-#ifndef STARCODER_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#error This file is NOT meant to be included outside of starcoder.cpp. Doing so is DANGEROUS. Be sure to know what you are doing before proceeding to #define STARCODER_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#endif
-#ifndef STARCODER_H
-#define STARCODER_H
-
-#include <string>
-#include <functional>
-#include <vector>
-#include <memory>
-#include "llmodel.h"
-
-struct StarcoderPrivate;
-class Starcoder : public LLModel {
-public:
-    Starcoder();
-    ~Starcoder();
-
-    bool supportsEmbedding() const override { return false; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
-    bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
-    size_t stateSize() const override;
-    size_t saveState(uint8_t *dest) const override;
-    size_t restoreState(const uint8_t *src) override;
-    void setThreadCount(int32_t n_threads) override;
-    int32_t threadCount() const override;
-
-private:
-    std::unique_ptr<StarcoderPrivate> d_ptr;
-
-protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    Token sampleToken(PromptContext &ctx) const override;
-    std::string tokenToString(Token) const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
-};
-
-#endif // STARCODER_H

gpt4all-bindings/cli/README.md

@@ -40,5 +40,5 @@ directory, if necessary.
 If you have already saved a model beforehand, specify its path with the `-m`/`--model` argument,
 for example:
 ```shell
-python app.py repl --model /home/user/my-gpt4all-models/GPT4All-13B-snoozy.ggmlv3.q4_0.bin
+python app.py repl --model /home/user/my-gpt4all-models/gpt4all-13b-snoozy-q4_0.gguf
 ```

gpt4all-bindings/cli/app.py

@@ -1,16 +1,17 @@
+#!/usr/bin/env python3
 """GPT4All CLI
 
 The GPT4All CLI is a self-contained script based on the `gpt4all` and `typer` packages. It offers a
 REPL to communicate with a language model similar to the chat GUI application, but more basic.
 """
 
+import importlib.metadata
 import io
-import pkg_resources  # should be present as a dependency of gpt4all
 import sys
-import typer
-
 from collections import namedtuple
 from typing_extensions import Annotated
+
+import typer
 from gpt4all import GPT4All
 
 
@@ -53,14 +54,18 @@ def repl(
     model: Annotated[
         str,
         typer.Option("--model", "-m", help="Model to use for chatbot"),
-    ] = "ggml-gpt4all-j-v1.3-groovy",
+    ] = "mistral-7b-instruct-v0.1.Q4_0.gguf",
     n_threads: Annotated[
         int,
         typer.Option("--n-threads", "-t", help="Number of threads to use for chatbot"),
     ] = None,
+    device: Annotated[
+        str,
+        typer.Option("--device", "-d", help="Device to use for chatbot, e.g. gpu, amd, nvidia, intel. Defaults to CPU."),
+    ] = None,
 ):
     """The CLI read-eval-print loop."""
-    gpt4all_instance = GPT4All(model)
+    gpt4all_instance = GPT4All(model, device=device)
 
     # if threads are passed, set them
     if n_threads is not None:
@@ -79,7 +84,7 @@ def repl(
 
     use_new_loop = False
     try:
-        version = pkg_resources.Environment()['gpt4all'][0].version
+        version = importlib.metadata.version('gpt4all')
         version_major = int(version.split('.')[0])
         if version_major >= 1:
             use_new_loop = True
@@ -115,6 +120,7 @@ def _old_loop(gpt4all_instance):
             n_predict=200,
             top_k=40,
             top_p=0.9,
+            min_p=0.0,
             temp=0.9,
             n_batch=9,
             repeat_penalty=1.1,
@@ -151,6 +157,7 @@ def _new_loop(gpt4all_instance):
                 temp=0.9,
                 top_k=40,
                 top_p=0.9,
+                min_p=0.0,
                 repeat_penalty=1.1,
                 repeat_last_n=64,
                 n_batch=9,

gpt4all-bindings/csharp/.editorconfig

@@ -41,6 +41,8 @@ insert_final_newline = true
 
 # IDE0055: Fix formatting
 dotnet_diagnostic.IDE0055.severity = error
+dotnet_diagnostic.CS1573.severity = suggestion
+dotnet_diagnostic.CS1591.severity = suggestion
 
 # Sort using and Import directives with System.* appearing first
 dotnet_sort_system_directives_first = true
@@ -343,4 +345,4 @@ dotnet_diagnostic.IDE2004.severity = warning
 [src/{VisualStudio}/**/*.{cs,vb}]
 # CA1822: Make member static
 # There is a risk of accidentally breaking an internal API that partners rely on though IVT.
-dotnet_code_quality.CA1822.api_surface = private
+dotnet_code_quality.CA1822.api_surface = private

gpt4all-bindings/csharp/Directory.Build.props

@@ -5,7 +5,7 @@
         <Company></Company>
         <Copyright></Copyright>
         <NeutralLanguage>en-US</NeutralLanguage>
-        <Version>0.6.3-alpha</Version>
+        <Version>0.6.4-alpha</Version>
         <VersionSuffix>$(VersionSuffix)</VersionSuffix>
         <Version Condition=" '$(VersionSuffix)' != '' ">$(Version)$(VersionSuffix)</Version>
         <TreatWarningsAsErrors>true</TreatWarningsAsErrors>

gpt4all-bindings/csharp/Gpt4All.Samples/Gpt4All.Samples.csproj

@@ -2,9 +2,10 @@
 
     <PropertyGroup>
         <OutputType>Exe</OutputType>
-        <TargetFramework>net7.0</TargetFramework>
+        <TargetFramework>net8.0</TargetFramework>
         <ImplicitUsings>enable</ImplicitUsings>
         <Nullable>enable</Nullable>
+        <GenerateDocumentationFile>true</GenerateDocumentationFile>
     </PropertyGroup>
 
     <ItemGroup>

gpt4all-bindings/csharp/Gpt4All.Tests/Gpt4All.Tests.csproj

@@ -1,10 +1,11 @@
 <Project Sdk="Microsoft.NET.Sdk">
 
     <PropertyGroup>
-        <TargetFramework>net7.0</TargetFramework>
+        <TargetFramework>net8.0</TargetFramework>
         <Nullable>enable</Nullable>
 
         <IsPackable>false</IsPackable>
+        <GenerateDocumentationFile>true</GenerateDocumentationFile>
     </PropertyGroup>
 
     <ItemGroup>

gpt4all-bindings/csharp/Gpt4All/Bindings/ILLModel.cs

@@ -5,8 +5,6 @@
 /// </summary>
 public interface ILLModel : IDisposable
 {
-    ModelType ModelType { get; }
-
     ulong GetStateSizeBytes();
 
     int GetThreadCount();

gpt4all-bindings/csharp/Gpt4All/Bindings/LLModel.cs

@@ -42,16 +42,12 @@ public record ModelRecalculatingEventArgs(bool IsRecalculating);
 public class LLModel : ILLModel
 {
     protected readonly IntPtr _handle;
-    private readonly ModelType _modelType;
     private readonly ILogger _logger;
     private bool _disposed;
 
-    public ModelType ModelType => _modelType;
-
-    internal LLModel(IntPtr handle, ModelType modelType, ILogger? logger = null)
+    internal LLModel(IntPtr handle, ILogger? logger = null)
     {
         _handle = handle;
-        _modelType = modelType;
         _logger = logger ?? NullLogger.Instance;
     }
 
@@ -59,10 +55,9 @@ public class LLModel : ILLModel
     /// Create a new model from a pointer
     /// </summary>
     /// <param name="handle">Pointer to underlying model</param>
-    /// <param name="modelType">The model type</param>
-    public static LLModel Create(IntPtr handle, ModelType modelType, ILogger? logger = null)
+    public static LLModel Create(IntPtr handle, ILogger? logger = null)
     {
-        return new LLModel(handle, modelType, logger: logger);
+        return new LLModel(handle, logger: logger);
     }
 
     /// <summary>
@@ -188,7 +183,7 @@ public class LLModel : ILLModel
     /// <returns>true if the model was loaded successfully, false otherwise.</returns>
     public bool Load(string modelPath)
     {
-        return NativeMethods.llmodel_loadModel(_handle, modelPath);
+        return NativeMethods.llmodel_loadModel(_handle, modelPath, 2048, 100);
     }
 
     protected void Destroy()
@@ -204,12 +199,7 @@ public class LLModel : ILLModel
             // dispose managed state
         }
 
-        switch (_modelType)
-        {
-            default:
-                Destroy();
-                break;
-        }
+        Destroy();
 
         _disposed = true;
     }

gpt4all-bindings/csharp/Gpt4All/Bindings/LLPromptContext.cs

@@ -64,6 +64,15 @@ public unsafe class LLModelPromptContext
         set => _ctx.top_p = value;
     }
 
+    /// <summary>
+    /// min p sampling probability threshold
+    /// </summary>
+    public float MinP
+    {
+        get => _ctx.min_p;
+        set => _ctx.min_p = value;
+    }
+
     /// <summary>
     /// temperature to adjust model's output distribution
     /// </summary>

gpt4all-bindings/csharp/Gpt4All/Bindings/NativeMethods.cs

@@ -29,6 +29,8 @@ public unsafe partial struct llmodel_prompt_context
 
     public float top_p;
 
+    public float min_p;
+
     public float temp;
 
     [NativeTypeName("int32_t")]
@@ -70,7 +72,9 @@ internal static unsafe partial class NativeMethods
     [return: MarshalAs(UnmanagedType.I1)]
     public static extern bool llmodel_loadModel(
         [NativeTypeName("llmodel_model")] IntPtr model,
-        [NativeTypeName("const char *")][MarshalAs(UnmanagedType.LPUTF8Str)] string model_path);
+        [NativeTypeName("const char *")][MarshalAs(UnmanagedType.LPUTF8Str)] string model_path,
+        [NativeTypeName("int32_t")] int n_ctx,
+        [NativeTypeName("int32_t")] int ngl);
 
     [DllImport("libllmodel", CallingConvention = CallingConvention.Cdecl, ExactSpelling = true)]
 

gpt4all-bindings/csharp/Gpt4All/Extensions/LLPromptContextExtensions.cs

@@ -16,6 +16,7 @@ internal static class LLPromptContextExtensions
             n_predict = {ctx.n_predict}
             top_k = {ctx.top_k}
             top_p = {ctx.top_p}
+            min_p = {ctx.min_p}
             temp = {ctx.temp}
             n_batch = {ctx.n_batch}
             repeat_penalty = {ctx.repeat_penalty}

gpt4all-bindings/csharp/Gpt4All/Extensions/PredictRequestOptionsExtensions.cs

@@ -12,6 +12,7 @@ public static class PredictRequestOptionsExtensions
             TokensSize = opts.TokensSize,
             TopK = opts.TopK,
             TopP = opts.TopP,
+            MinP = opts.MinP,
             PastNum = opts.PastConversationTokensNum,
             RepeatPenalty = opts.RepeatPenalty,
             Temperature = opts.Temperature,

gpt4all-bindings/csharp/Gpt4All/Gpt4All.csproj

@@ -1,9 +1,10 @@
 <Project Sdk="Microsoft.NET.Sdk">
     <PropertyGroup>
-        <TargetFramework>net6.0</TargetFramework>
         <ImplicitUsings>enable</ImplicitUsings>
         <Nullable>enable</Nullable>
         <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
+        <GenerateDocumentationFile>true</GenerateDocumentationFile>
+        <TargetFramework>net8.0</TargetFramework>
     </PropertyGroup>
     <ItemGroup>
         <!-- Windows -->

gpt4all-bindings/csharp/Gpt4All/Model/Gpt4AllModelFactory.cs

@@ -3,6 +3,7 @@ using Microsoft.Extensions.Logging.Abstractions;
 using Microsoft.Extensions.Logging;
 using Gpt4All.Bindings;
 using Gpt4All.LibraryLoader;
+using System.Runtime.InteropServices;
 
 namespace Gpt4All;
 
@@ -31,15 +32,18 @@ public class Gpt4AllModelFactory : IGpt4AllModelFactory
         }
     }
 
-    private IGpt4AllModel CreateModel(string modelPath)
+    private Gpt4All CreateModel(string modelPath)
     {
-        var modelType_ = ModelFileUtils.GetModelTypeFromModelFileHeader(modelPath);
-        _logger.LogInformation("Creating model path={ModelPath} type={ModelType}", modelPath, modelType_);
+        _logger.LogInformation("Creating model path={ModelPath}", modelPath);
         IntPtr error;
         var handle = NativeMethods.llmodel_model_create2(modelPath, "auto", out error);
+        if (error != IntPtr.Zero)
+        {
+            throw new Exception(Marshal.PtrToStringAnsi(error));
+        }
         _logger.LogDebug("Model created handle=0x{ModelHandle:X8}", handle);
         _logger.LogInformation("Model loading started");
-        var loadedSuccessfully = NativeMethods.llmodel_loadModel(handle, modelPath);
+        var loadedSuccessfully = NativeMethods.llmodel_loadModel(handle, modelPath, 2048, 100);
         _logger.LogInformation("Model loading completed success={ModelLoadSuccess}", loadedSuccessfully);
         if (!loadedSuccessfully)
         {
@@ -47,7 +51,7 @@ public class Gpt4AllModelFactory : IGpt4AllModelFactory
         }
 
         var logger = _loggerFactory.CreateLogger<LLModel>();
-        var underlyingModel = LLModel.Create(handle, modelType_, logger: logger);
+        var underlyingModel = LLModel.Create(handle, logger: logger);
 
         Debug.Assert(underlyingModel.IsLoaded());
 

gpt4all-bindings/csharp/Gpt4All/Model/ModelFileUtils.cs

@@ -1,24 +0,0 @@
-namespace Gpt4All;
-
-public static class ModelFileUtils
-{
-    private const uint GPTJ_MAGIC = 0x67676d6c;
-    private const uint LLAMA_MAGIC = 0x67676a74;
-    private const uint MPT_MAGIC = 0x67676d6d;
-
-    public static ModelType GetModelTypeFromModelFileHeader(string modelPath)
-    {
-        using var fileStream = new FileStream(modelPath, FileMode.Open);
-        using var binReader = new BinaryReader(fileStream);
-
-        var magic = binReader.ReadUInt32();
-
-        return magic switch
-        {
-            GPTJ_MAGIC => ModelType.GPTJ,
-            LLAMA_MAGIC => ModelType.LLAMA,
-            MPT_MAGIC => ModelType.MPT,
-            _ => throw new ArgumentOutOfRangeException($"Invalid model file. magic=0x{magic:X8}"),
-        };
-    }
-}

gpt4all-bindings/csharp/Gpt4All/Model/ModelOptions.cs

@@ -3,6 +3,4 @@
 public record ModelOptions
 {
     public int Threads { get; init; } = 4;
-
-    public ModelType ModelType { get; init; } = ModelType.GPTJ;
 }

gpt4all-bindings/csharp/Gpt4All/Model/ModelType.cs

@@ -1,11 +0,0 @@
-namespace Gpt4All;
-
-/// <summary>
-/// The supported model types
-/// </summary>
-public enum ModelType
-{
-    LLAMA = 0,
-    GPTJ,
-    MPT
-}

gpt4all-bindings/csharp/Gpt4All/Prediction/PredictRequestOptions.cs

@@ -16,6 +16,8 @@ public record PredictRequestOptions
 
     public float TopP { get; init; } = 0.9f;
 
+    public float MinP { get; init; } = 0.0f;
+
     public float Temperature { get; init; } = 0.1f;
 
     public int Batches { get; init; } = 8;

gpt4all-bindings/csharp/README.md

@@ -6,7 +6,10 @@ This package contains a set of C# bindings around the `llmodel` C-API.
 TBD
 
 ## Installation
-TBD NuGet
+
+Windows and Linux builds are available on NuGet: https://www.nuget.org/packages/Gpt4All
+
+macOS is WIP due to code signing issues, contributions are welcome.
 
 ## Project Structure
 ```
@@ -23,6 +26,12 @@ gpt4all-bindings/
           └── linux-x64
 ```
 
+## Prerequisites
+
+On Windows and Linux, building GPT4All requires the complete Vulkan SDK. You may download it from here: https://vulkan.lunarg.com/sdk/home
+
+macOS users do not need Vulkan, as GPT4All will use Metal instead.
+
 ## Local Build Instructions
 > **Note** 
 > Tested On:
@@ -54,7 +63,7 @@ chmod +x ./build_linux.sh
 1. Setup
 ```
 choco install mingw
-$env:Path += ";C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin"
+$env:Path += ";C:\ProgramData\mingw64\mingw64\bin"
 choco install -y cmake --installargs 'ADD_CMAKE_TO_PATH=System'
 ```
 2. Run the `./build_win-mingw.ps1` build script

gpt4all-bindings/csharp/build_linux.sh

@@ -1,3 +1,4 @@
+#!/bin/sh
 mkdir -p runtimes
 rm -rf runtimes/linux-x64
 mkdir -p runtimes/linux-x64/native
@@ -7,4 +8,3 @@ cmake --build runtimes/linux-x64/build --parallel --config Release
 cp runtimes/linux-x64/build/libllmodel.so  runtimes/linux-x64/native/libllmodel.so
 cp runtimes/linux-x64/build/libgptj*.so  runtimes/linux-x64/native/
 cp runtimes/linux-x64/build/libllama*.so  runtimes/linux-x64/native/
-cp runtimes/linux-x64/build/libmpt*.so  runtimes/linux-x64/native/

gpt4all-bindings/csharp/build_win-mingw.ps1

@@ -12,5 +12,5 @@ cmake -G "MinGW Makefiles" -S ..\..\gpt4all-backend -B $BUILD_DIR
 cmake --build $BUILD_DIR --parallel --config Release
 
 # copy native dlls
-cp "C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin\*dll" $LIBS_DIR
-cp "$BUILD_DIR\bin\*.dll" $LIBS_DIR
+cp "C:\ProgramData\mingw64\mingw64\bin\*dll" $LIBS_DIR
+cp "$BUILD_DIR\bin\*.dll" $LIBS_DIR

gpt4all-bindings/golang/Makefile

@@ -139,7 +139,7 @@ $(info I CXX:      $(CXXV))
 $(info )
 
 llmodel.o:
-	mkdir buildllm
+	[ -e buildllm ] || mkdir buildllm
 	cd buildllm && cmake ../../../gpt4all-backend/ $(CMAKEFLAGS) && make
 	cd buildllm && cp -rf CMakeFiles/llmodel.dir/llmodel_c.cpp.o ../llmodel_c.o
 	cd buildllm && cp -rf CMakeFiles/llmodel.dir/llmodel.cpp.o ../llmodel.o
@@ -150,7 +150,7 @@ clean:
 	rm -rf buildllm
 	rm -rf example/main
 
-binding.o: 
+binding.o: binding.cpp binding.h
 	$(CXX) $(CXXFLAGS) binding.cpp -o binding.o -c $(LDFLAGS)
 
 libgpt4all.a: binding.o llmodel.o

gpt4all-bindings/golang/README.md

@@ -24,7 +24,7 @@ func main() {
 		return true
 	})
 
-	_, err = model.Predict("Here are 4 steps to create a website:", gpt4all.SetTemperature(0.1))
+	_, err = model.Predict("Here are 4 steps to create a website:", "", "", gpt4all.SetTemperature(0.1))
 	if err != nil {
 		panic(err)
 	}
@@ -36,7 +36,7 @@ func main() {
 In order to use the bindings you will need to build `libgpt4all.a`:
 
 ```
-git clone https://github.com/nomic-ai/gpt4all
+git clone --recurse-submodules https://github.com/nomic-ai/gpt4all
 cd gpt4all/gpt4all-bindings/golang
 make libgpt4all.a
 ```

gpt4all-bindings/golang/binding.cpp

@@ -17,14 +17,13 @@
 
 void* load_model(const char *fname, int n_threads) {
     // load the model
-    llmodel_error new_error{};
+    const char *new_error;
     auto model = llmodel_model_create2(fname, "auto", &new_error);
-    if (model == nullptr ){
-        fprintf(stderr, "%s: error '%s'\n",
-                __func__, new_error.message);
+    if (model == nullptr) {
+        fprintf(stderr, "%s: error '%s'\n", __func__, new_error);
         return nullptr;
     }
-    if (!llmodel_loadModel(model, fname)) {
+    if (!llmodel_loadModel(model, fname, 2048, 100)) {
         llmodel_model_destroy(model);
         return nullptr;
     }
@@ -36,8 +35,9 @@ void* load_model(const char *fname, int n_threads) {
 std::string res = "";
 void * mm;
 
-void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, int top_k,
-                            float top_p, float temp, int n_batch,float ctx_erase)
+void model_prompt(const char *prompt, const char *prompt_template, int special, const char *fake_reply, 
+                    void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, 
+                    int top_k, float top_p, float min_p, float temp, int n_batch,float ctx_erase)
 {
     llmodel_model* model = (llmodel_model*) m;
 
@@ -70,6 +70,7 @@ void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n,
         .n_predict = 50,
         .top_k = 10,
         .top_p = 0.9,
+        .min_p = 0.0,
         .temp = 1.0,
         .n_batch = 1,
         .repeat_penalty = 1.2,
@@ -84,14 +85,15 @@ void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n,
     prompt_context->top_k = top_k;
     prompt_context->context_erase = ctx_erase;
     prompt_context->top_p = top_p;
+    prompt_context->min_p = min_p;
     prompt_context->temp = temp;
     prompt_context->n_batch = n_batch;    
 
-    llmodel_prompt(model, prompt,
+    llmodel_prompt(model, prompt, prompt_template,
                         lambda_prompt,
                         lambda_response,
                     lambda_recalculate,
-                    prompt_context );
+                    prompt_context, special, fake_reply);
 
     strcpy(result, res.c_str()); 
 

gpt4all-bindings/golang/binding.h

@@ -6,8 +6,9 @@ extern "C" {
 
 void* load_model(const char *fname, int n_threads);
 
-void model_prompt( const char *prompt, void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, int top_k,
-                            float top_p, float temp, int n_batch,float ctx_erase);
+void model_prompt(const char *prompt, const char *prompt_template, int special, const char *fake_reply, 
+                    void *m, char* result, int repeat_last_n, float repeat_penalty, int n_ctx, int tokens, 
+                    int top_k, float top_p, float min_p, float temp, int n_batch,float ctx_erase);
 
 void free_model(void *state_ptr);
 
@@ -15,4 +16,4 @@ extern unsigned char getTokenCallback(void *, char *);
 
 #ifdef __cplusplus
 }
-#endif
+#endif