cmake: fix Metal build after #2310 (#2350)

I don't understand why this is needed, but it works.

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
@@ -39,18 +42,18 @@ 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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
               hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
             fi
       - save_cache:  # this is the new step to save cache
-          key: macos-qt-cache_v2
+          key: macos-qt-cache-v3
           paths:
             - ~/Qt
       - run:
@@ -58,7 +61,7 @@ jobs:
           command: |
             mkdir build
             cd build
-            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.6/bin
+            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.7/bin
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
               -DCMAKE_GENERATOR:STRING=Ninja \
               -DBUILD_UNIVERSAL=ON \
@@ -88,23 +91,25 @@ 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 patchelf
+            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver qt.qt6.651.qtwaylandcompositor
             fi
       - save_cache:  # this is the new step to save cache
-          key: linux-qt-cache
+          key: linux-qt-cache-v2
           paths:
             - ~/Qt
       - run:
@@ -117,7 +122,8 @@ jobs:
           command: |
             set -eo pipefail
             export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
-            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.6/bin
+            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.7/bin
+            export PATH=$PATH:/usr/local/cuda/bin
             mkdir build
             cd build
             mkdir upload
@@ -142,16 +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:
@@ -159,6 +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: |
@@ -166,7 +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\4.6\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"
@@ -209,33 +220,34 @@ 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 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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.47 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver qt.qt6.651.qtwaylandcompositor
             fi
       - save_cache:  # this is the new step to save cache
-          key: linux-qt-cache
+          key: linux-qt-cache-v2
           paths:
             - ~/Qt
       - run:
           name: Build
           command: |
             export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
-            mkdir build
-            cd build
-            ~/Qt/Tools/CMake/bin/cmake -DCMAKE_BUILD_TYPE=Release -S ../gpt4all-chat -B .
-            ~/Qt/Tools/CMake/bin/cmake --build . --target all
+            export PATH=$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:
@@ -251,16 +263,16 @@ 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:
@@ -268,6 +280,11 @@ 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: |
@@ -287,17 +304,16 @@ jobs:
             $Env:INCLUDE = "${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"
-            mkdir build
-            cd build
+            $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 ."
-            & "C:\Qt\Tools\Ninja\ninja.exe"
+              "-S gpt4all-chat" `
+              "-B build"
+            & "C:\Qt\Tools\Ninja\ninja.exe" -C build
 
   build-gpt4all-chat-macos:
     macos:
@@ -311,52 +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
@@ -372,13 +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 "/*"
@@ -396,19 +410,19 @@ 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
+            cmake --build build --parallel
       - run:
           name: Build wheel
           command: |
@@ -435,13 +449,10 @@ 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_OSX_ARCHITECTURES="x86_64;arm64"
+            cmake --build build --parallel
       - run:
           name: Build wheel
           command: |
@@ -467,6 +478,11 @@ 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:
@@ -477,15 +493,13 @@ jobs:
       - run:
           name: Build C library
           command: |
-            git submodule init
-            git submodule update
+            git submodule update --init --recursive
             cd gpt4all-backend
-            mkdir build
-            cd build
-            $env:Path += ";C:\ProgramData\mingw64\mingw64\bin"
-            $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
+            $Env:Path += ";C:\ProgramData\mingw64\mingw64\bin"
+            $Env:Path += ";C:\VulkanSDK\1.3.261.1\bin"
+            $Env:VULKAN_SDK = "C:\VulkanSDK\1.3.261.1"
+            cmake -G "MinGW Makefiles" -B build -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=OFF
+            cmake --build build --parallel
       - run:
           name: Build wheel
           # TODO: As part of this task, we need to move mingw64 binaries into package.
@@ -540,11 +554,14 @@ jobs:
           command: |
             wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
             sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list            
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
             sudo apt-get update
-            sudo apt-get install -y cmake build-essential vulkan-sdk
+            sudo apt-get install -y cmake build-essential vulkan-sdk cuda-compiler-12-4 libcublas-dev-12-4 libnvidia-compute-550-server libmysqlclient21 libodbc2 libpq5
       - run:
           name: Build Libraries
           command: |
+            export PATH=$PATH:/usr/local/cuda/bin
             cd gpt4all-backend
             mkdir -p runtimes/build
             cd runtimes/build
@@ -609,6 +626,11 @@ 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: |
@@ -620,6 +642,7 @@ jobs:
             $Env:Path += ";$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
@@ -651,6 +674,11 @@ 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: |
@@ -660,6 +688,7 @@ 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
@@ -673,7 +702,7 @@ 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:
@@ -729,6 +758,10 @@ 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: |
@@ -776,7 +809,8 @@ 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:
@@ -818,7 +852,7 @@ 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
@@ -834,9 +868,9 @@ 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:
@@ -853,6 +887,7 @@ 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
@@ -883,6 +918,7 @@ 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
@@ -895,14 +931,14 @@ 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:
@@ -924,6 +960,7 @@ jobs:
               nvm install 18.16.0
               nvm use 18.16.0
       - run: node --version 
+      - run: corepack enable
       - run:           
           command: |
             npm install -g yarn
@@ -957,6 +994,7 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - run: 
           command: |
             cd gpt4all-bindings/typescript
@@ -971,9 +1009,12 @@ 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
@@ -1001,11 +1042,17 @@ jobs:
 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: << pipeline.parameters.run-chat-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-chat-workflow >>
     jobs:
       - hold:
           type: approval
@@ -1019,7 +1066,10 @@ workflows:
           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
@@ -1033,7 +1083,10 @@ 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:
@@ -1046,7 +1099,10 @@ 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
@@ -1081,15 +1137,20 @@ 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:
@@ -1132,21 +1193,21 @@ 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
 
 
@@ -1156,21 +1217,21 @@ 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:
@@ -1180,4 +1241,4 @@ workflows:
             - nuget-hold
             - build-csharp-windows
             - build-csharp-linux
-            - build-csharp-macos
+            #- build-csharp-macos

.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

@@ -1,55 +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: 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.

.gitmodules

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

@@ -1,83 +1,73 @@
 <h1 align="center">GPT4All</h1>
-
-<p align="center">Open-source large language models that run locally on your CPU and nearly any GPU</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 and Models</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.
+<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>
 
-> [!IMPORTANT]
-> GPT4All v2.5.0 and newer only supports models in GGUF format (.gguf). Models used with a previous version of GPT4All (.bin extension) will no longer work.
 
-GPT4All is an ecosystem to run **powerful** and **customized** large language models that work locally on consumer grade CPUs and any GPU. Note that your CPU needs to support [AVX or AVX2 instructions](https://en.wikipedia.org/wiki/Advanced_Vector_Extensions).
+## 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).
 
-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.
 
-### What's New ([Issue Tracker](https://github.com/orgs/nomic-ai/projects/2))
+
+### Installation
+
+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.
+
+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).
+
+
+### 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, Q6 quantizations in GGUF.
+    - [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 AMD, Intel, Samsung, Qualcomm and NVIDIA GPUs.
-- **August 15th, 2023**: GPT4All API launches allowing inference of local LLMs from docker containers.
-- **July 2023**: Stable support for LocalDocs, a GPT4All Plugin that allows you to privately and locally chat with your data.
+- **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.
+
+[Docker-based API server]: https://github.com/nomic-ai/gpt4all/tree/cef74c2be20f5b697055d5b8b506861c7b997fab/gpt4all-api
 
 
-### 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.
+### Building From Source
 
-Direct Installer Links:
+* Follow the instructions [here](gpt4all-chat/build_and_run.md) to build the GPT4All Chat UI from source.
 
-* [macOS](https://gpt4all.io/installers/gpt4all-installer-darwin.dmg)
-
-* [Windows](https://gpt4all.io/installers/gpt4all-installer-win64.exe)
-
-* [Ubuntu](https://gpt4all.io/installers/gpt4all-installer-linux.run)
-
-Find the most up-to-date information on the [GPT4All Website](https://gpt4all.io/)
-
-### Chat Client building and running
-
-* Follow the visual instructions on the chat client [build_and_run](gpt4all-chat/build_and_run.md) page
 
 ### 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!
@@ -87,6 +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">
@@ -101,6 +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

@@ -1,22 +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"
-    env_file:
-      - .env
-    environment:
-      - APP_ENVIRONMENT=dev
-      - WEB_CONCURRENCY=2
-      - LOGLEVEL=debug
-      - PORT=4891
-      - model=${MODEL_BIN} # using variable from .env file
-      - inference_mode=cpu
-    volumes:
-      - './gpt4all_api/app:/app'
-      - './gpt4all_api/models:/models' # models are mounted in the container
-    command: ["/start-reload.sh"]

gpt4all-api/gpt4all_api/Dockerfile.buildkit

@@ -1,17 +0,0 @@
-# syntax=docker/dockerfile:1.0.0-experimental
-FROM tiangolo/uvicorn-gunicorn:python3.11
-
-# 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
-

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

@@ -1,75 +0,0 @@
-import logging
-import time
-from typing import List
-from uuid import uuid4
-from fastapi import APIRouter
-from pydantic import BaseModel, Field
-from api_v1.settings import settings
-from fastapi.responses import StreamingResponse
-
-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(settings.model, description='The model to generate a completion from.')
-    messages: List[ChatCompletionMessage] = Field(..., description='Messages for the chat completion.')
-
-class ChatCompletionChoice(BaseModel):
-    message: ChatCompletionMessage
-    index: int
-    logprobs: float
-    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 based on the last message in the chat.
-    '''
-    # Example: Echo the last message content with some modification
-    if request.messages:
-        last_message = request.messages[-1].content
-        response_content = f"Echo: {last_message}"
-    else:
-        response_content = "No messages received."
-
-    # Create a chat message for the response
-    response_message = ChatCompletionMessage(role="system", content=response_content)
-
-    # Create a choice object with the response message
-    response_choice = ChatCompletionChoice(
-        message=response_message,
-        index=0,
-        logprobs=-1.0,  # Placeholder value
-        finish_reason="length"  # Placeholder value
-    )
-
-    # Create the response object
-    chat_response = ChatCompletionResponse(
-        id=str(uuid4()),
-        created=int(time.time()),
-        model=request.model,
-        choices=[response_choice],
-        usage=ChatCompletionUsage(prompt_tokens=0, completion_tokens=0, total_tokens=0),  # Placeholder values
-    )
-
-    return chat_response

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

@@ -1,39 +0,0 @@
-import requests
-from fastapi import APIRouter, HTTPException
-from pydantic import BaseModel, Field
-from typing import List, Dict
-
-# Define the router for the engines module
-router = APIRouter(prefix="/engines", tags=["Search Endpoints"])
-
-# Define the models for the engines module
-class ListEnginesResponse(BaseModel):
-    data: List[Dict] = Field(..., description="All available models.")
-
-class EngineResponse(BaseModel):
-    data: List[Dict] = Field(..., description="All available models.")
-
-
-# Define the routes for the engines module
-@router.get("/", response_model=ListEnginesResponse)
-async def list_engines():
-    try:
-        response = requests.get('https://raw.githubusercontent.com/nomic-ai/gpt4all/main/gpt4all-chat/metadata/models2.json')
-        response.raise_for_status()  # This will raise an HTTPError if the HTTP request returned an unsuccessful status code
-        engines = response.json()
-        return ListEnginesResponse(data=engines)
-    except requests.RequestException as e:
-        logger.error(f"Error fetching engine list: {e}")
-        raise HTTPException(status_code=500, detail="Error fetching engine list")
-
-# Define the routes for the engines module
-@router.get("/{engine_id}", response_model=EngineResponse)
-async def retrieve_engine(engine_id: str):
-    try:
-        # Implement logic to fetch a specific engine's details
-        # This is a placeholder, replace with your actual data retrieval logic
-        engine_details = {"id": engine_id, "name": "Engine Name", "description": "Engine Description"}
-        return EngineResponse(data=[engine_details])
-    except Exception as e:
-        logger.error(f"Error fetching engine details: {e}")
-        raise HTTPException(status_code=500, detail=f"Error fetching details for engine {engine_id}")

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

@@ -1,77 +0,0 @@
-"""
-Use the OpenAI python API to test gpt4all models.
-"""
-from typing import List, get_args
-import os
-from dotenv import load_dotenv
-
-import openai
-
-openai.api_base = "http://localhost:4891/v1"
-openai.api_key = "not needed for a local LLM"
-
-# Load the .env file
-env_path = 'gpt4all-api/gpt4all_api/.env'
-load_dotenv(dotenv_path=env_path)
-
-# Fetch MODEL_ID from .env file
-model_id = os.getenv('MODEL_BIN', 'default_model_id')
-embedding = os.getenv('EMBEDDING', 'default_embedding_model_id')
-print (model_id)
-print (embedding)
-
-def test_completion():
-    model = model_id
-    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 = model_id
-    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))
-
-# Modified test batch, problems with keyerror in response
-def test_batched_completion():
-    model = model_id  # replace with your specific model ID
-    prompt = "Who is Michael Jordan?"
-    responses = []
-    
-    # Loop to create completions one at a time
-    for _ in range(3):
-        response = openai.Completion.create(
-            model=model, prompt=prompt, max_tokens=50, temperature=0.28, top_p=0.95, n=1, echo=True, stream=False
-        )
-        responses.append(response)
-
-    # Assertions to check the responses
-    for response in responses:
-        assert len(response['choices'][0]['text']) > len(prompt)
-    
-    assert len(responses) == 3
-
-def test_embedding():
-    model = embedding
-    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/env

@@ -1,3 +0,0 @@
-# Add your GGUF compatible model LLM here. ie: MODEL_BIN="mistral-7b-instruct-v0.1.Q4_0", rename file ".env"
-# Make sure this LLM matches the model you placed inside the models folder
-MODEL_BIN=""

gpt4all-api/gpt4all_api/models/README.md

@@ -1 +0,0 @@
-### Drop GGUF compatible models here, make sure it matches MODEL_BIN on your .env file

gpt4all-api/gpt4all_api/requirements.txt

@@ -1,13 +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==0.28.0
-black
-isort
-python-dotenv

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 env 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)/venv ]; then $(PYTHON) -m venv $(ROOT_DIR)/venv; fi
-
-dependencies: venv
-	source $(ROOT_DIR)/venv/bin/activate; $(PYTHON) -m pip install -r $(ROOT_DIR)/$(APP_NAME)/requirements.txt
-
-clean: clean_testenv
-	# Remove existing environment
-	rm -rf $(ROOT_DIR)/venv;
-	rm -rf $(ROOT_DIR)/$(APP_NAME)/*.pyc;
-
-
-black:
-	source $(ROOT_DIR)/venv/bin/activate; black -l 120 -S --target-version py38 $(APP_NAME)
-
-isort:
-	source $(ROOT_DIR)/venv/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()
@@ -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)
@@ -97,15 +135,14 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
         LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
     prepare_target(llamamodel-mainline llama-mainline)
 
-    if (NOT LLAMA_METAL)
+    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()
 
-        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)
+    if (BUILD_VARIANT STREQUAL cuda)
+        set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
     endif()
 endforeach()
 
@@ -114,8 +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.cpp

@@ -1,902 +0,0 @@
-#define BERT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#include "bert_impl.h"
-#include "llmodel_shared.h"
-#include "ggml.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <map>
-#include <string>
-#include <vector>
-#include <iostream>
-#include <regex>
-#include <thread>
-#include <algorithm>
-#include <numeric>
-
-//#define DEBUG_BERT
-
-namespace {
-const char *modelType_ = "Bert";
-}
-
-typedef int32_t bert_vocab_id;
-
-// default hparams (all-MiniLM-L6-v2)
-struct bert_hparams
-{
-    int32_t n_vocab = 30522;
-    int32_t n_max_tokens = 512;
-    int32_t n_embd = 256;
-    int32_t n_intermediate = 1536;
-    int32_t n_head = 12;
-    int32_t n_layer = 6;
-};
-
-struct bert_layer
-{
-    // normalization
-    struct ggml_tensor *ln_att_w;
-    struct ggml_tensor *ln_att_b;
-
-    struct ggml_tensor *ln_out_w;
-    struct ggml_tensor *ln_out_b;
-
-    // attention
-    struct ggml_tensor *q_w;
-    struct ggml_tensor *q_b;
-    struct ggml_tensor *k_w;
-    struct ggml_tensor *k_b;
-    struct ggml_tensor *v_w;
-    struct ggml_tensor *v_b;
-
-    struct ggml_tensor *o_w;
-    struct ggml_tensor *o_b;
-
-    // ff
-    struct ggml_tensor *ff_i_w;
-    struct ggml_tensor *ff_i_b;
-
-    struct ggml_tensor *ff_o_w;
-    struct ggml_tensor *ff_o_b;
-};
-
-struct bert_vocab
-{
-    std::map<std::string, bert_vocab_id> token_to_id;
-    std::map<std::string, bert_vocab_id> subword_token_to_id;
-
-    std::map<bert_vocab_id, std::string> _id_to_token;
-    std::map<bert_vocab_id, std::string> _id_to_subword_token;
-};
-
-struct bert_model
-{
-    bert_hparams hparams;
-
-    // embeddings weights
-    struct ggml_tensor *word_embeddings;
-    struct ggml_tensor *token_type_embeddings;
-    struct ggml_tensor *position_embeddings;
-    struct ggml_tensor *ln_e_w;
-    struct ggml_tensor *ln_e_b;
-
-    std::vector<bert_layer> layers;
-
-    struct ggml_context *ctx;
-};
-
-// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
-struct bert_ctx
-{
-    bert_model model;
-    bert_vocab vocab;
-
-    size_t mem_per_token;
-    int64_t mem_per_input;
-    int32_t max_batch_n;
-    llm_buffer buf_compute;
-    llm_buffer work_buf;
-};
-
-int32_t bert_n_embd(bert_ctx * ctx)
-{
-    return ctx->model.hparams.n_embd;
-}
-
-int32_t bert_n_max_tokens(bert_ctx * ctx)
-{
-    return ctx->model.hparams.n_max_tokens;
-}
-
-const char* bert_vocab_id_to_token(bert_ctx * ctx, bert_vocab_id id) {
-    bert_vocab & vocab = ctx->vocab;
-    auto it = vocab._id_to_token.find(id);
-    if (it != vocab._id_to_token.end())
-    {
-        return it->second.c_str();
-    }
-    it = vocab._id_to_subword_token.find(id);
-    if (it != vocab._id_to_subword_token.end())
-    {
-        return it->second.c_str();
-    }
-    return "[UNK TOKEN from bert_vocab]";
-}
-
-//
-// Tokenizing
-//
-
-static size_t utf8_len(char src)
-{
-    const size_t lookup[] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4};
-    uint8_t highbits = static_cast<uint8_t>(src) >> 4;
-    return lookup[highbits];
-}
-
-std::string stripAccents(const std::string &inputString)
-{
-    std::string resultString;
-    std::map<std::string, char> accentMap = {{"À", 'A'},{"Á", 'A'},
-        {"Â", 'A'},{"Ã", 'A'},{"Ä", 'A'},{"Å", 'A'},{"à", 'a'},{"á", 'a'},
-        {"â", 'a'},{"ã", 'a'},{"ä", 'a'},{"å", 'a'},{"È", 'E'},{"É", 'E'},
-        {"Ê", 'E'},{"Ë", 'E'},{"è", 'e'},{"é", 'e'},{"ê", 'e'},{"ë", 'e'},
-        {"Ì", 'I'},{"Í", 'I'},{"Î", 'I'},{"Ï", 'I'},{"ì", 'i'},{"í", 'i'},
-        {"î", 'i'},{"ï", 'i'},{"Ò", 'O'},{"Ó", 'O'},{"Ô", 'O'},{"Õ", 'O'},
-        {"Ö", 'O'},{"ò", 'o'},{"ó", 'o'},{"ô", 'o'},{"õ", 'o'},{"ö", 'o'},
-        {"Ù", 'U'},{"Ú", 'U'},{"Û", 'U'},{"Ü", 'U'},{"ù", 'u'},{"ú", 'u'},
-        {"û", 'u'},{"ü", 'u'},{"Ý", 'Y'},{"ý", 'y'},{"Ç", 'C'},{"ç", 'c'},
-        {"Ñ", 'N'},{"ñ", 'n'},
-    };
-
-    for (size_t i = 0; i < inputString.length();)
-    {
-        int len = utf8_len(inputString[i]);
-        std::string curChar = inputString.substr(i, len);
-        auto iter = accentMap.find(curChar);
-        if (iter != accentMap.end())
-        {
-            resultString += iter->second;
-        }
-        else
-        {
-            resultString += curChar;
-        }
-        i += len;
-    }
-
-    return resultString;
-}
-
-std::string bert_normalize_prompt(const std::string &text)
-{
-    // TODO: handle chinese characters? https://github.com/huggingface/tokenizers/blob/ef5f50605ddf9f8caef1598c0e4853862b9707a7/tokenizers/src/normalizers/bert.rs#L98
-    std::string text2 = stripAccents(text);
-    for (size_t i = 0; i < text2.size(); i += utf8_len(text2[i]))
-    {
-        char c = text2[i];
-        if (c >= 'A' && c <= 'Z')
-            text2[i] = c - 'A' + 'a';
-    }
-    return text2;
-}
-
-std::vector<bert_vocab_id> bert_tokenize(
-    struct bert_ctx * ctx,
-    const char * text)
-{
-    const bert_vocab &vocab = ctx->vocab;
-
-    std::string str = text;
-
-    std::vector<std::string> words;
-    // first split the text into words
-    {
-        str = bert_normalize_prompt(str);
-
-        std::string pat = R"([[:punct:]]|[[:alpha:]]+|[[:digit:]]+)";
-
-        std::regex re(pat);
-        std::smatch m;
-
-        while (std::regex_search(str, m, re))
-        {
-            for (std::string x : m)
-            {
-                words.push_back(x);
-            }
-            str = m.suffix();
-        }
-    }
-
-    // find the longest tokens that form the words:
-    std::vector<bert_vocab_id> tokens;
-    int cls_tok_id = 101;
-    tokens.push_back(cls_tok_id);
-    for (const auto &word : words)
-    {
-        if (word.size() == 0)
-            continue;
-
-        int i = 0;
-        int n = word.size();
-        auto *token_map = &vocab.token_to_id;
-        while (i < n)
-        {
-            int j = n;
-            while (j > i)
-            {
-                auto it = token_map->find(word.substr(i, j - i));
-                if (it != token_map->end())
-                {
-                    tokens.push_back(it->second);
-                    i = j;
-                    token_map = &vocab.subword_token_to_id;
-                }
-                --j;
-            }
-            if (j == i)
-            {
-                fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
-                token_map = &vocab.subword_token_to_id;
-                ++i;
-            }
-        }
-    }
-
-    return tokens;
-}
-
-void bert_resize_ctx(bert_ctx * ctx, int32_t new_size) {
-    int64_t buf_size_new = ctx->mem_per_input * new_size;
-
-    // TODO: Max memory should be a param? Now just 1 GB
-    int64_t GB = 1 << 30;
-#if defined(DEBUG_BERT)
-    printf("%s: requested_buf_size %lldMB\n", __func__, buf_size_new / (1 << 20));
-#endif
-    if (buf_size_new > GB) {
-        int32_t adjusted_new_size = GB / ctx->mem_per_input;
-        if (adjusted_new_size < 1) adjusted_new_size = 1;
-#if defined(DEBUG_BERT)
-        printf("%s: requested batch size %d, actual new batch size %d\n", __func__, new_size, adjusted_new_size);
-#endif
-        new_size = adjusted_new_size;
-        buf_size_new = ctx->mem_per_input * new_size;
-    }
-    if (new_size > ctx->max_batch_n) {
-        ctx->buf_compute.resize(buf_size_new);
-        ctx->max_batch_n = new_size;
-    }
-}
-
-void bert_eval(
-    struct bert_ctx *ctx,
-    int32_t n_threads,
-    const bert_vocab_id *raw_tokens,
-    int32_t n_tokens,
-    float *embeddings)
-{
-    const bert_model& model = ctx->model;
-    bool mem_req_mode = !embeddings;
-
-    // batch_embeddings is nullptr for the initial memory requirements run
-    if (!mem_req_mode && 1 > ctx->max_batch_n)
-        bert_resize_ctx(ctx, 1);
-
-    const int N = n_tokens;
-    const auto &tokens = raw_tokens;
-
-    const auto &hparams = model.hparams;
-
-    const int n_embd = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_max_tokens = hparams.n_max_tokens;
-    const int n_head = hparams.n_head;
-
-    const int d_head = n_embd / n_head;
-
-    std::vector<float> result;
-    if (N > n_max_tokens)
-    {
-        fprintf(stderr, "Too many tokens, maximum is %d\n", n_max_tokens);
-        return;
-    }
-
-    auto & mem_per_token = ctx->mem_per_token;
-    auto & buf_compute   = ctx->buf_compute;
-
-    struct ggml_init_params params = {
-        .mem_size = buf_compute.size,
-        .mem_buffer = buf_compute.addr,
-        .no_alloc = false,
-    };
-
-    struct ggml_context *ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = {};
-
-    // Embeddings. word_embeddings + token_type_embeddings + position_embeddings
-    struct ggml_tensor *token_layer = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(token_layer->data, tokens, N * ggml_element_size(token_layer));
-
-    struct ggml_tensor *token_types = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    ggml_set_zero(token_types);
-
-    struct ggml_tensor *positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    for (int i = 0; i < N; i++)
-    {
-        ggml_set_i32_1d(positions, i, i);
-    }
-
-    struct ggml_tensor *inpL = ggml_get_rows(ctx0, model.word_embeddings, token_layer);
-
-    inpL = ggml_add(ctx0,
-                    ggml_get_rows(ctx0, model.token_type_embeddings, token_types),
-                    inpL);
-    inpL = ggml_add(ctx0,
-                    ggml_get_rows(ctx0, model.position_embeddings, positions),
-                    inpL);
-
-    // embd norm
-    {
-        inpL = ggml_norm(ctx0, inpL, 1e-5f);
-
-        inpL = ggml_add(ctx0,
-                        ggml_mul(ctx0,
-                                 ggml_repeat(ctx0, model.ln_e_w, inpL),
-                                 inpL),
-                        ggml_repeat(ctx0, model.ln_e_b, inpL));
-    }
-    // layers
-    for (int il = 0; il < n_layer; il++)
-    {
-        struct ggml_tensor *cur = inpL;
-
-        // self-attention
-        {
-            struct ggml_tensor *Qcur = cur;
-            Qcur = ggml_reshape_3d(ctx0,
-                                   ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, Qcur),
-                                            ggml_mul_mat(ctx0, model.layers[il].q_w, Qcur)),
-                                   d_head, n_head, N);
-            struct ggml_tensor *Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-
-            struct ggml_tensor *Kcur = cur;
-            Kcur = ggml_reshape_3d(ctx0,
-                                   ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].k_b, Kcur),
-                                            ggml_mul_mat(ctx0, model.layers[il].k_w, Kcur)),
-                                   d_head, n_head, N);
-            struct ggml_tensor *K = ggml_permute(ctx0, Kcur, 0, 2, 1, 3);
-
-            struct ggml_tensor *Vcur = cur;
-            Vcur = ggml_reshape_3d(ctx0,
-                                   ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].v_b, Vcur),
-                                            ggml_mul_mat(ctx0, model.layers[il].v_w, Vcur)),
-                                   d_head, n_head, N);
-            struct ggml_tensor *V = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
-
-            struct ggml_tensor *KQ = ggml_mul_mat(ctx0, K, Q);
-            // KQ = soft_max(KQ / sqrt(head width))
-            KQ = ggml_soft_max(ctx0,
-                               ggml_scale(ctx0,
-                                          KQ,
-                                          ggml_new_f32(ctx0, 1.0f / sqrt((float)d_head))));
-
-            V = ggml_cont(ctx0, ggml_transpose(ctx0, V));
-            struct ggml_tensor *KQV = ggml_mul_mat(ctx0, V, KQ);
-            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-
-            cur = ggml_cpy(ctx0,
-                           KQV,
-                           ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-        }
-        // attention output
-        cur = ggml_add(ctx0,
-                       ggml_repeat(ctx0, model.layers[il].o_b, cur),
-                       ggml_mul_mat(ctx0, model.layers[il].o_w, cur));
-
-        // re-add the layer input
-        cur = ggml_add(ctx0, cur, inpL);
-
-        // attention norm
-        {
-            cur = ggml_norm(ctx0, cur, 1e-5f);
-
-            cur = ggml_add(ctx0,
-                           ggml_mul(ctx0,
-                                    ggml_repeat(ctx0, model.layers[il].ln_att_w, cur),
-                                    cur),
-                           ggml_repeat(ctx0, model.layers[il].ln_att_b, cur));
-        }
-        struct ggml_tensor *att_output = cur;
-        // intermediate_output = self.intermediate(attention_output)
-        cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
-        cur = ggml_add(ctx0,
-                       ggml_repeat(ctx0, model.layers[il].ff_i_b, cur),
-                       cur);
-        cur = ggml_gelu(ctx0, cur);
-
-        // layer_output = self.output(intermediate_output, attention_output)
-        cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
-        cur = ggml_add(ctx0,
-                       ggml_repeat(ctx0, model.layers[il].ff_o_b, cur),
-                       cur);
-        // attentions bypass the intermediate layer
-        cur = ggml_add(ctx0, att_output, cur);
-
-        // output norm
-        {
-            cur = ggml_norm(ctx0, cur, 1e-5f);
-
-            cur = ggml_add(ctx0,
-                           ggml_mul(ctx0,
-                                    ggml_repeat(ctx0, model.layers[il].ln_out_w, cur),
-                                    cur),
-                           ggml_repeat(ctx0, model.layers[il].ln_out_b, cur));
-        }
-        inpL = cur;
-    }
-    inpL = ggml_cont(ctx0, ggml_transpose(ctx0, inpL));
-    // pooler
-    struct ggml_tensor *sum = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, N, 1);
-    ggml_set_f32(sum, 1.0f / N);
-    inpL = ggml_mul_mat(ctx0, inpL, sum);
-
-    ggml_tensor *output = inpL;
-    // run the computation
-    ggml_build_forward_expand(&gf, output);
-    //ggml_graph_compute_g4a()
-    ggml_graph_compute_g4a(ctx->work_buf, &gf, n_threads);
-    //ggml_graph_compute(ctx0, &gf);
-
-
-    // float *dat = ggml_get_data_f32(output);
-    // pretty_print_tensor(dat, output->ne, output->nb, output->n_dims - 1, "");
-
-    #ifdef GGML_PERF
-        // print timing information per ggml operation (for debugging purposes)
-        // requires GGML_PERF to be defined
-        ggml_graph_print(&gf);
-    #endif
-
-    if (!mem_req_mode) {
-        memcpy(embeddings, (float *)ggml_get_data(output), sizeof(float) * n_embd);
-    } else {
-        mem_per_token = ggml_used_mem(ctx0) / N;
-    }
-
-    // printf("used_mem = %zu KB \n", ggml_used_mem(ctx0) / 1024);
-    // printf("mem_per_token = %zu KB \n", mem_per_token / 1024);
-
-    ggml_free(ctx0);
-}
-
-//
-// Loading and setup
-//
-
-void bert_free(bert_ctx * ctx) {
-    delete ctx;
-}
-
-struct bert_ctx * bert_load_from_file(const char *fname)
-{
-#if defined(DEBUG_BERT)
-    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname);
-#endif
-
-    bert_ctx * new_bert = new bert_ctx;
-#if defined(GGML_USE_KOMPUTE)
-    new_bert->buf_compute.force_cpu = true;
-    new_bert->work_buf.force_cpu = true;
-#endif
-
-    bert_model & model = new_bert->model;
-    bert_vocab & vocab = new_bert->vocab;
-
-    struct gguf_init_params params = {
-        /*.no_alloc = */ false,
-        /*.ctx      = */ &model.ctx,
-    };
-    gguf_context *ggufctx = gguf_init_from_file(fname, params);
-    if (!ggufctx) {
-        fprintf(stderr, "%s: gguf_init_from_file() failed\n", __func__);
-        return nullptr;
-    }
-
-    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
-
-    // print some standard metadata
-    {
-        int keyidx;
-
-        keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
-        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-    }
-
-    // check required metadata
-    {
-        // check model architecture kv
-        int keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx == -1) {
-            fprintf(stderr, "%s: gguf model architecture not found!\n", __func__);
-            return nullptr;
-        }
-        if (strcmp(gguf_get_val_str(ggufctx, keyidx), "bert") != 0) {
-            fprintf(stderr, "%s: model architecture not supported!\n", __func__);
-            return nullptr;
-        }
-    }
-
-    // load hparams
-    {
-        auto &hparams = model.hparams;
-
-        bool ok = false;
-        int keyidx;
-
-        do {
-            keyidx = gguf_find_key(ggufctx, "bert.context_length");
-            if (keyidx == -1) { break; }
-            hparams.n_max_tokens = gguf_get_val_u32(ggufctx, keyidx);
-
-            keyidx = gguf_find_key(ggufctx, "bert.embedding_length");
-            if (keyidx == -1) { break; }
-            hparams.n_embd = gguf_get_val_u32(ggufctx, keyidx);
-
-            keyidx = gguf_find_key(ggufctx, "bert.feed_forward_length");
-            if (keyidx == -1) { break; }
-            hparams.n_intermediate = gguf_get_val_u32(ggufctx, keyidx);
-
-            keyidx = gguf_find_key(ggufctx, "bert.attention.head_count");
-            if (keyidx == -1) { break; }
-            hparams.n_head = gguf_get_val_u32(ggufctx, keyidx);
-
-            keyidx = gguf_find_key(ggufctx, "bert.block_count");
-            if (keyidx == -1) { break; }
-            hparams.n_layer = gguf_get_val_u32(ggufctx, keyidx);
-
-            ok = true;
-        } while (false);
-
-        if (!ok) {
-            fprintf(stderr, "%s: required hparam missing!\n", __func__);
-            return nullptr;
-        }
-
-#if defined(DEBUG_BERT)
-        printf("%s: n_max_tokens   = %d\n", __func__, hparams.n_max_tokens);
-        printf("%s: n_embd         = %d\n", __func__, hparams.n_embd);
-        printf("%s: n_intermediate = %d\n", __func__, hparams.n_intermediate);
-        printf("%s: n_head         = %d\n", __func__, hparams.n_head);
-        printf("%s: n_layer        = %d\n", __func__, hparams.n_layer);
-#endif
-    }
-
-    // load vocab
-    {
-        auto & hparams = model.hparams;
-
-        int keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.model");
-        if (keyidx == -1) {
-            fprintf(stderr, "%s: tokenizer model not found!\n", __func__);
-            return nullptr;
-        }
-        if (strcmp(gguf_get_val_str(ggufctx, keyidx), "bert") != 0) {
-            fprintf(stderr, "%s: tokenizer model not supported!\n", __func__);
-            return nullptr;
-        }
-
-        int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
-        if (tokens_keyidx == -1) {
-            fprintf(stderr, "%s: bert tokenizer vocab not found!\n", __func__);
-            return nullptr;
-        }
-
-        hparams.n_vocab = gguf_get_arr_n(ggufctx, tokens_keyidx);
-        printf("%s: bert 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);
-
-            if (word[0] == '#' && word[1] == '#')
-            {
-                vocab.subword_token_to_id[word.substr(2)] = i;
-                vocab._id_to_subword_token[i] = word;
-            }
-
-            if (vocab.token_to_id.count(word) == 0)
-            {
-                vocab.token_to_id[word] = i;
-                vocab._id_to_token[i] = word;
-            }
-        }
-    }
-
-    auto &ctx = model.ctx;
-
-#if defined(DEBUG_BERT)
-    printf("%s: ggml ctx size = %6.2f MB\n", __func__, ggml_get_mem_size(ctx) / (1024.0 * 1024.0));
-#endif
-
-    // prepare memory for the weights
-    {
-        const int n_layer = model.hparams.n_layer;
-        model.layers.resize(n_layer);
-
-        model.word_embeddings       = ggml_get_tensor(ctx, "token_embd.weight");
-        model.token_type_embeddings = ggml_get_tensor(ctx, "token_types.weight");
-        model.position_embeddings   = ggml_get_tensor(ctx, "position_embd.weight");
-        model.ln_e_w                = ggml_get_tensor(ctx, "output_norm.weight");
-        model.ln_e_b                = ggml_get_tensor(ctx, "output_norm.bias");
-
-        auto name = [](int i, std::string n) {
-            static std::string key;
-            key = "blk." + std::to_string(i) + "." + n;
-            return key.c_str();
-        };
-
-        for (int i = 0; i < n_layer; ++i)
-        {
-            auto &layer = model.layers[i];
-
-            layer.ln_att_w = ggml_get_tensor(ctx, name(i, "attn_norm.weight"));
-            layer.ln_att_b = ggml_get_tensor(ctx, name(i, "attn_norm.bias"));
-            layer.ln_out_w = ggml_get_tensor(ctx, name(i, "ffn_norm.weight"));
-            layer.ln_out_b = ggml_get_tensor(ctx, name(i, "ffn_norm.bias"));
-            layer.q_w      = ggml_get_tensor(ctx, name(i, "attn_q.weight"));
-            layer.q_b      = ggml_get_tensor(ctx, name(i, "attn_q.bias"));
-            layer.k_w      = ggml_get_tensor(ctx, name(i, "attn_k.weight"));
-            layer.k_b      = ggml_get_tensor(ctx, name(i, "attn_k.bias"));
-            layer.v_w      = ggml_get_tensor(ctx, name(i, "attn_v.weight"));
-            layer.v_b      = ggml_get_tensor(ctx, name(i, "attn_v.bias"));
-            layer.o_w      = ggml_get_tensor(ctx, name(i, "attn_output.weight"));
-            layer.o_b      = ggml_get_tensor(ctx, name(i, "attn_output.bias"));
-            layer.ff_i_w   = ggml_get_tensor(ctx, name(i, "ffn_up.weight"));
-            layer.ff_i_b   = ggml_get_tensor(ctx, name(i, "ffn_up.bias"));
-            layer.ff_o_w   = ggml_get_tensor(ctx, name(i, "ffn_down.weight"));
-            layer.ff_o_b   = ggml_get_tensor(ctx, name(i, "ffn_down.bias"));
-        }
-    }
-
-    // Calculate space requirements for setting up context buffers later
-    {
-        bert_vocab_id tokens[] = {0, 1, 2, 3};
-        // TODO: We set the initial buffer size to 16MB and hope it's enough. Maybe there is a better way to do this?
-        new_bert->buf_compute.resize(16 * 1024 * 1024);
-        bert_eval(new_bert, 1, tokens, 4, nullptr);
-        new_bert->max_batch_n = 0;
-
-        // TODO: Max tokens should be a param?
-        int32_t N = new_bert->model.hparams.n_max_tokens;
-        new_bert->mem_per_input = 2.2 * (new_bert->mem_per_token * N); // add 10% to account for ggml object overhead
-
-    }
-#if defined(DEBUG_BERT)
-    printf("%s: mem_per_token %ld KB, mem_per_input %ld MB\n", __func__, new_bert->mem_per_token / (1 << 10), new_bert->mem_per_input / (1 << 20));
-#endif
-
-    return new_bert;
-}
-
-struct BertPrivate {
-    const std::string modelPath;
-    bool modelLoaded;
-    bert_ctx *ctx = nullptr;
-    int64_t n_threads = 0;
-};
-
-Bert::Bert() : d_ptr(new BertPrivate) {
-    d_ptr->modelLoaded = false;
-}
-
-Bert::~Bert() {
-    bert_free(d_ptr->ctx);
-}
-
-bool Bert::loadModel(const std::string &modelPath)
-{
-    d_ptr->ctx = bert_load_from_file(modelPath.c_str());
-    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    d_ptr->modelLoaded = d_ptr->ctx != nullptr;
-    fflush(stdout);
-    return true;
-}
-
-bool Bert::isModelLoaded() const
-{
-    return d_ptr->modelLoaded;
-}
-
-size_t Bert::requiredMem(const std::string &/*modelPath*/)
-{
-    return 0;
-}
-
-size_t Bert::stateSize() const
-{
-    return 0;
-}
-
-size_t Bert::saveState(uint8_t */*dest*/) const
-{
-    return 0;
-}
-
-size_t Bert::restoreState(const uint8_t */*src*/)
-{
-    return 0;
-}
-
-void Bert::setThreadCount(int32_t n_threads)
-{
-    d_ptr->n_threads = n_threads;
-}
-
-int32_t Bert::threadCount() const
-{
-    return d_ptr->n_threads;
-}
-
-std::vector<float> Bert::embedding(const std::string &text)
-{
-    const int overlap = 32;
-    const LLModel::Token clsToken = 101;
-    const size_t contextLength = bert_n_max_tokens(d_ptr->ctx);
-    typedef std::vector<LLModel::Token> TokenString;
-    TokenString tokens = ::bert_tokenize(d_ptr->ctx, text.c_str());
-#if defined(DEBUG_BERT)
-    std::cerr << "embedding: " << tokens.size()
-              << " contextLength " << contextLength
-              << "\n";
-#endif
-    std::vector<double> embeddingsSum(bert_n_embd(d_ptr->ctx), 0);
-    int embeddingsSumTotal = 0;
-    size_t start_pos = 0;
-    bool isFirstChunk = true;
-    while (start_pos < tokens.size()) {
-        TokenString chunk;
-        if (!isFirstChunk)
-            chunk.push_back(clsToken);
-        const size_t l = isFirstChunk ? contextLength : contextLength - 1;
-        if (tokens.size() - start_pos > l) {
-            chunk.insert(chunk.end(), tokens.begin() + start_pos, tokens.begin() + start_pos + l);
-            start_pos = start_pos + contextLength - overlap;
-        } else {
-            chunk.insert(chunk.end(), tokens.begin() + start_pos, tokens.end());
-            start_pos = tokens.size();
-        }
-#if defined(DEBUG_BERT)
-        std::cerr << "chunk length: " << chunk.size()
-            << " embeddingsSumTotal " << embeddingsSumTotal
-            << " contextLength " << contextLength
-            << " start_pos " << start_pos
-            << "\n";
-#endif
-        embeddingsSumTotal++;
-        std::vector<float> embeddings(bert_n_embd(d_ptr->ctx));
-        bert_eval(d_ptr->ctx, d_ptr->n_threads, chunk.data(), chunk.size(), embeddings.data());
-        std::transform(embeddingsSum.begin(), embeddingsSum.end(), embeddings.begin(), embeddingsSum.begin(), std::plus<float>());
-        isFirstChunk = false;
-    }
-
-    std::transform(embeddingsSum.begin(), embeddingsSum.end(), embeddingsSum.begin(), [embeddingsSumTotal](float num){ return num / embeddingsSumTotal; });
-    double magnitude = std::sqrt(std::inner_product(embeddingsSum.begin(), embeddingsSum.end(), embeddingsSum.begin(), 0.0));
-    for (auto &value : embeddingsSum)
-        value /= magnitude;
-    std::vector<float> finalEmbeddings(embeddingsSum.begin(), embeddingsSum.end());
-    return finalEmbeddings;
-}
-
-std::vector<LLModel::Token> Bert::tokenize(PromptContext &, const std::string &str) const
-{
-    return ::bert_tokenize(d_ptr->ctx, str.c_str());
-}
-
-LLModel::Token Bert::sampleToken(PromptContext &/*promptCtx*/) const
-{
-    return 999 /*!*/;
-}
-
-std::string Bert::tokenToString(Token id) const
-{
-    return bert_vocab_id_to_token(d_ptr->ctx, id);
-}
-
-bool Bert::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
-{
-    std::vector<float> embeddings(bert_n_embd(d_ptr->ctx));
-    int32_t cls = 101;
-    const bool useCLS = tokens.front() != cls;
-    if (useCLS) {
-        std::vector<int32_t> myTokens;
-        myTokens.push_back(cls);
-        myTokens.insert(myTokens.end(), tokens.begin(), tokens.end());
-        bert_eval(d_ptr->ctx, d_ptr->n_threads, myTokens.data(), myTokens.size(), embeddings.data());
-    } else
-        bert_eval(d_ptr->ctx, d_ptr->n_threads, tokens.data(), tokens.size(), embeddings.data());
-    ctx.n_past = 0; // bert does not store any context
-    return true;
-}
-
-int32_t Bert::contextLength() const
-{
-    return bert_n_max_tokens(d_ptr->ctx);
-}
-
-const std::vector<LLModel::Token> &Bert::endTokens() const
-{
-    static const std::vector<LLModel::Token> out = { 102 /*sep*/};
-    return out;
-}
-
-std::string get_arch_name(gguf_context *ctx_gguf) {
-    std::string arch_name;
-    const int kid = gguf_find_key(ctx_gguf, "general.architecture");
-    enum gguf_type ktype = gguf_get_kv_type(ctx_gguf, kid);
-    if (ktype != GGUF_TYPE_STRING) {
-        throw std::runtime_error("ERROR: Can't get general architecture from gguf file.");
-    }
-    return gguf_get_val_str(ctx_gguf, kid);
-}
-
-#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(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);
-    if (!ctx_gguf)
-        return false;
-
-    bool isValid = gguf_get_version(ctx_gguf) <= 3;
-    isValid = isValid && get_arch_name(ctx_gguf) == "bert";
-
-    gguf_free(ctx_gguf);
-    return isValid;
-}
-
-DLL_EXPORT LLModel *construct() {
-    return new Bert;
-}
-}

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

@@ -53,6 +53,8 @@ public:
     }
 };
 #else
+#include <algorithm>
+#include <filesystem>
 #include <string>
 #include <exception>
 #include <stdexcept>
@@ -75,7 +77,9 @@ 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/gptj.cpp

@@ -343,7 +343,14 @@ bool gptj_eval(
     };
 
     struct ggml_context * ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = {};
+    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));
@@ -370,8 +377,14 @@ bool gptj_eval(
 
         // self-attention
         {
-            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), n_past, 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), n_past, n_rot, 0, 0);
+            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
             {
@@ -382,8 +395,8 @@ bool gptj_eval(
                         (   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));
+                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)
@@ -401,11 +414,7 @@ bool gptj_eval(
             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);
@@ -502,22 +511,22 @@ bool gptj_eval(
     // logits -> probs
     //inpL = ggml_soft_max(ctx0, inpL);
 
-    ggml_build_forward_expand(&gf, inpL);
+    ggml_build_forward_expand(gf, inpL);
 
     // run the computation
     {
         std::unique_ptr<uint8_t []> data;
-        auto plan = ggml_graph_plan(&gf, n_threads);
+        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);
+        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);
@@ -663,7 +672,9 @@ 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;
@@ -671,19 +682,24 @@ size_t GPTJ::requiredMem(const std::string &modelPath) {
     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;
 
     // load the model
-    if (!gptj_model_load(modelPath, *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;
 }
 
@@ -721,8 +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);
 }
 
@@ -767,13 +785,15 @@ const std::vector<LLModel::Token> &GPTJ::endTokens() const
     return fres;
 }
 
-std::string get_arch_name(gguf_context *ctx_gguf) {
-    std::string arch_name;
+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("ERROR: Can't get general architecture from gguf file.");
-    }
+    if (ktype != GGUF_TYPE_STRING)
+        throw std::runtime_error("key general.architecture has wrong type");
+
     return gguf_get_val_str(ctx_gguf, kid);
 }
 
@@ -796,21 +816,29 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(const char * fname) {
+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);
-    if (!ctx_gguf)
-        return false;
 
-    bool isValid = gguf_get_version(ctx_gguf) <= 3;
-    isValid = isValid && get_arch_name(ctx_gguf) == "gptj";
+    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
+        }
+    }
 
     gguf_free(ctx_gguf);
-    return isValid;
+    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 +4,68 @@
 #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, std::string *unavail_reason) 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;
+    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

@@ -2,48 +2,54 @@
 #include "dlhandle.h"
 #include "sysinfo.h"
 
-#include <iostream>
-#include <string>
-#include <vector>
-#include <fstream>
-#include <filesystem>
 #include <cassert>
 #include <cstdlib>
-#include <sstream>
+#include <filesystem>
+#include <fstream>
+#include <iostream>
+#include <memory>
+#include <optional>
 #include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
 #ifdef _MSC_VER
 #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_))) {
@@ -53,14 +59,17 @@ 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(const char*)>("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)
@@ -69,24 +78,26 @@ 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");
 }
 
 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;
 
-        std::string impl_name_re = "(bert|llama|gptj|llamamodel-mainline)";
-        if (requires_avxonly()) {
+        std::string impl_name_re = "(gptj|llamamodel-mainline)-(cpu|metal|kompute|vulkan|cuda)";
+        if (cpu_supports_avx2() == 0) {
             impl_name_re += "-avxonly";
-        } else {
-            impl_name_re += "-(default|metal)";
         }
         std::regex re(impl_name_re);
         auto search_in_directory = [&](const std::string& paths) {
@@ -105,9 +116,8 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
                     // 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 (...) {}
                 }
@@ -122,68 +132,149 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
     return *libs;
 }
 
+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;
 
-        if (!i.m_magicMatch(fname)) continue;
-        return &i;
+        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) {
-        std::cerr << "LLModel ERROR: Could not find any implementations for build variant: " << buildVariant << "\n";
-    }
-    return nullptr;
+    if (!buildVariantMatched)
+        return nullptr;
+    if (!archName)
+        throw UnsupportedModelError("Unsupported file format");
+
+    throw BadArchError(std::move(*archName));
 }
 
-LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::string buildVariant) {
-    if (!has_at_least_minimal_hardware()) {
-        std::cerr << "LLModel ERROR: CPU does not support AVX\n";
+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;
     }
 
-    // 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(modelPath.c_str(), "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;
-                }
-            }
-        }
-    #endif
-
-    if (!impl) {
-        //TODO: Auto-detect CUDA/OpenCL
-        if (buildVariant == "auto") {
-            if (requires_avxonly()) {
-                buildVariant = "avxonly";
-            } else {
-                buildVariant = "default";
-            }
-        }
-        impl = implementation(modelPath.c_str(), buildVariant);
-        if (!impl) 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));
     }
 
-    // Construct and return llmodel implementation
-    auto fres = impl->m_construct();
-    fres->m_implementation = impl;
-    return fres;
+    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;
+}
+
+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
+    return devices;
+}
+
+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) {
@@ -193,3 +284,11 @@ 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(const char *fname, 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)(const char *fname);
+        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;
@@ -50,73 +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;
         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*/, std::string *unavail_reason = nullptr) {
+    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 initializeGPUDevice(int /*device*/) { return false; }
-    virtual bool hasGPUDevice() { return false; }
-    virtual bool usingGPUDevice() { return false; }
-    static std::vector<GPUDevice> availableGPUDevices();
+
+    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
@@ -124,6 +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 {
@@ -11,8 +14,6 @@ struct LLModelWrapper {
     ~LLModelWrapper() { delete llModel; }
 };
 
-thread_local static std::string last_error_message;
-
 llmodel_model llmodel_model_create(const char *model_path) {
     const char *error;
     auto fres = llmodel_model_create2(model_path, "auto", &error);
@@ -22,98 +23,89 @@ llmodel_model llmodel_model_create(const char *model_path) {
     return fres;
 }
 
-llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, const char **error) {
-    auto wrapper = new LLModelWrapper;
+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);
-        if (!wrapper->llModel) {
-            last_error_message = "Model format not supported (no matching implementation found)";
-        }
+        llModel = LLModel::Implementation::construct(model_path, backend);
     } catch (const std::exception& e) {
-        last_error_message = e.what();
+        llmodel_set_error(error, e.what());
+        return nullptr;
     }
 
-    if (!wrapper->llModel) {
-        delete std::exchange(wrapper, nullptr);
-        if (error) {
-            *error = last_error_message.c_str();
-        }
-    }
-    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);
@@ -124,14 +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
@@ -146,6 +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;
@@ -153,38 +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();
 }
 
@@ -198,56 +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

@@ -39,6 +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
@@ -47,9 +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;
 };
@@ -81,6 +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
@@ -93,11 +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 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, const char **error);
+llmodel_model llmodel_model_create2(const char *model_path, const char *backend, const char **error);
 
 /**
  * Destroy a llmodel instance.
@@ -110,17 +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.
@@ -159,29 +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.
@@ -219,9 +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.
@@ -261,6 +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);
         }
@@ -98,7 +198,11 @@ void LLModel::prompt(const std::string &prompt,
         }
         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);
         }
@@ -169,34 +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

@@ -4,50 +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;
-    bool force_cpu = false;
-
-    llm_buffer() = default;
-
-    void resize(size_t size) {
-        free();
-
-        if (!ggml_vk_has_device() || force_cpu) {
-            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;
@@ -62,7 +18,6 @@ struct llm_buffer {
         delete[] addr;
     }
 };
-#endif
 
 struct llm_kv_cache {
     struct ggml_tensor * k;

gpt4all-backend/scripts/convert_gptj_to_gguf.py

@@ -27,7 +27,7 @@ from pathlib import Path
 
 import gguf
 import numpy as np
-from transformers import AutoTokenizer, GPTJConfig, GPTJForCausalLM
+from transformers import AutoConfig, AutoTokenizer, GPTJForCausalLM
 from transformers.models.gpt2 import tokenization_gpt2
 
 
@@ -63,7 +63,7 @@ gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
-config = GPTJConfig(dir_model)
+config = AutoConfig.from_pretrained(dir_model)
 
 block_count = config.n_layer
 gguf_writer.add_name("GPT-J")

gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py

@@ -1,168 +0,0 @@
-#!/usr/bin/env python3
-# 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"
-#
-
-from __future__ import annotations
-
-import json
-import struct
-import sys
-from pathlib import Path
-
-import gguf
-import numpy as np
-import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, MptConfig
-from transformers.models.gpt2 import tokenization_gpt2
-
-
-if not 3 <= len(sys.argv) < 5:
-    print("Usage: {} model-name dir-output [ftype]".format(Path(__file__).name))
-    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("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-    sys.exit(1)
-
-dir_model = Path(sys.argv[1])
-dir_out = Path(sys.argv[2])
-
-# make sure the output directory exists
-dir_out.mkdir(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 = int(sys.argv[3])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-        sys.exit(1)
-
-fname_out = dir_out / f"ggml-model-{dir_model.name}-{ftype_str[ftype]}.gguf"
-
-
-ARCH = gguf.MODEL_ARCH.MPT
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-config = AutoConfig.from_pretrained(dir_model, trust_remote_code=True)
-
-block_count = config.n_layers
-gguf_writer.add_name("MPT")
-gguf_writer.add_context_length(config.max_seq_len)
-gguf_writer.add_embedding_length(config.d_model)
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(4 * config.d_model)
-gguf_writer.add_head_count(config.n_heads)
-if kv_n_heads := config.attn_config.get('kv_n_heads'):
-    gguf_writer.add_head_count_kv(kv_n_heads)
-gguf_writer.add_max_alibi_bias(config.attn_config['alibi_bias_max'])
-gguf_writer.add_layer_norm_eps(MptConfig().layer_norm_epsilon)  # use default from upstream transformers
-gguf_writer.add_file_type(ftype)
-
-clip_qkv = config.attn_config['clip_qkv']
-if clip_qkv is not None:
-    gguf_writer.add_clamp_kqv(clip_qkv)
-
-print("gguf: get gpt2 tokenizer vocab")
-
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-special_ids = tokenizer.all_special_ids
-
-reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-added_tokens = tokenizer.get_added_vocab().values()
-byte_encoder = tokenization_gpt2.bytes_to_unicode()
-byte_decoder = {v: k for k, v in byte_encoder.items()}
-
-tokens: list[bytearray] = []
-toktypes: list[gguf.TokenType] = []
-
-# 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):
-    if i not in reverse_vocab:
-        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-        pad_token = f"[PAD{i}]".encode("utf8")
-        text = bytearray(pad_token)
-    elif i in added_tokens:
-        # these tokens are not encoded, for some reason
-        text = bytearray(reverse_vocab[i].encode('utf-8'))
-    else:
-        text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-
-    tokens.append(text)
-
-    # TODO(cebtenzzre): is there a better way to do this?
-    toktypes.append(gguf.TokenType.CONTROL if i in special_ids else gguf.TokenType.NORMAL)
-
-gguf_writer.add_tokenizer_model("gpt2")
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
-special_vocab.add_to_gguf(gguf_writer)
-
-print("gguf: get tensor metadata")
-
-print("Loading model:", dir_model)
-model = AutoModelForCausalLM.from_pretrained(
-    dir_model, config=config, torch_dtype=torch.float16 if ftype == 1 else torch.float32,
-    low_cpu_mem_usage=True, trust_remote_code=True,
-)
-print("Model loaded:", dir_model)
-
-tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
-
-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
-    # Keep token embeddings in fp32
-    if ftype == 1 and name[-7:] == ".weight" and n_dims == 2 and ".wte" not in name:
-        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_replit_v1_hf_to_gguf.py

@@ -1,145 +0,0 @@
-#!/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 sentencepiece import SentencePieceProcessor
-from transformers import AutoConfig, AutoModelForCausalLM, 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])
-
-# 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-replit-code-v1-3b-" + ftype_str[ftype] + ".gguf")
-
-
-ARCH = gguf.MODEL_ARCH.MPT
-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_layers
-gguf_writer.add_name("Replit")
-gguf_writer.add_context_length(config.max_seq_len)
-gguf_writer.add_embedding_length(config.d_model)
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(4 * config.d_model)
-gguf_writer.add_head_count(config.n_heads)
-gguf_writer.add_max_alibi_bias(config.attn_config.alibi_bias_max)
-gguf_writer.add_layer_norm_eps(config.layer_norm_epsilon)
-gguf_writer.add_file_type(ftype)
-
-clip_qkv = config.attn_config.clip_qkv
-if clip_qkv is not None:
-    gguf_writer.add_clamp_kqv(clip_qkv)
-
-print("gguf: get sentencepiece tokenizer vocab")
-
-tokenizer = SentencePieceProcessor(str(dir_model / "spiece.model"))
-#print(tokenizer.encode('I believe the meaning of life is'))
-
-tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
-
-for i in range(tokenizer.vocab_size()):
-    tokens.append(tokenizer.id_to_piece(i).encode('utf-8'))
-    scores.append(tokenizer.get_score(i))
-
-    toktype = gguf.TokenType.NORMAL
-    if tokenizer.is_unknown(i):
-        toktype = gguf.TokenType.UNKNOWN
-    elif tokenizer.is_control(i):
-        toktype = gguf.TokenType.CONTROL
-    elif tokenizer.is_unused(i):
-        toktype = gguf.TokenType.UNUSED
-    elif tokenizer.is_byte(i):
-        toktype = gguf.TokenType.BYTE
-
-    toktypes.append(toktype)
-
-gguf_writer.add_tokenizer_model("llama")  # sentencepiece
-gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
-
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
-special_vocab.add_to_gguf(gguf_writer)
-
-print("gguf: get tensor metadata")
-
-model = AutoModelForCausalLM.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)
-
-print(config)
-
-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 == 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-bindings/cli/app.py

@@ -59,9 +59,13 @@ def repl(
         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:
@@ -116,6 +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,
@@ -152,6 +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
 ```

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

@@ -23,7 +23,7 @@ void* load_model(const char *fname, int n_threads) {
         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;
     }
@@ -35,8 +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;
 
@@ -69,6 +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,
@@ -83,14 +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

gpt4all-bindings/golang/example/main.go

@@ -47,7 +47,7 @@ func main() {
 	for {
 		text := readMultiLineInput(reader)
 
-		_, err := l.Predict(text, gpt4all.SetTokens(tokens), gpt4all.SetTopK(90), gpt4all.SetTopP(0.86))
+		_, err := l.Predict(text, "", "", gpt4all.SetTokens(tokens), gpt4all.SetTopK(90), gpt4all.SetTopP(0.86))
 		if err != nil {
 			panic(err)
 		}

gpt4all-bindings/golang/gpt4all.go

@@ -6,8 +6,8 @@ package gpt4all
 // #cgo darwin CXXFLAGS: -std=c++17
 // #cgo LDFLAGS: -lgpt4all -lm -lstdc++ -ldl
 // 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);
 // extern unsigned char getTokenCallback(void *, char *);
 // void llmodel_set_implementation_search_path(const char *path);
@@ -47,7 +47,7 @@ func New(model string, opts ...ModelOption) (*Model, error) {
 	return gpt, nil
 }
 
-func (l *Model) Predict(text string, opts ...PredictOption) (string, error) {
+func (l *Model) Predict(text, template, fakeReplyText string, opts ...PredictOption) (string, error) {
 
 	po := NewPredictOptions(opts...)
 
@@ -55,10 +55,14 @@ func (l *Model) Predict(text string, opts ...PredictOption) (string, error) {
 	if po.Tokens == 0 {
 		po.Tokens = 99999999
 	}
+	templateInput := C.CString(template)
+	fakeReplyInput := C.CString(fakeReplyText)
 	out := make([]byte, po.Tokens)
 
-	C.model_prompt(input, l.state, (*C.char)(unsafe.Pointer(&out[0])), C.int(po.RepeatLastN), C.float(po.RepeatPenalty), C.int(po.ContextSize),
-		C.int(po.Tokens), C.int(po.TopK), C.float(po.TopP), C.float(po.Temperature), C.int(po.Batch), C.float(po.ContextErase))
+	C.model_prompt(input, templateInput, C.int(po.Special), fakeReplyInput, l.state, (*C.char)(unsafe.Pointer(&out[0])),
+		C.int(po.RepeatLastN), C.float(po.RepeatPenalty), C.int(po.ContextSize), C.int(po.Tokens),
+		C.int(po.TopK), C.float(po.TopP), C.float(po.MinP), C.float(po.Temperature), C.int(po.Batch),
+		C.float(po.ContextErase))
 
 	res := C.GoString((*C.char)(unsafe.Pointer(&out[0])))
 	res = strings.TrimPrefix(res, " ")

gpt4all-bindings/golang/options.go

@@ -1,8 +1,8 @@
 package gpt4all
 
 type PredictOptions struct {
-	ContextSize, RepeatLastN, Tokens, TopK, Batch  int
-	TopP, Temperature, ContextErase, RepeatPenalty float64
+	ContextSize, RepeatLastN, Tokens, TopK, Batch, Special int
+	TopP, MinP, Temperature, ContextErase, RepeatPenalty   float64
 }
 
 type PredictOption func(p *PredictOptions)
@@ -11,8 +11,10 @@ var DefaultOptions PredictOptions = PredictOptions{
 	Tokens:        200,
 	TopK:          10,
 	TopP:          0.90,
+	MinP:          0.0,
 	Temperature:   0.96,
 	Batch:         1,
+	Special:       0,
 	ContextErase:  0.55,
 	ContextSize:   1024,
 	RepeatLastN:   10,
@@ -50,6 +52,13 @@ func SetTopP(topp float64) PredictOption {
 	}
 }
 
+// SetMinP sets the value for min p sampling
+func SetMinP(minp float64) PredictOption {
+	return func(p *PredictOptions) {
+		p.MinP = minp
+	}
+}
+
 // SetRepeatPenalty sets the repeat penalty.
 func SetRepeatPenalty(ce float64) PredictOption {
 	return func(p *PredictOptions) {
@@ -85,6 +94,17 @@ func SetBatch(size int) PredictOption {
 	}
 }
 
+// SetSpecial is true if special tokens in the prompt should be processed, false otherwise.
+func SetSpecial(special bool) PredictOption {
+	return func(p *PredictOptions) {
+		if special {
+			p.Special = 1
+		} else {
+			p.Special = 0
+		}
+	}
+}
+
 // Create a new PredictOptions object with the given options.
 func NewPredictOptions(opts ...PredictOption) PredictOptions {
 	p := DefaultOptions

gpt4all-bindings/java/src/main/java/com/hexadevlabs/gpt4all/LLModel.java

@@ -32,6 +32,7 @@ public  class LLModel implements AutoCloseable {
             n_predict.set(128);
             top_k.set(40);
             top_p.set(0.95);
+            min_p.set(0.0);
             temp.set(0.28);
             n_batch.set(8);
             repeat_penalty.set(1.1);
@@ -71,6 +72,11 @@ public  class LLModel implements AutoCloseable {
                 return this;
             }
 
+            public Builder withMinP(float min_p) {
+                configToBuild.min_p.set(min_p);
+                return this;
+            }
+
             public Builder withTemp(float temp) {
                 configToBuild.temp.set(temp);
                 return this;
@@ -195,7 +201,7 @@ public  class LLModel implements AutoCloseable {
         if(model == null) {
             throw new IllegalStateException("Could not load, gpt4all backend returned error: " + error.getValue().getString(0));
         }
-        library.llmodel_loadModel(model, modelPathAbs);
+        library.llmodel_loadModel(model, modelPathAbs, 2048, 100);
 
         if(!library.llmodel_isModelLoaded(model)){
             throw new IllegalStateException("The model " + modelName + " could not be loaded");

gpt4all-bindings/java/src/main/java/com/hexadevlabs/gpt4all/LLModelLibrary.java

@@ -48,6 +48,7 @@ public interface LLModelLibrary {
         public final int32_t n_predict = new int32_t();
         public final int32_t top_k = new int32_t();
         public final Float top_p = new Float();
+        public final Float min_p = new Float();
         public final Float temp = new Float();
         public final int32_t n_batch = new int32_t();
         public final Float repeat_penalty = new Float();
@@ -61,7 +62,7 @@ public interface LLModelLibrary {
 
     Pointer llmodel_model_create2(String model_path, String build_variant, PointerByReference error);
     void llmodel_model_destroy(Pointer model);
-    boolean llmodel_loadModel(Pointer model, String model_path);
+    boolean llmodel_loadModel(Pointer model, String model_path, int n_ctx, int ngl);
     boolean llmodel_isModelLoaded(Pointer model);
     @u_int64_t long llmodel_get_state_size(Pointer model);
     @u_int64_t long llmodel_save_state_data(Pointer model, Pointer dest);

gpt4all-bindings/python/README.md

@@ -9,39 +9,52 @@ https://docs.gpt4all.io/gpt4all_python.html
 
 ## Installation
 
+The easiest way to install the Python bindings for GPT4All is to use pip:
+
 ```
 pip install gpt4all
 ```
 
-## Local Build Instructions
+This will download the latest version of the `gpt4all` package from PyPI.
+
+## Local Build
+
+As an alternative to downloading via pip, you may build the Python bindings from source.
 
 ### Prerequisites
 
-On Windows and Linux, building GPT4All requires the complete Vulkan SDK. You may download it from here: https://vulkan.lunarg.com/sdk/home
+You will need a compiler. On Windows, you should install Visual Studio with the C++ Development components. On macOS, you will need the full version of Xcode—Xcode Command Line Tools lacks certain required tools. On Linux, you will need a GCC or Clang toolchain with C++ support.
 
-macOS users do not need Vulkan, as GPT4All will use Metal instead.
+On Windows and Linux, building GPT4All with full GPU support requires the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home) and the latest [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
 
 ### Building the python bindings
 
-**NOTE**: If you are doing this on a Windows machine, you must build the GPT4All backend using [MinGW64](https://www.mingw-w64.org/) compiler.
-
-1. Setup `llmodel`
-
+1. Clone GPT4All and change directory:
 ```
 git clone --recurse-submodules https://github.com/nomic-ai/gpt4all.git
-cd gpt4all/gpt4all-backend/
-mkdir build
-cd build
-cmake ..
-cmake --build . --parallel  # optionally append: --config Release
+cd gpt4all/gpt4all-backend
 ```
-Confirm that `libllmodel.*` exists in `gpt4all-backend/build`.
 
-2. Setup Python package
+2. Build the backend.
 
+If you are using Windows and have Visual Studio installed:
+```
+cmake -B build
+cmake --build build --parallel --config RelWithDebInfo
+```
+
+For all other platforms:
+```
+cmake -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo
+cmake --build build --parallel
+```
+
+`RelWithDebInfo` is a good default, but you can also use `Release` or `Debug` depending on the situation.
+
+2. Install the Python package:
 ```
 cd ../../gpt4all-bindings/python
-pip3 install -e .
+pip install -e .
 ```
 
 ## Usage

gpt4all-bindings/python/docs/gpt4all_chat.md

@@ -7,7 +7,7 @@ It is optimized to run 7-13B parameter LLMs on the CPU's of any computer running
 ## Running LLMs on CPU
 The GPT4All Chat UI supports models from all newer versions of `llama.cpp` with `GGUF` models including the `Mistral`, `LLaMA2`, `LLaMA`, `OpenLLaMa`, `Falcon`, `MPT`, `Replit`, `Starcoder`, and `Bert` architectures
 
-GPT4All maintains an official list of recommended models located in [models2.json](https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models2.json). You can pull request new models to it and if accepted they will show up in the official download dialog.
+GPT4All maintains an official list of recommended models located in [models3.json](https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models3.json). You can pull request new models to it and if accepted they will show up in the official download dialog.
 
 #### Sideloading any GGUF model
 If a model is compatible with the gpt4all-backend, you can sideload it into GPT4All Chat by:
@@ -61,17 +61,7 @@ The general technique this plugin uses is called [Retrieval Augmented Generation
 These document chunks help your LLM respond to queries with knowledge about the contents of your data.
 The number of chunks and the size of each chunk can be configured in the LocalDocs plugin settings tab.
 
-LocalDocs supports the following file types:
-```json
-["txt", "doc", "docx", "pdf", "rtf", "odt", "html", "htm", "xls", "xlsx", "csv", "ods", "ppt", "pptx", "odp", "xml", "json", "log", "md", "org", "tex", "asc", "wks",
-"wpd", "wps", "wri", "xhtml", "xht", "xslt", "yaml", "yml", "dtd", "sgml", "tsv", "strings", "resx",
-"plist", "properties", "ini", "config", "bat", "sh", "ps1", "cmd", "awk", "sed", "vbs", "ics", "mht",
-"mhtml", "epub", "djvu", "azw", "azw3", "mobi", "fb2", "prc", "lit", "lrf", "tcr", "pdb", "oxps",
-"xps", "pages", "numbers", "key", "keynote", "abw", "zabw", "123", "wk1", "wk3", "wk4", "wk5", "wq1",
-"wq2", "xlw", "xlr", "dif", "slk", "sylk", "wb1", "wb2", "wb3", "qpw", "wdb", "wks", "wku", "wr1",
-"wrk", "xlk", "xlt", "xltm", "xltx", "xlsm", "xla", "xlam", "xll", "xld", "xlv", "xlw", "xlc", "xlm",
-"xlt", "xln"]
-```
+LocalDocs currently supports plain text files (`.txt`, `.md`, and `.rst`) and PDF files (`.pdf`).
 
 #### Troubleshooting and FAQ
 *My LocalDocs plugin isn't using my documents*

gpt4all-bindings/python/docs/gpt4all_cli.md

@@ -5,7 +5,7 @@ The GPT4All command-line interface (CLI) is a Python script which is built on to
 package. The source code, README, and local build instructions can be found
 [here][repo-bindings-cli].
 
-[docs-bindings-python]: gpt4all_python.html
+[docs-bindings-python]: gpt4all_python.md
 [repo-bindings-python]: https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/python
 [repo-bindings-cli]: https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/cli
 [typer]: https://typer.tiangolo.com/

gpt4all-bindings/python/docs/gpt4all_faq.md

@@ -61,12 +61,12 @@ or `allowDownload=true` (default), a model is automatically downloaded into `.ca
 unless it already exists.
 
 In case of connection issues or errors during the download, you might want to manually verify the model file's MD5
-checksum by comparing it with the one listed in [models2.json].
+checksum by comparing it with the one listed in [models3.json].
 
 As an alternative to the basic downloader built into the bindings, you can choose to download from the 
 <https://gpt4all.io/> website instead. Scroll down to 'Model Explorer' and pick your preferred model.
 
-[models2.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models2.json
+[models3.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models3.json
 
 #### I need the chat GUI and bindings to behave the same
 
@@ -93,7 +93,7 @@ The chat GUI and bindings are based on the same backend. You can make them behav
 - Next you'll have to compare the templates, adjusting them as necessary, based on how you're using the bindings.
     - Specifically, in Python:
         - With simple `generate()` calls, the input has to be surrounded with system and prompt templates.
-        - When using a chat session, it depends on whether the bindings are allowed to download [models2.json]. If yes,
+        - When using a chat session, it depends on whether the bindings are allowed to download [models3.json]. If yes,
           and in the chat GUI the default templates are used, it'll be handled automatically. If no, use
           `chat_session()` template parameters to customize them.
 

gpt4all-bindings/python/docs/gpt4all_modal.md

@@ -1,34 +0,0 @@
-# GPT4All with Modal Labs
-
-You can easily query any GPT4All model on [Modal Labs](https://modal.com/) infrastructure!
-## Example
-
-```python
-import modal
-
-def download_model():
-    import gpt4all
-    #you can use any model from https://gpt4all.io/models/models2.json
-    return gpt4all.GPT4All("ggml-gpt4all-j-v1.3-groovy.bin")
-
-image=modal.Image.debian_slim().pip_install("gpt4all").run_function(download_model)
-stub = modal.Stub("gpt4all", image=image)
-@stub.cls(keep_warm=1)
-class GPT4All:
-    def __enter__(self):
-        print("Downloading model")
-        self.gptj = download_model()
-        print("Loaded model")
-
-    @modal.method()
-    def generate(self):
-        messages = [{"role": "user", "content": "Name 3 colors"}]
-        completion = self.gptj.chat_completion(messages)
-        print(f"Completion: {completion}")
-
-@stub.local_entrypoint()
-def main():
-    model = GPT4All()
-    for i in range(10):
-        model.generate.call()
-```

gpt4all-bindings/python/docs/gpt4all_nodejs.md

@@ -1,44 +1,126 @@
 # GPT4All Node.js API
 
+Native Node.js LLM bindings for all.
+
 ```sh
-yarn add gpt4all@alpha
+yarn add gpt4all@latest
 
-npm install gpt4all@alpha
+npm install gpt4all@latest
+
+pnpm install gpt4all@latest
 
-pnpm install gpt4all@alpha
 ```
 
-The original [GPT4All typescript bindings](https://github.com/nomic-ai/gpt4all-ts) are now out of date.
+## Contents
 
-*   New bindings created by [jacoobes](https://github.com/jacoobes), [limez](https://github.com/iimez) and the [nomic ai community](https://home.nomic.ai), for all to use.
-*   The nodejs api has made strides to mirror the python api. It is not 100% mirrored, but many pieces of the api resemble its python counterpart.
-*   Everything should work out the box.
 *   See [API Reference](#api-reference)
+*   See [Examples](#api-example)
+*   See [Developing](#develop)
+*   GPT4ALL nodejs bindings created by [jacoobes](https://github.com/jacoobes), [limez](https://github.com/iimez) and the [nomic ai community](https://home.nomic.ai), for all to use.
 
-### Chat Completion (alpha)
+## Api Example
+
+### Chat Completion
 
 ```js
-import { createCompletion, loadModel } from '../src/gpt4all.js'
+import { LLModel, createCompletion, DEFAULT_DIRECTORY, DEFAULT_LIBRARIES_DIRECTORY, loadModel } from '../src/gpt4all.js'
 
-const model = await loadModel('ggml-vicuna-7b-1.1-q4_2', { verbose: true });
+const model = await loadModel( 'mistral-7b-openorca.gguf2.Q4_0.gguf', { verbose: true, device: 'gpu' });
 
-const response = await createCompletion(model, [
-    { role : 'system', content: 'You are meant to be annoying and unhelpful.'  },
-    { role : 'user', content: 'What is 1 + 1?'  } 
-]);
+const completion1 = await createCompletion(model, 'What is 1 + 1?', { verbose: true, })
+console.log(completion1.message)
+
+const completion2 = await createCompletion(model, 'And if we add two?', {  verbose: true  })
+console.log(completion2.message)
+
+model.dispose()
+```
+
+### Embedding
+
+```js
+import { loadModel, createEmbedding } from '../src/gpt4all.js'
+
+const embedder = await loadModel("all-MiniLM-L6-v2-f16.gguf", { verbose: true, type: 'embedding'})
+
+console.log(createEmbedding(embedder, "Maybe Minecraft was the friends we made along the way"));
+```
+
+### Chat Sessions
+
+```js
+import { loadModel, createCompletion } from "../src/gpt4all.js";
+
+const model = await loadModel("orca-mini-3b-gguf2-q4_0.gguf", {
+    verbose: true,
+    device: "gpu",
+});
+
+const chat = await model.createChatSession();
+
+await createCompletion(
+    chat,
+    "Why are bananas rather blue than bread at night sometimes?",
+    {
+        verbose: true,
+    }
+);
+await createCompletion(chat, "Are you sure?", { verbose: true, });
 
 ```
 
-### Embedding (alpha)
+### Streaming responses
 
 ```js
-import { createEmbedding, loadModel } from '../src/gpt4all.js'
+import gpt from "../src/gpt4all.js";
 
-const model = await loadModel('ggml-all-MiniLM-L6-v2-f16', { verbose: true });
+const model = await gpt.loadModel("mistral-7b-openorca.gguf2.Q4_0.gguf", {
+    device: "gpu",
+});
 
-const fltArray = createEmbedding(model, "Pain is inevitable, suffering optional");
+process.stdout.write("### Stream:");
+const stream = gpt.createCompletionStream(model, "How are you?");
+stream.tokens.on("data", (data) => {
+    process.stdout.write(data);
+});
+//wait till stream finishes. We cannot continue until this one is done.
+await stream.result;
+process.stdout.write("\n");
+
+process.stdout.write("### Stream with pipe:");
+const stream2 = gpt.createCompletionStream(
+    model,
+    "Please say something nice about node streams."
+);
+stream2.tokens.pipe(process.stdout);
+await stream2.result;
+process.stdout.write("\n");
+
+console.log("done");
+model.dispose();
 ```
 
+### Async Generators
+
+```js
+import gpt from "../src/gpt4all.js";
+
+const model = await gpt.loadModel("mistral-7b-openorca.gguf2.Q4_0.gguf", {
+    device: "gpu",
+});
+
+process.stdout.write("### Generator:");
+const gen = gpt.createCompletionGenerator(model, "Redstone in Minecraft is Turing Complete. Let that sink in. (let it in!)");
+for await (const chunk of gen) {
+    process.stdout.write(chunk);
+}
+
+process.stdout.write("\n");
+model.dispose();
+```
+
+## Develop
+
 ### Build Instructions
 
 *   binding.gyp is compile config
@@ -82,8 +164,6 @@ yarn
 git submodule update --init --depth 1 --recursive
 ```
 
-**AS OF NEW BACKEND** to build the backend,
-
 ```sh
 yarn build:backend
 ```
@@ -130,21 +210,27 @@ yarn test
 
 *   why your model may be spewing bull 💩
     *   The downloaded model is broken (just reinstall or download from official site)
-    *   That's it so far
+*   Your model is hanging after a call to generate tokens.
+    *   Is `nPast` set too high? This may cause your model to hang (03/16/2024), Linux Mint, Ubuntu 22.04
+*   Your GPU usage is still high after node.js exits.
+    *   Make sure to call `model.dispose()`!!!
 
 ### Roadmap
 
-This package is in active development, and breaking changes may happen until the api stabilizes. Here's what's the todo list:
+This package has been stabilizing over time development, and breaking changes may happen until the api stabilizes. Here's what's the todo list:
 
+*   \[ ] Purely offline. Per the gui, which can be run completely offline, the bindings should be as well.
+*   \[ ] NPM bundle size reduction via optionalDependencies strategy (need help)
+    *   Should include prebuilds to avoid painful node-gyp errors
+*   \[x] createChatSession ( the python equivalent to create\_chat\_session )
+*   \[x] generateTokens, the new name for createTokenStream. As of 3.2.0, this is released but not 100% tested. Check spec/generator.mjs!
+*   \[x] ~~createTokenStream, an async iterator that streams each token emitted from the model. Planning on following this [example](https://github.com/nodejs/node-addon-examples/tree/main/threadsafe-async-iterator)~~ May not implement unless someone else can complete
 *   \[x] prompt models via a threadsafe function in order to have proper non blocking behavior in nodejs
-*   \[ ] ~~createTokenStream, an async iterator that streams each token emitted from the model. Planning on following this [example](https://github.com/nodejs/node-addon-examples/tree/main/threadsafe-async-iterator)~~ May not implement unless someone else can complete
+*   \[x] generateTokens is the new name for this^
 *   \[x] proper unit testing (integrate with circle ci)
 *   \[x] publish to npm under alpha tag `gpt4all@alpha`
 *   \[x] have more people test on other platforms (mac tester needed)
 *   \[x] switch to new pluggable backend
-*   \[ ] NPM bundle size reduction via optionalDependencies strategy (need help)
-    *   Should include prebuilds to avoid painful node-gyp errors
-*   \[ ] createChatSession ( the python equivalent to create\_chat\_session )
 
 ### API Reference
 
@@ -152,126 +238,219 @@ This package is in active development, and breaking changes may happen until the
 
 ##### Table of Contents
 
-*   [ModelType](#modeltype)
-*   [ModelFile](#modelfile)
-    *   [gptj](#gptj)
-    *   [llama](#llama)
-    *   [mpt](#mpt)
-    *   [replit](#replit)
 *   [type](#type)
+*   [TokenCallback](#tokencallback)
+*   [ChatSessionOptions](#chatsessionoptions)
+    *   [systemPrompt](#systemprompt)
+    *   [messages](#messages)
+*   [initialize](#initialize)
+    *   [Parameters](#parameters)
+*   [generate](#generate)
+    *   [Parameters](#parameters-1)
+*   [InferenceModel](#inferencemodel)
+    *   [createChatSession](#createchatsession)
+        *   [Parameters](#parameters-2)
+    *   [generate](#generate-1)
+        *   [Parameters](#parameters-3)
+    *   [dispose](#dispose)
+*   [EmbeddingModel](#embeddingmodel)
+    *   [dispose](#dispose-1)
+*   [InferenceResult](#inferenceresult)
 *   [LLModel](#llmodel)
     *   [constructor](#constructor)
-        *   [Parameters](#parameters)
+        *   [Parameters](#parameters-4)
     *   [type](#type-1)
     *   [name](#name)
     *   [stateSize](#statesize)
     *   [threadCount](#threadcount)
     *   [setThreadCount](#setthreadcount)
-        *   [Parameters](#parameters-1)
-    *   [raw\_prompt](#raw_prompt)
-        *   [Parameters](#parameters-2)
+        *   [Parameters](#parameters-5)
+    *   [infer](#infer)
+        *   [Parameters](#parameters-6)
     *   [embed](#embed)
-        *   [Parameters](#parameters-3)
+        *   [Parameters](#parameters-7)
     *   [isModelLoaded](#ismodelloaded)
     *   [setLibraryPath](#setlibrarypath)
-        *   [Parameters](#parameters-4)
+        *   [Parameters](#parameters-8)
     *   [getLibraryPath](#getlibrarypath)
-*   [loadModel](#loadmodel)
-    *   [Parameters](#parameters-5)
-*   [createCompletion](#createcompletion)
-    *   [Parameters](#parameters-6)
-*   [createEmbedding](#createembedding)
-    *   [Parameters](#parameters-7)
-*   [CompletionOptions](#completionoptions)
+    *   [initGpuByString](#initgpubystring)
+        *   [Parameters](#parameters-9)
+    *   [hasGpuDevice](#hasgpudevice)
+    *   [listGpu](#listgpu)
+        *   [Parameters](#parameters-10)
+    *   [dispose](#dispose-2)
+*   [GpuDevice](#gpudevice)
+    *   [type](#type-2)
+*   [LoadModelOptions](#loadmodeloptions)
+    *   [modelPath](#modelpath)
+    *   [librariesPath](#librariespath)
+    *   [modelConfigFile](#modelconfigfile)
+    *   [allowDownload](#allowdownload)
     *   [verbose](#verbose)
-    *   [systemPromptTemplate](#systemprompttemplate)
-    *   [promptTemplate](#prompttemplate)
-    *   [promptHeader](#promptheader)
-    *   [promptFooter](#promptfooter)
-*   [PromptMessage](#promptmessage)
+    *   [device](#device)
+    *   [nCtx](#nctx)
+    *   [ngl](#ngl)
+*   [loadModel](#loadmodel)
+    *   [Parameters](#parameters-11)
+*   [InferenceProvider](#inferenceprovider)
+*   [createCompletion](#createcompletion)
+    *   [Parameters](#parameters-12)
+*   [createCompletionStream](#createcompletionstream)
+    *   [Parameters](#parameters-13)
+*   [createCompletionGenerator](#createcompletiongenerator)
+    *   [Parameters](#parameters-14)
+*   [createEmbedding](#createembedding)
+    *   [Parameters](#parameters-15)
+*   [CompletionOptions](#completionoptions)
+    *   [verbose](#verbose-1)
+    *   [onToken](#ontoken)
+*   [Message](#message)
     *   [role](#role)
     *   [content](#content)
 *   [prompt\_tokens](#prompt_tokens)
 *   [completion\_tokens](#completion_tokens)
 *   [total\_tokens](#total_tokens)
+*   [n\_past\_tokens](#n_past_tokens)
 *   [CompletionReturn](#completionreturn)
     *   [model](#model)
     *   [usage](#usage)
-    *   [choices](#choices)
-*   [CompletionChoice](#completionchoice)
-    *   [message](#message)
+    *   [message](#message-1)
+*   [CompletionStreamReturn](#completionstreamreturn)
 *   [LLModelPromptContext](#llmodelpromptcontext)
     *   [logitsSize](#logitssize)
     *   [tokensSize](#tokenssize)
     *   [nPast](#npast)
-    *   [nCtx](#nctx)
     *   [nPredict](#npredict)
+    *   [promptTemplate](#prompttemplate)
+    *   [nCtx](#nctx-1)
     *   [topK](#topk)
     *   [topP](#topp)
-    *   [temp](#temp)
+    *   [minP](#minp)
+    *   [temperature](#temperature)
     *   [nBatch](#nbatch)
     *   [repeatPenalty](#repeatpenalty)
     *   [repeatLastN](#repeatlastn)
     *   [contextErase](#contexterase)
-*   [createTokenStream](#createtokenstream)
-    *   [Parameters](#parameters-8)
 *   [DEFAULT\_DIRECTORY](#default_directory)
 *   [DEFAULT\_LIBRARIES\_DIRECTORY](#default_libraries_directory)
 *   [DEFAULT\_MODEL\_CONFIG](#default_model_config)
-*   [DEFAULT\_PROMT\_CONTEXT](#default_promt_context)
+*   [DEFAULT\_PROMPT\_CONTEXT](#default_prompt_context)
 *   [DEFAULT\_MODEL\_LIST\_URL](#default_model_list_url)
 *   [downloadModel](#downloadmodel)
-    *   [Parameters](#parameters-9)
+    *   [Parameters](#parameters-16)
     *   [Examples](#examples)
 *   [DownloadModelOptions](#downloadmodeloptions)
-    *   [modelPath](#modelpath)
-    *   [verbose](#verbose-1)
+    *   [modelPath](#modelpath-1)
+    *   [verbose](#verbose-2)
     *   [url](#url)
     *   [md5sum](#md5sum)
 *   [DownloadController](#downloadcontroller)
     *   [cancel](#cancel)
     *   [promise](#promise)
 
-#### ModelType
-
-Type of the model
-
-Type: (`"gptj"` | `"llama"` | `"mpt"` | `"replit"`)
-
-#### ModelFile
-
-Full list of models available
-@deprecated These model names are outdated and this type will not be maintained, please use a string literal instead
-
-##### gptj
-
-List of GPT-J Models
-
-Type: (`"ggml-gpt4all-j-v1.3-groovy.bin"` | `"ggml-gpt4all-j-v1.2-jazzy.bin"` | `"ggml-gpt4all-j-v1.1-breezy.bin"` | `"ggml-gpt4all-j.bin"`)
-
-##### llama
-
-List Llama Models
-
-Type: (`"ggml-gpt4all-l13b-snoozy.bin"` | `"ggml-vicuna-7b-1.1-q4_2.bin"` | `"ggml-vicuna-13b-1.1-q4_2.bin"` | `"ggml-wizardLM-7B.q4_2.bin"` | `"ggml-stable-vicuna-13B.q4_2.bin"` | `"ggml-nous-gpt4-vicuna-13b.bin"` | `"ggml-v3-13b-hermes-q5_1.bin"`)
-
-##### mpt
-
-List of MPT Models
-
-Type: (`"ggml-mpt-7b-base.bin"` | `"ggml-mpt-7b-chat.bin"` | `"ggml-mpt-7b-instruct.bin"`)
-
-##### replit
-
-List of Replit Models
-
-Type: `"ggml-replit-code-v1-3b.bin"`
-
 #### type
 
 Model architecture. This argument currently does not have any functionality and is just used as descriptive identifier for user.
 
-Type: [ModelType](#modeltype)
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+#### TokenCallback
+
+Callback for controlling token generation. Return false to stop token generation.
+
+Type: function (tokenId: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number), token: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String), total: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)): [boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)
+
+#### ChatSessionOptions
+
+**Extends Partial\<LLModelPromptContext>**
+
+Options for the chat session.
+
+##### systemPrompt
+
+System prompt to ingest on initialization.
+
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+##### messages
+
+Messages to ingest on initialization.
+
+Type: [Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Array)<[Message](#message)>
+
+#### initialize
+
+Ingests system prompt and initial messages.
+Sets this chat session as the active chat session of the model.
+
+##### Parameters
+
+*   `options` **[ChatSessionOptions](#chatsessionoptions)** The options for the chat session.
+
+Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)\<void>**&#x20;
+
+#### generate
+
+Prompts the model in chat-session context.
+
+##### Parameters
+
+*   `prompt` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The prompt input.
+*   `options` **[CompletionOptions](#completionoptions)?** Prompt context and other options.
+*   `callback` **[TokenCallback](#tokencallback)?** Token generation callback.
+
+<!---->
+
+*   Throws **[Error](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Error)** If the chat session is not the active chat session of the model.
+
+Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)<[CompletionReturn](#completionreturn)>** The model's response to the prompt.
+
+#### InferenceModel
+
+InferenceModel represents an LLM which can make chat predictions, similar to GPT transformers.
+
+##### createChatSession
+
+Create a chat session with the model.
+
+###### Parameters
+
+*   `options` **[ChatSessionOptions](#chatsessionoptions)?** The options for the chat session.
+
+Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)\<ChatSession>** The chat session.
+
+##### generate
+
+Prompts the model with a given input and optional parameters.
+
+###### Parameters
+
+*   `prompt` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)**&#x20;
+*   `options` **[CompletionOptions](#completionoptions)?** Prompt context and other options.
+*   `callback` **[TokenCallback](#tokencallback)?** Token generation callback.
+*   `input`  The prompt input.
+
+Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)<[CompletionReturn](#completionreturn)>** The model's response to the prompt.
+
+##### dispose
+
+delete and cleanup the native model
+
+Returns **void**&#x20;
+
+#### EmbeddingModel
+
+EmbeddingModel represents an LLM which can create embeddings, which are float arrays
+
+##### dispose
+
+delete and cleanup the native model
+
+Returns **void**&#x20;
+
+#### InferenceResult
+
+Shape of LLModel's inference result.
 
 #### LLModel
 
@@ -292,9 +471,9 @@ Initialize a new LLModel.
 
 ##### type
 
-either 'gpt', mpt', or 'llama' or undefined
+undefined or user supplied
 
-Returns **([ModelType](#modeltype) | [undefined](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/undefined))**&#x20;
+Returns **([string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String) | [undefined](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/undefined))**&#x20;
 
 ##### name
 
@@ -326,7 +505,7 @@ Set the number of threads used for model inference.
 
 Returns **void**&#x20;
 
-##### raw\_prompt
+##### infer
 
 Prompt the model with a given input and optional parameters.
 This is the raw output from model.
@@ -334,23 +513,20 @@ Use the prompt function exported for a value
 
 ###### Parameters
 
-*   `q` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The prompt input.
-*   `params` **Partial<[LLModelPromptContext](#llmodelpromptcontext)>** Optional parameters for the prompt context.
-*   `callback` **function (res: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)): void**&#x20;
+*   `prompt` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The prompt input.
+*   `promptContext` **Partial<[LLModelPromptContext](#llmodelpromptcontext)>** Optional parameters for the prompt context.
+*   `callback` **[TokenCallback](#tokencallback)?** optional callback to control token generation.
 
-Returns **void** The result of the model prompt.
+Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)<[InferenceResult](#inferenceresult)>** The result of the model prompt.
 
 ##### embed
 
 Embed text with the model. Keep in mind that
-not all models can embed text, (only bert can embed as of 07/16/2023 (mm/dd/yyyy))
 Use the prompt function exported for a value
 
 ###### Parameters
 
-*   `text` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)**&#x20;
-*   `q`  The prompt input.
-*   `params`  Optional parameters for the prompt context.
+*   `text` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The prompt input.
 
 Returns **[Float32Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Float32Array)** The result of the model prompt.
 
@@ -376,6 +552,114 @@ Where to get the pluggable backend libraries
 
 Returns **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)**&#x20;
 
+##### initGpuByString
+
+Initiate a GPU by a string identifier.
+
+###### Parameters
+
+*   `memory_required` **[number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)** Should be in the range size\_t or will throw
+*   `device_name` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** 'amd' | 'nvidia' | 'intel' | 'gpu' | gpu name.
+    read LoadModelOptions.device for more information
+
+Returns **[boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)**&#x20;
+
+##### hasGpuDevice
+
+From C documentation
+
+Returns **[boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)** True if a GPU device is successfully initialized, false otherwise.
+
+##### listGpu
+
+GPUs that are usable for this LLModel
+
+###### Parameters
+
+*   `nCtx` **[number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)** Maximum size of context window
+
+<!---->
+
+*   Throws **any** if hasGpuDevice returns false (i think)
+
+Returns **[Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Array)<[GpuDevice](#gpudevice)>**&#x20;
+
+##### dispose
+
+delete and cleanup the native model
+
+Returns **void**&#x20;
+
+#### GpuDevice
+
+an object that contains gpu data on this machine.
+
+##### type
+
+same as VkPhysicalDeviceType
+
+Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
+
+#### LoadModelOptions
+
+Options that configure a model's behavior.
+
+##### modelPath
+
+Where to look for model files.
+
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+##### librariesPath
+
+Where to look for the backend libraries.
+
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+##### modelConfigFile
+
+The path to the model configuration file, useful for offline usage or custom model configurations.
+
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+##### allowDownload
+
+Whether to allow downloading the model if it is not present at the specified path.
+
+Type: [boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)
+
+##### verbose
+
+Enable verbose logging.
+
+Type: [boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)
+
+##### device
+
+The processing unit on which the model will run. It can be set to
+
+*   "cpu": Model will run on the central processing unit.
+*   "gpu": Model will run on the best available graphics processing unit, irrespective of its vendor.
+*   "amd", "nvidia", "intel": Model will run on the best available GPU from the specified vendor.
+*   "gpu name": Model will run on the GPU that matches the name if it's available.
+    Note: If a GPU device lacks sufficient RAM to accommodate the model, an error will be thrown, and the GPT4All
+    instance will be rendered invalid. It's advised to ensure the device has enough memory before initiating the
+    model.
+
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+##### nCtx
+
+The Maximum window size of this model
+
+Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
+
+##### ngl
+
+Number of gpu layers needed
+
+Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
+
 #### loadModel
 
 Loads a machine learning model with the specified name. The defacto way to create a model.
@@ -384,9 +668,13 @@ By default this will download a model from the official GPT4ALL website, if a mo
 ##### Parameters
 
 *   `modelName` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The name of the model to load.
-*   `options` **(LoadModelOptions | [undefined](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/undefined))?** (Optional) Additional options for loading the model.
+*   `options` **([LoadModelOptions](#loadmodeloptions) | [undefined](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/undefined))?** (Optional) Additional options for loading the model.
 
-Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)<(InferenceModel | EmbeddingModel)>** A promise that resolves to an instance of the loaded LLModel.
+Returns **[Promise](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Promise)<([InferenceModel](#inferencemodel) | [EmbeddingModel](#embeddingmodel))>** A promise that resolves to an instance of the loaded LLModel.
+
+#### InferenceProvider
+
+Interface for inference, implemented by InferenceModel and ChatSession.
 
 #### createCompletion
 
@@ -394,12 +682,36 @@ The nodejs equivalent to python binding's chat\_completion
 
 ##### Parameters
 
-*   `model` **InferenceModel** The language model object.
-*   `messages` **[Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Array)<[PromptMessage](#promptmessage)>** The array of messages for the conversation.
+*   `provider` **[InferenceProvider](#inferenceprovider)** The inference model object or chat session
+*   `message` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The user input message
 *   `options` **[CompletionOptions](#completionoptions)** The options for creating the completion.
 
 Returns **[CompletionReturn](#completionreturn)** The completion result.
 
+#### createCompletionStream
+
+Streaming variant of createCompletion, returns a stream of tokens and a promise that resolves to the completion result.
+
+##### Parameters
+
+*   `provider` **[InferenceProvider](#inferenceprovider)** The inference model object or chat session
+*   `message` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The user input message.
+*   `options` **[CompletionOptions](#completionoptions)** The options for creating the completion.
+
+Returns **[CompletionStreamReturn](#completionstreamreturn)** An object of token stream and the completion result promise.
+
+#### createCompletionGenerator
+
+Creates an async generator of tokens
+
+##### Parameters
+
+*   `provider` **[InferenceProvider](#inferenceprovider)** The inference model object or chat session
+*   `message` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The user input message.
+*   `options` **[CompletionOptions](#completionoptions)** The options for creating the completion.
+
+Returns **AsyncGenerator<[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)>** The stream of generated tokens
+
 #### createEmbedding
 
 The nodejs moral equivalent to python binding's Embed4All().embed()
@@ -407,7 +719,7 @@ meow
 
 ##### Parameters
 
-*   `model` **EmbeddingModel** The language model object.
+*   `model` **[EmbeddingModel](#embeddingmodel)** The language model object.
 *   `text` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** text to embed
 
 Returns **[Float32Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Float32Array)** The completion result.
@@ -424,34 +736,15 @@ Indicates if verbose logging is enabled.
 
 Type: [boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)
 
-##### systemPromptTemplate
+##### onToken
 
-Template for the system message. Will be put before the conversation with %1 being replaced by all system messages.
-Note that if this is not defined, system messages will not be included in the prompt.
+Callback for controlling token generation. Return false to stop processing.
 
-Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+Type: [TokenCallback](#tokencallback)
 
-##### promptTemplate
+#### Message
 
-Template for user messages, with %1 being replaced by the message.
-
-Type: [boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Boolean)
-
-##### promptHeader
-
-The initial instruction for the model, on top of the prompt
-
-Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
-
-##### promptFooter
-
-The last instruction for the model, appended to the end of the prompt.
-
-Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
-
-#### PromptMessage
-
-A message in the conversation, identical to OpenAI's chat message.
+A message in the conversation.
 
 ##### role
 
@@ -467,7 +760,7 @@ Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Globa
 
 #### prompt\_tokens
 
-The number of tokens used in the prompt.
+The number of tokens used in the prompt. Currently not available and always 0.
 
 Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
 
@@ -479,13 +772,19 @@ Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Globa
 
 #### total\_tokens
 
-The total number of tokens used.
+The total number of tokens used. Currently not available and always 0.
+
+Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
+
+#### n\_past\_tokens
+
+Number of tokens used in the conversation.
 
 Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
 
 #### CompletionReturn
 
-The result of the completion, similar to OpenAI's format.
+The result of a completion.
 
 ##### model
 
@@ -497,23 +796,17 @@ Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Globa
 
 Token usage report.
 
-Type: {prompt\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number), completion\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number), total\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)}
-
-##### choices
-
-The generated completions.
-
-Type: [Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Array)<[CompletionChoice](#completionchoice)>
-
-#### CompletionChoice
-
-A completion choice, similar to OpenAI's format.
+Type: {prompt\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number), completion\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number), total\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number), n\_past\_tokens: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)}
 
 ##### message
 
-Response message
+The generated completion.
 
-Type: [PromptMessage](#promptmessage)
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+#### CompletionStreamReturn
+
+The result of a streamed completion, containing a stream of tokens and a promise that resolves to the completion result.
 
 #### LLModelPromptContext
 
@@ -534,18 +827,29 @@ Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Globa
 ##### nPast
 
 The number of tokens in the past conversation.
-
-Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
-
-##### nCtx
-
-The number of tokens possible in the context window.
+This controls how far back the model looks when generating completions.
 
 Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
 
 ##### nPredict
 
-The number of tokens to predict.
+The maximum number of tokens to predict.
+
+Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
+
+##### promptTemplate
+
+Template for user / assistant message pairs.
+%1 is required and will be replaced by the user input.
+%2 is optional and will be replaced by the assistant response.
+
+Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
+
+##### nCtx
+
+The context window size. Do not use, it has no effect. See loadModel options.
+THIS IS DEPRECATED!!!
+Use loadModel's nCtx option instead.
 
 Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
 
@@ -568,12 +872,16 @@ above a threshold P. This method, also known as nucleus sampling, finds a balanc
 and quality by considering both token probabilities and the number of tokens available for sampling.
 When using a higher value for top-P (eg., 0.95), the generated text becomes more diverse.
 On the other hand, a lower value (eg., 0.1) produces more focused and conservative text.
-The default value is 0.4, which is aimed to be the middle ground between focus and diversity, but
-for more creative tasks a higher top-p value will be beneficial, about 0.5-0.9 is a good range for that.
 
 Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
 
-##### temp
+##### minP
+
+The minimum probability of a token to be considered.
+
+Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
+
+##### temperature
 
 The temperature to adjust the model's output distribution.
 Temperature is like a knob that adjusts how creative or focused the output becomes. Higher temperatures
@@ -618,18 +926,6 @@ The percentage of context to erase if the context window is exceeded.
 
 Type: [number](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Number)
 
-#### createTokenStream
-
-TODO: Help wanted to implement this
-
-##### Parameters
-
-*   `llmodel` **[LLModel](#llmodel)**&#x20;
-*   `messages` **[Array](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/Array)<[PromptMessage](#promptmessage)>**&#x20;
-*   `options` **[CompletionOptions](#completionoptions)**&#x20;
-
-Returns **function (ll: [LLModel](#llmodel)): AsyncGenerator<[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)>**&#x20;
-
 #### DEFAULT\_DIRECTORY
 
 From python api:
@@ -652,7 +948,7 @@ Default model configuration.
 
 Type: ModelConfig
 
-#### DEFAULT\_PROMT\_CONTEXT
+#### DEFAULT\_PROMPT\_CONTEXT
 
 Default prompt context.
 
@@ -672,7 +968,7 @@ By default this downloads without waiting. use the controller returned to alter
 ##### Parameters
 
 *   `modelName` **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** The model to be downloaded.
-*   `options` **DownloadOptions** to pass into the downloader. Default is { location: (cwd), verbose: false }.
+*   `options` **[DownloadModelOptions](#downloadmodeloptions)** to pass into the downloader. Default is { location: (cwd), verbose: false }.
 
 ##### Examples
 

gpt4all-bindings/python/docs/gpt4all_python.md

@@ -8,30 +8,22 @@ The source code and local build instructions can be found [here](https://github.
 pip install gpt4all
 ```
 
-=== "GPT4All Example"
-    ``` py
-    from gpt4all import GPT4All
-    model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
-    output = model.generate("The capital of France is ", max_tokens=3)
-    print(output)
-    ```
-=== "Output"
-    ```
-    1. Paris
-    ```
+``` py
+from gpt4all import GPT4All
+model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
+```
 
 This will:
 
 - Instantiate `GPT4All`,  which is the primary public API to your large language model (LLM).
 - Automatically download the given model to `~/.cache/gpt4all/` if not already present.
-- Through `model.generate(...)` the model starts working on a response. There are various ways to
-  steer that process. Here, `max_tokens` sets an upper limit, i.e. a hard cut-off point to the output.
+
+Read further to see how to chat with this model.
 
 
 ### Chatting with GPT4All
-Local LLMs can be optimized for chat conversations by reusing previous computational history.
-
-Use the GPT4All `chat_session` context manager to hold chat conversations with the model.
+To start chatting with a local LLM, you will need to start a chat session. Within a chat session, the model will be
+prompted with the appropriate template, and history will be preserved between successive calls to `generate()`.
 
 === "GPT4All Example"
     ``` py
@@ -72,15 +64,19 @@ Use the GPT4All `chat_session` context manager to hold chat conversations with t
     ]
     ```
 
-When using GPT4All models in the `chat_session` context:
+When using GPT4All models in the `chat_session()` context:
 
 - Consecutive chat exchanges are taken into account and not discarded until the session ends; as long as the model has capacity.
-- Internal K/V caches are preserved from previous conversation history, speeding up inference.
-- The model is given a system and prompt template which make it chatty. Depending on `allow_download=True` (default),
-  it will obtain the latest version of [models2.json] from the repository, which contains specifically tailored templates
-  for models. Conversely, if it is not allowed to download, it falls back to default templates instead.
+- A system prompt is inserted into the beginning of the model's context.
+- Each prompt passed to `generate()` is wrapped in the appropriate prompt template. If you pass `allow_download=False`
+  to GPT4All or are using a model that is not from the official models list, you must pass a prompt template using the
+  `prompt_template` parameter of `chat_session()`.
 
-[models2.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models2.json
+NOTE: If you do not use `chat_session()`, calls to `generate()` will not be wrapped in a prompt template. This will
+cause the model to *continue* the prompt instead of *answering* it. When in doubt, use a chat session, as many newer
+models are designed to be used exclusively with a prompt template.
+
+[models3.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models3.json
 
 
 ### Streaming Generations
@@ -91,13 +87,14 @@ To interact with GPT4All responses as the model generates, use the `streaming=Tr
     from gpt4all import GPT4All
     model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
     tokens = []
-    for token in model.generate("The capital of France is", max_tokens=20, streaming=True):
-        tokens.append(token)
+    with model.chat_session():
+        for token in model.generate("What is the capital of France?", streaming=True):
+            tokens.append(token)
     print(tokens)
     ```
 === "Output"
     ```
-    [' Paris', ' is', ' a', ' city', ' that', ' has', ' been', ' a', ' major', ' cultural', ' and', ' economic', ' center', ' for', ' over', ' ', '2', ',', '0', '0']
+    [' The', ' capital', ' of', ' France', ' is', ' Paris', '.']
     ```
 
 
@@ -131,20 +128,11 @@ generation; be sure to review all their descriptions.
 The model folder can be set with the `model_path` parameter when creating a `GPT4All` instance. The example below is
 is the same as if it weren't provided; that is, `~/.cache/gpt4all/` is the default folder.
 
-=== "GPT4All Model Folder Example"
-    ``` py
-    from pathlib import Path
-    from gpt4all import GPT4All
-    model = GPT4All(model_name='orca-mini-3b-gguf2-q4_0.gguf',
-                    model_path=(Path.home() / '.cache' / 'gpt4all'),
-                    allow_download=False)
-    response = model.generate('my favorite 3 fruits are:', temp=0)
-    print(response)
-    ```
-=== "Output"
-    ```
-    My favorite three fruits are apples, bananas and oranges.
-    ```
+``` py
+from pathlib import Path
+from gpt4all import GPT4All
+model = GPT4All(model_name='orca-mini-3b-gguf2-q4_0.gguf', model_path=Path.home() / '.cache' / 'gpt4all')
+```
 
 If you want to point it at the chat GUI's default folder, it should be:
 === "macOS"
@@ -179,22 +167,20 @@ Alternatively, you could also change the module's default model directory:
 
 ``` py
 from pathlib import Path
-import gpt4all.gpt4all
-gpt4all.gpt4all.DEFAULT_MODEL_DIRECTORY = Path.home() / 'my' / 'models-directory'
-from gpt4all import GPT4All
+from gpt4all import GPT4All, gpt4all
+gpt4all.DEFAULT_MODEL_DIRECTORY = Path.home() / 'my' / 'models-directory'
 model = GPT4All('orca-mini-3b-gguf2-q4_0.gguf')
-...
 ```
 
 
 ### Managing Templates
-Session templates can be customized when starting a `chat_session` context:
+When using a `chat_session()`, you may customize the system prompt, and set the prompt template if necessary:
 
 === "GPT4All Custom Session Templates Example"
     ``` py
     from gpt4all import GPT4All
     model = GPT4All('wizardlm-13b-v1.2.Q4_0.gguf')
-    system_template = 'A chat between a curious user and an artificial intelligence assistant.'
+    system_template = 'A chat between a curious user and an artificial intelligence assistant.\n'
     # many models use triple hash '###' for keywords, Vicunas are simpler:
     prompt_template = 'USER: {0}\nASSISTANT: '
     with model.chat_session(system_template, prompt_template):
@@ -218,193 +204,38 @@ Session templates can be customized when starting a `chat_session` context:
     particles, making the sky appear blue to our eyes.
     ```
 
-To do the same outside a session, the input has to be formatted manually. For example:
-
-=== "GPT4All Templates Outside a Session Example"
-    ``` py
-    model = GPT4All('wizardlm-13b-v1.2.Q4_0.gguf')
-    system_template = 'A chat between a curious user and an artificial intelligence assistant.'
-    prompt_template = 'USER: {0}\nASSISTANT: '
-    prompts = ['name 3 colors', 'now name 3 fruits', 'what were the 3 colors in your earlier response?']
-    first_input = system_template + prompt_template.format(prompts[0])
-    response = model.generate(first_input, temp=0)
-    print(response)
-    for prompt in prompts[1:]:
-        response = model.generate(prompt_template.format(prompt), temp=0)
-        print(response)
-    ```
-=== "Output"
-    ```
-    1) Red
-    2) Blue
-    3) Green
-
-    1. Apple
-    2. Banana
-    3. Orange
-
-    The colors in my previous response are blue, green and red.
-    ```
-
-Ultimately, the method `GPT4All._format_chat_prompt_template()` is responsible for formatting templates. It can be
-customized in a subclass. As an example:
-
-=== "Custom Subclass"
-    ``` py
-    from itertools import cycle
-    from gpt4all import GPT4All
-
-    class RotatingTemplateGPT4All(GPT4All):
-        def __init__(self, *args, **kwargs):
-            super().__init__(*args, **kwargs)
-            self._templates = [
-                "Respond like a pirate.",
-                "Respond like a politician.",
-                "Respond like a philosopher.",
-                "Respond like a Klingon.",
-            ]
-            self._cycling_templates = cycle(self._templates)
-
-        def _format_chat_prompt_template(
-            self,
-            messages: list,
-            default_prompt_header: str = "",
-            default_prompt_footer: str = "",
-        ) -> str:
-            full_prompt = default_prompt_header + "\n\n" if default_prompt_header != "" else ""
-            for message in messages:
-                if message["role"] == "user":
-                    user_message = f"USER: {message['content']} {next(self._cycling_templates)}\n"
-                    full_prompt += user_message
-                if message["role"] == "assistant":
-                    assistant_message = f"ASSISTANT: {message['content']}\n"
-                    full_prompt += assistant_message
-            full_prompt += "\n\n" + default_prompt_footer if default_prompt_footer != "" else ""
-            print(full_prompt)
-            return full_prompt
-    ```
-=== "GPT4All Custom Subclass Example"
-    ``` py
-    model = RotatingTemplateGPT4All('wizardlm-13b-v1.2.Q4_0.gguf')
-    with model.chat_session():  # starting a session is optional in this example
-        response1 = model.generate("hi, who are you?")
-        print(response1)
-        print()
-        response2 = model.generate("what can you tell me about snakes?")
-        print(response2)
-        print()
-        response3 = model.generate("what's your opinion on Chess?")
-        print(response3)
-        print()
-        response4 = model.generate("tell me about ancient Rome.")
-        print(response4)
-    ```
-=== "Possible Output"
-    ```
-    USER: hi, who are you? Respond like a pirate.
-
-    Pirate: Ahoy there mateys! I be Cap'n Jack Sparrow of the Black Pearl.
-
-    USER: what can you tell me about snakes? Respond like a politician.
-
-    Politician: Snakes have been making headlines lately due to their ability to
-    slither into tight spaces and evade capture, much like myself during my last
-    election campaign. However, I believe that with proper education and
-    understanding of these creatures, we can work together towards creating a
-    safer environment for both humans and snakes alike.
-
-    USER: what's your opinion on Chess? Respond like a philosopher.
-
-    Philosopher: The game of chess is often used as an analogy to illustrate the
-    complexities of life and decision-making processes. However, I believe that it
-    can also be seen as a reflection of our own consciousness and subconscious mind.
-    Just as each piece on the board has its unique role to play in shaping the
-    outcome of the game, we too have different roles to fulfill in creating our own
-    personal narrative.
-
-    USER: tell me about ancient Rome. Respond like a Klingon.
-
-    Klingon: Ancient Rome was once a great empire that ruled over much of Europe and
-    the Mediterranean region. However, just as the Empire fell due to internal strife
-    and external threats, so too did my own house come crashing down when I failed to
-    protect our homeworld from invading forces.
-    ```
-
-
-### Introspection
-A less apparent feature is the capacity to log the final prompt that gets sent to the model. It relies on
-[Python's logging facilities][py-logging] implemented in the `pyllmodel` module at the `INFO` level. You can activate it
-for example with a `basicConfig`, which displays it on the standard error stream. It's worth mentioning that Python's
-logging infrastructure offers [many more customization options][py-logging-cookbook].
-
-[py-logging]: https://docs.python.org/3/howto/logging.html
-[py-logging-cookbook]: https://docs.python.org/3/howto/logging-cookbook.html
-
-=== "GPT4All Prompt Logging Example"
-    ``` py
-    import logging
-    from gpt4all import GPT4All
-    logging.basicConfig(level=logging.INFO)
-    model = GPT4All('nous-hermes-llama2-13b.Q4_0.gguf')
-    with model.chat_session('You are a geography expert.\nBe terse.',
-                            '### Instruction:\n{0}\n### Response:\n'):
-        response = model.generate('who are you?', temp=0)
-        print(response)
-        response = model.generate('what are your favorite 3 mountains?', temp=0)
-        print(response)
-    ```
-=== "Output"
-    ```
-    INFO:gpt4all.pyllmodel:LLModel.prompt_model -- prompt:
-    You are a geography expert.
-    Be terse.
-
-    ### Instruction:
-    who are you?
-    ### Response:
-
-    ===/LLModel.prompt_model -- prompt/===
-    I am an AI-powered chatbot designed to assist users with their queries related to geographical information.
-    INFO:gpt4all.pyllmodel:LLModel.prompt_model -- prompt:
-    ### Instruction:
-    what are your favorite 3 mountains?
-    ### Response:
-
-    ===/LLModel.prompt_model -- prompt/===
-    1) Mount Everest - Located in the Himalayas, it is the highest mountain on Earth and a significant challenge for mountaineers.
-    2) Kangchenjunga - This mountain is located in the Himalayas and is the third-highest peak in the world after Mount Everest and K2.
-    3) Lhotse - Located in the Himalayas, it is the fourth highest mountain on Earth and offers a challenging climb for experienced mountaineers.
-    ```
-
 
 ### Without Online Connectivity
-To prevent GPT4All from accessing online resources, instantiate it with `allow_download=False`. This will disable both
-downloading missing models and [models2.json], which contains information about them. As a result, predefined templates
-are used instead of model-specific system and prompt templates:
+To prevent GPT4All from accessing online resources, instantiate it with `allow_download=False`. When using this flag,
+there will be no default system prompt by default, and you must specify the prompt template yourself.
 
-=== "GPT4All Default Templates Example"
+You can retrieve a model's default system prompt and prompt template with an online instance of GPT4All:
+
+=== "Prompt Template Retrieval"
     ``` py
     from gpt4all import GPT4All
-    model = GPT4All('ggml-mpt-7b-chat.bin', allow_download=False)
-    # when downloads are disabled, it will use the default templates:
-    print("default system template:", repr(model.config['systemPrompt']))
-    print("default prompt template:", repr(model.config['promptTemplate']))
-    print()
-    # even when inside a session:
-    with model.chat_session():
-        assert model.current_chat_session[0]['role'] == 'system'
-        print("session system template:", repr(model.current_chat_session[0]['content']))
-        print("session prompt template:", repr(model._current_prompt_template))
+    model = GPT4All('orca-mini-3b-gguf2-q4_0.gguf')
+    print(repr(model.config['systemPrompt']))
+    print(repr(model.config['promptTemplate']))
     ```
 === "Output"
-    ```
-    default system template: ''
-    default prompt template: '### Human: \n{0}\n### Assistant:\n'
-
-    session system template: ''
-    session prompt template: '### Human: \n{0}\n### Assistant:\n'
+    ```py
+    '### System:\nYou are an AI assistant that follows instruction extremely well. Help as much as you can.\n\n'
+    '### User:\n{0}\n### Response:\n'
     ```
 
+Then you can pass them explicitly when creating an offline instance:
+
+``` py
+from gpt4all import GPT4All
+model = GPT4All('orca-mini-3b-gguf2-q4_0.gguf', allow_download=False)
+
+system_prompt = '### System:\nYou are an AI assistant that follows instruction extremely well. Help as much as you can.\n\n'
+prompt_template = '### User:\n{0}\n\n### Response:\n'
+
+with model.chat_session(system_prompt=system_prompt, prompt_template=prompt_template):
+    ...
+```
 
 ### Interrupting Generation
 The simplest way to stop generation is to set a fixed upper limit with the `max_tokens` parameter.

gpt4all-bindings/python/docs/gpt4all_python_embedding.md

@@ -1,18 +1,41 @@
 # Embeddings
-GPT4All supports generating high quality embeddings of arbitrary length documents of text using a CPU optimized contrastively trained [Sentence Transformer](https://www.sbert.net/). These embeddings are comparable in quality for many tasks with OpenAI.
+GPT4All supports generating high quality embeddings of arbitrary length text using any embedding model supported by llama.cpp.
+
+An embedding is a vector representation of a piece of text. Embeddings are useful for tasks such as retrieval for
+question answering (including retrieval augmented generation or *RAG*), semantic similarity search, classification, and
+topic clustering.
+
+## Supported Embedding Models
+
+The following models have built-in support in Embed4All:
+
+| Name               | Embed4All `model_name`                               | Context Length | Embedding Length | File Size |
+|--------------------|------------------------------------------------------|---------------:|-----------------:|----------:|
+| [SBert]            | all‑MiniLM‑L6‑v2.gguf2.f16.gguf |            512 |              384 |    44 MiB |
+| [Nomic Embed v1]   | nomic‑embed‑text‑v1.f16.gguf    |           2048 |              768 |   262 MiB |
+| [Nomic Embed v1.5] | nomic‑embed‑text‑v1.5.f16.gguf  |           2048 |           64-768 |   262 MiB |
+
+The context length is the maximum number of word pieces, or *tokens*, that a model can embed at once. Embedding texts
+longer than a model's context length requires some kind of strategy; see [Embedding Longer Texts] for more information.
+
+The embedding length is the size of the vector returned by `Embed4All.embed`.
+
+[SBert]: https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
+[Nomic Embed v1]: https://huggingface.co/nomic-ai/nomic-embed-text-v1
+[Nomic Embed v1.5]: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
+[Embedding Longer Texts]: #embedding-longer-texts
 
 ## Quickstart
-
 ```bash
 pip install gpt4all
 ```
 
-### Generating embeddings
-The embedding model will automatically be downloaded if not installed.
+### Generating Embeddings
+By default, embeddings will be generated on the CPU using all-MiniLM-L6-v2.
 
 === "Embed4All Example"
-    ``` py
-    from gpt4all import GPT4All, Embed4All
+    ```py
+    from gpt4all import Embed4All
     text = 'The quick brown fox jumps over the lazy dog'
     embedder = Embed4All()
     output = embedder.embed(text)
@@ -22,13 +45,131 @@ The embedding model will automatically be downloaded if not installed.
     ```
     [0.034696947783231735, -0.07192722707986832, 0.06923297047615051, ...]
     ```
-### Speed of embedding generation
-The following table lists the generation speed for text document captured on an Intel i913900HX CPU with DDR5 5600 running with 8 threads under stable load.
 
-| Tokens          | 128  | 512  | 2048 | 8129 | 16,384 |
-| --------------- | ---- | ---- | ---- | ---- | ---- |
-| Wall time (s)   | .02  | .08  | .24  | .96  | 1.9  |
-| Tokens / Second | 6508 | 6431 | 8622 | 8509 | 8369 |
+You can also use the GPU to accelerate the embedding model by specifying the `device` parameter. See the [GPT4All
+constructor] for more information.
+
+=== "GPU Example"
+    ```py
+    from gpt4all import Embed4All
+    text = 'The quick brown fox jumps over the lazy dog'
+    embedder = Embed4All(device='gpu')
+    output = embedder.embed(text)
+    print(output)
+    ```
+=== "Output"
+    ```
+    [0.034696947783231735, -0.07192722707986832, 0.06923297047615051, ...]
+    ```
+
+[GPT4All constructor]: gpt4all_python.md#gpt4all.gpt4all.GPT4All.__init__
+
+### Nomic Embed
+
+Embed4All has built-in support for Nomic's open-source embedding model, [Nomic Embed]. When using this model, you must
+specify the task type using the `prefix` argument. This may be one of `search_query`, `search_document`,
+`classification`, or `clustering`. For retrieval applications, you should prepend `search_document` for all of your
+documents and `search_query` for your queries. See the [Nomic Embedding Guide] for more info.
+
+=== "Nomic Embed Example"
+    ```py
+    from gpt4all import Embed4All
+    text = 'Who is Laurens van der Maaten?'
+    embedder = Embed4All('nomic-embed-text-v1.f16.gguf')
+    output = embedder.embed(text, prefix='search_query')
+    print(output)
+    ```
+=== "Output"
+    ```
+    [-0.013357644900679588, 0.027070969343185425, -0.0232995692640543, ...]
+    ```
+
+[Nomic Embed]: https://blog.nomic.ai/posts/nomic-embed-text-v1
+[Nomic Embedding Guide]: https://docs.nomic.ai/atlas/guides/embeddings#embedding-task-types
+
+### Embedding Longer Texts
+
+Embed4All accepts a parameter called `long_text_mode`. This controls the behavior of Embed4All for texts longer than the
+context length of the embedding model.
+
+In the default mode of "mean", Embed4All will break long inputs into chunks and average their embeddings to compute the
+final result.
+
+To change this behavior, you can set the `long_text_mode` parameter to "truncate", which will truncate the input to the
+sequence length of the model before generating a single embedding.
+
+=== "Truncation Example"
+    ```py
+    from gpt4all import Embed4All
+    text = 'The ' * 512 + 'The quick brown fox jumps over the lazy dog'
+    embedder = Embed4All()
+    output = embedder.embed(text, long_text_mode="mean")
+    print(output)
+    print()
+    output = embedder.embed(text, long_text_mode="truncate")
+    print(output)
+    ```
+=== "Output"
+    ```
+    [0.0039850445464253426, 0.04558328539133072, 0.0035536508075892925, ...]
+
+    [-0.009771130047738552, 0.034792833030223846, -0.013273917138576508, ...]
+    ```
+
+
+### Batching
+
+You can send multiple texts to Embed4All in a single call. This can give faster results when individual texts are
+significantly smaller than `n_ctx` tokens. (`n_ctx` defaults to 2048.)
+
+=== "Batching Example"
+    ```py
+    from gpt4all import Embed4All
+    texts = ['The quick brown fox jumps over the lazy dog', 'Foo bar baz']
+    embedder = Embed4All()
+    output = embedder.embed(texts)
+    print(output[0])
+    print()
+    print(output[1])
+    ```
+=== "Output"
+    ```
+    [0.03551332652568817, 0.06137588247656822, 0.05281158909201622, ...]
+
+    [-0.03879690542817116, 0.00013223080895841122, 0.023148687556385994, ...]
+    ```
+
+The number of texts that can be embedded in one pass of the model is proportional to the `n_ctx` parameter of Embed4All.
+Increasing it may increase batched embedding throughput if you have a fast GPU, at the cost of VRAM.
+```py
+embedder = Embed4All(n_ctx=4096, device='gpu')
+```
+
+
+### Resizable Dimensionality
+
+The embedding dimension of Nomic Embed v1.5 can be resized using the `dimensionality` parameter. This parameter supports
+any value between 64 and 768.
+
+Shorter embeddings use less storage, memory, and bandwidth with a small performance cost. See the [blog post] for more
+info.
+
+[blog post]: https://blog.nomic.ai/posts/nomic-embed-matryoshka
+
+=== "Matryoshka Example"
+    ```py
+    from gpt4all import Embed4All
+    text = 'The quick brown fox jumps over the lazy dog'
+    embedder = Embed4All('nomic-embed-text-v1.5.f16.gguf')
+    output = embedder.embed(text, dimensionality=64)
+    print(len(output))
+    print(output)
+    ```
+=== "Output"
+    ```
+    64
+    [-0.03567073494195938, 0.1301717758178711, -0.4333043396472931, ...]
+    ```
 
 
 ### API documentation

gpt4all-bindings/python/docs/index.md

@@ -26,7 +26,6 @@ is organized as a monorepo with the following structure:
 - **gpt4all-backend**: The GPT4All backend maintains and exposes a universal, performance optimized C API for running inference with multi-billion parameter Transformer Decoders.
 This C API is then bound to any higher level programming language such as C++, Python, Go, etc.
 - **gpt4all-bindings**: GPT4All bindings contain a variety of high-level programming languages that implement the C API. Each directory is a bound programming language. The [CLI](gpt4all_cli.md) is included here, as well.
-- **gpt4all-api**: The GPT4All API (under initial development) exposes REST API endpoints for gathering completions and embeddings from large language models.
 - **gpt4all-chat**: GPT4All Chat is an OS native chat application that runs on macOS, Windows and Linux. It is the easiest way to run local, privacy aware chat assistants on everyday hardware. You can download it on the [GPT4All Website](https://gpt4all.io) and read its source code in the monorepo.
 
 Explore detailed documentation for the backend, bindings and chat client in the sidebar.
@@ -38,7 +37,7 @@ The GPT4All software ecosystem is compatible with the following Transformer arch
 - `MPT` (including `Replit`)
 - `GPT-J`
 
-You can find an exhaustive list of supported models on the [website](https://gpt4all.io) or in the [models directory](https://raw.githubusercontent.com/nomic-ai/gpt4all/main/gpt4all-chat/metadata/models2.json)
+You can find an exhaustive list of supported models on the [website](https://gpt4all.io) or in the [models directory](https://raw.githubusercontent.com/nomic-ai/gpt4all/main/gpt4all-chat/metadata/models3.json)
 
 
 GPT4All models are artifacts produced through a process known as neural network quantization.

gpt4all-bindings/python/gpt4all/__init__.py

@@ -1,2 +1 @@
-from .gpt4all import Embed4All, GPT4All  # noqa
-from .pyllmodel import LLModel  # noqa
+from .gpt4all import CancellationError as CancellationError, Embed4All as Embed4All, GPT4All as GPT4All

gpt4all-bindings/python/gpt4all/_pyllmodel.py

@@ -1,34 +1,42 @@
-import atexit
+from __future__ import annotations
+
 import ctypes
-import importlib.resources
-import logging
 import os
 import platform
 import re
 import subprocess
 import sys
 import threading
-from contextlib import ExitStack
+from enum import Enum
 from queue import Queue
-from typing import Callable, Iterable, List
+from typing import TYPE_CHECKING, Any, Callable, Generic, Iterable, Literal, NoReturn, TypeVar, overload
 
-logger: logging.Logger = logging.getLogger(__name__)
+if sys.version_info >= (3, 9):
+    import importlib.resources as importlib_resources
+else:
+    import importlib_resources
 
+if (3, 9) <= sys.version_info < (3, 11):
+    # python 3.9 broke generic TypedDict, python 3.11 fixed it
+    from typing_extensions import TypedDict
+else:
+    from typing import TypedDict
+
+if TYPE_CHECKING:
+    from typing_extensions import TypeAlias
+
+EmbeddingsType = TypeVar('EmbeddingsType', bound='list[Any]')
 
-file_manager = ExitStack()
-atexit.register(file_manager.close)  # clean up files on exit
 
 # TODO: provide a config file to make this more robust
-MODEL_LIB_PATH = file_manager.enter_context(importlib.resources.as_file(
-    importlib.resources.files("gpt4all") / "llmodel_DO_NOT_MODIFY" / "build",
-))
+MODEL_LIB_PATH = importlib_resources.files("gpt4all") / "llmodel_DO_NOT_MODIFY" / "build"
 
 
 def load_llmodel_library():
     ext = {"Darwin": "dylib", "Linux": "so", "Windows": "dll"}[platform.system()]
 
     try:
-        # Linux, Windows, MinGW
+        # macOS, Linux, MinGW
         lib = ctypes.CDLL(str(MODEL_LIB_PATH / f"libllmodel.{ext}"))
     except FileNotFoundError:
         if ext != 'dll':
@@ -53,6 +61,7 @@ class LLModelPromptContext(ctypes.Structure):
         ("n_predict", ctypes.c_int32),
         ("top_k", ctypes.c_int32),
         ("top_p", ctypes.c_float),
+        ("min_p", ctypes.c_float),
         ("temp", ctypes.c_float),
         ("n_batch", ctypes.c_int32),
         ("repeat_penalty", ctypes.c_float),
@@ -62,6 +71,7 @@ class LLModelPromptContext(ctypes.Structure):
 
 class LLModelGPUDevice(ctypes.Structure):
     _fields_ = [
+        ("backend", ctypes.c_char_p),
         ("index", ctypes.c_int32),
         ("type", ctypes.c_int32),
         ("heapSize", ctypes.c_size_t),
@@ -79,9 +89,9 @@ llmodel.llmodel_model_create2.restype = ctypes.c_void_p
 llmodel.llmodel_model_destroy.argtypes = [ctypes.c_void_p]
 llmodel.llmodel_model_destroy.restype = None
 
-llmodel.llmodel_loadModel.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
+llmodel.llmodel_loadModel.argtypes = [ctypes.c_void_p, ctypes.c_char_p, ctypes.c_int]
 llmodel.llmodel_loadModel.restype = ctypes.c_bool
-llmodel.llmodel_required_mem.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
+llmodel.llmodel_required_mem.argtypes = [ctypes.c_void_p, ctypes.c_char_p, ctypes.c_int]
 llmodel.llmodel_required_mem.restype = ctypes.c_size_t
 llmodel.llmodel_isModelLoaded.argtypes = [ctypes.c_void_p]
 llmodel.llmodel_isModelLoaded.restype = ctypes.c_bool
@@ -89,25 +99,36 @@ llmodel.llmodel_isModelLoaded.restype = ctypes.c_bool
 PromptCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_int32)
 ResponseCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_int32, ctypes.c_char_p)
 RecalculateCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_bool)
+EmbCancelCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.POINTER(ctypes.c_uint), ctypes.c_uint, ctypes.c_char_p)
 
 llmodel.llmodel_prompt.argtypes = [
     ctypes.c_void_p,
     ctypes.c_char_p,
+    ctypes.c_char_p,
     PromptCallback,
     ResponseCallback,
     RecalculateCallback,
     ctypes.POINTER(LLModelPromptContext),
+    ctypes.c_bool,
+    ctypes.c_char_p,
 ]
 
 llmodel.llmodel_prompt.restype = None
 
-llmodel.llmodel_embedding.argtypes = [
+llmodel.llmodel_embed.argtypes = [
     ctypes.c_void_p,
-    ctypes.c_char_p,
+    ctypes.POINTER(ctypes.c_char_p),
     ctypes.POINTER(ctypes.c_size_t),
+    ctypes.c_char_p,
+    ctypes.c_int,
+    ctypes.POINTER(ctypes.c_size_t),
+    ctypes.c_bool,
+    ctypes.c_bool,
+    EmbCancelCallback,
+    ctypes.POINTER(ctypes.c_char_p),
 ]
 
-llmodel.llmodel_embedding.restype = ctypes.POINTER(ctypes.c_float)
+llmodel.llmodel_embed.restype = ctypes.POINTER(ctypes.c_float)
 
 llmodel.llmodel_free_embedding.argtypes = [ctypes.POINTER(ctypes.c_float)]
 llmodel.llmodel_free_embedding.restype = None
@@ -121,9 +142,9 @@ llmodel.llmodel_set_implementation_search_path.restype = None
 llmodel.llmodel_threadCount.argtypes = [ctypes.c_void_p]
 llmodel.llmodel_threadCount.restype = ctypes.c_int32
 
-llmodel.llmodel_set_implementation_search_path(str(MODEL_LIB_PATH).replace("\\", r"\\").encode("utf-8"))
+llmodel.llmodel_set_implementation_search_path(str(MODEL_LIB_PATH).encode())
 
-llmodel.llmodel_available_gpu_devices.argtypes = [ctypes.c_void_p, ctypes.c_size_t, ctypes.POINTER(ctypes.c_int32)]
+llmodel.llmodel_available_gpu_devices.argtypes = [ctypes.c_size_t, ctypes.POINTER(ctypes.c_int32)]
 llmodel.llmodel_available_gpu_devices.restype = ctypes.POINTER(LLModelGPUDevice)
 
 llmodel.llmodel_gpu_init_gpu_device_by_string.argtypes = [ctypes.c_void_p, ctypes.c_size_t, ctypes.c_char_p]
@@ -138,20 +159,33 @@ llmodel.llmodel_gpu_init_gpu_device_by_int.restype = ctypes.c_bool
 llmodel.llmodel_has_gpu_device.argtypes = [ctypes.c_void_p]
 llmodel.llmodel_has_gpu_device.restype = ctypes.c_bool
 
+llmodel.llmodel_model_backend_name.argtypes = [ctypes.c_void_p]
+llmodel.llmodel_model_backend_name.restype = ctypes.c_char_p
+
+llmodel.llmodel_model_gpu_device_name.argtypes = [ctypes.c_void_p]
+llmodel.llmodel_model_gpu_device_name.restype = ctypes.c_char_p
+
 ResponseCallbackType = Callable[[int, str], bool]
 RawResponseCallbackType = Callable[[int, bytes], bool]
+EmbCancelCallbackType: TypeAlias = 'Callable[[list[int], str], bool]'
 
 
 def empty_response_callback(token_id: int, response: str) -> bool:
     return True
 
 
-def _create_model(model_path: bytes) -> ctypes.c_void_p:
-    err = ctypes.c_char_p()
-    model = llmodel.llmodel_model_create2(model_path, b"auto", ctypes.byref(err))
-    if model is None:
-        raise ValueError(f"Unable to instantiate model: {err.decode()}")
-    return model
+# Symbol to terminate from generator
+class Sentinel(Enum):
+    TERMINATING_SYMBOL = 0
+
+
+class EmbedResult(Generic[EmbeddingsType], TypedDict):
+    embeddings: EmbeddingsType
+    n_prompt_tokens: int
+
+
+class CancellationError(Exception):
+    """raised when embedding is canceled"""
 
 
 class LLModel:
@@ -159,118 +193,117 @@ class LLModel:
     Base class and universal wrapper for GPT4All language models
     built around llmodel C-API.
 
-    Attributes
+    Parameters
     ----------
-    model: llmodel_model
-        Ctype pointer to underlying model
-    model_name: str
-        Model name
+    model_path : str
+        Path to the model.
+    n_ctx : int
+        Maximum size of context window
+    ngl : int
+        Number of GPU layers to use (Vulkan)
+    backend : str
+        Backend to use. One of 'auto', 'cpu', 'metal', 'kompute', or 'cuda'.
     """
 
-    def __init__(self):
-        self.model = None
-        self.model_name = None
-        self.context = None
-        self.llmodel_lib = llmodel
-
+    def __init__(self, model_path: str, n_ctx: int, ngl: int, backend: str):
+        self.model_path = model_path.encode()
+        self.n_ctx = n_ctx
+        self.ngl = ngl
+        self.context: LLModelPromptContext | None = None
         self.buffer = bytearray()
         self.buff_expecting_cont_bytes: int = 0
 
-    def __del__(self):
+        # Construct a model implementation
+        err = ctypes.c_char_p()
+        model = llmodel.llmodel_model_create2(self.model_path, backend.encode(), ctypes.byref(err))
+        if model is None:
+            s = err.value
+            raise RuntimeError(f"Unable to instantiate model: {'null' if s is None else s.decode()}")
+        self.model: ctypes.c_void_p | None = model
+
+    def __del__(self, llmodel=llmodel):
+        if hasattr(self, 'model'):
+            self.close()
+
+    def close(self) -> None:
         if self.model is not None:
-            self.llmodel_lib.llmodel_model_destroy(self.model)
+            llmodel.llmodel_model_destroy(self.model)
+            self.model = None
 
-    def memory_needed(self, model_path: str) -> int:
-        model_path_enc = model_path.encode("utf-8")
-        self.model = _create_model(model_path_enc)
-        return llmodel.llmodel_required_mem(self.model, model_path_enc)
+    def _raise_closed(self) -> NoReturn:
+        raise ValueError("Attempted operation on a closed LLModel")
 
-    def list_gpu(self, model_path: str) -> list:
+    @property
+    def backend(self) -> Literal["cpu", "kompute", "cuda", "metal"]:
+        if self.model is None:
+            self._raise_closed()
+        return llmodel.llmodel_model_backend_name(self.model).decode()
+
+    @property
+    def device(self) -> str | None:
+        if self.model is None:
+            self._raise_closed()
+        dev = llmodel.llmodel_model_gpu_device_name(self.model)
+        return None if dev is None else dev.decode()
+
+    @staticmethod
+    def list_gpus(mem_required: int = 0) -> list[str]:
         """
-        Lists available GPU devices that satisfy the model's memory requirements.
+        List the names of the available GPU devices with at least `mem_required` bytes of VRAM.
 
-        Parameters
-        ----------
-        model_path : str
-            Path to the model.
+        Args:
+            mem_required: The minimum amount of VRAM, in bytes
 
-        Returns
-        -------
-        list
-            A list of LLModelGPUDevice structures representing available GPU devices.
+        Returns:
+            A list of strings representing the names of the available GPU devices.
         """
-        if self.model is not None:
-            model_path_enc = model_path.encode("utf-8")
-            mem_required = llmodel.llmodel_required_mem(self.model, model_path_enc)
-        else:
-            mem_required = self.memory_needed(model_path)
         num_devices = ctypes.c_int32(0)
-        devices_ptr = self.llmodel_lib.llmodel_available_gpu_devices(self.model, mem_required, ctypes.byref(num_devices))
+        devices_ptr = llmodel.llmodel_available_gpu_devices(mem_required, ctypes.byref(num_devices))
         if not devices_ptr:
             raise ValueError("Unable to retrieve available GPU devices")
-        devices = [devices_ptr[i] for i in range(num_devices.value)]
-        return devices
+        return [f'{d.backend.decode()}:{d.name.decode()}' for d in devices_ptr[:num_devices.value]]
 
-    def init_gpu(self, model_path: str, device: str):
-        if self.model is not None:
-            model_path_enc = model_path.encode("utf-8")
-            mem_required = llmodel.llmodel_required_mem(self.model, model_path_enc)
-        else:
-            mem_required = self.memory_needed(model_path)
-        device_enc = device.encode("utf-8")
-        success = self.llmodel_lib.llmodel_gpu_init_gpu_device_by_string(self.model, mem_required, device_enc)
-        if not success:
-            # Retrieve all GPUs without considering memory requirements.
-            num_devices = ctypes.c_int32(0)
-            all_devices_ptr = self.llmodel_lib.llmodel_available_gpu_devices(self.model, 0, ctypes.byref(num_devices))
-            if not all_devices_ptr:
-                raise ValueError("Unable to retrieve list of all GPU devices")
-            all_gpus = [all_devices_ptr[i].name.decode('utf-8') for i in range(num_devices.value)]
+    def init_gpu(self, device: str):
+        if self.model is None:
+            self._raise_closed()
 
-            # Retrieve GPUs that meet the memory requirements using list_gpu
-            available_gpus = [device.name.decode('utf-8') for device in self.list_gpu(model_path)]
+        mem_required = llmodel.llmodel_required_mem(self.model, self.model_path, self.n_ctx, self.ngl)
 
-            # Identify GPUs that are unavailable due to insufficient memory or features
-            unavailable_gpus = set(all_gpus) - set(available_gpus)
+        if llmodel.llmodel_gpu_init_gpu_device_by_string(self.model, mem_required, device.encode()):
+            return
 
-            # Formulate the error message
-            error_msg = "Unable to initialize model on GPU: '{}'.".format(device)
-            error_msg += "\nAvailable GPUs: {}.".format(available_gpus)
-            error_msg += "\nUnavailable GPUs due to insufficient memory or features: {}.".format(unavailable_gpus)
-            raise ValueError(error_msg)
+        all_gpus = self.list_gpus()
+        available_gpus = self.list_gpus(mem_required)
+        unavailable_gpus = set(all_gpus).difference(available_gpus)
 
-    def load_model(self, model_path: str) -> bool:
+        error_msg = "Unable to initialize model on GPU: {!r}".format(device)
+        error_msg += "\nAvailable GPUs: {}".format(available_gpus)
+        error_msg += "\nUnavailable GPUs due to insufficient memory or features: {}".format(unavailable_gpus)
+        raise ValueError(error_msg)
+
+    def load_model(self) -> bool:
         """
         Load model from a file.
 
-        Parameters
-        ----------
-        model_path : str
-            Model filepath
-
         Returns
         -------
         True if model loaded successfully, False otherwise
         """
-        model_path_enc = model_path.encode("utf-8")
-        self.model = _create_model(model_path_enc)
+        if self.model is None:
+            self._raise_closed()
 
-        llmodel.llmodel_loadModel(self.model, model_path_enc)
-
-        filename = os.path.basename(model_path)
-        self.model_name = os.path.splitext(filename)[0]
-
-        if llmodel.llmodel_isModelLoaded(self.model):
-            return True
-        else:
-            return False
+        return llmodel.llmodel_loadModel(self.model, self.model_path, self.n_ctx, self.ngl)
 
     def set_thread_count(self, n_threads):
+        if self.model is None:
+            self._raise_closed()
         if not llmodel.llmodel_isModelLoaded(self.model):
             raise Exception("Model not loaded")
         llmodel.llmodel_setThreadCount(self.model, n_threads)
 
     def thread_count(self):
+        if self.model is None:
+            self._raise_closed()
         if not llmodel.llmodel_isModelLoaded(self.model):
             raise Exception("Model not loaded")
         return llmodel.llmodel_threadCount(self.model)
@@ -280,6 +313,7 @@ class LLModel:
         n_predict: int = 4096,
         top_k: int = 40,
         top_p: float = 0.9,
+        min_p: float = 0.0,
         temp: float = 0.1,
         n_batch: int = 8,
         repeat_penalty: float = 1.2,
@@ -288,7 +322,7 @@ class LLModel:
         reset_context: bool = False,
     ):
         if self.context is None:
-            self.context = LLModelPromptContext(
+            context = LLModelPromptContext(
                 logits_size=0,
                 tokens_size=0,
                 n_past=0,
@@ -296,48 +330,112 @@ class LLModel:
                 n_predict=n_predict,
                 top_k=top_k,
                 top_p=top_p,
+                min_p=min_p,
                 temp=temp,
                 n_batch=n_batch,
                 repeat_penalty=repeat_penalty,
                 repeat_last_n=repeat_last_n,
                 context_erase=context_erase,
             )
-        elif reset_context:
-            self.context.n_past = 0
+            self.context = context
+        else:
+            context = self.context
+            if reset_context:
+                self.context.n_past = 0
 
         self.context.n_predict = n_predict
         self.context.top_k = top_k
         self.context.top_p = top_p
+        self.context.min_p = min_p
         self.context.temp = temp
         self.context.n_batch = n_batch
         self.context.repeat_penalty = repeat_penalty
         self.context.repeat_last_n = repeat_last_n
         self.context.context_erase = context_erase
 
-    def generate_embedding(self, text: str) -> List[float]:
-        if not text:
-            raise ValueError("Text must not be None or empty")
+    @overload
+    def generate_embeddings(
+        self, text: str, prefix: str | None, dimensionality: int, do_mean: bool, atlas: bool,
+        cancel_cb: EmbCancelCallbackType | None,
+    ) -> EmbedResult[list[float]]: ...
+    @overload
+    def generate_embeddings(
+        self, text: list[str], prefix: str | None, dimensionality: int, do_mean: bool, atlas: bool,
+        cancel_cb: EmbCancelCallbackType | None,
+    ) -> EmbedResult[list[list[float]]]: ...
+    @overload
+    def generate_embeddings(
+        self, text: str | list[str], prefix: str | None, dimensionality: int, do_mean: bool, atlas: bool,
+        cancel_cb: EmbCancelCallbackType | None,
+    ) -> EmbedResult[list[Any]]: ...
 
+    def generate_embeddings(
+        self, text: str | list[str], prefix: str | None, dimensionality: int, do_mean: bool, atlas: bool,
+        cancel_cb: EmbCancelCallbackType | None,
+    ) -> EmbedResult[list[Any]]:
+        if not text:
+            raise ValueError("text must not be None or empty")
+
+        if self.model is None:
+            self._raise_closed()
+
+        if single_text := isinstance(text, str):
+            text = [text]
+
+        # prepare input
         embedding_size = ctypes.c_size_t()
-        c_text = ctypes.c_char_p(text.encode('utf-8'))
-        embedding_ptr = llmodel.llmodel_embedding(self.model, c_text, ctypes.byref(embedding_size))
-        embedding_array = [embedding_ptr[i] for i in range(embedding_size.value)]
+        token_count = ctypes.c_size_t()
+        error = ctypes.c_char_p()
+        c_prefix = ctypes.c_char_p() if prefix is None else prefix.encode()
+        c_texts = (ctypes.c_char_p * (len(text) + 1))()
+        for i, t in enumerate(text):
+            c_texts[i] = t.encode()
+
+        def wrap_cancel_cb(batch_sizes: Any, n_batch: int, backend: bytes) -> bool:
+            assert cancel_cb is not None
+            return cancel_cb(batch_sizes[:n_batch], backend.decode())
+
+        cancel_cb_wrapper = EmbCancelCallback() if cancel_cb is None else EmbCancelCallback(wrap_cancel_cb)
+
+        # generate the embeddings
+        embedding_ptr = llmodel.llmodel_embed(
+            self.model, c_texts, ctypes.byref(embedding_size), c_prefix, dimensionality, ctypes.byref(token_count),
+            do_mean, atlas, cancel_cb_wrapper, ctypes.byref(error),
+        )
+
+        if not embedding_ptr:
+            msg = "(unknown error)" if error.value is None else error.value.decode()
+            if msg == "operation was canceled":
+                raise CancellationError(msg)
+            raise RuntimeError(f'Failed to generate embeddings: {msg}')
+
+        # extract output
+        n_embd = embedding_size.value // len(text)
+        embedding_array = [
+            embedding_ptr[i:i + n_embd]
+            for i in range(0, embedding_size.value, n_embd)
+        ]
         llmodel.llmodel_free_embedding(embedding_ptr)
-        return list(embedding_array)
+
+        embeddings = embedding_array[0] if single_text else embedding_array
+        return {'embeddings': embeddings, 'n_prompt_tokens': token_count.value}
 
     def prompt_model(
         self,
         prompt: str,
+        prompt_template: str,
         callback: ResponseCallbackType,
         n_predict: int = 4096,
         top_k: int = 40,
         top_p: float = 0.9,
+        min_p: float = 0.0,
         temp: float = 0.1,
         n_batch: int = 8,
         repeat_penalty: float = 1.2,
         repeat_last_n: int = 10,
         context_erase: float = 0.75,
         reset_context: bool = False,
+        special: bool = False,
     ):
         """
         Generate response from model from a prompt.
@@ -354,23 +452,17 @@ class LLModel:
         None
         """
 
+        if self.model is None:
+            self._raise_closed()
+
         self.buffer.clear()
         self.buff_expecting_cont_bytes = 0
 
-        logger.info(
-            "LLModel.prompt_model -- prompt:\n"
-            + "%s\n"
-            + "===/LLModel.prompt_model -- prompt/===",
-            prompt,
-        )
-
-        prompt_bytes = prompt.encode("utf-8")
-        prompt_ptr = ctypes.c_char_p(prompt_bytes)
-
         self._set_context(
             n_predict=n_predict,
             top_k=top_k,
             top_p=top_p,
+            min_p=min_p,
             temp=temp,
             n_batch=n_batch,
             repeat_penalty=repeat_penalty,
@@ -381,21 +473,24 @@ class LLModel:
 
         llmodel.llmodel_prompt(
             self.model,
-            prompt_ptr,
+            ctypes.c_char_p(prompt.encode()),
+            ctypes.c_char_p(prompt_template.encode()),
             PromptCallback(self._prompt_callback),
             ResponseCallback(self._callback_decoder(callback)),
             RecalculateCallback(self._recalculate_callback),
             self.context,
+            special,
+            ctypes.c_char_p(),
         )
 
 
     def prompt_model_streaming(
-        self, prompt: str, callback: ResponseCallbackType = empty_response_callback, **kwargs
+        self, prompt: str, prompt_template: str, callback: ResponseCallbackType = empty_response_callback, **kwargs
     ) -> Iterable[str]:
-        # Symbol to terminate from generator
-        TERMINATING_SYMBOL = object()
+        if self.model is None:
+            self._raise_closed()
 
-        output_queue: Queue = Queue()
+        output_queue: Queue[str | Sentinel] = Queue()
 
         # Put response tokens into an output queue
         def _generator_callback_wrapper(callback: ResponseCallbackType) -> ResponseCallbackType:
@@ -410,15 +505,15 @@ class LLModel:
 
             return _generator_callback
 
-        def run_llmodel_prompt(prompt: str, callback: ResponseCallbackType, **kwargs):
-            self.prompt_model(prompt, callback, **kwargs)
-            output_queue.put(TERMINATING_SYMBOL)
+        def run_llmodel_prompt(prompt: str, prompt_template: str, callback: ResponseCallbackType, **kwargs):
+            self.prompt_model(prompt, prompt_template, callback, **kwargs)
+            output_queue.put(Sentinel.TERMINATING_SYMBOL)
 
         # Kick off llmodel_prompt in separate thread so we can return generator
         # immediately
         thread = threading.Thread(
             target=run_llmodel_prompt,
-            args=(prompt, _generator_callback_wrapper(callback)),
+            args=(prompt, prompt_template, _generator_callback_wrapper(callback)),
             kwargs=kwargs,
         )
         thread.start()
@@ -426,7 +521,7 @@ class LLModel:
         # Generator
         while True:
             response = output_queue.get()
-            if response is TERMINATING_SYMBOL:
+            if isinstance(response, Sentinel):
                 break
             yield response
 
@@ -449,7 +544,7 @@ class LLModel:
                 else: 
                     # beginning of a byte sequence
                     if len(self.buffer) > 0:
-                        decoded.append(self.buffer.decode('utf-8', 'replace'))
+                        decoded.append(self.buffer.decode(errors='replace'))
 
                         self.buffer.clear()
 
@@ -458,7 +553,7 @@ class LLModel:
 
                 if self.buff_expecting_cont_bytes <= 0: 
                     # received the whole sequence or an out of place continuation byte
-                    decoded.append(self.buffer.decode('utf-8', 'replace'))
+                    decoded.append(self.buffer.decode(errors='replace'))
 
                     self.buffer.clear()
                     self.buff_expecting_cont_bytes = 0

gpt4all-bindings/python/gpt4all/gpt4all.py

@@ -3,30 +3,39 @@ Python only API for running all GPT4All models.
 """
 from __future__ import annotations
 
+import hashlib
 import os
+import platform
+import re
 import sys
 import time
+import warnings
 from contextlib import contextmanager
 from pathlib import Path
-from typing import Any, Dict, Iterable, List, Optional, Union
+from types import TracebackType
+from typing import TYPE_CHECKING, Any, Iterable, Literal, Protocol, overload
 
 import requests
 from requests.exceptions import ChunkedEncodingError
 from tqdm import tqdm
 from urllib3.exceptions import IncompleteRead, ProtocolError
 
-from . import pyllmodel
+from ._pyllmodel import (CancellationError as CancellationError, EmbCancelCallbackType, EmbedResult as EmbedResult,
+                         LLModel, ResponseCallbackType, empty_response_callback)
+
+if TYPE_CHECKING:
+    from typing_extensions import Self, TypeAlias
+
+if sys.platform == 'darwin':
+    import fcntl
 
 # TODO: move to config
-DEFAULT_MODEL_DIRECTORY = os.path.join(str(Path.home()), ".cache", "gpt4all").replace("\\", "\\\\")
+DEFAULT_MODEL_DIRECTORY = Path.home() / ".cache" / "gpt4all"
 
-DEFAULT_MODEL_CONFIG = {
-    "systemPrompt": "",
-    "promptTemplate": "### Human: \n{0}\n### Assistant:\n",
-}
+DEFAULT_PROMPT_TEMPLATE = "### Human:\n{0}\n\n### Assistant:\n"
 
-ConfigType = Dict[str, str]
-MessageType = Dict[str, str]
+ConfigType: TypeAlias = 'dict[str, Any]'
+MessageType: TypeAlias = 'dict[str, str]'
 
 
 class Embed4All:
@@ -34,26 +43,121 @@ class Embed4All:
     Python class that handles embeddings for GPT4All.
     """
 
-    def __init__(self, model_name: Optional[str] = None, n_threads: Optional[int] = None, **kwargs):
+    MIN_DIMENSIONALITY = 64
+
+    def __init__(self, model_name: str | None = None, *, n_threads: int | None = None, device: str | None = None, **kwargs: Any):
         """
         Constructor
 
         Args:
             n_threads: number of CPU threads used by GPT4All. Default is None, then the number of threads are determined automatically.
+            device: The processing unit on which the embedding model will run. See the `GPT4All` constructor for more info.
+            kwargs: Remaining keyword arguments are passed to the `GPT4All` constructor.
         """
-        self.gpt4all = GPT4All(model_name or 'all-MiniLM-L6-v2-f16.gguf', n_threads=n_threads, **kwargs)
+        if model_name is None:
+            model_name = 'all-MiniLM-L6-v2.gguf2.f16.gguf'
+        self.gpt4all = GPT4All(model_name, n_threads=n_threads, device=device, **kwargs)
 
-    def embed(self, text: str) -> List[float]:
+    def __enter__(self) -> Self:
+        return self
+
+    def __exit__(
+        self, typ: type[BaseException] | None, value: BaseException | None, tb: TracebackType | None,
+    ) -> None:
+        self.close()
+
+    def close(self) -> None:
+        """Delete the model instance and free associated system resources."""
+        self.gpt4all.close()
+
+    # return_dict=False
+    @overload
+    def embed(
+        self, text: str, *, prefix: str | None = ..., dimensionality: int | None = ..., long_text_mode: str = ...,
+        return_dict: Literal[False] = ..., atlas: bool = ..., cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> list[float]: ...
+    @overload
+    def embed(
+        self, text: list[str], *, prefix: str | None = ..., dimensionality: int | None = ..., long_text_mode: str = ...,
+        return_dict: Literal[False] = ..., atlas: bool = ..., cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> list[list[float]]: ...
+    @overload
+    def embed(
+        self, text: str | list[str], *, prefix: str | None = ..., dimensionality: int | None = ...,
+        long_text_mode: str = ..., return_dict: Literal[False] = ..., atlas: bool = ...,
+        cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> list[Any]: ...
+
+    # return_dict=True
+    @overload
+    def embed(
+        self, text: str, *, prefix: str | None = ..., dimensionality: int | None = ..., long_text_mode: str = ...,
+        return_dict: Literal[True], atlas: bool = ..., cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> EmbedResult[list[float]]: ...
+    @overload
+    def embed(
+        self, text: list[str], *, prefix: str | None = ..., dimensionality: int | None = ..., long_text_mode: str = ...,
+        return_dict: Literal[True], atlas: bool = ..., cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> EmbedResult[list[list[float]]]: ...
+    @overload
+    def embed(
+        self, text: str | list[str], *, prefix: str | None = ..., dimensionality: int | None = ...,
+        long_text_mode: str = ..., return_dict: Literal[True], atlas: bool = ...,
+        cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> EmbedResult[list[Any]]: ...
+
+    # return type unknown
+    @overload
+    def embed(
+        self, text: str | list[str], *, prefix: str | None = ..., dimensionality: int | None = ...,
+        long_text_mode: str = ..., return_dict: bool = ..., atlas: bool = ...,
+        cancel_cb: EmbCancelCallbackType | None = ...,
+    ) -> Any: ...
+
+    def embed(
+        self, text: str | list[str], *, prefix: str | None = None, dimensionality: int | None = None,
+        long_text_mode: str = "mean", return_dict: bool = False, atlas: bool = False,
+        cancel_cb: EmbCancelCallbackType | None = None,
+    ) -> Any:
         """
-        Generate an embedding.
+        Generate one or more embeddings.
 
         Args:
-            text: The text document to generate an embedding for.
+            text: A text or list of texts to generate embeddings for.
+            prefix: The model-specific prefix representing the embedding task, without the trailing colon. For Nomic
+                Embed, this can be `search_query`, `search_document`, `classification`, or `clustering`. Defaults to
+                `search_document` or equivalent if known; otherwise, you must explicitly pass a prefix or an empty
+                string if none applies.
+            dimensionality: The embedding dimension, for use with Matryoshka-capable models. Defaults to full-size.
+            long_text_mode: How to handle texts longer than the model can accept. One of `mean` or `truncate`.
+            return_dict: Return the result as a dict that includes the number of prompt tokens processed.
+            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.
+            cancel_cb: Called with arguments (batch_sizes, backend_name). Return true to cancel embedding.
 
         Returns:
-            An embedding of your document of text.
+            With return_dict=False, an embedding or list of embeddings of your text(s).
+            With return_dict=True, a dict with keys 'embeddings' and 'n_prompt_tokens'.
+
+        Raises:
+            CancellationError: If cancel_cb returned True and embedding was canceled.
         """
-        return self.gpt4all.model.generate_embedding(text)
+        if dimensionality is None:
+            dimensionality = -1
+        else:
+            if dimensionality <= 0:
+                raise ValueError(f'Dimensionality must be None or a positive integer, got {dimensionality}')
+            if dimensionality < self.MIN_DIMENSIONALITY:
+                warnings.warn(
+                    f'Dimensionality {dimensionality} is less than the suggested minimum of {self.MIN_DIMENSIONALITY}.'
+                    ' Performance may be degraded.'
+                )
+        try:
+            do_mean = {"mean": True, "truncate": False}[long_text_mode]
+        except KeyError:
+            raise ValueError(f"Long text mode must be one of 'mean' or 'truncate', got {long_text_mode!r}")
+        result = self.gpt4all.model.generate_embeddings(text, prefix, dimensionality, do_mean, atlas, cancel_cb)
+        return result if return_dict else result['embeddings']
 
 
 class GPT4All:
@@ -64,11 +168,14 @@ class GPT4All:
     def __init__(
         self,
         model_name: str,
-        model_path: Optional[Union[str, os.PathLike[str]]] = None,
-        model_type: Optional[str] = None,
+        *,
+        model_path: str | os.PathLike[str] | None = None,
+        model_type: str | None = None,
         allow_download: bool = True,
-        n_threads: Optional[int] = None,
-        device: Optional[str] = "cpu",
+        n_threads: int | None = None,
+        device: str | None = None,
+        n_ctx: int = 2048,
+        ngl: int = 100,
         verbose: bool = False,
     ):
         """
@@ -84,46 +191,100 @@ class GPT4All:
             n_threads: number of CPU threads used by GPT4All. Default is None, then the number of threads are determined automatically.
             device: The processing unit on which the GPT4All model will run. It can be set to:
                 - "cpu": Model will run on the central processing unit.
-                - "gpu": Model will run on the best available graphics processing unit, irrespective of its vendor.
-                - "amd", "nvidia", "intel": Model will run on the best available GPU from the specified vendor.
-                Alternatively, a specific GPU name can also be provided, and the model will run on the GPU that matches the name if it's available.
-                Default is "cpu".
+                - "gpu": Use Metal on ARM64 macOS, otherwise the same as "kompute".
+                - "kompute": Use the best GPU provided by the Kompute backend.
+                - "cuda": Use the best GPU provided by the CUDA backend.
+                - "amd", "nvidia": Use the best GPU provided by the Kompute backend from this vendor.
+                - A specific device name from the list returned by `GPT4All.list_gpus()`.
+                Default is Metal on ARM64 macOS, "cpu" otherwise.
 
                 Note: If a selected GPU device does not have sufficient RAM to accommodate the model, an error will be thrown, and the GPT4All instance will be rendered invalid. It's advised to ensure the device has enough memory before initiating the model.
+            n_ctx: Maximum size of context window
+            ngl: Number of GPU layers to use (Vulkan)
+            verbose: If True, print debug messages.
         """
+
         self.model_type = model_type
-        self.model = pyllmodel.LLModel()
+        self._history: list[MessageType] | None = None
+        self._current_prompt_template: str = "{0}"
+
+        device_init = None
+        if sys.platform == 'darwin':
+            if device is None:
+                backend = 'auto'  # 'auto' is effectively 'metal' due to currently non-functional fallback
+            elif device == 'cpu':
+                backend = 'cpu'
+            else:
+                if platform.machine() != 'arm64' or device != 'gpu':
+                    raise ValueError(f'Unknown device for this platform: {device}')
+                backend = 'metal'
+        else:
+            backend = 'kompute'
+            if device is None or device == 'cpu':
+                pass  # use kompute with no device
+            elif device in ('cuda', 'kompute'):
+                backend = device
+                device_init = 'gpu'
+            elif device.startswith('cuda:'):
+                backend = 'cuda'
+                device_init = device.removeprefix('cuda:')
+            else:
+                device_init = device.removeprefix('kompute:')
+
         # Retrieve model and download if allowed
         self.config: ConfigType = self.retrieve_model(model_name, model_path=model_path, allow_download=allow_download, verbose=verbose)
-        if device is not None:
-            if device != "cpu":
-                self.model.init_gpu(model_path=self.config["path"], device=device)
-        self.model.load_model(self.config["path"])
+        self.model = LLModel(self.config["path"], n_ctx, ngl, backend)
+        if device_init is not None:
+            self.model.init_gpu(device_init)
+        self.model.load_model()
         # Set n_threads
         if n_threads is not None:
             self.model.set_thread_count(n_threads)
 
-        self._is_chat_session_activated: bool = False
-        self.current_chat_session: List[MessageType] = empty_chat_session()
-        self._current_prompt_template: str = "{0}"
+    def __enter__(self) -> Self:
+        return self
+
+    def __exit__(
+        self, typ: type[BaseException] | None, value: BaseException | None, tb: TracebackType | None,
+    ) -> None:
+        self.close()
+
+    def close(self) -> None:
+        """Delete the model instance and free associated system resources."""
+        self.model.close()
+
+    @property
+    def backend(self) -> Literal["cpu", "kompute", "cuda", "metal"]:
+        """The name of the llama.cpp backend currently in use. One of "cpu", "kompute", "cuda", or "metal"."""
+        return self.model.backend
+
+    @property
+    def device(self) -> str | None:
+        """The name of the GPU device currently in use, or None for backends other than Kompute or CUDA."""
+        return self.model.device
+
+    @property
+    def current_chat_session(self) -> list[MessageType] | None:
+        return None if self._history is None else list(self._history)
 
     @staticmethod
-    def list_models() -> List[ConfigType]:
+    def list_models() -> list[ConfigType]:
         """
-        Fetch model list from https://gpt4all.io/models/models2.json.
+        Fetch model list from https://gpt4all.io/models/models3.json.
 
         Returns:
             Model list in JSON format.
         """
-        resp = requests.get("https://gpt4all.io/models/models2.json")
+        resp = requests.get("https://gpt4all.io/models/models3.json")
         if resp.status_code != 200:
             raise ValueError(f'Request failed: HTTP {resp.status_code} {resp.reason}')
         return resp.json()
 
-    @staticmethod
+    @classmethod
     def retrieve_model(
+        cls,
         model_name: str,
-        model_path: Optional[Union[str, os.PathLike[str]]] = None,
+        model_path: str | os.PathLike[str] | None = None,
         allow_download: bool = True,
         verbose: bool = False,
     ) -> ConfigType:
@@ -144,59 +305,57 @@ class GPT4All:
         model_filename = append_extension_if_missing(model_name)
 
         # get the config for the model
-        config: ConfigType = DEFAULT_MODEL_CONFIG
+        config: ConfigType = {}
         if allow_download:
-            available_models = GPT4All.list_models()
+            available_models = cls.list_models()
 
             for m in available_models:
                 if model_filename == m["filename"]:
+                    tmpl = m.get("promptTemplate", DEFAULT_PROMPT_TEMPLATE)
+                    # change to Python-style formatting
+                    m["promptTemplate"] = tmpl.replace("%1", "{0}", 1).replace("%2", "{1}", 1)
                     config.update(m)
-                    config["systemPrompt"] = config["systemPrompt"].strip()
-                    config["promptTemplate"] = config["promptTemplate"].replace(
-                        "%1", "{0}", 1
-                    )  # change to Python-style formatting
                     break
 
         # Validate download directory
         if model_path is None:
             try:
                 os.makedirs(DEFAULT_MODEL_DIRECTORY, exist_ok=True)
-            except OSError as exc:
-                raise ValueError(
-                    f"Failed to create model download directory at {DEFAULT_MODEL_DIRECTORY}: {exc}. "
-                    "Please specify model_path."
-                )
+            except OSError as e:
+                raise RuntimeError("Failed to create model download directory") from e
             model_path = DEFAULT_MODEL_DIRECTORY
         else:
-            model_path = str(model_path).replace("\\", "\\\\")
+            model_path = Path(model_path)
 
-        if not os.path.exists(model_path):
-            raise ValueError(f"Invalid model directory: {model_path}")
+        if not model_path.exists():
+            raise FileNotFoundError(f"Model directory does not exist: {model_path!r}")
 
-        model_dest = os.path.join(model_path, model_filename).replace("\\", "\\\\")
-        if os.path.exists(model_dest):
-            config.pop("url", None)
-            config["path"] = model_dest
+        model_dest = model_path / model_filename
+        if model_dest.exists():
+            config["path"] = str(model_dest)
             if verbose:
-                print("Found model file at", model_dest, file=sys.stderr)
-
-        # If model file does not exist, download
+                print(f"Found model file at {str(model_dest)!r}", file=sys.stderr)
         elif allow_download:
-            url = config.pop("url", None)
-
-            config["path"] = GPT4All.download_model(model_filename, model_path, verbose=verbose, url=url)
+            # If model file does not exist, download
+            filesize = config.get("filesize")
+            config["path"] = str(cls.download_model(
+                model_filename, model_path, verbose=verbose, url=config.get("url"),
+                expected_size=None if filesize is None else int(filesize), expected_md5=config.get("md5sum"),
+            ))
         else:
-            raise ValueError("Failed to retrieve model")
+            raise FileNotFoundError(f"Model file does not exist: {model_dest!r}")
 
         return config
 
     @staticmethod
     def download_model(
         model_filename: str,
-        model_path: Union[str, os.PathLike[str]],
+        model_path: str | os.PathLike[str],
         verbose: bool = True,
-        url: Optional[str] = None,
-    ) -> str:
+        url: str | None = None,
+        expected_size: int | None = None,
+        expected_md5: str | None = None,
+    ) -> str | os.PathLike[str]:
         """
         Download model from https://gpt4all.io.
 
@@ -205,30 +364,30 @@ class GPT4All:
             model_path: Path to download model to.
             verbose: If True (default), print debug messages.
             url: the models remote url (e.g. may be hosted on HF)
+            expected_size: The expected size of the download.
+            expected_md5: The expected MD5 hash of the download.
 
         Returns:
             Model file destination.
         """
 
-        def get_download_url(model_filename):
-            if url:
-                return url
-            return f"https://gpt4all.io/models/gguf/{model_filename}"
-
         # Download model
-        download_path = os.path.join(model_path, model_filename).replace("\\", "\\\\")
-        download_url = get_download_url(model_filename)
+        if url is None:
+            url = f"https://gpt4all.io/models/gguf/{model_filename}"
 
         def make_request(offset=None):
             headers = {}
             if offset:
                 print(f"\nDownload interrupted, resuming from byte position {offset}", file=sys.stderr)
                 headers['Range'] = f'bytes={offset}-'  # resume incomplete response
-            response = requests.get(download_url, stream=True, headers=headers)
+                headers["Accept-Encoding"] = "identity"  # Content-Encoding changes meaning of ranges
+            response = requests.get(url, stream=True, headers=headers)
             if response.status_code not in (200, 206):
                 raise ValueError(f'Request failed: HTTP {response.status_code} {response.reason}')
             if offset and (response.status_code != 206 or str(offset) not in response.headers.get('Content-Range', '')):
                 raise ValueError('Connection was interrupted and server does not support range requests')
+            if (enc := response.headers.get("Content-Encoding")) is not None:
+                raise ValueError(f"Expected identity Content-Encoding, got {enc}")
             return response
 
         response = make_request()
@@ -236,60 +395,109 @@ class GPT4All:
         total_size_in_bytes = int(response.headers.get("content-length", 0))
         block_size = 2**20  # 1 MB
 
-        with open(download_path, "wb") as file, \
-                tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True) as progress_bar:
+        partial_path = Path(model_path) / (model_filename + ".part")
+
+        with open(partial_path, "w+b") as partf:
             try:
-                while True:
-                    last_progress = progress_bar.n
-                    try:
-                        for data in response.iter_content(block_size):
-                            file.write(data)
-                            progress_bar.update(len(data))
-                    except ChunkedEncodingError as cee:
-                        if cee.args and isinstance(pe := cee.args[0], ProtocolError):
-                            if len(pe.args) >= 2 and isinstance(ir := pe.args[1], IncompleteRead):
-                                assert progress_bar.n <= ir.partial  # urllib3 may be ahead of us but never behind
-                                # the socket was closed during a read - retry
-                                response = make_request(progress_bar.n)
-                                continue
-                        raise
-                    if total_size_in_bytes != 0 and progress_bar.n < total_size_in_bytes:
-                        if progress_bar.n == last_progress:
-                            raise RuntimeError('Download not making progress, aborting.')
-                        # server closed connection prematurely - retry
-                        response = make_request(progress_bar.n)
-                        continue
-                    break
-            except Exception:
+                with tqdm(desc="Downloading", total=total_size_in_bytes, unit="iB", unit_scale=True) as progress_bar:
+                    while True:
+                        last_progress = progress_bar.n
+                        try:
+                            for data in response.iter_content(block_size):
+                                partf.write(data)
+                                progress_bar.update(len(data))
+                        except ChunkedEncodingError as cee:
+                            if cee.args and isinstance(pe := cee.args[0], ProtocolError):
+                                if len(pe.args) >= 2 and isinstance(ir := pe.args[1], IncompleteRead):
+                                    assert progress_bar.n <= ir.partial  # urllib3 may be ahead of us but never behind
+                                    # the socket was closed during a read - retry
+                                    response = make_request(progress_bar.n)
+                                    continue
+                            raise
+                        if total_size_in_bytes != 0 and progress_bar.n < total_size_in_bytes:
+                            if progress_bar.n == last_progress:
+                                raise RuntimeError("Download not making progress, aborting.")
+                            # server closed connection prematurely - retry
+                            response = make_request(progress_bar.n)
+                            continue
+                        break
+
+                # verify file integrity
+                file_size = partf.tell()
+                if expected_size is not None and file_size != expected_size:
+                    raise ValueError(f"Expected file size of {expected_size} bytes, got {file_size}")
+                if expected_md5 is not None:
+                    partf.seek(0)
+                    hsh = hashlib.md5()
+                    with tqdm(desc="Verifying", total=file_size, unit="iB", unit_scale=True) as bar:
+                        while chunk := partf.read(block_size):
+                            hsh.update(chunk)
+                            bar.update(len(chunk))
+                    if hsh.hexdigest() != expected_md5.lower():
+                        raise ValueError(f"Expected MD5 hash of {expected_md5!r}, got {hsh.hexdigest()!r}")
+            except:
                 if verbose:
                     print("Cleaning up the interrupted download...", file=sys.stderr)
                 try:
-                    os.remove(download_path)
+                    os.remove(partial_path)
                 except OSError:
                     pass
                 raise
 
-        if os.name == 'nt':
-            time.sleep(2)  # Sleep for a little bit so Windows can remove file lock
+            # flush buffers and sync the inode
+            partf.flush()
+            _fsync(partf)
+
+        # move to final destination
+        download_path = Path(model_path) / model_filename
+        try:
+            os.rename(partial_path, download_path)
+        except FileExistsError:
+            try:
+                os.remove(partial_path)
+            except OSError:
+                pass
+            raise
 
         if verbose:
-            print("Model downloaded at:", download_path, file=sys.stderr)
+            print(f"Model downloaded to {str(download_path)!r}", file=sys.stderr)
         return download_path
 
+    @overload
+    def generate(
+        self, prompt: str, *, max_tokens: int = ..., temp: float = ..., top_k: int = ..., top_p: float = ...,
+        min_p: float = ..., repeat_penalty: float = ..., repeat_last_n: int = ..., n_batch: int = ...,
+        n_predict: int | None = ..., streaming: Literal[False] = ..., callback: ResponseCallbackType = ...,
+    ) -> str: ...
+    @overload
+    def generate(
+        self, prompt: str, *, max_tokens: int = ..., temp: float = ..., top_k: int = ..., top_p: float = ...,
+        min_p: float = ..., repeat_penalty: float = ..., repeat_last_n: int = ..., n_batch: int = ...,
+        n_predict: int | None = ..., streaming: Literal[True], callback: ResponseCallbackType = ...,
+    ) -> Iterable[str]: ...
+    @overload
+    def generate(
+        self, prompt: str, *, max_tokens: int = ..., temp: float = ..., top_k: int = ..., top_p: float = ...,
+        min_p: float = ..., repeat_penalty: float = ..., repeat_last_n: int = ..., n_batch: int = ...,
+        n_predict: int | None = ..., streaming: bool, callback: ResponseCallbackType = ...,
+    ) -> Any: ...
+
     def generate(
         self,
         prompt: str,
+        *,
         max_tokens: int = 200,
         temp: float = 0.7,
         top_k: int = 40,
         top_p: float = 0.4,
+        min_p: float = 0.0,
         repeat_penalty: float = 1.18,
         repeat_last_n: int = 64,
         n_batch: int = 8,
-        n_predict: Optional[int] = None,
+        n_predict: int | None = None,
         streaming: bool = False,
-        callback: pyllmodel.ResponseCallbackType = pyllmodel.empty_response_callback,
-    ) -> Union[str, Iterable[str]]:
+        callback: ResponseCallbackType = empty_response_callback,
+    ) -> Any:
         """
         Generate outputs from any GPT4All model.
 
@@ -299,6 +507,7 @@ class GPT4All:
             temp: The model temperature. Larger values increase creativity but decrease factuality.
             top_k: Randomly sample from the top_k most likely tokens at each generation step. Set this to 1 for greedy decoding.
             top_p: Randomly sample at each generation step from the top most likely tokens whose probabilities add up to top_p.
+            min_p: Randomly sample at each generation step from the top most likely tokens whose probabilities are at least min_p.
             repeat_penalty: Penalize the model for repetition. Higher values result in less repetition.
             repeat_last_n: How far in the models generation history to apply the repeat penalty.
             n_batch: Number of prompt tokens processed in parallel. Larger values decrease latency but increase resource requirements.
@@ -311,44 +520,61 @@ class GPT4All:
         """
 
         # Preparing the model request
-        generate_kwargs: Dict[str, Any] = dict(
+        generate_kwargs: dict[str, Any] = dict(
             temp=temp,
             top_k=top_k,
             top_p=top_p,
+            min_p=min_p,
             repeat_penalty=repeat_penalty,
             repeat_last_n=repeat_last_n,
             n_batch=n_batch,
             n_predict=n_predict if n_predict is not None else max_tokens,
         )
 
-        if self._is_chat_session_activated:
+        if self._history is not None:
             # check if there is only one message, i.e. system prompt:
-            generate_kwargs["reset_context"] = len(self.current_chat_session) == 1
-            self.current_chat_session.append({"role": "user", "content": prompt})
+            reset = len(self._history) == 1
+            self._history.append({"role": "user", "content": prompt})
 
-            prompt = self._format_chat_prompt_template(
-                messages=self.current_chat_session[-1:],
-                default_prompt_header=self.current_chat_session[0]["content"]
-                if generate_kwargs["reset_context"]
-                else "",
-            )
+            fct_func = self._format_chat_prompt_template.__func__  # type: ignore[attr-defined]
+            if fct_func is GPT4All._format_chat_prompt_template:
+                if reset:
+                    # ingest system prompt
+                    # use "%1%2" and not "%1" to avoid implicit whitespace
+                    self.model.prompt_model(self._history[0]["content"], "%1%2",
+                                            empty_response_callback,
+                                            n_batch=n_batch, n_predict=0, reset_context=True, special=True)
+                prompt_template = self._current_prompt_template.format("%1", "%2")
+            else:
+                warnings.warn(
+                    "_format_chat_prompt_template is deprecated. Please use a chat session with a prompt template.",
+                    DeprecationWarning,
+                )
+                # special tokens won't be processed
+                prompt = self._format_chat_prompt_template(
+                    self._history[-1:],
+                    self._history[0]["content"] if reset else "",
+                )
+                prompt_template = "%1"
+                generate_kwargs["reset_context"] = reset
         else:
+            prompt_template = "%1"
             generate_kwargs["reset_context"] = True
 
         # Prepare the callback, process the model response
-        output_collector: List[MessageType]
+        output_collector: list[MessageType]
         output_collector = [
             {"content": ""}
-        ]  # placeholder for the self.current_chat_session if chat session is not activated
+        ]  # placeholder for the self._history if chat session is not activated
 
-        if self._is_chat_session_activated:
-            self.current_chat_session.append({"role": "assistant", "content": ""})
-            output_collector = self.current_chat_session
+        if self._history is not None:
+            self._history.append({"role": "assistant", "content": ""})
+            output_collector = self._history
 
         def _callback_wrapper(
-            callback: pyllmodel.ResponseCallbackType,
-            output_collector: List[MessageType],
-        ) -> pyllmodel.ResponseCallbackType:
+            callback: ResponseCallbackType,
+            output_collector: list[MessageType],
+        ) -> ResponseCallbackType:
             def _callback(token_id: int, response: str) -> bool:
                 nonlocal callback, output_collector
 
@@ -361,14 +587,16 @@ class GPT4All:
         # Send the request to the model
         if streaming:
             return self.model.prompt_model_streaming(
-                prompt=prompt,
-                callback=_callback_wrapper(callback, output_collector),
+                prompt,
+                prompt_template,
+                _callback_wrapper(callback, output_collector),
                 **generate_kwargs,
             )
 
         self.model.prompt_model(
-            prompt=prompt,
-            callback=_callback_wrapper(callback, output_collector),
+            prompt,
+            prompt_template,
+            _callback_wrapper(callback, output_collector),
             **generate_kwargs,
         )
 
@@ -377,8 +605,8 @@ class GPT4All:
     @contextmanager
     def chat_session(
         self,
-        system_prompt: str = "",
-        prompt_template: str = "",
+        system_prompt: str | None = None,
+        prompt_template: str | None = None,
     ):
         """
         Context manager to hold an inference optimized chat session with a GPT4All model.
@@ -387,27 +615,51 @@ class GPT4All:
             system_prompt: An initial instruction for the model.
             prompt_template: Template for the prompts with {0} being replaced by the user message.
         """
-        # Code to acquire resource, e.g.:
-        self._is_chat_session_activated = True
-        self.current_chat_session = empty_chat_session(system_prompt or self.config["systemPrompt"])
-        self._current_prompt_template = prompt_template or self.config["promptTemplate"]
+
+        if system_prompt is None:
+            system_prompt = self.config.get("systemPrompt", "")
+
+        if prompt_template is None:
+            if (tmpl := self.config.get("promptTemplate")) is None:
+                warnings.warn("Use of a sideloaded model or allow_download=False without specifying a prompt template "
+                              "is deprecated. Defaulting to Alpaca.", DeprecationWarning)
+                tmpl = DEFAULT_PROMPT_TEMPLATE
+            prompt_template = tmpl
+
+        if re.search(r"%1(?![0-9])", prompt_template):
+            raise ValueError("Prompt template containing a literal '%1' is not supported. For a prompt "
+                             "placeholder, please use '{0}' instead.")
+
+        self._history = [{"role": "system", "content": system_prompt}]
+        self._current_prompt_template = prompt_template
         try:
             yield self
         finally:
-            # Code to release resource, e.g.:
-            self._is_chat_session_activated = False
-            self.current_chat_session = empty_chat_session()
+            self._history = None
             self._current_prompt_template = "{0}"
 
+    @staticmethod
+    def list_gpus() -> list[str]:
+        """
+        List the names of the available GPU devices.
+
+        Returns:
+            A list of strings representing the names of the available GPU devices.
+        """
+        return LLModel.list_gpus()
+
     def _format_chat_prompt_template(
         self,
-        messages: List[MessageType],
+        messages: list[MessageType],
         default_prompt_header: str = "",
         default_prompt_footer: str = "",
     ) -> str:
         """
         Helper method for building a prompt from list of messages using the self._current_prompt_template as a template for each message.
 
+        Warning:
+            This function was deprecated in version 2.3.0, and will be removed in a future release.
+
         Args:
             messages:  List of dictionaries. Each dictionary should have a "role" key
                 with value of "system", "assistant", or "user" and a "content" key with a
@@ -419,24 +671,6 @@ class GPT4All:
             Formatted prompt.
         """
 
-        if isinstance(default_prompt_header, bool):
-            import warnings
-
-            warnings.warn(
-                "Using True/False for the 'default_prompt_header' is deprecated. Use a string instead.",
-                DeprecationWarning,
-            )
-            default_prompt_header = ""
-
-        if isinstance(default_prompt_footer, bool):
-            import warnings
-
-            warnings.warn(
-                "Using True/False for the 'default_prompt_footer' is deprecated. Use a string instead.",
-                DeprecationWarning,
-            )
-            default_prompt_footer = ""
-
         full_prompt = default_prompt_header + "\n\n" if default_prompt_header != "" else ""
 
         for message in messages:
@@ -452,11 +686,23 @@ class GPT4All:
         return full_prompt
 
 
-def empty_chat_session(system_prompt: str = "") -> List[MessageType]:
-    return [{"role": "system", "content": system_prompt}]
-
-
 def append_extension_if_missing(model_name):
     if not model_name.endswith((".bin", ".gguf")):
         model_name += ".gguf"
     return model_name
+
+
+class _HasFileno(Protocol):
+    def fileno(self) -> int: ...
+
+
+def _fsync(fd: int | _HasFileno) -> None:
+    if sys.platform == 'darwin':
+        # Apple's fsync does not flush the drive write cache
+        try:
+            fcntl.fcntl(fd, fcntl.F_FULLFSYNC)
+        except OSError:
+            pass  # fall back to fsync
+        else:
+            return
+    os.fsync(fd)

gpt4all-bindings/python/gpt4all/tests/test_gpt4all.py

@@ -28,12 +28,8 @@ def test_inference():
     assert len(tokens) > 0
 
     with model.chat_session():
-        tokens = list(model.generate(prompt='hello', top_k=1, streaming=True))
-        model.current_chat_session.append({'role': 'assistant', 'content': ''.join(tokens)})
-
-        tokens = list(model.generate(prompt='write me a poem about dogs', top_k=1, streaming=True))
-        model.current_chat_session.append({'role': 'assistant', 'content': ''.join(tokens)})
-
+        model.generate(prompt='hello', top_k=1, streaming=True)
+        model.generate(prompt='write me a poem about dogs', top_k=1, streaming=True)
         print(model.current_chat_session)
 
 
@@ -115,13 +111,13 @@ def test_empty_embedding():
         output = embedder.embed(text)
 
 def test_download_model(tmp_path: Path):
-    import gpt4all.gpt4all
-    old_default_dir = gpt4all.gpt4all.DEFAULT_MODEL_DIRECTORY
-    gpt4all.gpt4all.DEFAULT_MODEL_DIRECTORY = tmp_path  # temporary pytest directory to ensure a download happens
+    from gpt4all import gpt4all
+    old_default_dir = gpt4all.DEFAULT_MODEL_DIRECTORY
+    gpt4all.DEFAULT_MODEL_DIRECTORY = tmp_path  # temporary pytest directory to ensure a download happens
     try:
         model = GPT4All(model_name='ggml-all-MiniLM-L6-v2-f16.bin')
         model_path = tmp_path / model.config['filename']
         assert model_path.absolute() == Path(model.config['path']).absolute()
         assert model_path.stat().st_size == int(model.config['filesize'])
     finally:
-        gpt4all.gpt4all.DEFAULT_MODEL_DIRECTORY = old_default_dir
+        gpt4all.DEFAULT_MODEL_DIRECTORY = old_default_dir

gpt4all-bindings/python/mkdocs.yml

@@ -14,10 +14,8 @@ nav:
       - 'GPT4All in Python':
         - 'Generation': 'gpt4all_python.md'
         - 'Embedding': 'gpt4all_python_embedding.md'
-      - 'GPT4ALL in NodeJs': 'gpt4all_typescript.md'
+      - 'GPT4ALL in NodeJs': 'gpt4all_nodejs.md'
       - 'gpt4all_cli.md'
-#    - 'Tutorials':
-#      - 'gpt4all_modal.md'
     - 'Wiki':
       - 'gpt4all_faq.md'
 
@@ -44,8 +42,8 @@ markdown_extensions:
   - pymdownx.tabbed:
       alternate_style: true
   - pymdownx.emoji:
-      emoji_index: !!python/name:materialx.emoji.twemoji
-      emoji_generator: !!python/name:materialx.emoji.to_svg
+      emoji_index: !!python/name:material.extensions.emoji.twemoji
+      emoji_generator: !!python/name:material.extensions.emoji.to_svg
       options:
         custom_icons:
           - docs/overrides/.icons

gpt4all-bindings/python/setup.py

@@ -1,12 +1,13 @@
 from setuptools import setup, find_packages
 import os
+import pathlib
 import platform
 import shutil
 
 package_name = "gpt4all"
 
 # Define the location of your prebuilt C library files
-SRC_CLIB_DIRECtORY = os.path.join("..", "..", "gpt4all-backend")
+SRC_CLIB_DIRECTORY = os.path.join("..", "..", "gpt4all-backend")
 SRC_CLIB_BUILD_DIRECTORY = os.path.join("..", "..", "gpt4all-backend", "build") 
 
 LIB_NAME = "llmodel"
@@ -44,7 +45,7 @@ def copy_prebuilt_C_lib(src_dir, dest_dir, dest_build_dir):
                 d = os.path.join(dest_dir, item)
                 shutil.copy2(s, d)
                 files_copied += 1
-            if item.endswith(lib_ext) or item.endswith('.metal'):
+            if item.endswith(lib_ext) or item.endswith('.metallib'):
                 s = os.path.join(dirpath, item)
                 d = os.path.join(dest_build_dir, item)
                 shutil.copy2(s, d)
@@ -55,17 +56,29 @@ def copy_prebuilt_C_lib(src_dir, dest_dir, dest_build_dir):
 
 # NOTE: You must provide correct path to the prebuilt llmodel C library. 
 # Specifically, the llmodel.h and C shared library are needed.
-copy_prebuilt_C_lib(SRC_CLIB_DIRECtORY,
+copy_prebuilt_C_lib(SRC_CLIB_DIRECTORY,
                     DEST_CLIB_DIRECTORY,
                     DEST_CLIB_BUILD_DIRECTORY)
 
+
+def get_long_description():
+    with open(pathlib.Path(__file__).parent / "README.md", encoding="utf-8") as fp:
+        return fp.read()
+
+
 setup(
     name=package_name,
-    version="2.0.2",
+    version="2.7.0",
     description="Python bindings for GPT4All",
+    long_description=get_long_description(),
+    long_description_content_type="text/markdown",
     author="Nomic and the Open Source Community",
     author_email="support@nomic.ai",
-    url="https://pypi.org/project/gpt4all/",
+    url="https://gpt4all.io/",
+    project_urls={
+        "Documentation": "https://docs.gpt4all.io/gpt4all_python.html",
+        "Source code": "https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/python",
+    },
     classifiers = [
         "Programming Language :: Python :: 3",
         "License :: OSI Approved :: MIT License",
@@ -73,7 +86,12 @@ setup(
     ],
     python_requires='>=3.8',
     packages=find_packages(),
-    install_requires=['requests', 'tqdm'],
+    install_requires=[
+        'requests',
+        'tqdm',
+        'importlib_resources; python_version < "3.9"',
+        'typing-extensions>=4.3.0; python_version >= "3.9" and python_version < "3.11"',
+    ],
     extras_require={
         'dev': [
             'pytest',
@@ -85,7 +103,8 @@ setup(
             'mkdocstrings[python]',
             'mkdocs-jupyter',
             'black',
-            'isort'
+            'isort',
+            'typing-extensions>=3.10',
         ]
     },
     package_data={'llmodel': [os.path.join(DEST_CLIB_DIRECTORY, "*")]},