Fix state issues with reloading model.

Signed-off-by: Adam Treat <treat.adam@gmail.com>

.circleci/config.yml

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

.circleci/continue_config.yml

@@ -5,6 +5,9 @@ orbs:
   node: circleci/node@5.1
 
 parameters:
+  run-all-workflows:
+    type: boolean
+    default: false
   run-default-workflow:
     type: boolean
     default: false
@@ -27,7 +30,176 @@ jobs:
       - image: circleci/python:3.7
     steps:
       - run: echo "CircleCI pipeline triggered"
-
+  build-offline-chat-installer-macos:
+    macos:
+      xcode: 14.0.0
+    steps:
+      - checkout
+      - run:
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
+      - restore_cache:  # this is the new step to restore cache
+          keys:
+            - macos-qt-cache_v2
+      - 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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.46 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+              hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
+            fi
+      - save_cache:  # this is the new step to save cache
+          key: macos-qt-cache_v2
+          paths:
+            - ~/Qt
+      - run:
+          name: Build
+          command: |
+            mkdir build
+            cd build
+            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.6/bin
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
+              -DCMAKE_GENERATOR:STRING=Ninja \
+              -DBUILD_UNIVERSAL=ON \
+              -DMACDEPLOYQT=~/Qt/6.5.1/macos/bin/macdeployqt \
+              -DGPT4ALL_OFFLINE_INSTALLER=ON \
+              -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.5.1/macos/lib/cmake/Qt6 \
+              -DCMAKE_MAKE_PROGRAM:FILEPATH=~/Qt/Tools/Ninja/ninja \
+              -S ../gpt4all-chat \
+              -B .
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target all
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target install
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target package
+            mkdir upload
+            cp gpt4all-installer-* upload
+      - store_artifacts:
+          path: build/upload
+  build-offline-chat-installer-linux:
+    machine:
+      image: ubuntu-2204:2023.04.2
+    steps:
+      - checkout
+      - run:
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
+      - restore_cache:  # this is the new step to restore cache
+          keys:
+            - linux-qt-cache
+      - 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
+      - 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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.46 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver qt.qt6.651.qtwaylandcompositor
+            fi
+      - save_cache:  # this is the new step to save cache
+          key: linux-qt-cache
+          paths:
+            - ~/Qt
+      - run:
+          name: Build linuxdeployqt
+          command: |
+            git clone https://github.com/nomic-ai/linuxdeployqt
+            cd linuxdeployqt && qmake && sudo make install
+      - run:
+          name: Build
+          command: |
+            set -eo pipefail
+            export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
+            export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.6/bin
+            mkdir build
+            cd build
+            mkdir upload
+            ~/Qt/Tools/CMake/bin/cmake -DGPT4ALL_OFFLINE_INSTALLER=ON -DCMAKE_BUILD_TYPE=Release -S ../gpt4all-chat -B .
+            ~/Qt/Tools/CMake/bin/cmake --build . --target all
+            ~/Qt/Tools/CMake/bin/cmake --build . --target install
+            ~/Qt/Tools/CMake/bin/cmake --build . --target package
+            cp gpt4all-installer-* upload
+      - store_artifacts:
+          path: build/upload
+  build-offline-chat-installer-windows:
+    machine:
+      image: 'windows-server-2019-vs2019:2022.08.1'
+      resource_class: windows.large
+      shell: powershell.exe -ExecutionPolicy Bypass
+    steps:
+      - checkout
+      - run:
+          name: Update Submodules
+          command: |
+            git submodule sync
+            git submodule update --init --recursive
+      - restore_cache:  # this is the new step to restore cache
+          keys:
+            - windows-qt-cache
+      - 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
+            }
+      - save_cache:  # this is the new step to save cache
+          key: windows-qt-cache
+          paths:
+            - C:\Qt
+      - run:
+          name: Install VulkanSDK
+          command: |
+            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
+      - run:
+          name: Build
+          command: |
+            $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Windows Kits\10\bin\x64"
+            $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Windows Kits\10\bin\10.0.22000.0\x64"
+            $Env:PATH = "${Env:PATH};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX64\x64"
+            $Env:PATH = "${Env:PATH};C:\VulkanSDK\1.3.261.1\bin"
+            $Env:PATH = "${Env:PATH};C:\Qt\Tools\QtInstallerFramework\4.6\bin"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\ucrt\x64"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22000.0\um\x64"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\lib\x64"
+            $Env:LIB = "${Env:LIB};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\lib\x64"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\ucrt"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\um"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\shared"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\winrt"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Windows Kits\10\include\10.0.22000.0\cppwinrt"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\VS\include"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include"
+            $Env:INCLUDE = "${Env:INCLUDE};C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include"
+            mkdir build
+            cd build
+            & "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" `
+              "-DGPT4ALL_OFFLINE_INSTALLER=ON" `
+              "-S ..\gpt4all-chat" `
+              "-B ."
+            & "C:\Qt\Tools\Ninja\ninja.exe"
+            & "C:\Qt\Tools\Ninja\ninja.exe" install
+            & "C:\Qt\Tools\Ninja\ninja.exe" package
+            mkdir upload
+            copy gpt4all-installer-win64.exe upload
+      - store_artifacts:
+          path: build/upload
   build-gpt4all-chat-linux:
     machine:
       image: ubuntu-2204:2023.04.2
@@ -63,10 +235,8 @@ jobs:
           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
+            ~/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:
@@ -118,17 +288,16 @@ 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:
@@ -159,16 +328,15 @@ jobs:
       - 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
@@ -178,6 +346,7 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - node/install-packages:
           pkg-manager: yarn
           app-dir: gpt4all-bindings/typescript
@@ -232,18 +401,17 @@ 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
-            cmake ..
-            cmake --build . --parallel
+            cmake -B build
+            cmake --build build --parallel
       - run:
           name: Build wheel
           command: |
             cd gpt4all-bindings/python/
             python setup.py bdist_wheel --plat-name=manylinux1_x86_64
+      - store_artifacts:
+          path: gpt4all-bindings/python/dist
       - persist_to_workspace:
           root: gpt4all-bindings/python/dist
           paths:
@@ -263,18 +431,17 @@ 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: |
             cd gpt4all-bindings/python
-            python setup.py bdist_wheel --plat-name=macosx_10_9_universal2
+            python setup.py bdist_wheel --plat-name=macosx_10_15_universal2
+      - store_artifacts:
+          path: gpt4all-bindings/python/dist
       - persist_to_workspace:
           root: gpt4all-bindings/python/dist
           paths:
@@ -288,9 +455,6 @@ jobs:
       - run:
           name: Install MinGW64
           command: choco install -y mingw --force --no-progress
-      - run:
-          name: Add MinGW64 to PATH
-          command: $env:Path += ";C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin"
       - run:
           name: Install VulkanSDK
           command: |
@@ -306,14 +470,13 @@ 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:\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.
@@ -323,9 +486,11 @@ jobs:
             cd gpt4all
             mkdir llmodel_DO_NOT_MODIFY
             mkdir llmodel_DO_NOT_MODIFY/build/
-            cp 'C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin\*dll' 'llmodel_DO_NOT_MODIFY/build/'
+            cp 'C:\ProgramData\mingw64\mingw64\bin\*dll' 'llmodel_DO_NOT_MODIFY/build/'
             cd ..
             python setup.py bdist_wheel --plat-name=win_amd64
+      - store_artifacts:
+          path: gpt4all-bindings/python/dist
       - persist_to_workspace:
           root: gpt4all-bindings/python/dist
           paths:
@@ -442,7 +607,7 @@ jobs:
       - run:
           name: Build Libraries
           command: |
-            $MinGWBin = "C:\ProgramData\chocolatey\lib\mingw\tools\install\mingw64\bin"
+            $MinGWBin = "C:\ProgramData\mingw64\mingw64\bin"
             $Env:Path += ";$MinGwBin"
             $Env:Path += ";C:\Program Files\CMake\bin"
             $Env:Path += ";C:\VulkanSDK\1.3.261.1\bin"
@@ -499,7 +664,7 @@ jobs:
 
   build-csharp-linux:
     docker:
-      - image: mcr.microsoft.com/dotnet/sdk:7.0-jammy # Ubuntu 22.04
+      - image: mcr.microsoft.com/dotnet/sdk:8.0
     steps:
       - checkout
       - attach_workspace:
@@ -555,6 +720,10 @@ jobs:
             - gpt4all-csharp-nuget-packages-win
       - attach_workspace:
           at: C:\Users\circleci\workspace
+      - run:
+          name: "Install .NET"
+          command: |
+            choco install -y dotnet-8.0-sdk
       - run:
           name: "Prepare Native Libs"
           command: |
@@ -602,7 +771,8 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install --cask dotnet-sdk
+            brew tap isen-ng/dotnet-sdk-versions
+            brew install --cask dotnet-sdk8-0-100
       - attach_workspace:
           at: /tmp/workspace
       - run:
@@ -644,7 +814,7 @@ jobs:
 
   store-and-upload-nupkgs:
     docker:
-      - image: mcr.microsoft.com/dotnet/sdk:6.0-jammy # Ubuntu 22.04
+      - image: mcr.microsoft.com/dotnet/sdk:8.0
     steps:
       - attach_workspace:
           at: /tmp/workspace
@@ -660,9 +830,9 @@ jobs:
             cp /tmp/workspace/runtimes/linux-x64/*.so runtimes/linux-x64/native/
             mkdir -p runtimes/win-x64/native
             cp /tmp/workspace/runtimes/win-x64/*.dll runtimes/win-x64/native/
-            mkdir -p runtimes/osx/native
-            cp /tmp/workspace/runtimes/osx-x64/*.dylib runtimes/osx/native/
-            cp /tmp/workspace/runtimes/osx-x64/*.metal runtimes/osx/native/
+            #mkdir -p runtimes/osx/native
+            #cp /tmp/workspace/runtimes/osx-x64/*.dylib runtimes/osx/native/
+            #cp /tmp/workspace/runtimes/osx-x64/*.metal runtimes/osx/native/
             dotnet pack ./Gpt4All/Gpt4All.csproj -p:IncludeSymbols=true -p:SymbolPackageFormat=snupkg -c Release
             dotnet nuget push ./Gpt4All/bin/Release/Gpt4All.*.nupkg -s $NUGET_URL -k $NUGET_TOKEN --skip-duplicate
       - store_artifacts:
@@ -679,9 +849,11 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - node/install-packages:
           app-dir: gpt4all-bindings/typescript
           pkg-manager: yarn
+          override-ci-command: yarn install
       - run:
           command: | 
             cd gpt4all-bindings/typescript
@@ -708,9 +880,11 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - node/install-packages:
           app-dir: gpt4all-bindings/typescript
           pkg-manager: yarn
+          override-ci-command: yarn install
       - run:
           command: | 
             cd gpt4all-bindings/typescript
@@ -719,14 +893,14 @@ jobs:
           name: "Persisting all necessary things to workspace"
           command: |  
             mkdir -p gpt4all-backend/prebuilds/darwin-x64
-            mkdir -p gpt4all-backend/runtimes/darwin-x64
-            cp /tmp/gpt4all-backend/runtimes/osx-x64/*-*.* gpt4all-backend/runtimes/darwin-x64
+            mkdir -p gpt4all-backend/runtimes/darwin
+            cp /tmp/gpt4all-backend/runtimes/osx-x64/*-*.* gpt4all-backend/runtimes/darwin
             cp gpt4all-bindings/typescript/prebuilds/darwin-x64/*.node gpt4all-backend/prebuilds/darwin-x64
       - persist_to_workspace:
           root: gpt4all-backend
           paths:
             - prebuilds/darwin-x64/*.node
-            - runtimes/darwin-x64/*-*.*
+            - runtimes/darwin/*-*.*
 
   build-nodejs-windows: 
     executor:
@@ -748,6 +922,7 @@ jobs:
               nvm install 18.16.0
               nvm use 18.16.0
       - run: node --version 
+      - run: corepack enable
       - run:           
           command: |
             npm install -g yarn
@@ -781,6 +956,7 @@ jobs:
           install-yarn: true
           node-version: "18.16"
       - run: node --version
+      - run: corepack enable
       - run: 
           command: |
             cd gpt4all-bindings/typescript
@@ -795,9 +971,12 @@ jobs:
             cp /tmp/gpt4all-backend/runtimes/linux-x64/*-*.so runtimes/linux-x64/native/ 
             cp /tmp/gpt4all-backend/prebuilds/linux-x64/*.node prebuilds/linux-x64/    
 
-            mkdir -p runtimes/darwin-x64/native
+            # darwin has univeral runtime libraries
+            mkdir -p runtimes/darwin/native
             mkdir -p prebuilds/darwin-x64/
-            cp /tmp/gpt4all-backend/runtimes/darwin-x64/*-*.* runtimes/darwin-x64/native/
+
+            cp /tmp/gpt4all-backend/runtimes/darwin/*-*.* runtimes/darwin/native/
+
             cp /tmp/gpt4all-backend/prebuilds/darwin-x64/*.node prebuilds/darwin-x64/    
             
             # Fallback build if user is not on above prebuilds
@@ -820,16 +999,39 @@ jobs:
           command: |
             cd gpt4all-bindings/typescript
             npm set //registry.npmjs.org/:_authToken=$NPM_TOKEN
-            npm publish --access public --tag alpha
+            npm publish
 
 workflows:
   version: 2
   default:
-    when: << pipeline.parameters.run-default-workflow >>
+    when:
+      or:
+       - << pipeline.parameters.run-all-workflows >>
+       - << pipeline.parameters.run-default-workflow >>
     jobs:
       - default-job
+  build-chat-offline-installers:
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-chat-workflow >>
+    jobs:
+      - hold:
+          type: approval
+      - build-offline-chat-installer-macos:
+          requires:
+            - hold
+      - build-offline-chat-installer-windows:
+          requires:
+            - hold
+      - build-offline-chat-installer-linux:
+          requires:
+            - hold
   build-and-test-gpt4all-chat:
-    when: << pipeline.parameters.run-chat-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-chat-workflow >>
     jobs:
       - hold:
           type: approval
@@ -843,7 +1045,10 @@ workflows:
           requires:
             - hold
   deploy-docs:
-    when: << pipeline.parameters.run-python-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-python-workflow >>
     jobs:
       - build-ts-docs:
           filters:
@@ -856,7 +1061,10 @@ workflows:
               only:
                 - main
   build-py-deploy:
-    when: << pipeline.parameters.run-python-workflow >>
+    when:
+      or:
+        - << pipeline.parameters.run-all-workflows >>
+        - << pipeline.parameters.run-python-workflow >>
     jobs:
       - pypi-hold:
           type: approval
@@ -891,10 +1099,11 @@ 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
@@ -990,4 +1199,4 @@ workflows:
             - nuget-hold
             - build-csharp-windows
             - build-csharp-linux
-            - build-csharp-macos
+            #- build-csharp-macos

.codespellrc

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

.github/ISSUE_TEMPLATE/bindings-bug.md

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

.github/ISSUE_TEMPLATE/bug-report.yml

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

.github/ISSUE_TEMPLATE/chat-bug.md

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

.github/ISSUE_TEMPLATE/config.yml

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

.github/ISSUE_TEMPLATE/documentation.md

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

.github/ISSUE_TEMPLATE/documentation.yml

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

.github/ISSUE_TEMPLATE/feature-request.md

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

.github/ISSUE_TEMPLATE/feature-request.yml

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

.github/ISSUE_TEMPLATE/other-bug.md

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

.github/ISSUE_TEMPLATE/other.yml

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

.gitignore

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

.gitmodules

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

README.md

@@ -1,9 +1,9 @@
 <h1 align="center">GPT4All</h1>
 
-<p align="center">Open-source assistant-style large language models that run locally on your CPU</p>
+<p align="center">Open-source large language models that run locally on your CPU and nearly any GPU</p>
 
 <p align="center">
-<a href="https://gpt4all.io">GPT4All Website</a>
+<a href="https://gpt4all.io">GPT4All Website and Models</a>
 </p>
 
 <p align="center">
@@ -30,13 +30,24 @@ Run on an M1 macOS Device (not sped up!)
 </p>
 
 ## GPT4All: An ecosystem of open-source on-edge large language models.
-GPT4All is an ecosystem to train and deploy **powerful** and **customized** large language models that run locally on consumer grade CPUs. Note that your CPU needs to support [AVX or AVX2 instructions](https://en.wikipedia.org/wiki/Advanced_Vector_Extensions).
+
+> [!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).
 
 Learn more in the [documentation](https://docs.gpt4all.io).
 
-The goal is simple - be the best instruction tuned assistant-style language model that any person or enterprise can freely use, distribute and build on.
+A GPT4All model is a 3GB - 8GB file that you can download and plug into the GPT4All open-source ecosystem software. **Nomic AI** supports and maintains this software ecosystem to enforce quality and security alongside spearheading the effort to allow any person or enterprise to easily train and deploy their own on-edge large language models.
 
-A GPT4All model is a 3GB - 8GB file that you can download and plug into the GPT4All 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. 
+### What's New ([Issue Tracker](https://github.com/orgs/nomic-ai/projects/2))
+- **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.
+    - 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.
 
 
 ### Chat Client

gpt4all-api/README.md

@@ -43,7 +43,7 @@ Run
 ```bash
 docker compose up --build
 ```
-and edit files in the `api` directory. The api will hot-reload on changes.
+and edit files in the `app` directory. The api will hot-reload on changes.
 
 You can run the unit tests with
 

gpt4all-api/docker-compose.yaml

@@ -7,13 +7,16 @@ services:
     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=ggml-mpt-7b-chat.bin
+      - 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

@@ -1,8 +1,6 @@
 # syntax=docker/dockerfile:1.0.0-experimental
 FROM tiangolo/uvicorn-gunicorn:python3.11
 
-ARG MODEL_BIN=ggml-mpt-7b-chat.bin
-
 # Put first so anytime this file changes other cached layers are invalidated.
 COPY gpt4all_api/requirements.txt /requirements.txt
 
@@ -17,7 +15,3 @@ COPY gpt4all_api/app /app
 
 RUN mkdir -p /models
 
-# Include the following line to bake a model into the image and not have to download it on API start.
-RUN wget -q --show-progress=off https://gpt4all.io/models/${MODEL_BIN} -P /models \
-  && md5sum /models/${MODEL_BIN}
-

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

@@ -1,39 +1,35 @@
 import logging
 import time
-from typing import Dict, List
-
-from api_v1.settings import settings
-from fastapi import APIRouter, Depends, Response, Security, status
+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(..., description='The model to generate a completion from.')
-    messages: List[ChatCompletionMessage] = Field(..., description='The model to generate a completion from.')
-
+    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'
@@ -42,20 +38,38 @@ class ChatCompletionResponse(BaseModel):
     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.
+    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."
 
-    return ChatCompletionResponse(
-        id='asdf',
-        created=time.time(),
-        model=request.model,
-        choices=[{}],
-        usage={'prompt_tokens': 0, 'completion_tokens': 0, 'total_tokens': 0},
+    # 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/engines.py

@@ -1,40 +1,39 @@
-import logging
-from typing import Dict, List
-
-from api_v1.settings import settings
-from fastapi import APIRouter, Depends, Response, Security, status
+import requests
+from fastapi import APIRouter, HTTPException
 from pydantic import BaseModel, Field
+from typing import List, Dict
 
-logger = logging.getLogger(__name__)
-logger.setLevel(logging.DEBUG)
-
-### This should follow https://github.com/openai/openai-openapi/blob/master/openapi.yaml
-
+# 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.")
 
 
-router = APIRouter(prefix="/engines", tags=["Search Endpoints"])
-
-
+# Define the routes for the engines module
 @router.get("/", response_model=ListEnginesResponse)
 async def list_engines():
-    '''
-    List all available GPT4All models from
-    https://raw.githubusercontent.com/nomic-ai/gpt4all/main/gpt4all-chat/metadata/models.json
-    '''
-    raise NotImplementedError()
-    return ListEnginesResponse(data=[])
-
+    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):
-    ''' '''
-
-    raise NotImplementedError()
-    return EngineResponse()
+    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/tests/test_endpoints.py

@@ -2,16 +2,26 @@
 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 = "ggml-mpt-7b-chat.bin"
+    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
@@ -19,7 +29,7 @@ def test_completion():
     assert len(response['choices'][0]['text']) > len(prompt)
 
 def test_streaming_completion():
-    model = "ggml-mpt-7b-chat.bin"
+    model = model_id
     prompt = "Who is Michael Jordan?"
     tokens = []
     for resp in openai.Completion.create(
@@ -36,19 +46,27 @@ def test_streaming_completion():
     assert (len(tokens) > 0)
     assert (len("".join(tokens)) > len(prompt))
 
-
+# Modified test batch, problems with keyerror in response
 def test_batched_completion():
-    model = "ggml-mpt-7b-chat.bin"
+    model = model_id  # replace with your specific model ID
     prompt = "Who is Michael Jordan?"
-    response = openai.Completion.create(
-        model=model, prompt=[prompt] * 3, max_tokens=50, temperature=0.28, top_p=0.95, n=1, echo=True, stream=False
-    )
-    assert len(response['choices'][0]['text']) > len(prompt)
-    assert len(response['choices']) == 3
+    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 = "ggml-all-MiniLM-L6-v2-f16.bin"
+    model = embedding
     prompt = "Who is Michael Jordan?"
     response = openai.Embedding.create(model=model, input=prompt)
     output = response["data"][0]["embedding"]
@@ -56,4 +74,4 @@ def test_embedding():
 
     assert response["model"] == model
     assert isinstance(output, list)
-    assert all(isinstance(x, args) for x in output)
+    assert all(isinstance(x, args) for x in output)

gpt4all-api/gpt4all_api/env

@@ -0,0 +1,3 @@
+# 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

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

gpt4all-api/gpt4all_api/requirements.txt

@@ -7,6 +7,7 @@ fastapi>=0.95.0
 Jinja2>=3.0
 gpt4all>=1.0.0
 pytest
-openai
+openai==0.28.0
 black
-isort
+isort
+python-dotenv

gpt4all-api/makefile

@@ -14,7 +14,7 @@ testenv_gpu: clean_testenv test_build
 	docker compose -f docker-compose.yaml -f docker-compose.gpu.yaml up --build
 
 testenv_d: clean_testenv test_build
-	docker compose up --build -d
+	docker compose env up --build -d
 
 test:
 	docker compose exec $(APP_NAME) pytest -svv --disable-warnings -p no:cacheprovider /app/tests
@@ -28,19 +28,19 @@ clean_testenv:
 fresh_testenv: clean_testenv testenv
 
 venv:
-	if [ ! -d $(ROOT_DIR)/env ]; then $(PYTHON) -m venv $(ROOT_DIR)/env; fi
+	if [ ! -d $(ROOT_DIR)/venv ]; then $(PYTHON) -m venv $(ROOT_DIR)/venv; fi
 
 dependencies: venv
-	source $(ROOT_DIR)/env/bin/activate; $(PYTHON) -m pip install -r $(ROOT_DIR)/$(APP_NAME)/requirements.txt
+	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)/env;
+	rm -rf $(ROOT_DIR)/venv;
 	rm -rf $(ROOT_DIR)/$(APP_NAME)/*.pyc;
 
 
 black:
-	source $(ROOT_DIR)/env/bin/activate; black -l 120 -S --target-version py38 $(APP_NAME)
+	source $(ROOT_DIR)/venv/bin/activate; black -l 120 -S --target-version py38 $(APP_NAME)
 
 isort:
-	source $(ROOT_DIR)/env/bin/activate; isort  --ignore-whitespace --atomic -w 120 $(APP_NAME)
+	source $(ROOT_DIR)/venv/bin/activate; isort  --ignore-whitespace --atomic -w 120 $(APP_NAME)

gpt4all-backend/CMakeLists.txt

@@ -20,7 +20,7 @@ endif()
 include_directories("${CMAKE_CURRENT_BINARY_DIR}")
 
 set(LLMODEL_VERSION_MAJOR 0)
-set(LLMODEL_VERSION_MINOR 4)
+set(LLMODEL_VERSION_MINOR 5)
 set(LLMODEL_VERSION_PATCH 0)
 set(LLMODEL_VERSION "${LLMODEL_VERSION_MAJOR}.${LLMODEL_VERSION_MINOR}.${LLMODEL_VERSION_PATCH}")
 project(llmodel VERSION ${LLMODEL_VERSION} LANGUAGES CXX C)
@@ -39,10 +39,6 @@ else()
     message(STATUS "Interprocedural optimization support detected")
 endif()
 
-if(NOT APPLE)
-  set(LLAMA_KOMPUTE YES)
-endif()
-
 include(llama.cpp.cmake)
 
 set(BUILD_VARIANTS default avxonly)
@@ -97,35 +93,15 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
         LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
     prepare_target(llamamodel-mainline llama-mainline)
 
-    add_library(replit-mainline-${BUILD_VARIANT} SHARED
-    replit.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-    target_compile_definitions(replit-mainline-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-    prepare_target(replit-mainline llama-mainline)
-
     if (NOT LLAMA_METAL)
-# FIXME: These need to be forward ported to latest ggml
-#        add_library(gptj-${BUILD_VARIANT} SHARED
-#            gptj.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-#        prepare_target(gptj ggml-230511)
-
-        add_library(falcon-${BUILD_VARIANT} SHARED
-            falcon.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-        target_compile_definitions(falcon-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-        prepare_target(falcon llama-mainline)
-# FIXME: These need to be forward ported to latest ggml
-#        add_library(mpt-${BUILD_VARIANT} SHARED
-#            mpt.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-#        prepare_target(mpt ggml-230511)
+        add_library(gptj-${BUILD_VARIANT} SHARED
+            gptj.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
+        prepare_target(gptj llama-mainline)
 
         add_library(bert-${BUILD_VARIANT} SHARED
             bert.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
         target_compile_definitions(bert-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
         prepare_target(bert llama-mainline)
-
-        add_library(starcoder-${BUILD_VARIANT} SHARED
-            starcoder.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
-        target_compile_definitions(starcoder-${BUILD_VARIANT} PRIVATE LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-        prepare_target(starcoder llama-mainline)
     endif()
 endforeach()
 

gpt4all-backend/bert.cpp

@@ -4,10 +4,10 @@
 #include "ggml.h"
 
 #include <cassert>
+#include <cinttypes>
 #include <cmath>
 #include <cstdio>
 #include <cstring>
-#include <fstream>
 #include <map>
 #include <string>
 #include <vector>
@@ -34,7 +34,6 @@ struct bert_hparams
     int32_t n_intermediate = 1536;
     int32_t n_head = 12;
     int32_t n_layer = 6;
-    int32_t f16 = 1;
 };
 
 struct bert_layer
@@ -88,7 +87,6 @@ struct bert_model
     std::vector<bert_layer> layers;
 
     struct ggml_context *ctx;
-    std::map<std::string, struct ggml_tensor *> tensors;
 };
 
 // Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
@@ -319,7 +317,7 @@ void bert_eval(
     };
 
     struct ggml_context *ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = {};
+    struct ggml_cgraph *gf = ggml_new_graph(ctx0);
 
     // Embeddings. word_embeddings + token_type_embeddings + position_embeddings
     struct ggml_tensor *token_layer = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
@@ -345,7 +343,7 @@ void bert_eval(
 
     // embd norm
     {
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, 1e-5f);
 
         inpL = ggml_add(ctx0,
                         ggml_mul(ctx0,
@@ -383,10 +381,9 @@ void bert_eval(
 
             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))));
+            KQ = ggml_soft_max(
+                ctx0, ggml_scale(ctx0, KQ, 1.0f / sqrt((float)d_head))
+            );
 
             V = ggml_cont(ctx0, ggml_transpose(ctx0, V));
             struct ggml_tensor *KQV = ggml_mul_mat(ctx0, V, KQ);
@@ -406,7 +403,7 @@ void bert_eval(
 
         // attention norm
         {
-            cur = ggml_norm(ctx0, cur);
+            cur = ggml_norm(ctx0, cur, 1e-5f);
 
             cur = ggml_add(ctx0,
                            ggml_mul(ctx0,
@@ -432,7 +429,7 @@ void bert_eval(
 
         // output norm
         {
-            cur = ggml_norm(ctx0, cur);
+            cur = ggml_norm(ctx0, cur, 1e-5f);
 
             cur = ggml_add(ctx0,
                            ggml_mul(ctx0,
@@ -450,10 +447,10 @@ void bert_eval(
 
     ggml_tensor *output = inpL;
     // run the computation
-    ggml_build_forward_expand(&gf, output);
+    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);
+    ggml_graph_compute_g4a(ctx->work_buf, gf, n_threads);
+    //ggml_graph_compute(ctx0, gf);
 
 
     // float *dat = ggml_get_data_f32(output);
@@ -462,7 +459,7 @@ void bert_eval(
     #ifdef GGML_PERF
         // print timing information per ggml operation (for debugging purposes)
         // requires GGML_PERF to be defined
-        ggml_graph_print(&gf);
+        ggml_graph_print(gf);
     #endif
 
     if (!mem_req_mode) {
@@ -482,7 +479,6 @@ void bert_eval(
 //
 
 void bert_free(bert_ctx * ctx) {
-    ggml_free(ctx->model.ctx);
     delete ctx;
 }
 
@@ -492,63 +488,131 @@ struct bert_ctx * bert_load_from_file(const char *fname)
     printf("%s: loading model from '%s' - please wait ...\n", __func__, fname);
 #endif
 
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin)
-    {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname);
+    bert_ctx * new_bert = new bert_ctx;
+
+    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;
     }
 
-    // verify magic
+    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
     {
-        uint32_t magic;
-        fin.read((char *)&magic, sizeof(magic));
-        if (magic != 0x62657274)
-        {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname);
+        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;
         }
     }
 
-    bert_ctx * new_bert = new bert_ctx;
-    bert_model & model = new_bert->model;
-    bert_vocab & vocab = new_bert->vocab;
-
     // load hparams
     {
         auto &hparams = model.hparams;
 
-        fin.read((char *)&hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *)&hparams.n_max_tokens, sizeof(hparams.n_max_tokens));
-        fin.read((char *)&hparams.n_embd, sizeof(hparams.n_embd));
-        fin.read((char *)&hparams.n_intermediate, sizeof(hparams.n_intermediate));
-        fin.read((char *)&hparams.n_head, sizeof(hparams.n_head));
-        fin.read((char *)&hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *)&hparams.f16, sizeof(hparams.f16));
+        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_vocab = %d\n", __func__, hparams.n_vocab);
         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);
-        printf("%s: f16     = %d\n", __func__, hparams.f16);
+        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
     {
-        int32_t n_vocab = model.hparams.n_vocab;
+        auto & hparams = model.hparams;
 
-        std::string word;
-        for (int i = 0; i < n_vocab; i++)
-        {
-            uint32_t len;
-            fin.read((char *)&len, sizeof(len));
+        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;
+        }
 
-            word.resize(len);
-            fin.read((char *)word.data(), len);
+        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] == '#')
             {
@@ -564,290 +628,52 @@ struct bert_ctx * bert_load_from_file(const char *fname)
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
-    // in order to save memory and also to speed up the computation
-    ggml_type wtype = GGML_TYPE_COUNT;
-    switch (model.hparams.f16)
-    {
-    case 0:
-        wtype = GGML_TYPE_F32;
-        break;
-    case 1:
-        wtype = GGML_TYPE_F16;
-        break;
-    case 2:
-        wtype = GGML_TYPE_Q4_0;
-        break;
-    case 3:
-        wtype = GGML_TYPE_Q4_1;
-        break;
-    default:
-    {
-        fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
-                __func__, fname, model.hparams.f16);
-        bert_free(new_bert);
-        return nullptr;
-    }
-    }
-
     auto &ctx = model.ctx;
 
-    size_t model_mem_req = 0;
-
-    {
-        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_intermediate = hparams.n_intermediate;
-        const int n_vocab = hparams.n_vocab;
-
-        // Calculate size requirements
-
-        model_mem_req += n_embd * n_vocab * ggml_type_sizef(wtype); // word_embeddings
-        model_mem_req += n_embd * 2 * ggml_type_sizef(wtype); // token_type_embeddings
-        model_mem_req += n_embd * n_max_tokens * ggml_type_sizef(wtype); // position_embeddings
-
-        model_mem_req += 2 * n_embd * ggml_type_sizef(GGML_TYPE_F32); // ln_e_*
-
-        model_mem_req += 4 * n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32)); // ln_*
-
-        model_mem_req += 4 * n_layer * (n_embd * n_embd * ggml_type_sizef(wtype)); // kqvo weights
-        model_mem_req += 4 * n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32)); // kqvo bias
-
-        model_mem_req += 2 * n_layer * (n_embd * n_intermediate * ggml_type_sizef(wtype)); // ff_*_w
-        model_mem_req += n_layer * (n_intermediate * ggml_type_sizef(GGML_TYPE_F32)); // ff_i_b
-        model_mem_req += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32)); // ff_o_b
-
-        model_mem_req += (5 + 16 * n_layer) * ggml_tensor_overhead(); // object overhead
-
 #if defined(DEBUG_BERT)
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, model_mem_req / (1024.0 * 1024.0));
+    printf("%s: ggml ctx size = %6.2f MB\n", __func__, ggml_get_mem_size(ctx) / (1024.0 * 1024.0));
 #endif
-    }
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            .mem_size = model_mem_req,
-            .mem_buffer = NULL,
-            .no_alloc = false,
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx)
-        {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            bert_free(new_bert);
-            return nullptr;
-        }
-    }
 
     // prepare memory for the weights
     {
-        const auto &hparams = model.hparams;
-
-        const int n_embd = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_intermediate = hparams.n_intermediate;
-        const int n_max_tokens = hparams.n_max_tokens;
-        const int n_vocab = hparams.n_vocab;
-
+        const int n_layer = model.hparams.n_layer;
         model.layers.resize(n_layer);
 
-        model.word_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
-        model.token_type_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, 2);
-        model.position_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_max_tokens);
+        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");
 
-        model.ln_e_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-        model.ln_e_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-        // map by name
-        model.tensors["embeddings.word_embeddings.weight"] = model.word_embeddings;
-        model.tensors["embeddings.token_type_embeddings.weight"] = model.token_type_embeddings;
-        model.tensors["embeddings.position_embeddings.weight"] = model.position_embeddings;
-
-        model.tensors["embeddings.LayerNorm.weight"] = model.ln_e_w;
-        model.tensors["embeddings.LayerNorm.bias"] = model.ln_e_b;
+        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_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.ln_att_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.ln_out_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.ln_out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-            layer.q_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
-            layer.q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.k_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
-            layer.k_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.v_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
-            layer.v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-            layer.o_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
-            layer.o_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-            layer.ff_i_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_intermediate);
-            layer.ff_i_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_intermediate);
-
-            layer.ff_o_w = ggml_new_tensor_2d(ctx, wtype, n_intermediate, n_embd);
-            layer.ff_o_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-            // map by name
-
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.self.query.weight"] = layer.q_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.self.query.bias"] = layer.q_b;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.self.key.weight"] = layer.k_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.self.key.bias"] = layer.k_b;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.self.value.weight"] = layer.v_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.self.value.bias"] = layer.v_b;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.output.LayerNorm.weight"] = layer.ln_att_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.output.LayerNorm.bias"] = layer.ln_att_b;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.output.dense.weight"] = layer.o_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".attention.output.dense.bias"] = layer.o_b;
-
-            model.tensors["encoder.layer." + std::to_string(i) + ".intermediate.dense.weight"] = layer.ff_i_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".intermediate.dense.bias"] = layer.ff_i_b;
-
-            model.tensors["encoder.layer." + std::to_string(i) + ".output.LayerNorm.weight"] = layer.ln_out_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".output.LayerNorm.bias"] = layer.ln_out_b;
-            model.tensors["encoder.layer." + std::to_string(i) + ".output.dense.weight"] = layer.ff_o_w;
-            model.tensors["encoder.layer." + std::to_string(i) + ".output.dense.bias"] = layer.ff_o_b;
+            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"));
         }
     }
 
-    // load weights
-    {
-        int n_tensors = 0;
-#if defined(DEBUG_BERT)
-        size_t total_size = 0;
-#endif
-
-#if defined(DEBUG_BERT)
-        printf("%s: ", __func__);
-#endif
-
-        while (true)
-        {
-            int32_t n_dims;
-            int32_t length;
-            int32_t ftype;
-
-            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ftype), sizeof(ftype));
-
-            if (fin.eof())
-            {
-                break;
-            }
-
-            int64_t nelements = 1;
-            int64_t ne[2] = {1, 1};
-            for (int i = 0; i < n_dims; ++i)
-            {
-                int32_t ne_cur;
-                fin.read(reinterpret_cast<char *>(&ne_cur), sizeof(ne_cur));
-                ne[i] = ne_cur;
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-
-            if (model.tensors.find(name.data()) == model.tensors.end())
-            {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                bert_free(new_bert);
-                return nullptr;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements)
-            {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                bert_free(new_bert);
-                return nullptr;
-            }
-
-            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1])
-            {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%ld, %ld], expected [%ld, %ld]\n",
-                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
-                bert_free(new_bert);
-                return nullptr;
-            }
-
-#if defined(DEBUG_BERT)
-            static const char *ftype_str[] = {
-                "f32",
-                "f16",
-                "q4_0",
-                "q4_1",
-            };
-            printf("%24s - [%5ld, %5ld], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor) / 1024.0 / 1024.0, ggml_nbytes(tensor));
-#endif
-
-            size_t bpe = 0;
-
-            switch (ftype)
-            {
-            case 0:
-                bpe = ggml_type_size(GGML_TYPE_F32);
-                break;
-            case 1:
-                bpe = ggml_type_size(GGML_TYPE_F16);
-                break;
-            case 2:
-                bpe = ggml_type_size(GGML_TYPE_Q4_0);
-                assert(ne[0] % 64 == 0);
-                break;
-            case 3:
-                bpe = ggml_type_size(GGML_TYPE_Q4_1);
-                assert(ne[0] % 64 == 0);
-                break;
-            default:
-            {
-                fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
-                bert_free(new_bert);
-                return nullptr;
-            }
-            };
-
-            if ((nelements * bpe) / ggml_blck_size(tensor->type) != ggml_nbytes(tensor))
-            {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %lu\n",
-                        __func__, name.data(), ggml_nbytes(tensor), nelements * bpe);
-                bert_free(new_bert);
-                return nullptr;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-#if defined(DEBUG_BERT)
-            // printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
-            total_size += ggml_nbytes(tensor);
-#endif
-
-            if (++n_tensors % 8 == 0)
-            {
-#if defined(DEBUG_BERT)
-                printf(".");
-                fflush(stdout);
-#endif
-            }
-        }
-
-#if defined(DEBUG_BERT)
-        printf(" done\n");
-        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size / 1024.0 / 1024.0, n_tensors);
-#endif
-    }
-
-    fin.close();
-
     // Calculate space requirements for setting up context buffers later
     {
         bert_vocab_id tokens[] = {0, 1, 2, 3};
@@ -883,12 +709,20 @@ Bert::~Bert() {
     bert_free(d_ptr->ctx);
 }
 
-bool Bert::loadModel(const std::string &modelPath)
+bool Bert::loadModel(const std::string &modelPath, int n_ctx, int ngl)
 {
-    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;
+    (void)n_ctx;
+    (void)ngl;
+    d_ptr->modelLoaded = false;
+
+    auto * ctx = bert_load_from_file(modelPath.c_str());
     fflush(stdout);
+    if (!ctx)
+        return false;
+
+    d_ptr->ctx = ctx;
+    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    d_ptr->modelLoaded = true;
     return true;
 }
 
@@ -897,8 +731,11 @@ bool Bert::isModelLoaded() const
     return d_ptr->modelLoaded;
 }
 
-size_t Bert::requiredMem(const std::string &/*modelPath*/)
+size_t Bert::requiredMem(const std::string &modelPath, int n_ctx, int ngl)
 {
+    (void)modelPath;
+    (void)n_ctx;
+    (void)ngl;
     return 0;
 }
 
@@ -977,8 +814,10 @@ std::vector<float> Bert::embedding(const std::string &text)
     return finalEmbeddings;
 }
 
-std::vector<LLModel::Token> Bert::tokenize(PromptContext &, const std::string &str) const
+std::vector<LLModel::Token> Bert::tokenize(PromptContext &ctx, const std::string &str, bool special) const
 {
+    (void)ctx;
+    (void)special;
     return ::bert_tokenize(d_ptr->ctx, str.c_str());
 }
 
@@ -1019,6 +858,16 @@ const std::vector<LLModel::Token> &Bert::endTokens() const
     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
@@ -1038,13 +887,21 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(std::istream& f) {
-    uint32_t magic = 0;
-    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
-    if (magic != 0x62657274) {
-         return false;
-    }
-    return true;
+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() {

gpt4all-backend/bert_impl.h

@@ -18,9 +18,9 @@ public:
 
     bool supportsEmbedding() const override { return true; }
     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;
@@ -33,12 +33,13 @@ private:
     std::unique_ptr<BertPrivate> 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 true; }
 };
 
 #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/falcon.cpp

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

gpt4all-backend/falcon_impl.h

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

gpt4all-backend/gptj.cpp

@@ -9,7 +9,6 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
-#include <fstream>
 #include <map>
 #include <string>
 #include <vector>
@@ -42,7 +41,7 @@ struct gptj_hparams {
     int32_t n_head  = 16;
     int32_t n_layer = 28;
     int32_t n_rot   = 64;
-    int32_t f16     = 1;
+    float norm_eps  = 1e-5;
 };
 
 struct gptj_layer {
@@ -128,216 +127,149 @@ static bool kv_cache_init(
     return true;
 }
 
-// load the model's weights from a stream
-bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & model, gpt_vocab & vocab, size_t * mem_req = nullptr) {
+// load the model's weights from a file path
+bool gptj_model_load(const std::string &fname, gptj_model & model, gpt_vocab & vocab, size_t * mem_req = nullptr) {
     printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
     if(mem_req != nullptr) {
         *mem_req = 0;
     }
 
-    // verify magic
-    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        if (magic != 0x67676d6c) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
-        }
+    // create the ggml context
+    struct gguf_init_params params = {
+        /*.no_alloc = */ false,
+        /*.ctx      = */ &model.ctx,
+    };
+
+    gguf_context *ggufctx = gguf_init_from_file(fname.c_str(), params);
+    if (!ggufctx) {
+        fprintf(stderr, "%s: gguf_init_from_file() failed\n", __func__);
+        return false;
     }
 
     // load hparams
     {
         auto & hparams = model.hparams;
 
-        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
-        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
-        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
-        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
-        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+        bool ok = false;
+        int keyidx;
+
+        do {
+            keyidx = gguf_find_key(ggufctx, "gptj.context_length");
+            if (keyidx == -1) { break; }
+            hparams.n_ctx = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.embedding_length");
+            if (keyidx == -1) { break; }
+            hparams.n_embd = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.attention.head_count");
+            if (keyidx == -1) { break; }
+            hparams.n_head = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.block_count");
+            if (keyidx == -1) { break; }
+            hparams.n_layer = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.rope.dimension_count");
+            if (keyidx == -1) { break; }
+            hparams.n_rot = gguf_get_val_u32(ggufctx, keyidx);
+
+            keyidx = gguf_find_key(ggufctx, "gptj.attention.layer_norm_epsilon");
+            if (keyidx == -1) { break; }
+            hparams.norm_eps = gguf_get_val_f32(ggufctx, keyidx);
+
+            ok = true;
+        } while (false);
+
+        if (!ok) {
+            fprintf(stderr, "%s: required hparam missing!\n", __func__);
+            return false;
+        }
 
-        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
         printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
         printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
         printf("%s: n_head  = %d\n", __func__, hparams.n_head);
         printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
         printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
-        printf("%s: f16     = %d\n", __func__, hparams.f16);
     }
 
     // load vocab
     {
-        int32_t n_vocab = 0;
-        fin.read((char *) &n_vocab, sizeof(n_vocab));
+        auto & hparams = model.hparams;
 
-        if (n_vocab != model.hparams.n_vocab) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
-                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+        int keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.model");
+        if (keyidx == -1) {
+            fprintf(stderr, "%s: tokenizer model not found!\n", __func__);
+            return false;
+        }
+        if (strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
+            fprintf(stderr, "%s: tokenizer model not supported!\n", __func__);
             return false;
         }
 
-        std::string word;
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
+        int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
+        if (tokens_keyidx == -1) {
+            fprintf(stderr, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            return false;
+        }
 
-            word.resize(len);
-            fin.read((char *) word.data(), len);
+        hparams.n_vocab = gguf_get_arr_n(ggufctx, tokens_keyidx);
+        printf("%s: gpt2 tokenizer vocab = %d\n", __func__, int(hparams.n_vocab));
 
+        for (int i = 0; i < hparams.n_vocab; i++) {
+            std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
             vocab.token_to_id[word] = i;
             vocab.id_to_token[i] = word;
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
-    // in order to save memory and also to speed up the computation
-    ggml_type wtype = GGML_TYPE_COUNT;
-    switch (model.hparams.f16) {
-        case 0: wtype = GGML_TYPE_F32;  break;
-        case 1: wtype = GGML_TYPE_F16;  break;
-        case 2: wtype = GGML_TYPE_Q4_0; break;
-        case 3: wtype = GGML_TYPE_Q4_1; break;
-        case 5: wtype = GGML_TYPE_Q4_2; break;
-        default:
-                {
-                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
-                            __func__, fname.c_str(), model.hparams.f16);
-                    return false;
-                }
-    }
-
     auto & ctx = model.ctx;
 
-    size_t ctx_size = 0;
-
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);         // lmh_g
-        ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
-
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
-
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
-
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
-
-        ctx_size += (5 + 10*n_layer)*256; // object overhead
-
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
+    size_t ctx_size = ggml_get_mem_size(ctx);
+    printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size / (1024.0 * 1024.0));
 
     if (mem_req != nullptr) {
-        *mem_req += ctx_size;
-        const int n_embd  = model.hparams.n_embd;
-        const int n_layer = model.hparams.n_layer;
-
-        const int64_t n_mem      = (int64_t)n_layer*model.hparams.n_ctx;
-        const int64_t n_elements = n_embd*n_mem;
-
-        *mem_req += (2u*n_elements*ggml_type_size(wtype) + 2_MiB);
+        *mem_req = ctx_size;
+        gguf_free(ggufctx);
         return false;
     }
 
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-            .no_alloc = false
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
     // prepare memory for the weights
     {
         const auto & hparams = model.hparams;
+        model.layers.resize(hparams.n_layer);
 
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_vocab = hparams.n_vocab;
+        model.wte    = ggml_get_tensor(ctx, "token_embd.weight");
 
-        model.layers.resize(n_layer);
+        model.ln_f_g = ggml_get_tensor(ctx, "output_norm.weight");
+        model.ln_f_b = ggml_get_tensor(ctx, "output_norm.bias");
 
-        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.lmh_g  = ggml_get_tensor(ctx, "output.weight");
+        model.lmh_b  = ggml_get_tensor(ctx, "output.bias");
 
-        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        auto name = [](int i, std::string n) {
+            static std::string key;
+            key = "blk." + std::to_string(i) + "." + n;
+            return key.c_str();
+        };
 
-        model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
-        model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
-
-        // map by name
-        model.tensors["transformer.wte.weight"] = model.wte;
-
-        model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
-        model.tensors["transformer.ln_f.bias"]   = model.ln_f_b;
-
-        model.tensors["lm_head.weight"] = model.lmh_g;
-        model.tensors["lm_head.bias"]   = model.lmh_b;
-
-        for (int i = 0; i < n_layer; ++i) {
+        for (int i = 0; i < hparams.n_layer; ++i) {
             auto & layer = model.layers[i];
 
-            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_g          = ggml_get_tensor(ctx, name(i, "attn_norm.weight"));
+            layer.ln_1_b          = ggml_get_tensor(ctx, name(i, "attn_norm.bias"));
 
-            layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
-            layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
-            layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_q_proj_w = ggml_get_tensor(ctx, name(i, "attn_q.weight"));
+            layer.c_attn_k_proj_w = ggml_get_tensor(ctx, name(i, "attn_k.weight"));
+            layer.c_attn_v_proj_w = ggml_get_tensor(ctx, name(i, "attn_v.weight"));
 
-            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_proj_w   = ggml_get_tensor(ctx, name(i, "attn_output.weight"));
 
-            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
-            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+            layer.c_mlp_fc_w      = ggml_get_tensor(ctx, name(i, "ffn_up.weight"));
+            layer.c_mlp_fc_b      = ggml_get_tensor(ctx, name(i, "ffn_up.bias"));
 
-            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
-            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-
-            // map by name
-            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"]          = layer.ln_1_g;
-            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"]            = layer.ln_1_b;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"]   = layer.c_attn_q_proj_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"]   = layer.c_attn_k_proj_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"]   = layer.c_attn_v_proj_w;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"]     = layer.c_mlp_fc_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"]       = layer.c_mlp_fc_b;
-
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"]    = layer.c_mlp_proj_w;
-            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"]      = layer.c_mlp_proj_b;
+            layer.c_mlp_proj_w    = ggml_get_tensor(ctx, name(i, "ffn_down.weight"));
+            layer.c_mlp_proj_b    = ggml_get_tensor(ctx, name(i, "ffn_down.bias"));
         }
     }
 
@@ -354,113 +286,12 @@ bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & m
         printf("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
     }
 
-    // load weights
-    {
-        int n_tensors = 0;
-        size_t total_size = 0;
-
-        printf("%s: ", __func__);
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            int32_t ftype;
-
-            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                return false;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                return false;
-            }
-
-            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
-                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            if (0) {
-                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
-                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
-            }
-
-            size_t bpe = 0;
-
-            switch (ftype) {
-                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
-                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
-                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
-                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
-                default:
-                        {
-                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
-                            return false;
-                        }
-            };
-
-            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
-                return false;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
-            total_size += ggml_nbytes(tensor);
-            if (++n_tensors % 8 == 0) {
-                printf(".");
-                fflush(stdout);
-            }
-        }
-
-        printf(" done\n");
-
-        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
-    }
-
     model.scr0_buf.resize(256u * 1024 * 1024);
     model.scr1_buf.resize(256u * 1024 * 1024);
 
     return true;
 }
 
-// load the model's weights from a file path
-bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
-
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
-    bool loaded = gptj_model_load(fname, fin, model, vocab);
-    fin.close();
-    return loaded;
-}
-
 // evaluate the transformer
 //
 //   - model:     the model
@@ -512,8 +343,14 @@ bool gptj_eval(
     };
 
     struct ggml_context * ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    int * data = (int *) KQ_pos->data;
+    for (int i = 0; i < N; ++i) {
+        data[i] = n_past + i;
+    }
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -526,7 +363,7 @@ bool gptj_eval(
         ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });
         // norm
         {
-            cur = ggml_norm(ctx0, inpL);
+            cur = ggml_norm(ctx0, inpL, model.hparams.norm_eps);
 
             // cur = ln_1_g*cur + ln_1_b
             cur = ggml_add(ctx0,
@@ -540,48 +377,44 @@ bool gptj_eval(
 
         // self-attention
         {
-            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
-            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
-            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur);
+            struct ggml_tensor * Qcur = ggml_rope(
+                ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N),
+                KQ_pos, n_rot, 0, 0
+            );
+            struct ggml_tensor * Kcur = ggml_rope(
+                ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N),
+                KQ_pos, n_rot, 0, 0
+            );
 
             // store key and value to memory
             {
-                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, model.kv_self.v, N*n_embd, (ggml_element_size(model.kv_self.v)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur));
 
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.kv_self.v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.kv_self.v),
+                        (il*n_ctx)*ggml_element_size(model.kv_self.v)*n_embd + n_past*ggml_element_size(model.kv_self.v));
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
             }
 
             // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_rope(ctx0,
-                            ggml_cpy(ctx0,
-                                Qcur,
-                                ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
-                            n_past, n_rot, 0),
-                        0, 2, 1, 3);
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
 
             // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
-                        ggml_rope(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
-                                n_embd/n_head, n_head, n_past + N),
-                            n_past, n_rot, 1),
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrt(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
@@ -590,17 +423,15 @@ bool gptj_eval(
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
 
             // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
-            struct ggml_tensor * V_trans =
-                ggml_cpy(ctx0,
-                        ggml_permute(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, model.kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd),
-                                n_embd/n_head, n_head, n_past + N),
-                            1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, model.kv_self.v->type, n_past + N, n_embd/n_head, n_head));
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.kv_self.v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.kv_self.v),
+                        n_ctx*ggml_element_size(model.kv_self.v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd);
 
             // KQV = transpose(V) * KQ_soft_max
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
@@ -656,7 +487,7 @@ bool gptj_eval(
 
     // norm
     {
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, model.hparams.norm_eps);
 
         // inpL = ln_f_g*inpL + ln_f_b
         inpL = ggml_add(ctx0,
@@ -680,13 +511,22 @@ bool gptj_eval(
     // logits -> probs
     //inpL = ggml_soft_max(ctx0, inpL);
 
+    ggml_build_forward_expand(gf, inpL);
+
     // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    {
+        std::unique_ptr<uint8_t []> data;
+        auto plan = ggml_graph_plan(gf, n_threads);
+        if (plan.work_size > 0) {
+            data.reset(new uint8_t[plan.work_size]);
+            plan.work_data = data.get();
+        }
+        ggml_graph_compute(gf, &plan);
+    }
 
     //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //    ggml_graph_print   (gf);
+    //    ggml_graph_dump_dot(gf, NULL, "gpt-2.dot");
     //}
 
     //embd_w.resize(n_vocab*N);
@@ -832,30 +672,34 @@ GPTJ::GPTJ()
     d_ptr->modelLoaded = false;
 }
 
-size_t GPTJ::requiredMem(const std::string &modelPath) {
+size_t GPTJ::requiredMem(const std::string &modelPath, int n_ctx, int ngl) {
+    (void)n_ctx;
+    (void)ngl;
     gptj_model dummy_model;
     gpt_vocab dummy_vocab;
     size_t mem_req;
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-    gptj_model_load(modelPath, fin, dummy_model, dummy_vocab, &mem_req);
+    gptj_model_load(modelPath, dummy_model, dummy_vocab, &mem_req);
     return mem_req;
 }
 
-bool GPTJ::loadModel(const std::string &modelPath) {
+bool GPTJ::loadModel(const std::string &modelPath, int n_ctx, int ngl) {
+    (void)n_ctx;
+    (void)ngl;
+    d_ptr->modelLoaded = false;
+
     std::mt19937 rng(time(NULL));
     d_ptr->rng = rng;
 
-    auto fin = std::ifstream(modelPath, std::ios::binary);
-
     // load the model
-    if (!gptj_model_load(modelPath, fin, *d_ptr->model, d_ptr->vocab)) {
+    bool ok = gptj_model_load(modelPath, *d_ptr->model, d_ptr->vocab);
+    fflush(stdout);
+    if (!ok) {
         std::cerr << "GPT-J ERROR: failed to load model from " <<  modelPath;
         return false;
     }
 
     d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
     d_ptr->modelLoaded = true;
-    fflush(stdout);
     return true;
 }
 
@@ -893,8 +737,10 @@ size_t GPTJ::restoreState(const uint8_t *src)
     return gptj_set_state_data(d_ptr->model, &d_ptr->rng, src);
 }
 
-std::vector<LLModel::Token> GPTJ::tokenize(PromptContext &, const std::string &str) const
+std::vector<LLModel::Token> GPTJ::tokenize(PromptContext &ctx, const std::string &str, bool special) const
 {
+    (void)ctx;
+    (void)special;
     return ::gpt_tokenize(d_ptr->vocab, str);
 }
 
@@ -939,6 +785,16 @@ const std::vector<LLModel::Token> &GPTJ::endTokens() const
     return fres;
 }
 
+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
@@ -958,15 +814,21 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(std::istream& f) {
-    uint32_t magic = 0;
-    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
-    gptj_hparams hparams;
-    f.read(reinterpret_cast<char*>(&hparams), sizeof(hparams));
-    if (!(hparams.n_vocab >= 50300 && hparams.n_vocab <= 50400)) {
-        return false; // not a gptj.
-    }
-    return magic == 0x67676d6c;
+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) == "gptj";
+
+    gguf_free(ctx_gguf);
+    return isValid;
 }
 
 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/llama.cpp.cmake

@@ -38,6 +38,12 @@ else()
     endif()
 endif()
 
+if (APPLE)
+    set(LLAMA_KOMPUTE_DEFAULT OFF)
+else()
+    set(LLAMA_KOMPUTE_DEFAULT ON)
+endif()
+
 
 #
 # Option list
@@ -77,7 +83,7 @@ option(LLAMA_OPENBLAS               "llama: use OpenBLAS"
 #option(LLAMA_CUBLAS                 "llama: use cuBLAS"                                     OFF)
 #option(LLAMA_CLBLAST                "llama: use CLBlast"                                    OFF)
 #option(LLAMA_METAL                  "llama: use Metal"                                      OFF)
-#option(LLAMA_K_QUANTS               "llama: use k-quants"                                   ON)
+option(LLAMA_KOMPUTE                "llama: use Kompute"                                    ${LLAMA_KOMPUTE_DEFAULT})
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_DMMV_Y "1" CACHE STRING  "llama: y block size for dmmv CUDA kernels")
@@ -154,6 +160,12 @@ if (LLAMA_OPENBLAS)
 endif()
 
 if (LLAMA_KOMPUTE)
+    set(LLAMA_DIR ${CMAKE_CURRENT_SOURCE_DIR}/llama.cpp-mainline)
+    if (NOT EXISTS "${LLAMA_DIR}/kompute/CMakeLists.txt")
+        message(FATAL_ERROR "Kompute not found")
+    endif()
+    message(STATUS "Kompute found")
+
     add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
     find_package(Vulkan COMPONENTS glslc REQUIRED)
     find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
@@ -161,8 +173,6 @@ if (LLAMA_KOMPUTE)
         message(FATAL_ERROR "glslc not found")
     endif()
 
-    set(LLAMA_DIR ${CMAKE_CURRENT_SOURCE_DIR}/llama.cpp-mainline)
-
     function(compile_shader)
       set(options)
       set(oneValueArgs)
@@ -174,6 +184,10 @@ if (LLAMA_KOMPUTE)
         add_custom_command(
             OUTPUT ${spv_file}
             DEPENDS ${LLAMA_DIR}/${source}
+              ${LLAMA_DIR}/kompute-shaders/common.comp
+              ${LLAMA_DIR}/kompute-shaders/op_getrows.comp
+              ${LLAMA_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+              ${LLAMA_DIR}/kompute-shaders/op_mul_mv_q_n.comp
             COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${LLAMA_DIR}/${source}
             COMMENT "Compiling ${source} to ${source}.spv"
         )
@@ -185,96 +199,118 @@ if (LLAMA_KOMPUTE)
         string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
         set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
         message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        add_custom_command(
-          OUTPUT ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${spv_file} >> ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-          COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-          DEPENDS ${spv_file} xxd
-          COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-        )
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+            )
+        else()
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+            )
+        endif()
       endforeach()
     endfunction()
 
-    if (EXISTS "${LLAMA_DIR}/kompute/CMakeLists.txt")
-        message(STATUS "Kompute found")
-        add_subdirectory(${LLAMA_DIR}/kompute)
+    set(KOMPUTE_OPT_LOG_LEVEL Critical CACHE STRING "Kompute log level")
+    add_subdirectory(${LLAMA_DIR}/kompute)
 
-        # Compile our shaders
-        compile_shader(SOURCES
-          kompute/op_scale.comp
-          kompute/op_add.comp
-          kompute/op_addrow.comp
-          kompute/op_mul.comp
-          kompute/op_mulrow.comp
-          kompute/op_silu.comp
-          kompute/op_relu.comp
-          kompute/op_gelu.comp
-          kompute/op_softmax.comp
-          kompute/op_norm.comp
-          kompute/op_rmsnorm.comp
-          kompute/op_diagmask.comp
-          kompute/op_mul_mat_f16.comp
-          kompute/op_mul_mat_q4_0.comp
-          kompute/op_mul_mat_q4_1.comp
-          kompute/op_getrows_f16.comp
-          kompute/op_getrows_q4_0.comp
-          kompute/op_getrows_q4_1.comp
-          kompute/op_rope.comp
-          kompute/op_cpy_f16_f16.comp
-          kompute/op_cpy_f16_f32.comp
-          kompute/op_cpy_f32_f16.comp
-          kompute/op_cpy_f32_f32.comp
-        )
+    # Compile our shaders
+    compile_shader(SOURCES
+      kompute-shaders/op_scale.comp
+      kompute-shaders/op_scale_8.comp
+      kompute-shaders/op_add.comp
+      kompute-shaders/op_addrow.comp
+      kompute-shaders/op_mul.comp
+      kompute-shaders/op_silu.comp
+      kompute-shaders/op_relu.comp
+      kompute-shaders/op_gelu.comp
+      kompute-shaders/op_softmax.comp
+      kompute-shaders/op_norm.comp
+      kompute-shaders/op_rmsnorm.comp
+      kompute-shaders/op_diagmask.comp
+      kompute-shaders/op_mul_mat_mat_f32.comp
+      kompute-shaders/op_mul_mat_f16.comp
+      kompute-shaders/op_mul_mat_q8_0.comp
+      kompute-shaders/op_mul_mat_q4_0.comp
+      kompute-shaders/op_mul_mat_q4_1.comp
+      kompute-shaders/op_mul_mat_q6_k.comp
+      kompute-shaders/op_getrows_f16.comp
+      kompute-shaders/op_getrows_q4_0.comp
+      kompute-shaders/op_getrows_q4_1.comp
+      kompute-shaders/op_getrows_q6_k.comp
+      kompute-shaders/op_rope_f16.comp
+      kompute-shaders/op_rope_f32.comp
+      kompute-shaders/op_cpy_f16_f16.comp
+      kompute-shaders/op_cpy_f16_f32.comp
+      kompute-shaders/op_cpy_f32_f16.comp
+      kompute-shaders/op_cpy_f32_f32.comp
+    )
 
-        # Create a custom target for our generated shaders
-        add_custom_target(generated_shaders DEPENDS
-          shaderop_scale.h
-          shaderop_add.h
-          shaderop_addrow.h
-          shaderop_mul.h
-          shaderop_mulrow.h
-          shaderop_silu.h
-          shaderop_relu.h
-          shaderop_gelu.h
-          shaderop_softmax.h
-          shaderop_norm.h
-          shaderop_rmsnorm.h
-          shaderop_diagmask.h
-          shaderop_mul_mat_f16.h
-          shaderop_mul_mat_q4_0.h
-          shaderop_mul_mat_q4_1.h
-          shaderop_getrows_f16.h
-          shaderop_getrows_q4_0.h
-          shaderop_getrows_q4_1.h
-          shaderop_rope.h
-          shaderop_cpy_f16_f16.h
-          shaderop_cpy_f16_f32.h
-          shaderop_cpy_f32_f16.h
-          shaderop_cpy_f32_f32.h
-        )
+    # Create a custom target for our generated shaders
+    add_custom_target(generated_shaders DEPENDS
+      shaderop_scale.h
+      shaderop_scale_8.h
+      shaderop_add.h
+      shaderop_addrow.h
+      shaderop_mul.h
+      shaderop_silu.h
+      shaderop_relu.h
+      shaderop_gelu.h
+      shaderop_softmax.h
+      shaderop_norm.h
+      shaderop_rmsnorm.h
+      shaderop_diagmask.h
+      shaderop_mul_mat_mat_f32.h
+      shaderop_mul_mat_f16.h
+      shaderop_mul_mat_q8_0.h
+      shaderop_mul_mat_q4_0.h
+      shaderop_mul_mat_q4_1.h
+      shaderop_mul_mat_q6_k.h
+      shaderop_getrows_f16.h
+      shaderop_getrows_q4_0.h
+      shaderop_getrows_q4_1.h
+      shaderop_getrows_q6_k.h
+      shaderop_rope_f16.h
+      shaderop_rope_f32.h
+      shaderop_cpy_f16_f16.h
+      shaderop_cpy_f16_f32.h
+      shaderop_cpy_f32_f16.h
+      shaderop_cpy_f32_f32.h
+    )
 
-        # Create a custom command that depends on the generated_shaders
-        add_custom_command(
-            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan.stamp
-            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan.stamp
-            DEPENDS generated_shaders
-            COMMENT "Ensuring shaders are generated before compiling ggml-vulkan.cpp"
-        )
+    # Create a custom command that depends on the generated_shaders
+    add_custom_command(
+        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        DEPENDS generated_shaders
+        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+    )
 
-        # Add the stamp to the main sources to ensure dependency tracking
-        set(GGML_SOURCES_KOMPUTE ${LLAMA_DIR}/ggml-vulkan.cpp ${LLAMA_DIR}/ggml-vulkan.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan.stamp)
-        add_compile_definitions(GGML_USE_KOMPUTE)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
-    else()
-        message(WARNING "Kompute not found")
-    endif()
+    # Add the stamp to the main sources to ensure dependency tracking
+    set(GGML_SOURCES_KOMPUTE ${LLAMA_DIR}/ggml-kompute.cpp ${LLAMA_DIR}/ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+    add_compile_definitions(GGML_USE_KOMPUTE)
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
+    set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
 endif()
 
 if (LLAMA_ALL_WARNINGS)
@@ -330,6 +366,13 @@ endif()
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+if (MSVC)
+  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+else ()
+  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+endif ()
+
 if (NOT MSVC)
     if (LLAMA_STATIC)
         add_link_options(-static)
@@ -345,6 +388,139 @@ if (NOT MSVC)
     endif()
 endif()
 
+if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
+    message(STATUS "ARM detected")
+    if (MSVC)
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+        add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+    else()
+        include(CheckCXXCompilerFlag)
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            add_compile_options(-mfp16-format=ieee)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            # Raspberry Pi 2
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            add_compile_options(-mno-unaligned-access)
+        endif()
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
+    message(STATUS "x86 detected")
+    if (MSVC)
+        if (LLAMA_AVX512)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (LLAMA_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (LLAMA_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+        elseif (LLAMA_AVX2)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX2>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX2>)
+        elseif (LLAMA_AVX)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
+        endif()
+    else()
+        if (LLAMA_F16C)
+            add_compile_options(-mf16c)
+        endif()
+        if (LLAMA_FMA)
+            add_compile_options(-mfma)
+        endif()
+        if (LLAMA_AVX)
+            add_compile_options(-mavx)
+        endif()
+        if (LLAMA_AVX2)
+            add_compile_options(-mavx2)
+        endif()
+        if (LLAMA_AVX512)
+            add_compile_options(-mavx512f)
+            add_compile_options(-mavx512bw)
+        endif()
+        if (LLAMA_AVX512_VBMI)
+            add_compile_options(-mavx512vbmi)
+        endif()
+        if (LLAMA_AVX512_VNNI)
+            add_compile_options(-mavx512vnni)
+        endif()
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    message(STATUS "PowerPC detected")
+    add_compile_options(-mcpu=native -mtune=native)
+    #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+else()
+    message(STATUS "Unknown architecture")
+endif()
+
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (
+    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
+    CMAKE_SYSTEM_NAME MATCHES "iOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "tvOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
+)
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
 function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
     message(STATUS "Configuring ggml implementation target llama${SUFFIX} in ${CMAKE_CURRENT_SOURCE_DIR}/${DIRECTORY}")
 
@@ -396,33 +572,26 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
         endif()
     endif()
 
-    set(GGML_SOURCES_QUANT_K )
-    set(GGML_METAL_SOURCES )
-    if (LLAMA_K_QUANTS)
-        set(GGML_SOURCES_QUANT_K
-            ${DIRECTORY}/k_quants.h
-            ${DIRECTORY}/k_quants.c)
+    set(GGML_METAL_SOURCES)
+    if (LLAMA_METAL)
+        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+        find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
 
-        if (LLAMA_METAL)
-            find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-            find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-            find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-            find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
+        set(GGML_METAL_SOURCES ${DIRECTORY}/ggml-metal.m ${DIRECTORY}/ggml-metal.h)
+        # get full path to the file
+        #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
 
-            set(GGML_METAL_SOURCES ${DIRECTORY}/ggml-metal.m ${DIRECTORY}/ggml-metal.h)
-            # get full path to the file
-            #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+        # copy ggml-metal.metal to bin directory
+        configure_file(${DIRECTORY}/ggml-metal.metal bin/ggml-metal.metal COPYONLY)
 
-            # copy ggml-metal.metal to bin directory
-            configure_file(${DIRECTORY}/ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
-                ${FOUNDATION_LIBRARY}
-                ${METAL_FRAMEWORK}
-                ${METALKIT_FRAMEWORK}
-                ${METALPERFORMANCE_FRAMEWORK}
-            )
-        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+            ${FOUNDATION_LIBRARY}
+            ${METAL_FRAMEWORK}
+            ${METALKIT_FRAMEWORK}
+            ${METALPERFORMANCE_FRAMEWORK}
+        )
     endif()
 
     add_library(ggml${SUFFIX} OBJECT
@@ -430,16 +599,15 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
                 ${DIRECTORY}/ggml.h
                 ${DIRECTORY}/ggml-alloc.c
                 ${DIRECTORY}/ggml-alloc.h
-                ${GGML_SOURCES_QUANT_K}
+                ${DIRECTORY}/ggml-backend.c
+                ${DIRECTORY}/ggml-backend.h
+                ${DIRECTORY}/ggml-quants.h
+                ${DIRECTORY}/ggml-quants.c
                 ${GGML_SOURCES_CUDA}
                 ${GGML_METAL_SOURCES}
                 ${GGML_OPENCL_SOURCES}
                 ${GGML_SOURCES_KOMPUTE})
 
-    if (LLAMA_K_QUANTS)
-        target_compile_definitions(ggml${SUFFIX} PUBLIC GGML_USE_K_QUANTS)
-    endif()
-
     if (LLAMA_METAL AND GGML_METAL_SOURCES)
         target_compile_definitions(ggml${SUFFIX} PUBLIC GGML_USE_METAL GGML_METAL_NDEBUG)
     endif()
@@ -452,15 +620,14 @@ function(include_ggml DIRECTORY SUFFIX WITH_LLAMA)
 
     if (WITH_LLAMA)
         # Backwards compatibility with old llama.cpp versions
-        set(LLAMA_UTIL_SOURCE_FILE llama-util.h)
+#        set(LLAMA_UTIL_SOURCE_FILE llama-util.h)
         if (NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${DIRECTORY}/${LLAMA_UTIL_SOURCE_FILE})
             set(LLAMA_UTIL_SOURCE_FILE llama_util.h)
         endif()
 
         add_library(llama${SUFFIX} STATIC
                     ${DIRECTORY}/llama.cpp
-                    ${DIRECTORY}/llama.h
-                    ${DIRECTORY}/${LLAMA_UTIL_SOURCE_FILE})
+                    ${DIRECTORY}/llama.h)
 
         if (LLAMA_METAL AND GGML_METAL_SOURCES)
             target_compile_definitions(llama${SUFFIX} PUBLIC GGML_USE_METAL GGML_METAL_NDEBUG)

gpt4all-backend/llamamodel.cpp

@@ -6,58 +6,58 @@
 #include <cstdio>
 #include <cstring>
 #include <fstream>
-#include <map>
-#include <string>
-#include <vector>
+#include <iomanip>
 #include <iostream>
-#if defined(_WIN32) && defined(_MSC_VER)
-    #define WIN32_LEAN_AND_MEAN
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-    #include <windows.h>
-    #include <io.h>
-    #include <stdio.h>
-#else
-    #include <unistd.h>
-#endif
+#include <map>
 #include <random>
+#include <sstream>
+#include <stdexcept>
+#include <string>
 #include <thread>
 #include <unordered_set>
+#include <vector>
 
 #include <llama.h>
 #include <ggml.h>
-
 #ifdef GGML_USE_KOMPUTE
-#include "ggml-vulkan.h"
+#include <ggml-kompute.h>
 #endif
 
-namespace {
-const char *modelType_ = "LLaMA";
+using namespace std::string_literals;
+
+// Maximum supported GGUF version
+static constexpr int GGUF_VER_MAX = 3;
+
+static const char * const modelType_ = "LLaMA";
+
+static bool llama_verbose() {
+    const char* var = getenv("GPT4ALL_VERBOSE_LLAMACPP");
+    return var && *var;
+}
+
+static void llama_log_callback(enum ggml_log_level level, const char *text, void *userdata) {
+    (void)userdata;
+    if (llama_verbose() || level <= GGML_LOG_LEVEL_ERROR) {
+        fputs(text, stderr);
+    }
 }
 
 struct gpt_params {
     int32_t seed          = -1;   // RNG seed
     int32_t n_keep        = 0;    // number of tokens to keep from initial prompt
-#if LLAMA_DATE <= 230511
-    int32_t n_parts       = -1;   // amount of model parts (-1 = determine from model dimensions)
-#endif
 
-#if LLAMA_DATE >= 230519
     // sampling parameters
     float   tfs_z         = 1.0f; // 1.0 = disabled
     float   typical_p     = 1.0f; // 1.0 = disabled
-#endif
 
     std::string prompt = "";
 
-    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
+    enum ggml_type kv_type = GGML_TYPE_F16; // use f16 instead of f32 for memory kv
 
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
 };
 
-#if LLAMA_DATE >= 230519
 static int llama_sample_top_p_top_k(
         llama_context *ctx,
         const llama_token *last_n_tokens_data,
@@ -65,9 +65,10 @@ static int llama_sample_top_p_top_k(
         int top_k,
         float top_p,
         float temp,
-        float repeat_penalty) {
-    auto logits = llama_get_logits(ctx);
-    auto n_vocab = llama_n_vocab(ctx);
+        float repeat_penalty,
+        int32_t pos) {
+    auto logits = llama_get_logits_ith(ctx, pos);
+    auto n_vocab = llama_n_vocab(llama_get_model(ctx));
     // Populate initial list of all candidates
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -76,23 +77,76 @@ static int llama_sample_top_p_top_k(
     }
     llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
     // Sample repeat penalty
-    llama_sample_repetition_penalty(nullptr, &candidates_p, last_n_tokens_data, last_n_tokens_size, repeat_penalty);
+    llama_sample_repetition_penalties(nullptr, &candidates_p, last_n_tokens_data, last_n_tokens_size, repeat_penalty, 0.0f, 0.0f);
     // Temperature sampling
     llama_sample_top_k(ctx, &candidates_p, top_k, 1);
     llama_sample_tail_free(ctx, &candidates_p, 1.0f, 1);
     llama_sample_typical(ctx, &candidates_p, 1.0f, 1);
     llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-    llama_sample_temperature(ctx, &candidates_p, temp);
+    llama_sample_temp(ctx, &candidates_p, temp);
     return llama_sample_token(ctx, &candidates_p);
 }
-#endif
+
+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);
+}
+
+static gguf_context *load_gguf(const char *fname) {
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ nullptr,
+    };
+    gguf_context *ctx = gguf_init_from_file(fname, params);
+    if (!ctx) {
+        std::cerr << __func__ << ": gguf_init_from_file failed\n";
+        return nullptr;
+    }
+
+    int gguf_ver = gguf_get_version(ctx);
+    if (gguf_ver > GGUF_VER_MAX) {
+        std::cerr << __func__ << ": unsupported gguf version: " << gguf_ver << "\n";
+        gguf_free(ctx);
+        return nullptr;
+    }
+
+    return ctx;
+}
+
+static int32_t get_arch_key_u32(std::string const &modelPath, std::string const &archKey) {
+    auto * ctx = load_gguf(modelPath.c_str());
+    auto arch = get_arch_name(ctx);
+
+    int32_t value = -1;
+    if (ctx) {
+        auto key = arch + "." + archKey;
+        int keyidx = gguf_find_key(ctx, key.c_str());
+        if (keyidx != -1) {
+            value = gguf_get_val_u32(ctx, keyidx);
+        } else {
+            std::cerr << __func__ << ": " << key << "not found in " << modelPath << "\n";
+        }
+    }
+
+    gguf_free(ctx);
+    return value;
+}
 
 struct LLamaPrivate {
     const std::string modelPath;
     bool modelLoaded;
+    int device = -1;
+    llama_model *model = nullptr;
     llama_context *ctx = nullptr;
-    llama_context_params params;
+    llama_model_params model_params;
+    llama_context_params ctx_params;
     int64_t n_threads = 0;
+    std::vector<LLModel::Token> end_tokens;
 };
 
 LLamaModel::LLamaModel()
@@ -111,7 +165,9 @@ struct llama_file_hparams {
     enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16;
 };
 
-size_t LLamaModel::requiredMem(const std::string &modelPath) {
+size_t LLamaModel::requiredMem(const std::string &modelPath, int n_ctx, int ngl) {
+    // TODO(cebtenzzre): update to GGUF
+    (void)ngl; // FIXME(cetenzzre): use this value
     auto fin = std::ifstream(modelPath, std::ios::binary);
     fin.seekg(0, std::ios_base::end);
     size_t filesize = fin.tellg();
@@ -128,64 +184,154 @@ size_t LLamaModel::requiredMem(const std::string &modelPath) {
     fin.read(reinterpret_cast<char*>(&hparams.n_layer), sizeof(hparams.n_layer));
     fin.read(reinterpret_cast<char*>(&hparams.n_rot), sizeof(hparams.n_rot));
     fin.read(reinterpret_cast<char*>(&hparams.ftype), sizeof(hparams.ftype));
-    const size_t n_ctx = 2048;
     const size_t kvcache_element_size = 2; // fp16
     const size_t est_kvcache_size = hparams.n_embd * hparams.n_layer * 2u * n_ctx * kvcache_element_size;
     return filesize + est_kvcache_size;
 }
 
-bool LLamaModel::loadModel(const std::string &modelPath)
+bool LLamaModel::isModelBlacklisted(const std::string &modelPath) {
+    auto * ctx = load_gguf(modelPath.c_str());
+    if (!ctx) {
+        std::cerr << __func__ << ": failed to load " << modelPath << "\n";
+        return false;
+    }
+
+    auto get_key = [ctx, &modelPath](const char *name) {
+        int keyidx = gguf_find_key(ctx, name);
+        if (keyidx == -1) {
+            throw std::logic_error(name + " not found in "s + modelPath);
+        }
+        return keyidx;
+    };
+
+    bool res = false;
+    try {
+        std::string name(gguf_get_val_str(ctx, get_key("general.name")));
+        int token_idx = get_key("tokenizer.ggml.tokens");
+        int n_vocab = gguf_get_arr_n(ctx, token_idx);
+
+        // check for known bad models
+        if (name == "open-orca_mistral-7b-openorca"
+            && n_vocab == 32002
+            && gguf_get_arr_str(ctx, token_idx, 32000) == "<dummy32000>"s // should be <|im_end|>
+        ) {
+            res = true;
+        }
+    } catch (const std::logic_error &e) {
+        std::cerr << __func__ << ": " << e.what() << "\n";
+    }
+
+    gguf_free(ctx);
+    return res;
+}
+
+bool LLamaModel::loadModel(const std::string &modelPath, int n_ctx, int ngl)
 {
-    // load the model
-    d_ptr->params = llama_context_default_params();
+    d_ptr->modelLoaded = false;
+
+    // clean up after previous loadModel()
+    if (d_ptr->model) {
+        llama_free_model(d_ptr->model);
+        d_ptr->model = nullptr;
+    }
+    if (d_ptr->ctx) {
+        llama_free(d_ptr->ctx);
+        d_ptr->ctx = nullptr;
+    }
+
+    if (n_ctx < 8) {
+        std::cerr << "warning: minimum context size is 8, using minimum size.\n";
+        n_ctx = 8;
+    }
+
+    // -- load the model --
 
     gpt_params params;
-    d_ptr->params.n_ctx      = 2048;
-    d_ptr->params.seed       = params.seed;
-    d_ptr->params.f16_kv     = params.memory_f16;
-    d_ptr->params.use_mmap   = params.use_mmap;
+
+    d_ptr->model_params = llama_model_default_params();
+
+    d_ptr->model_params.use_mmap  = params.use_mmap;
 #if defined (__APPLE__)
-    d_ptr->params.use_mlock  = true;
+    d_ptr->model_params.use_mlock = true;
 #else
-    d_ptr->params.use_mlock  = params.use_mlock;
-#endif
-#if LLAMA_DATE <= 230511
-    d_ptr->params.n_parts  = params.n_parts;
+    d_ptr->model_params.use_mlock = params.use_mlock;
 #endif
+
+    d_ptr->model_params.progress_callback = &LLModel::staticProgressCallback;
+    d_ptr->model_params.progress_callback_user_data = this;
+
 #ifdef GGML_USE_METAL
-    std::cerr << "llama.cpp: using Metal" << std::endl;
-    // metal always runs the whole model if n_gpu_layers is not 0, at least
-    // currently
-    d_ptr->params.n_gpu_layers = 1;
-#endif
-#ifdef GGML_USE_KOMPUTE
-    if (ggml_vk_has_device()) {
-        // vulkan always runs the whole model if n_gpu_layers is not 0, at least
-        // currently
-        d_ptr->params.n_gpu_layers = 1;
+    if (llama_verbose()) {
+        std::cerr << "llama.cpp: using Metal" << std::endl;
+    }
+
+    // always fully offload on Metal
+    // TODO(cebtenzzre): use this parameter to allow using more than 53% of system RAM to load a model
+    d_ptr->model_params.n_gpu_layers = 100;
+#elif defined(GGML_USE_KOMPUTE)
+    if (d_ptr->device != -1) {
+        d_ptr->model_params.main_gpu = d_ptr->device;
+        d_ptr->model_params.n_gpu_layers = ngl;
     }
 #endif
 
-    d_ptr->ctx = llama_init_from_file(modelPath.c_str(), d_ptr->params);
-    if (!d_ptr->ctx) {
+    d_ptr->model = llama_load_model_from_file_gpt4all(modelPath.c_str(), &d_ptr->model_params);
+    if (!d_ptr->model) {
+        fflush(stdout);
+        d_ptr->device = -1;
         std::cerr << "LLAMA ERROR: failed to load model from " <<  modelPath << std::endl;
         return false;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(d_ptr->model);
+    if (n_ctx > n_ctx_train) {
+        std::cerr << "warning: model was trained on only " << n_ctx_train << " context tokens ("
+                  << n_ctx << " specified)\n";
+    }
+
+    // -- initialize the context --
+
+    d_ptr->ctx_params = llama_context_default_params();
+
+    d_ptr->ctx_params.n_ctx   = n_ctx;
+    d_ptr->ctx_params.seed    = params.seed;
+    d_ptr->ctx_params.type_k  = params.kv_type;
+    d_ptr->ctx_params.type_v  = params.kv_type;
+
+    // The new batch API provides space for n_vocab*n_tokens logits. Tell llama.cpp early
+    // that we want this many logits so the state serializes consistently.
+    d_ptr->ctx_params.logits_all = true;
+
+    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    d_ptr->ctx_params.n_threads       = d_ptr->n_threads;
+    d_ptr->ctx_params.n_threads_batch = d_ptr->n_threads;
+
+    d_ptr->ctx = llama_new_context_with_model(d_ptr->model, d_ptr->ctx_params);
+    if (!d_ptr->ctx) {
+        fflush(stdout);
+        std::cerr << "LLAMA ERROR: failed to init context for model " <<  modelPath << std::endl;
+        llama_free_model(d_ptr->model);
+        d_ptr->model = nullptr;
+        d_ptr->device = -1;
+        return false;
+    }
+
+    d_ptr->end_tokens = {llama_token_eos(d_ptr->model)};
+
 #ifdef GGML_USE_KOMPUTE
-    if (ggml_vk_has_device()) {
+    if (usingGPUDevice() && ggml_vk_has_device()) {
         std::cerr << "llama.cpp: using Vulkan on " << ggml_vk_current_device().name << std::endl;
     }
 #endif
 
-    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    fflush(stdout);
     d_ptr->modelLoaded = true;
-    fflush(stderr);
     return true;
 }
 
 void LLamaModel::setThreadCount(int32_t n_threads) {
     d_ptr->n_threads = n_threads;
+    llama_set_n_threads(d_ptr->ctx, n_threads, n_threads);
 }
 
 int32_t LLamaModel::threadCount() const {
@@ -194,9 +340,10 @@ int32_t LLamaModel::threadCount() const {
 
 LLamaModel::~LLamaModel()
 {
-    if(d_ptr->ctx) {
+    if (d_ptr->ctx) {
         llama_free(d_ptr->ctx);
     }
+    llama_free_model(d_ptr->model);
 }
 
 bool LLamaModel::isModelLoaded() const
@@ -220,18 +367,20 @@ size_t LLamaModel::restoreState(const uint8_t *src)
     return llama_set_state_data(d_ptr->ctx, const_cast<uint8_t*>(src));
 }
 
-std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, const std::string &str) const
+std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, const std::string &str, bool special) const
 {
-    const bool useBOS = ctx.n_past == 0 && (ctx.tokens.empty() || ctx.tokens.front() != llama_token_bos());
-    std::vector<LLModel::Token> fres(str.size()+4);
-    auto fres_len = llama_tokenize(d_ptr->ctx, str.c_str(), fres.data(), fres.size(), useBOS);
+    const bool wantBOS = ctx.n_past == 0 && ctx.tokens.empty();
+    const bool useBOS = wantBOS && shouldAddBOS();
+    auto strCat = wantBOS && !special ? " " + str : str; // insert leading space ourselves, llama.cpp fork doesn't anymore
+    std::vector<LLModel::Token> fres(strCat.size()+4);
+    auto fres_len = llama_tokenize(d_ptr->model, strCat.c_str(), strCat.length(), fres.data(), fres.size(), useBOS, special);
     fres.resize(fres_len);
     return fres;
 }
 
 std::string LLamaModel::tokenToString(Token id) const
 {
-    return llama_token_to_str(d_ptr->ctx, id);
+    return llama_token_to_piece(d_ptr->ctx, id);
 }
 
 LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
@@ -240,21 +389,32 @@ LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
     return llama_sample_top_p_top_k(d_ptr->ctx,
         promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
         n_prev_toks, promptCtx.top_k, promptCtx.top_p, promptCtx.temp,
-        promptCtx.repeat_penalty);
+        promptCtx.repeat_penalty, promptCtx.n_last_batch_tokens - 1);
 }
 
 bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
 {
-    // When we recalculate context we could have erased the original BOS token... we need to replace it
-    const bool useBOS = ctx.n_past == 0 && (ctx.tokens.empty() || ctx.tokens.front() != llama_token_bos());
-    if (useBOS) {
-        std::vector<int32_t> myTokens;
-        myTokens.push_back(llama_token_bos());
-        myTokens.insert(myTokens.end(), tokens.begin(), tokens.end());
-        ctx.n_past += 1;
-        return llama_eval(d_ptr->ctx, myTokens.data(), myTokens.size(), ctx.n_past, d_ptr->n_threads) == 0;
-    } else
-        return llama_eval(d_ptr->ctx, tokens.data(), tokens.size(), ctx.n_past, d_ptr->n_threads) == 0;
+    llama_kv_cache_seq_rm(d_ptr->ctx, 0, ctx.n_past, -1);
+
+    llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
+
+    batch.n_tokens = tokens.size();
+    ctx.n_last_batch_tokens = tokens.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token   [i] = tokens[i];
+        batch.pos     [i] = ctx.n_past + i;
+        batch.n_seq_id[i] = 1;
+        batch.seq_id  [i][0] = 0;
+        batch.logits  [i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    int res = llama_decode(d_ptr->ctx, batch);
+    llama_batch_free(batch);
+    return res == 0;
 }
 
 int32_t LLamaModel::contextLength() const
@@ -264,66 +424,84 @@ int32_t LLamaModel::contextLength() const
 
 const std::vector<LLModel::Token> &LLamaModel::endTokens() const
 {
-    static const std::vector<LLModel::Token> fres = {llama_token_eos()};
-    return fres;
+    return d_ptr->end_tokens;
 }
 
-#if defined(GGML_USE_KOMPUTE)
-#include "ggml-vulkan.h"
-#endif
-
-std::vector<LLModel::GPUDevice> LLamaModel::availableGPUDevices(size_t memoryRequired)
+bool LLamaModel::shouldAddBOS() const
 {
-#if defined(GGML_USE_KOMPUTE)
-    std::vector<ggml_vk_device> vkDevices = ggml_vk_available_devices(memoryRequired);
+    int add_bos = llama_add_bos_token(d_ptr->model);
+    return add_bos != -1 ? bool(add_bos) : llama_vocab_type(d_ptr->model) == LLAMA_VOCAB_TYPE_SPM;
+}
 
-    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;
+int32_t LLamaModel::maxContextLength(std::string const &modelPath) const
+{
+    return get_arch_key_u32(modelPath, "context_length");
+}
 
-        devices.push_back(device);
+int32_t LLamaModel::layerCount(std::string const &modelPath) const
+{
+    return get_arch_key_u32(modelPath, "block_count");
+}
+
+std::vector<LLModel::GPUDevice> LLamaModel::availableGPUDevices(size_t memoryRequired) const
+{
+#ifdef GGML_USE_KOMPUTE
+    size_t count = 0;
+    auto * vkDevices = ggml_vk_available_devices(memoryRequired, &count);
+
+    if (vkDevices) {
+        std::vector<LLModel::GPUDevice> devices;
+        devices.reserve(count);
+
+        for (size_t i = 0; i < count; ++i) {
+            auto & dev = vkDevices[i];
+            devices.emplace_back(
+                /* index    = */ dev.index,
+                /* type     = */ dev.type,
+                /* heapSize = */ dev.heapSize,
+                /* name     = */ dev.name,
+                /* vendor   = */ dev.vendor
+            );
+            ggml_vk_device_destroy(&dev);
+        }
+
+        free(vkDevices);
+        return devices;
     }
-
-    return devices;
 #else
-    return std::vector<LLModel::GPUDevice>();
+    std::cerr << __func__ << ": built without Kompute\n";
 #endif
+
+    return {};
 }
 
-bool LLamaModel::initializeGPUDevice(size_t memoryRequired, const std::string& device)
+bool LLamaModel::initializeGPUDevice(size_t memoryRequired, const std::string &name) const
 {
 #if defined(GGML_USE_KOMPUTE)
-    return ggml_vk_init_device(memoryRequired, device);
+    ggml_vk_device device;
+    bool ok = ggml_vk_get_device(&device, memoryRequired, name.c_str());
+    if (ok) {
+        d_ptr->device = device.index;
+        return true;
+    }
 #else
+    (void)memoryRequired;
+    (void)name;
+#endif
     return false;
-#endif
 }
 
-bool LLamaModel::initializeGPUDevice(const LLModel::GPUDevice &device)
+bool LLamaModel::initializeGPUDevice(int device, std::string *unavail_reason) const
 {
 #if defined(GGML_USE_KOMPUTE)
-    ggml_vk_device vkDevice;
-    vkDevice.index = device.index;
-    vkDevice.type = device.type;
-    vkDevice.heapSize = device.heapSize;
-    vkDevice.name = device.name;
-    vkDevice.vendor = device.vendor;
-    return ggml_vk_init_device(vkDevice);
-#else
-    return false;
-#endif
-}
-
-bool LLamaModel::initializeGPUDevice(int device)
-{
-#if defined(GGML_USE_KOMPUTE)
-    return ggml_vk_init_device(device);
+    (void)unavail_reason;
+    d_ptr->device = device;
+    return true;
 #else
+    (void)device;
+    if (unavail_reason) {
+        *unavail_reason = "built without Kompute";
+    }
     return false;
 #endif
 }
@@ -331,7 +509,18 @@ bool LLamaModel::initializeGPUDevice(int device)
 bool LLamaModel::hasGPUDevice()
 {
 #if defined(GGML_USE_KOMPUTE)
-    return ggml_vk_has_device();
+    return d_ptr->device != -1;
+#else
+    return false;
+#endif
+}
+
+bool LLamaModel::usingGPUDevice()
+{
+#if defined(GGML_USE_KOMPUTE)
+    return hasGPUDevice() && d_ptr->model_params.n_gpu_layers > 0;
+#elif defined(GGML_USE_METAL)
+    return true;
 #else
     return false;
 #endif
@@ -356,42 +545,31 @@ DLL_EXPORT const char *get_build_variant() {
     return GGML_BUILD_VARIANT;
 }
 
-DLL_EXPORT bool magic_match(std::istream& f) {
-    // Check magic
-    uint32_t magic = 0;
-    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
-    if (magic != 0x67676a74) return false;
-    // Check version
-    uint32_t version = 0;
-    f.read(reinterpret_cast<char*>(&version), sizeof(version));
-    if (!(version LLAMA_VERSIONS)) {
-        return false;
+DLL_EXPORT bool magic_match(const char *fname) {
+    auto * ctx = load_gguf(fname);
+    auto arch = get_arch_name(ctx);
+
+    bool valid = true;
+
+    static const std::vector<const char *> known_arches {
+        "baichuan", "bloom", "codeshell", "falcon", "gemma", "gpt2", "llama", "mpt", "orion", "persimmon", "phi2",
+        "plamo", "qwen", "qwen2", "refact", "stablelm", "starcoder"
+    };
+
+    if (std::find(known_arches.begin(), known_arches.end(), arch) == known_arches.end()) {
+        // not supported by this version of llama.cpp
+        if (!(arch == "gptj" || arch == "bert")) { // we support these via other modules
+            std::cerr << __func__ << ": unsupported model architecture: " << arch << "\n";
+        }
+        valid = false;
     }
-    llama_file_hparams hparams;
-    f.read(reinterpret_cast<char*>(&hparams), sizeof(hparams));
-    if (!(hparams.n_vocab >= 32000 && hparams.n_vocab <= 32100)) {
-        return false; // not a llama.
-    }
-#ifdef GGML_USE_METAL
-    // Check quant supported on metal
-    // skip fields
-    switch(hparams.ftype) {
-        // currently supported on Metal https://github.com/ggerganov/llama.cpp/blob/ae9663f1887513e152839e91f61c513075a19422/ggml-metal.m#L51-L55
-        case LLAMA_FTYPE_MOSTLY_F16:
-        case LLAMA_FTYPE_MOSTLY_Q2_K:
-        case LLAMA_FTYPE_MOSTLY_Q4_0:
-        case LLAMA_FTYPE_MOSTLY_Q6_K:
-        case LLAMA_FTYPE_MOSTLY_Q4_K_S:
-        case LLAMA_FTYPE_MOSTLY_Q4_K_M:
-            return true;
-        default: // unsupported quant-type for Metal
-            return false;
-    }
-#endif
-    return true;
+
+    gguf_free(ctx);
+    return valid;
 }
 
 DLL_EXPORT LLModel *construct() {
+    llama_log_set(llama_log_callback, nullptr);
     return new LLamaModel;
 }
 }

gpt4all-backend/llamamodel_impl.h

@@ -4,8 +4,9 @@
 #ifndef LLAMAMODEL_H
 #define LLAMAMODEL_H
 
-#include <string>
 #include <functional>
+#include <memory>
+#include <string>
 #include <vector>
 #include "llmodel.h"
 
@@ -17,30 +18,35 @@ 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 isModelBlacklisted(const std::string &modelPath) override;
     bool isModelLoaded() const override;
-    size_t requiredMem(const std::string &modelPath) override;
+    size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) override;
     size_t stateSize() const override;
     size_t saveState(uint8_t *dest) const override;
     size_t restoreState(const uint8_t *src) override;
     void setThreadCount(int32_t n_threads) override;
     int32_t threadCount() const override;
-    std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired) override;
-    bool initializeGPUDevice(size_t memoryRequired, const std::string& device) override;
-    bool initializeGPUDevice(const GPUDevice &device) override;
-    bool initializeGPUDevice(int device) override;
+    std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired) const override;
+    bool initializeGPUDevice(size_t memoryRequired, const std::string &name) const override;
+    bool initializeGPUDevice(int device, std::string *unavail_reason) const override;
     bool hasGPUDevice() override;
+    bool usingGPUDevice() override;
 
 private:
-    LLamaPrivate *d_ptr;
+    std::unique_ptr<LLamaPrivate> d_ptr;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    std::string tokenToString(Token) const override;
-    Token sampleToken(PromptContext& ctx) const override;
-    bool evalTokens(PromptContext& ctx, const std::vector<int32_t> &tokens) const override;
+    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;
 };
 
 #endif // LLAMAMODEL_H

gpt4all-backend/llmodel.cpp

@@ -2,14 +2,17 @@
 #include "dlhandle.h"
 #include "sysinfo.h"
 
-#include <iostream>
-#include <string>
-#include <vector>
-#include <fstream>
-#include <filesystem>
 #include <cassert>
 #include <cstdlib>
+#include <filesystem>
+#include <fstream>
+#include <iostream>
+#include <memory>
+#include <regex>
 #include <sstream>
+#include <string>
+#include <vector>
+
 #ifdef _MSC_VER
 #include <intrin.h>
 #endif
@@ -52,7 +55,7 @@ LLModel::Implementation::Implementation(Dlhandle &&dlhandle_)
     auto get_build_variant = m_dlhandle->get<const char *()>("get_build_variant");
     assert(get_build_variant);
     m_buildVariant = get_build_variant();
-    m_magicMatch = m_dlhandle->get<bool(std::ifstream&)>("magic_match");
+    m_magicMatch = m_dlhandle->get<bool(const char*)>("magic_match");
     assert(m_magicMatch);
     m_construct = m_dlhandle->get<LLModel *()>("construct");
     assert(m_construct);
@@ -81,6 +84,13 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
     static auto* libs = new std::vector<Implementation>([] () {
         std::vector<Implementation> fres;
 
+        std::string impl_name_re = "(bert|gptj|llamamodel-mainline)";
+        if (requires_avxonly()) {
+            impl_name_re += "-avxonly";
+        } else {
+            impl_name_re += "-(default|metal)";
+        }
+        std::regex re(impl_name_re);
         auto search_in_directory = [&](const std::string& paths) {
             std::stringstream ss(paths);
             std::string path;
@@ -90,7 +100,10 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
                 // Iterate over all libraries
                 for (const auto& f : std::filesystem::directory_iterator(fs_path)) {
                     const std::filesystem::path& p = f.path();
+
                     if (p.extension() != LIB_FILE_EXT) continue;
+                    if (!std::regex_search(p.stem().string(), re)) continue;
+
                     // Add to list if model implementation
                     try {
                         Dlhandle dl(p.string());
@@ -111,35 +124,43 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
     return *libs;
 }
 
-const LLModel::Implementation* LLModel::Implementation::implementation(std::ifstream& f, const std::string& buildVariant) {
+const LLModel::Implementation* LLModel::Implementation::implementation(const char *fname, const std::string& buildVariant) {
+    bool buildVariantMatched = false;
     for (const auto& i : implementationList()) {
-        f.seekg(0);
-        if (!i.m_magicMatch(f)) continue;
         if (buildVariant != i.m_buildVariant) continue;
+        buildVariantMatched = true;
+
+        if (!i.m_magicMatch(fname)) continue;
         return &i;
     }
+
+    if (!buildVariantMatched) {
+        std::cerr << "LLModel ERROR: Could not find any implementations for build variant: " << buildVariant << "\n";
+    }
     return nullptr;
 }
 
-LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::string buildVariant) {
-
-    if (!has_at_least_minimal_hardware())
+LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::string buildVariant, int n_ctx) {
+    if (!has_at_least_minimal_hardware()) {
+        std::cerr << "LLModel ERROR: CPU does not support AVX\n";
         return nullptr;
+    }
 
-    // Read magic
-    std::ifstream f(modelPath, std::ios::binary);
-    if (!f) return nullptr;
     // Get correct implementation
     const Implementation* impl = nullptr;
 
     #if defined(__APPLE__) && defined(__arm64__) // FIXME: See if metal works for intel macs
         if (buildVariant == "auto") {
             size_t total_mem = getSystemTotalRAMInBytes();
-            impl = implementation(f, "metal");
+            impl = implementation(modelPath.c_str(), "metal");
             if(impl) {
                 LLModel* metalimpl = impl->m_construct();
                 metalimpl->m_implementation = impl;
-                size_t req_mem = metalimpl->requiredMem(modelPath);
+                /* 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. */
+                size_t req_mem = metalimpl->requiredMem(modelPath, n_ctx, 100);
                 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) {
@@ -150,6 +171,8 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
                 }
             }
         }
+    #else
+        (void)n_ctx;
     #endif
 
     if (!impl) {
@@ -161,10 +184,9 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
                 buildVariant = "default";
             }
         }
-        impl = implementation(f, buildVariant);
+        impl = implementation(modelPath.c_str(), buildVariant);
         if (!impl) return nullptr;
     }
-    f.close();
 
     // Construct and return llmodel implementation
     auto fres = impl->m_construct();
@@ -172,6 +194,40 @@ LLModel *LLModel::Implementation::construct(const std::string &modelPath, std::s
     return fres;
 }
 
+LLModel *LLModel::Implementation::constructDefaultLlama() {
+    static std::unique_ptr<LLModel> llama([]() -> LLModel * {
+        const LLModel::Implementation *impl = nullptr;
+        for (const auto &i : implementationList()) {
+            if (i.m_buildVariant == "metal" || i.m_modelType != "LLaMA") continue;
+            impl = &i;
+        }
+        if (!impl) {
+            std::cerr << "LLModel ERROR: Could not find CPU LLaMA implementation\n";
+            return nullptr;
+        }
+        auto fres = impl->m_construct();
+        fres->m_implementation = impl;
+        return fres;
+    }());
+    return llama.get();
+}
+
+std::vector<LLModel::GPUDevice> LLModel::Implementation::availableGPUDevices() {
+    auto * llama = constructDefaultLlama();
+    if (llama) { return llama->availableGPUDevices(0); }
+    return {};
+}
+
+int32_t LLModel::Implementation::maxContextLength(const std::string &modelPath) {
+    auto * llama = constructDefaultLlama();
+    return llama ? llama->maxContextLength(modelPath) : -1;
+}
+
+int32_t LLModel::Implementation::layerCount(const std::string &modelPath) {
+    auto * llama = constructDefaultLlama();
+    return llama ? llama->layerCount(modelPath) : -1;
+}
+
 void LLModel::Implementation::setImplementationsSearchPath(const std::string& path) {
     s_implementations_search_path = path;
 }

gpt4all-backend/llmodel.h

@@ -15,28 +15,44 @@ class Dlhandle;
 class LLModel {
 public:
     using Token = int32_t;
+
+    struct GPUDevice {
+        int index;
+        int type;
+        size_t heapSize;
+        std::string name;
+        std::string vendor;
+
+        GPUDevice(int index, int type, size_t heapSize, std::string name, std::string vendor):
+            index(index), type(type), heapSize(heapSize), name(std::move(name)), vendor(std::move(vendor)) {}
+    };
+
     class Implementation {
     public:
-        Implementation(Dlhandle&&);
-        Implementation(const Implementation&) = delete;
-        Implementation(Implementation&&);
+        Implementation(Dlhandle &&);
+        Implementation(const Implementation &) = delete;
+        Implementation(Implementation &&);
         ~Implementation();
 
         std::string_view modelType() const { return m_modelType; }
         std::string_view buildVariant() const { return m_buildVariant; }
 
-        static bool isImplementation(const Dlhandle&);
-        static const std::vector<Implementation>& implementationList();
-        static const Implementation *implementation(std::ifstream& f, const std::string& buildVariant);
-        static LLModel *construct(const std::string &modelPath, std::string buildVariant = "auto");
-        static void setImplementationsSearchPath(const std::string& path);
-        static const std::string& implementationsSearchPath();
+        static bool isImplementation(const Dlhandle &dl);
+        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", int n_ctx = 2048);
+        static std::vector<GPUDevice> availableGPUDevices();
+        static int32_t maxContextLength(const std::string &modelPath);
+        static int32_t layerCount(const std::string &modelPath);
+        static void setImplementationsSearchPath(const std::string &path);
+        static const std::string &implementationsSearchPath();
 
     private:
-        bool (*m_magicMatch)(std::ifstream& f);
+        static LLModel *constructDefaultLlama();
+
+        bool (*m_magicMatch)(const char *fname);
         LLModel *(*m_construct)();
 
-    private:
         std::string_view m_modelType;
         std::string_view m_buildVariant;
         Dlhandle *m_dlhandle;
@@ -54,62 +70,91 @@ public:
         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) { (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);
 
-    virtual void setThreadCount(int32_t /*n_threads*/) {}
+    virtual void setThreadCount(int32_t n_threads) { (void)n_threads; }
     virtual int32_t threadCount() const { return 1; }
 
-    const Implementation& implementation() const {
+    const Implementation &implementation() const {
         return *m_implementation;
     }
 
-    virtual std::vector<GPUDevice> availableGPUDevices(size_t /*memoryRequired*/) { return std::vector<GPUDevice>(); }
-    virtual bool initializeGPUDevice(size_t /*memoryRequired*/, const std::string& /*device*/) { return false; }
-    virtual bool initializeGPUDevice(const GPUDevice &/*device*/) { return false; }
-    virtual bool initializeGPUDevice(int /*device*/) { return false; }
+    virtual std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired) const {
+        (void)memoryRequired;
+        return {};
+    }
+
+    virtual bool initializeGPUDevice(size_t memoryRequired, const std::string &name) const {
+        (void)memoryRequired;
+        (void)name;
+        return false;
+    }
+
+    virtual bool initializeGPUDevice(int device, std::string *unavail_reason = nullptr) const {
+        (void)device;
+        if (unavail_reason) {
+            *unavail_reason = "model has no GPU support";
+        }
+        return false;
+    }
+
     virtual bool hasGPUDevice() { return false; }
+    virtual bool usingGPUDevice() { return false; }
+
+    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
@@ -117,6 +162,24 @@ 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 +1,9 @@
 #include "llmodel_c.h"
 #include "llmodel.h"
 
-#include <cstring>
 #include <cerrno>
+#include <cstring>
+#include <iostream>
 #include <utility>
 
 struct LLModelWrapper {
@@ -11,45 +12,33 @@ struct LLModelWrapper {
     ~LLModelWrapper() { delete llModel; }
 };
 
-
 thread_local static std::string last_error_message;
 
-
 llmodel_model llmodel_model_create(const char *model_path) {
-    auto fres = llmodel_model_create2(model_path, "auto", nullptr);
+    const char *error;
+    auto fres = llmodel_model_create2(model_path, "auto", &error);
     if (!fres) {
-        fprintf(stderr, "Invalid model file\n");
+        fprintf(stderr, "Unable to instantiate model: %s\n", error);
     }
     return fres;
 }
 
-llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, llmodel_error *error) {
+llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, const char **error) {
     auto wrapper = new LLModelWrapper;
-    int error_code = 0;
 
     try {
         wrapper->llModel = LLModel::Implementation::construct(model_path, build_variant);
+        if (!wrapper->llModel) {
+            last_error_message = "Model format not supported (no matching implementation found)";
+        }
     } catch (const std::exception& e) {
-        error_code = EINVAL;
         last_error_message = e.what();
     }
 
     if (!wrapper->llModel) {
         delete std::exchange(wrapper, nullptr);
-        // Get errno and error message if none
-        if (error_code == 0) {
-            if (errno != 0) {
-                error_code = errno;
-                last_error_message = std::strerror(error_code);
-            } else {
-                error_code = ENOTSUP;
-                last_error_message = "Model format not supported (no matching implementation found)";
-            }
-        }
-        // Set error argument
         if (error) {
-            error->message = last_error_message.c_str();
-            error->code = error_code;
+            *error = last_error_message.c_str();
         }
     }
     return reinterpret_cast<llmodel_model*>(wrapper);
@@ -59,16 +48,23 @@ void llmodel_model_destroy(llmodel_model model) {
     delete reinterpret_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);
+    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);
+
+    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)
@@ -112,10 +108,12 @@ bool recalculate_wrapper(bool is_recalculating, void *user_data) {
 }
 
 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)
 {
     LLModelWrapper *wrapper = reinterpret_cast<LLModelWrapper*>(model);
 
@@ -143,7 +141,7 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     wrapper->promptContext.contextErase = ctx->context_erase;
 
     // Call the C++ prompt method
-    wrapper->llModel->prompt(prompt, prompt_func, response_func, recalc_func, wrapper->promptContext);
+    wrapper->llModel->prompt(prompt, prompt_template, prompt_func, response_func, recalc_func, wrapper->promptContext, special);
 
     // Update the C context by giving access to the wrappers raw pointers to std::vector data
     // which involves no copies
@@ -242,14 +240,8 @@ bool llmodel_gpu_init_gpu_device_by_string(llmodel_model model, size_t memoryReq
 
 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);
+    return wrapper->llModel->initializeGPUDevice(device->index);
 }
 
 bool llmodel_gpu_init_gpu_device_by_int(llmodel_model model, int device)

gpt4all-backend/llmodel_c.h

@@ -23,17 +23,6 @@ extern "C" {
  */
 typedef void *llmodel_model;
 
-/**
- * Structure containing any errors that may eventually occur
- */
-struct llmodel_error {
-    const char *message;  // Human readable error description; Thread-local; guaranteed to survive until next llmodel C API call
-    int code;             // errno; 0 if none
-};
-#ifndef __cplusplus
-typedef struct llmodel_error llmodel_error;
-#endif
-
 /**
  * llmodel_prompt_context structure for holding the prompt context.
  * NOTE: The implementation takes care of all the memory handling of the raw logits pointer and the
@@ -105,10 +94,10 @@ DEPRECATED llmodel_model llmodel_model_create(const char *model_path);
  * Recognises correct model type from file at model_path
  * @param model_path A string representing the path to the model file; will only be used to detect model type.
  * @param build_variant A string representing the implementation to use (auto, default, avxonly, ...),
- * @param error A pointer to a llmodel_error; will only be set on error.
+ * @param error A pointer to a string; will only be set on error.
  * @return A pointer to the llmodel_model instance; NULL on error.
  */
-llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, llmodel_error *error);
+llmodel_model llmodel_model_create2(const char *model_path, const char *build_variant, const char **error);
 
 /**
  * Destroy a llmodel instance.
@@ -121,17 +110,21 @@ void llmodel_model_destroy(llmodel_model model);
  * Estimate RAM requirement for a model file
  * @param model A pointer to the llmodel_model instance.
  * @param model_path A string representing the path to the model file.
+ * @param n_ctx Maximum size of context window
+ * @param ngl Number of GPU layers to use (Vulkan)
  * @return size greater than 0 if the model was parsed successfully, 0 if file could not be parsed.
  */
-size_t llmodel_required_mem(llmodel_model model, const char *model_path);
+size_t llmodel_required_mem(llmodel_model model, const char *model_path, int n_ctx, int ngl);
 
 /**
  * Load a model from a file.
  * @param model A pointer to the llmodel_model instance.
  * @param model_path A string representing the path to the model file.
+ * @param n_ctx Maximum size of context window
+ * @param ngl Number of GPU layers to use (Vulkan)
  * @return true if the model was loaded successfully, false otherwise.
  */
-bool llmodel_loadModel(llmodel_model model, const char *model_path);
+bool llmodel_loadModel(llmodel_model model, const char *model_path, int n_ctx, int ngl);
 
 /**
  * Check if a model is loaded.
@@ -170,16 +163,20 @@ 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 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);
 
 /**
  * Generate an embedding using the model.

gpt4all-backend/llmodel_shared.cpp

@@ -2,11 +2,20 @@
 
 #include <cassert>
 #include <iostream>
+#include <regex>
 #include <unordered_set>
 
+// 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);
@@ -26,11 +35,36 @@ 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";
@@ -38,15 +72,86 @@ 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
+    if (placeholders.size() >= 2) {
+        // template: end of assistant prompt
+        size_t start = placeholders[1].position() + placeholders[1].length();
+        auto asstSuffix = promptTemplate.substr(start);
+        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();
 
@@ -69,11 +174,6 @@ 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);
         }
@@ -88,13 +188,17 @@ void LLModel::prompt(const std::string &prompt,
             if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
                 promptCtx.tokens.erase(promptCtx.tokens.begin());
             promptCtx.tokens.push_back(batch.at(t));
+            promptCtx.n_past += 1;
             if (!promptCallback(batch.at(t)))
                 return;
         }
-        promptCtx.n_past += batch.size();
         i = batch_end;
     }
+}
 
+void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
+                               std::function<bool(bool)> recalculateCallback,
+                               PromptContext &promptCtx) {
     std::string cachedResponse;
     std::vector<Token> cachedTokens;
     std::unordered_set<std::string> reversePrompts
@@ -108,11 +212,6 @@ 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);
         }
@@ -122,8 +221,6 @@ void LLModel::prompt(const std::string &prompt,
             return;
         }
 
-        promptCtx.n_past += 1;
-
         // display text
         for (const auto token : endTokens()) {
             if (id == token) return;
@@ -158,6 +255,7 @@ void LLModel::prompt(const std::string &prompt,
             if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
                 promptCtx.tokens.erase(promptCtx.tokens.begin());
             promptCtx.tokens.push_back(t);
+            promptCtx.n_past += 1;
             //TODO: Conversion to std::string can be avoided here...
             if (!responseCallback(t, std::string(tokenToString(t))))
                 return;
@@ -166,8 +264,9 @@ void LLModel::prompt(const std::string &prompt,
     }
 }
 
-std::vector<float> LLModel::embedding(const std::string &/*text*/)
+std::vector<float> LLModel::embedding(const std::string &text)
 {
+    (void)text;
     if (!supportsCompletion()) {
         std::string errorMessage = "ERROR: this model does not support generating embeddings!\n";
         std::cerr << implementation().modelType() << errorMessage;

gpt4all-backend/llmodel_shared.h

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

gpt4all-backend/mpt.cpp

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

gpt4all-backend/mpt_impl.h

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

gpt4all-backend/replit_impl.h

@@ -1,43 +0,0 @@
-#ifndef REPLIT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#error This file is NOT meant to be included outside of replit.cpp. Doing so is DANGEROUS. Be sure to know what you are doing before proceeding to #define REPLIT_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
-#endif
-#ifndef REPLIT_H
-#define REPLIT_H
-
-#include <string>
-#include <functional>
-#include <vector>
-#include "llmodel.h"
-
-#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
-
-struct ReplitPrivate;
-class Replit : public LLModel {
-public:
-    Replit();
-    ~Replit();
-
-    bool supportsEmbedding() const override { return false; }
-    bool supportsCompletion() const override { return true; }
-    bool loadModel(const std::string &modelPath) override;
-    bool isModelLoaded() const override;
-    size_t requiredMem(const std::string & modelPath) override;
-    size_t stateSize() const override;
-    size_t saveState(uint8_t *dest) const override;
-    size_t restoreState(const uint8_t *src) override;
-    void setThreadCount(int32_t n_threads) override;
-    int32_t threadCount() const override;
-
-private:
-    ReplitPrivate *d_ptr;
-
-protected:
-    std::vector<Token> tokenize(PromptContext &, const std::string&) const override;
-    std::string tokenToString(Token) const override;
-    Token sampleToken(PromptContext &ctx) const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    int32_t contextLength() const override;
-    const std::vector<Token>& endTokens() const override;
-};
-
-#endif // REPLIT_H

gpt4all-backend/scripts/convert_bert_hf_to_ggml.py

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

gpt4all-backend/scripts/convert_bert_hf_to_gguf.py

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

gpt4all-backend/scripts/convert_falcon_hf_to_ggml.py

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

gpt4all-backend/scripts/convert_gptj_to_gguf.py

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

gpt4all-backend/scripts/convert_mpt_hf_to_ggml.py

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

gpt4all-backend/scripts/convert_replit_hf_to_ggml.py

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

gpt4all-backend/starcoder_impl.h

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

gpt4all-bindings/cli/README.md

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

gpt4all-bindings/cli/app.py

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

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/NativeMethods.cs

@@ -70,7 +70,9 @@ 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/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/README.md

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

gpt4all-bindings/csharp/build_linux.sh

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

gpt4all-bindings/csharp/build_win-mingw.ps1

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

gpt4all-bindings/golang/Makefile

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

gpt4all-bindings/golang/README.md

@@ -36,7 +36,7 @@ func main() {
 In order to use the bindings you will need to build `libgpt4all.a`:
 
 ```
-git clone https://github.com/nomic-ai/gpt4all
+git clone --recurse-submodules https://github.com/nomic-ai/gpt4all
 cd gpt4all/gpt4all-bindings/golang
 make libgpt4all.a
 ```

gpt4all-bindings/golang/binding.cpp

@@ -17,14 +17,13 @@
 
 void* load_model(const char *fname, int n_threads) {
     // load the model
-    llmodel_error new_error{};
+    const char *new_error;
     auto model = llmodel_model_create2(fname, "auto", &new_error);
-    if (model == nullptr ){
-        fprintf(stderr, "%s: error '%s'\n",
-                __func__, new_error.message);
+    if (model == nullptr) {
+        fprintf(stderr, "%s: error '%s'\n", __func__, new_error);
         return nullptr;
     }
-    if (!llmodel_loadModel(model, fname)) {
+    if (!llmodel_loadModel(model, fname, 2048, 100)) {
         llmodel_model_destroy(model);
         return nullptr;
     }

gpt4all-bindings/java/README.md

@@ -22,7 +22,7 @@ implementation 'com.hexadevlabs:gpt4all-java-binding:1.1.5'
 
 To add the library dependency for another build system see [Maven Central Java bindings](https://central.sonatype.com/artifact/com.hexadevlabs/gpt4all-java-binding/).
 
-To download model binary weights file use a URL such as [`https://gpt4all.io/models/ggml-gpt4all-j-v1.3-groovy.bin`](https://gpt4all.io/models/ggml-gpt4all-j-v1.3-groovy.bin).
+To download model binary weights file use a URL such as [`https://gpt4all.io/models/gguf/gpt4all-13b-snoozy-q4_0.gguf`](https://gpt4all.io/models/gguf/gpt4all-13b-snoozy-q4_0.gguf).
 
 For information about other models available see the [model file list](https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-chat#manual-download-of-models).
 
@@ -123,4 +123,4 @@ If this is the case you can easily download and install the latest x64 Microsoft
    - Falcon model support included.
 4. Version **1.1.5**:
    - Add a check for model file readability before loading model.
-   
+   

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

@@ -1,6 +1,7 @@
 package com.hexadevlabs.gpt4all;
 
 import jnr.ffi.Pointer;
+import jnr.ffi.byref.PointerByReference;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -8,9 +9,8 @@ import java.io.ByteArrayOutputStream;
 import java.nio.charset.StandardCharsets;
 import java.nio.file.Files;
 import java.nio.file.Path;
-import java.util.HashMap;
-import java.util.List;
-import java.util.Map;
+import java.util.*;
+import java.util.stream.Collectors;
 
 public  class LLModel implements AutoCloseable {
 
@@ -177,7 +177,7 @@ public  class LLModel implements AutoCloseable {
         modelName = modelPath.getFileName().toString();
         String modelPathAbs = modelPath.toAbsolutePath().toString();
 
-        LLModelLibrary.LLModelError error = new LLModelLibrary.LLModelError(jnr.ffi.Runtime.getSystemRuntime());
+        PointerByReference error = new PointerByReference();
 
         // Check if model file exists
         if(!Files.exists(modelPath)){
@@ -193,9 +193,9 @@ public  class LLModel implements AutoCloseable {
         model = library.llmodel_model_create2(modelPathAbs, "auto", error);
 
         if(model == null) {
-            throw new IllegalStateException("Could not load, gpt4all backend returned error: " + error.message);
+            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");
@@ -306,6 +306,197 @@ public  class LLModel implements AutoCloseable {
         };
     }
 
+    /**
+     * The array of messages for the conversation.
+     */
+    public static class Messages {
+
+        private final List<PromptMessage> messages = new ArrayList<>();
+
+        public Messages(PromptMessage...messages) {
+            this.messages.addAll(Arrays.asList(messages));
+        }
+
+        public Messages(List<PromptMessage> messages) {
+            this.messages.addAll(messages);
+        }
+
+        public Messages addPromptMessage(PromptMessage promptMessage) {
+            this.messages.add(promptMessage);
+            return this;
+        }
+
+        List<PromptMessage> toList() {
+            return Collections.unmodifiableList(this.messages);
+        }
+
+        List<Map<String, String>> toListMap() {
+            return messages.stream()
+                    .map(PromptMessage::toMap).collect(Collectors.toList());
+        }
+
+    }
+
+    /**
+     * A message in the conversation, identical to OpenAI's chat message.
+     */
+    public static class PromptMessage {
+
+        private static final String ROLE = "role";
+        private static final String CONTENT = "content";
+
+        private final Map<String, String> message = new HashMap<>();
+
+        public PromptMessage() {
+        }
+
+        public PromptMessage(Role role, String content) {
+            addRole(role);
+            addContent(content);
+        }
+
+        public PromptMessage addRole(Role role) {
+            return this.addParameter(ROLE, role.type());
+        }
+
+        public PromptMessage addContent(String content) {
+            return this.addParameter(CONTENT, content);
+        }
+
+        public PromptMessage addParameter(String key, String value) {
+            this.message.put(key, value);
+            return this;
+        }
+
+        public String content() {
+            return this.parameter(CONTENT);
+        }
+
+        public Role role() {
+            String role = this.parameter(ROLE);
+            return Role.from(role);
+        }
+
+        public String parameter(String key) {
+            return this.message.get(key);
+        }
+
+        Map<String, String> toMap() {
+            return Collections.unmodifiableMap(this.message);
+        }
+
+    }
+
+    public enum Role {
+
+        SYSTEM("system"), ASSISTANT("assistant"), USER("user");
+
+        private final String type;
+
+        String type() {
+            return this.type;
+        }
+
+        static Role from(String type) {
+
+            if (type == null) {
+                return null;
+            }
+
+            switch (type) {
+                case "system": return SYSTEM;
+                case "assistant": return ASSISTANT;
+                case "user": return USER;
+                default: throw new IllegalArgumentException(
+                        String.format("You passed %s type but only %s are supported",
+                                type, Arrays.toString(Role.values())
+                        )
+                );
+            }
+        }
+
+        Role(String type) {
+            this.type = type;
+        }
+
+        @Override
+        public String toString() {
+            return type();
+        }
+    }
+
+    /**
+     * The result of the completion, similar to OpenAI's format.
+     */
+    public static class CompletionReturn {
+        private String model;
+        private Usage usage;
+        private Choices choices;
+
+        public CompletionReturn(String model, Usage usage, Choices choices) {
+            this.model = model;
+            this.usage = usage;
+            this.choices = choices;
+        }
+
+        public Choices choices() {
+            return choices;
+        }
+
+        public String model() {
+            return model;
+        }
+
+        public Usage usage() {
+            return usage;
+        }
+    }
+
+    /**
+     * The generated completions.
+     */
+    public static class Choices {
+
+        private final List<CompletionChoice> choices = new ArrayList<>();
+
+        public Choices(List<CompletionChoice> choices) {
+            this.choices.addAll(choices);
+        }
+
+        public Choices(CompletionChoice...completionChoices){
+            this.choices.addAll(Arrays.asList(completionChoices));
+        }
+
+        public Choices addCompletionChoice(CompletionChoice completionChoice) {
+            this.choices.add(completionChoice);
+            return this;
+        }
+
+        public CompletionChoice first() {
+            return this.choices.get(0);
+        }
+
+        public int totalChoices() {
+            return this.choices.size();
+        }
+
+        public CompletionChoice get(int index) {
+            return this.choices.get(index);
+        }
+
+        public List<CompletionChoice> choices() {
+            return Collections.unmodifiableList(choices);
+        }
+    }
+
+    /**
+     * A completion choice, similar to OpenAI's format.
+     */
+    public static class CompletionChoice extends PromptMessage {
+        public CompletionChoice(Role role, String content) {
+            super(role, content);
+        }
+    }
 
     public static class ChatCompletionResponse {
         public String model;
@@ -323,6 +514,41 @@ public  class LLModel implements AutoCloseable {
         // Getters and setters
     }
 
+    public CompletionReturn chatCompletionResponse(Messages messages,
+                                                                GenerationConfig generationConfig) {
+        return chatCompletion(messages, generationConfig, false, false);
+    }
+
+    /**
+     * chatCompletion formats the existing chat conversation into a template to be
+     * easier to process for chat UIs. It is not absolutely necessary as generate method
+     * may be directly used to make generations with gpt models.
+     *
+     * @param messages object to create theMessages to send to GPT model
+     * @param generationConfig How to decode/process the generation.
+     * @param streamToStdOut Send tokens as they are calculated Standard output.
+     * @param outputFullPromptToStdOut Should full prompt built out of messages be sent to Standard output.
+     * @return CompletionReturn contains stats and generated Text.
+     */
+    public CompletionReturn chatCompletion(Messages messages,
+                                                  GenerationConfig generationConfig, boolean streamToStdOut,
+                                                  boolean outputFullPromptToStdOut) {
+
+        String fullPrompt = buildPrompt(messages.toListMap());
+
+        if(outputFullPromptToStdOut)
+            System.out.print(fullPrompt);
+
+        String generatedText = generate(fullPrompt, generationConfig, streamToStdOut);
+
+        final CompletionChoice promptMessage = new CompletionChoice(Role.ASSISTANT, generatedText);
+        final Choices choices = new Choices(promptMessage);
+
+        final Usage usage = getUsage(fullPrompt, generatedText);
+        return new CompletionReturn(this.modelName, usage, choices);
+
+    }
+
     public ChatCompletionResponse chatCompletion(List<Map<String, String>> messages,
                                                               GenerationConfig generationConfig) {
         return chatCompletion(messages, generationConfig, false, false);
@@ -352,19 +578,23 @@ public  class LLModel implements AutoCloseable {
         ChatCompletionResponse  response = new ChatCompletionResponse();
         response.model = this.modelName;
 
-        Usage usage = new Usage();
-        usage.promptTokens = fullPrompt.length();
-        usage.completionTokens = generatedText.length();
-        usage.totalTokens = fullPrompt.length() + generatedText.length();
-        response.usage =  usage;
+        response.usage = getUsage(fullPrompt, generatedText);
 
         Map<String, String> message = new HashMap<>();
         message.put("role", "assistant");
         message.put("content", generatedText);
 
         response.choices = List.of(message);
-
         return response;
+
+    }
+
+    private Usage getUsage(String fullPrompt, String generatedText) {
+        Usage usage = new Usage();
+        usage.promptTokens = fullPrompt.length();
+        usage.completionTokens = generatedText.length();
+        usage.totalTokens = fullPrompt.length() + generatedText.length();
+        return usage;
     }
 
     protected static String buildPrompt(List<Map<String, String>> messages) {
@@ -402,4 +632,4 @@ public  class LLModel implements AutoCloseable {
         library.llmodel_model_destroy(model);
     }
 
-}
+}

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

@@ -1,6 +1,7 @@
 package com.hexadevlabs.gpt4all;
 
 import jnr.ffi.Pointer;
+import jnr.ffi.byref.PointerByReference;
 import jnr.ffi.Struct;
 import jnr.ffi.annotations.Delegate;
 import jnr.ffi.annotations.Encoding;
@@ -58,9 +59,9 @@ public interface LLModelLibrary {
         }
     }
 
-    Pointer llmodel_model_create2(String model_path, String build_variant, @Out LLModelError llmodel_error);
+    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/java/src/test/java/com/hexadevlabs/gpt4all/BasicTests.java

@@ -28,6 +28,33 @@ import static org.mockito.Mockito.*;
 @ExtendWith(MockitoExtension.class)
 public class BasicTests {
 
+    @Test
+    public void simplePromptWithObject(){
+
+        LLModel model = Mockito.spy(new LLModel());
+
+        LLModel.GenerationConfig config =
+                LLModel.config()
+                        .withNPredict(20)
+                        .build();
+
+        // The generate method will return "4"
+        doReturn("4").when( model ).generate(anyString(), eq(config), eq(true));
+
+        LLModel.PromptMessage promptMessage1 = new LLModel.PromptMessage(LLModel.Role.SYSTEM, "You are a helpful assistant");
+        LLModel.PromptMessage promptMessage2 = new LLModel.PromptMessage(LLModel.Role.USER, "Add 2+2");
+
+        LLModel.Messages messages = new LLModel.Messages(promptMessage1, promptMessage2);
+
+        LLModel.CompletionReturn response = model.chatCompletion(
+                messages, config, true, true);
+
+        assertTrue( response.choices().first().content().contains("4") );
+
+        // Verifies the prompt and response are certain length.
+        assertEquals( 224 , response.usage().totalTokens );
+    }
+
     @Test
     public void simplePrompt(){
 

gpt4all-bindings/python/README.md

@@ -9,31 +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.
 
-**NOTE**: If you are doing this on a Windows machine, you must build the GPT4All backend using [MinGW64](https://www.mingw-w64.org/) compiler.
+## Local Build
 
-1. Setup `llmodel`
+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
+
+macOS users do not need Vulkan, as GPT4All will use Metal instead.
+
+### Building the python bindings
+
+1. Clone GPT4All and change directory:
 ```
-git clone --recurse-submodules git@github.com:nomic-ai/gpt4all.git
-cd gpt4all/gpt4all-backend/
-mkdir build
-cd build
-cmake ..
-cmake --build . --parallel  # optionally append: --config Release
+git clone --recurse-submodules https://github.com/nomic-ai/gpt4all.git
+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
@@ -42,7 +63,7 @@ Test it out! In a Python script or console:
 
 ```python
 from gpt4all import GPT4All
-model = GPT4All("orca-mini-3b.ggmlv3.q4_0.bin")
+model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
 output = model.generate("The capital of France is ", max_tokens=3)
 print(output)
 ```
@@ -51,7 +72,7 @@ print(output)
 GPU Usage
 ```python
 from gpt4all import GPT4All
-model = GPT4All("orca-mini-3b.ggmlv3.q4_0.bin", device='gpu') # device='amd', device='intel'
+model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf", device='gpu') # device='amd', device='intel'
 output = model.generate("The capital of France is ", max_tokens=3)
 print(output)
 ```

gpt4all-bindings/python/docs/gpt4all_chat.md

@@ -5,48 +5,46 @@ The [GPT4All Chat Client](https://gpt4all.io) lets you easily interact with any
 It is optimized to run 7-13B parameter LLMs on the CPU's of any computer running OSX/Windows/Linux.
 
 ## Running LLMs on CPU
-The GPT4All Chat UI supports models from all newer versions of `GGML`, `llama.cpp` including the `LLaMA`, `MPT`, `replit`,  `GPT-J` and `falcon` architectures
+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 [models.json](https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models.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 [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.
 
-#### Sideloading any GGML model
+#### Sideloading any GGUF model
 If a model is compatible with the gpt4all-backend, you can sideload it into GPT4All Chat by:
 
-1. Downloading your model in GGML format. It should be a 3-8 GB file similar to the ones [here](https://huggingface.co/TheBloke/Samantha-7B-GGML/tree/main).
-2. Identifying your GPT4All model downloads folder. This is the path listed at the bottom of the downloads dialog(Three lines in top left>Downloads).
-3. Placing your downloaded model inside the GPT4All's model downloads folder.
+1. Downloading your model in GGUF format. It should be a 3-8 GB file similar to the ones [here](https://huggingface.co/TheBloke/Orca-2-7B-GGUF/tree/main).
+2. Identifying your GPT4All model downloads folder. This is the path listed at the bottom of the downloads dialog.
+3. Placing your downloaded model inside GPT4All's model downloads folder.
 4. Restarting your GPT4ALL app. Your model should appear in the model selection list.
 
 ## Plugins
 GPT4All Chat Plugins allow you to expand the capabilities of Local LLMs.
 
-### LocalDocs Beta Plugin (Chat With Your Data)
-LocalDocs is a GPT4All plugin that allows you to chat with your local files and data.
+### LocalDocs Plugin (Chat With Your Data)
+LocalDocs is a GPT4All feature that allows you to chat with your local files and data.
 It allows you to utilize powerful local LLMs to chat with private data without any data leaving your computer or server.
-When using LocalDocs, your LLM will cite the sources that most likely contributed to a given output. Note, even an LLM equipped with LocalDocs can hallucinate. If the LocalDocs plugin decides to utilize your documents to help answer a prompt, you will see references appear below the response.
+When using LocalDocs, your LLM will cite the sources that most likely contributed to a given output. Note, even an LLM equipped with LocalDocs can hallucinate. The LocalDocs plugin will utilize your documents to help answer prompts and you will see references appear below the response.
 
 <p align="center">
-  <img width="70%" src="https://github.com/nomic-ai/gpt4all/assets/13879686/f70f40b4-9684-46d8-b388-ca186f63d13e">
-</p>
-<p align="center">
-GPT4All-Snoozy with LocalDocs. Try GPT4All-Groovy for a faster experience!
+  <img width="70%" src="https://github.com/nomic-ai/gpt4all/assets/10168/fe5dd3c0-b3cc-4701-98d3-0280dfbcf26f">
 </p>
 
 #### Enabling LocalDocs
 1. Install the latest version of GPT4All Chat from [GPT4All Website](https://gpt4all.io).
 2. Go to `Settings > LocalDocs tab`.
-3. Configure a collection (folder) on your computer that contains the files your LLM should have access to. You can alter the contents of the folder/directory at anytime. As you
+3. Download the SBert model
+4. Configure a collection (folder) on your computer that contains the files your LLM should have access to. You can alter the contents of the folder/directory at anytime. As you
 add more files to your collection, your LLM will dynamically be able to access them.
-4. Spin up a chat session with any LLM (including external ones like ChatGPT but warning data will leave your machine!)
-5. At the top right, click the database icon and select which collection you want your LLM to know about during your chat session.
+5. Spin up a chat session with any LLM (including external ones like ChatGPT but warning data will leave your machine!)
+6. At the top right, click the database icon and select which collection you want your LLM to know about during your chat session.
+7. You can begin searching with your localdocs even before the collection has completed indexing, but note the search will not include those parts of the collection yet to be indexed.
 
 #### LocalDocs Capabilities
-LocalDocs allows your LLM to have context about the contents of your documentation collection. Not all prompts/question will utilize your document
-collection for context. If LocalDocs was used in your LLMs response, you will see references to the document snippets that LocalDocs used.
+LocalDocs allows your LLM to have context about the contents of your documentation collection.
 
 LocalDocs **can**:
 
-- Query your documents based upon your prompt / question. If your documents contain answers that may help answer your question/prompt LocalDocs will try to utilize snippets of your documents to provide context.
+- Query your documents based upon your prompt / question. Your documents will be searched for snippets that can be used to provide context for an answer. The most relevant snippets will be inserted into your prompts context, but it will be up to the underlying model to decide how best to use the provided context.
 
 LocalDocs **cannot**:
 
@@ -62,32 +60,17 @@ 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.
-For indexing speed purposes, LocalDocs uses pre-deep-learning n-gram and TF-IDF based retrieval when deciding
-what document chunks your LLM should use as context. You'll find its of comparable quality
-with embedding based retrieval approaches but magnitudes faster to ingest data.
 
-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*
 
 - Make sure LocalDocs is enabled for your chat session (the DB icon on the top-right should have a border)
-- Try to modify your prompt to be more specific and use terminology that is in your document. This will increase the likelihood that LocalDocs matches document snippets for your question.
 - If your document collection is large, wait 1-2 minutes for it to finish indexing.
 
 
 #### LocalDocs Roadmap
-- Embedding based semantic search for retrieval. 
 - Customize model fine-tuned with retrieval in the loop.
 - Plugin compatibility with chat client server mode.
 

gpt4all-bindings/python/docs/gpt4all_cli.md

@@ -166,7 +166,7 @@ If you want to use a different model, you can do so with the `-m`/`--model` para
 model file name is provided, it will again check in `.cache/gpt4all/` and might start downloading.
 If instead given a path to an existing model, the command could for example look like this:
 ```shell
-python app.py repl --model /home/user/my-gpt4all-models/GPT4All-13B-snoozy.ggmlv3.q4_0.bin
+python app.py repl --model /home/user/my-gpt4all-models/gpt4all-13b-snoozy-q4_0.gguf
 ```
 
 When you're done and want to end a session, simply type `/exit`.

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 [models.json].
+checksum by comparing it with the one listed in [models2.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.
 
-[models.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models.json
+[models2.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models2.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 [models.json]. If yes,
+        - When using a chat session, it depends on whether the bindings are allowed to download [models2.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

@@ -8,7 +8,7 @@ import modal
 
 def download_model():
     import gpt4all
-    #you can use any model from https://gpt4all.io/models/models.json
+    #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)
@@ -31,4 +31,4 @@ def main():
     model = GPT4All()
     for i in range(10):
         model.generate.call()
-```
+```

gpt4all-bindings/python/docs/gpt4all_nodejs.md

@@ -1,11 +1,14 @@
 # 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.
@@ -15,12 +18,12 @@ The original [GPT4All typescript bindings](https://github.com/nomic-ai/gpt4all-t
 *   Everything should work out the box.
 *   See [API Reference](#api-reference)
 
-### Chat Completion (alpha)
+### Chat Completion
 
 ```js
 import { createCompletion, 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.Q4_0.gguf', { verbose: true });
 
 const response = await createCompletion(model, [
     { role : 'system', content: 'You are meant to be annoying and unhelpful.'  },
@@ -29,7 +32,7 @@ const response = await createCompletion(model, [
 
 ```
 
-### Embedding (alpha)
+### Embedding
 
 ```js
 import { createEmbedding, loadModel } from '../src/gpt4all.js'
@@ -58,6 +61,8 @@ const fltArray = createEmbedding(model, "Pain is inevitable, suffering optional"
 *   (win) msvc version 143
     *   Can be obtained with visual studio 2022 build tools
 *   python 3
+*   On Windows and Linux, building GPT4All requires the complete Vulkan SDK. You may download it from here: https://vulkan.lunarg.com/sdk/home
+*   macOS users do not need Vulkan, as GPT4All will use Metal instead.
 
 ### Build (from source)
 
@@ -80,8 +85,6 @@ yarn
 git submodule update --init --depth 1 --recursive
 ```
 
-**AS OF NEW BACKEND** to build the backend,
-
 ```sh
 yarn build:backend
 ```
@@ -150,13 +153,16 @@ 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)
+*   [InferenceModel](#inferencemodel)
+    *   [dispose](#dispose)
+*   [EmbeddingModel](#embeddingmodel)
+    *   [dispose](#dispose-1)
 *   [LLModel](#llmodel)
     *   [constructor](#constructor)
         *   [Parameters](#parameters)
@@ -174,12 +180,20 @@ This package is in active development, and breaking changes may happen until the
     *   [setLibraryPath](#setlibrarypath)
         *   [Parameters](#parameters-4)
     *   [getLibraryPath](#getlibrarypath)
+    *   [initGpuByString](#initgpubystring)
+        *   [Parameters](#parameters-5)
+    *   [hasGpuDevice](#hasgpudevice)
+    *   [listGpu](#listgpu)
+    *   [dispose](#dispose-2)
+*   [GpuDevice](#gpudevice)
+    *   [type](#type-2)
+*   [LoadModelOptions](#loadmodeloptions)
 *   [loadModel](#loadmodel)
-    *   [Parameters](#parameters-5)
-*   [createCompletion](#createcompletion)
     *   [Parameters](#parameters-6)
-*   [createEmbedding](#createembedding)
+*   [createCompletion](#createcompletion)
     *   [Parameters](#parameters-7)
+*   [createEmbedding](#createembedding)
+    *   [Parameters](#parameters-8)
 *   [CompletionOptions](#completionoptions)
     *   [verbose](#verbose)
     *   [systemPromptTemplate](#systemprompttemplate)
@@ -212,14 +226,14 @@ This package is in active development, and breaking changes may happen until the
     *   [repeatLastN](#repeatlastn)
     *   [contextErase](#contexterase)
 *   [createTokenStream](#createtokenstream)
-    *   [Parameters](#parameters-8)
+    *   [Parameters](#parameters-9)
 *   [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-10)
     *   [Examples](#examples)
 *   [DownloadModelOptions](#downloadmodeloptions)
     *   [modelPath](#modelpath)
@@ -230,16 +244,10 @@ This package is in active development, and breaking changes may happen until the
     *   [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
+DEPRECATED!! These model names are outdated and this type will not be maintained, please use a string literal instead
 
 ##### gptj
 
@@ -269,7 +277,27 @@ Type: `"ggml-replit-code-v1-3b.bin"`
 
 Model architecture. This argument currently does not have any functionality and is just used as descriptive identifier for user.
 
-Type: [ModelType](#modeltype)
+Type: ModelType
+
+#### InferenceModel
+
+InferenceModel represents an LLM which can make chat predictions, similar to GPT transformers.
+
+##### dispose
+
+delete and cleanup the native model
+
+Returns **void** 
+
+#### EmbeddingModel
+
+EmbeddingModel represents an LLM which can create embeddings, which are float arrays
+
+##### dispose
+
+delete and cleanup the native model
+
+Returns **void** 
 
 #### LLModel
 
@@ -292,7 +320,7 @@ Initialize a new LLModel.
 
 either 'gpt', mpt', or 'llama' or undefined
 
-Returns **([ModelType](#modeltype) | [undefined](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/undefined))** 
+Returns **(ModelType | [undefined](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/undefined))** 
 
 ##### name
 
@@ -374,6 +402,52 @@ Where to get the pluggable backend libraries
 
 Returns **[string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)** 
 
+##### 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)** 
+
+##### 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
+
+*   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.
+
 #### loadModel
 
 Loads a machine learning model with the specified name. The defacto way to create a model.
@@ -382,9 +456,9 @@ 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.
 
 #### createCompletion
 
@@ -392,7 +466,7 @@ The nodejs equivalent to python binding's chat\_completion
 
 ##### Parameters
 
-*   `model` **InferenceModel** The language model object.
+*   `model` **[InferenceModel](#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.
 *   `options` **[CompletionOptions](#completionoptions)** The options for creating the completion.
 
@@ -405,7 +479,7 @@ 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.
@@ -650,7 +724,7 @@ Default model configuration.
 
 Type: ModelConfig
 
-#### DEFAULT\_PROMT\_CONTEXT
+#### DEFAULT\_PROMPT\_CONTEXT
 
 Default prompt context.
 
@@ -703,7 +777,7 @@ Type: [boolean](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Glob
 
 ##### url
 
-Remote download url. Defaults to `https://gpt4all.io/models/<modelName>`
+Remote download url. Defaults to `https://gpt4all.io/models/gguf/<modelName>`
 
 Type: [string](https://developer.mozilla.org/docs/Web/JavaScript/Reference/Global_Objects/String)
 

gpt4all-bindings/python/docs/gpt4all_python.md

@@ -11,7 +11,7 @@ pip install gpt4all
 === "GPT4All Example"
     ``` py
     from gpt4all import GPT4All
-    model = GPT4All("orca-mini-3b.ggmlv3.q4_0.bin")
+    model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
     output = model.generate("The capital of France is ", max_tokens=3)
     print(output)
     ```
@@ -35,7 +35,7 @@ Use the GPT4All `chat_session` context manager to hold chat conversations with t
 
 === "GPT4All Example"
     ``` py
-    model = GPT4All(model_name='orca-mini-3b.ggmlv3.q4_0.bin')
+    model = GPT4All(model_name='orca-mini-3b-gguf2-q4_0.gguf')
     with model.chat_session():
         response1 = model.generate(prompt='hello', temp=0)
         response2 = model.generate(prompt='write me a short poem', temp=0)
@@ -77,10 +77,10 @@ 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 [models.json] from the repository, which contains specifically tailored templates
+  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.
 
-[models.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models.json
+[models2.json]: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/metadata/models2.json
 
 
 ### Streaming Generations
@@ -89,7 +89,7 @@ To interact with GPT4All responses as the model generates, use the `streaming=Tr
 === "GPT4All Streaming Example"
     ``` py
     from gpt4all import GPT4All
-    model = GPT4All("orca-mini-3b.ggmlv3.q4_0.bin")
+    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)
@@ -135,7 +135,7 @@ is the same as if it weren't provided; that is, `~/.cache/gpt4all/` is the defau
     ``` py
     from pathlib import Path
     from gpt4all import GPT4All
-    model = GPT4All(model_name='orca-mini-3b.ggmlv3.q4_0.bin',
+    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)
@@ -152,7 +152,7 @@ If you want to point it at the chat GUI's default folder, it should be:
     from pathlib import Path
     from gpt4all import GPT4All
 
-    model_name = 'orca-mini-3b.ggmlv3.q4_0.bin'
+    model_name = 'orca-mini-3b-gguf2-q4_0.gguf'
     model_path = Path.home() / 'Library' / 'Application Support' / 'nomic.ai' / 'GPT4All'
     model = GPT4All(model_name, model_path)
     ```
@@ -161,7 +161,7 @@ If you want to point it at the chat GUI's default folder, it should be:
     from pathlib import Path
     from gpt4all import GPT4All
     import os
-    model_name = 'orca-mini-3b.ggmlv3.q4_0.bin'
+    model_name = 'orca-mini-3b-gguf2-q4_0.gguf'
     model_path = Path(os.environ['LOCALAPPDATA']) / 'nomic.ai' / 'GPT4All'
     model = GPT4All(model_name, model_path)
     ```
@@ -170,7 +170,7 @@ If you want to point it at the chat GUI's default folder, it should be:
     from pathlib import Path
     from gpt4all import GPT4All
 
-    model_name = 'orca-mini-3b.ggmlv3.q4_0.bin'
+    model_name = 'orca-mini-3b-gguf2-q4_0.gguf'
     model_path = Path.home() / '.local' / 'share' / 'nomic.ai' / 'GPT4All'
     model = GPT4All(model_name, model_path)
     ```
@@ -182,7 +182,7 @@ from pathlib import Path
 import gpt4all.gpt4all
 gpt4all.gpt4all.DEFAULT_MODEL_DIRECTORY = Path.home() / 'my' / 'models-directory'
 from gpt4all import GPT4All
-model = GPT4All('orca-mini-3b.ggmlv3.q4_0.bin')
+model = GPT4All('orca-mini-3b-gguf2-q4_0.gguf')
 ...
 ```
 
@@ -193,7 +193,7 @@ Session templates can be customized when starting a `chat_session` context:
 === "GPT4All Custom Session Templates Example"
     ``` py
     from gpt4all import GPT4All
-    model = GPT4All('ggml-Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin')
+    model = GPT4All('wizardlm-13b-v1.2.Q4_0.gguf')
     system_template = 'A chat between a curious user and an artificial intelligence assistant.'
     # many models use triple hash '###' for keywords, Vicunas are simpler:
     prompt_template = 'USER: {0}\nASSISTANT: '
@@ -222,7 +222,7 @@ To do the same outside a session, the input has to be formatted manually. For ex
 
 === "GPT4All Templates Outside a Session Example"
     ``` py
-    model = GPT4All('ggml-Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin')
+    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?']
@@ -246,90 +246,6 @@ To do the same outside a session, the input has to be formatted manually. For ex
     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('ggml-Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin')
-    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
@@ -345,7 +261,7 @@ logging infrastructure offers [many more customization options][py-logging-cookb
     import logging
     from gpt4all import GPT4All
     logging.basicConfig(level=logging.INFO)
-    model = GPT4All('nous-hermes-13b.ggmlv3.q4_0.bin')
+    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)
@@ -379,7 +295,7 @@ logging infrastructure offers [many more customization options][py-logging-cookb
 
 ### Without Online Connectivity
 To prevent GPT4All from accessing online resources, instantiate it with `allow_download=False`. This will disable both
-downloading missing models and [models.json], which contains information about them. As a result, predefined templates
+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:
 
 === "GPT4All Default Templates Example"
@@ -414,7 +330,7 @@ If you know exactly when a model should stop responding, you can add a custom ca
 === "GPT4All Custom Stop Callback"
     ``` py
     from gpt4all import GPT4All
-    model = GPT4All('orca-mini-3b.ggmlv3.q4_0.bin')
+    model = GPT4All('orca-mini-3b-gguf2-q4_0.gguf')
 
     def stop_on_token_callback(token_id, token_string):
         # one sentence is enough:

gpt4all-bindings/python/docs/index.md

@@ -9,7 +9,7 @@ GPT4All software is optimized to run inference of 3-13 billion parameter large l
 === "GPT4All Example"
     ``` py
     from gpt4all import GPT4All
-    model = GPT4All("orca-mini-3b.ggmlv3.q4_0.bin")
+    model = GPT4All("orca-mini-3b-gguf2-q4_0.gguf")
     output = model.generate("The capital of France is ", max_tokens=3)
     print(output)
     ```
@@ -38,7 +38,7 @@ 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/models.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/models2.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 Embed4All as Embed4All, GPT4All as GPT4All

gpt4all-bindings/python/gpt4all/_pyllmodel.py

@@ -1,57 +1,43 @@
+from __future__ import annotations
+
 import ctypes
+import importlib.resources
 import logging
 import os
 import platform
-from queue import Queue
 import re
 import subprocess
 import sys
 import threading
+from enum import Enum
+from queue import Queue
 from typing import Callable, Iterable, List
 
-import pkg_resources
-
 logger: logging.Logger = logging.getLogger(__name__)
 
 
 # TODO: provide a config file to make this more robust
-LLMODEL_PATH = os.path.join("llmodel_DO_NOT_MODIFY", "build").replace("\\", "\\\\")
-MODEL_LIB_PATH = str(pkg_resources.resource_filename("gpt4all", LLMODEL_PATH)).replace("\\", "\\\\")
+MODEL_LIB_PATH = importlib.resources.files("gpt4all") / "llmodel_DO_NOT_MODIFY" / "build"
 
 
 def load_llmodel_library():
-    system = platform.system()
+    ext = {"Darwin": "dylib", "Linux": "so", "Windows": "dll"}[platform.system()]
 
-    def get_c_shared_lib_extension():
-        if system == "Darwin":
-            return "dylib"
-        elif system == "Linux":
-            return "so"
-        elif system == "Windows":
-            return "dll"
-        else:
-            raise Exception("Operating System not supported")
+    try:
+        # Linux, Windows, MinGW
+        lib = ctypes.CDLL(str(MODEL_LIB_PATH / f"libllmodel.{ext}"))
+    except FileNotFoundError:
+        if ext != 'dll':
+            raise
+        # MSVC
+        lib = ctypes.CDLL(str(MODEL_LIB_PATH / "llmodel.dll"))
 
-    c_lib_ext = get_c_shared_lib_extension()
-
-    llmodel_file = "libllmodel" + "." + c_lib_ext
-
-    llmodel_dir = str(pkg_resources.resource_filename("gpt4all", os.path.join(LLMODEL_PATH, llmodel_file))).replace(
-        "\\", "\\\\"
-    )
-
-    llmodel_lib = ctypes.CDLL(llmodel_dir)
-
-    return llmodel_lib
+    return lib
 
 
 llmodel = load_llmodel_library()
 
 
-class LLModelError(ctypes.Structure):
-    _fields_ = [("message", ctypes.c_char_p), ("code", ctypes.c_int32)]
-
-
 class LLModelPromptContext(ctypes.Structure):
     _fields_ = [
         ("logits", ctypes.POINTER(ctypes.c_float)),
@@ -83,15 +69,15 @@ class LLModelGPUDevice(ctypes.Structure):
 llmodel.llmodel_model_create.argtypes = [ctypes.c_char_p]
 llmodel.llmodel_model_create.restype = ctypes.c_void_p
 
-llmodel.llmodel_model_create2.argtypes = [ctypes.c_char_p, ctypes.c_char_p, ctypes.POINTER(LLModelError)]
+llmodel.llmodel_model_create2.argtypes = [ctypes.c_char_p, ctypes.c_char_p, ctypes.POINTER(ctypes.c_char_p)]
 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
@@ -103,10 +89,12 @@ RecalculateCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_bool)
 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,
 ]
 
 llmodel.llmodel_prompt.restype = None
@@ -131,7 +119,7 @@ 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(MODEL_LIB_PATH.encode("utf-8"))
+llmodel.llmodel_set_implementation_search_path(str(MODEL_LIB_PATH).replace("\\", r"\\").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.restype = ctypes.POINTER(LLModelGPUDevice)
@@ -156,123 +144,87 @@ def empty_response_callback(token_id: int, response: str) -> bool:
     return True
 
 
+# Symbol to terminate from generator
+class Sentinel(Enum):
+    TERMINATING_SYMBOL = 0
+
+
 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)
     """
 
-    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):
+        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
 
+        # Construct a model implementation
+        err = ctypes.c_char_p()
+        model = llmodel.llmodel_model_create2(self.model_path, b"auto", ctypes.byref(err))
+        if model is None:
+            s = err.value
+            raise ValueError(f"Unable to instantiate model: {'null' if s is None else s.decode()}")
+        self.model = model
+
     def __del__(self):
-        if self.model is not None:
-            self.llmodel_lib.llmodel_model_destroy(self.model)
+        if hasattr(self, 'model'):
+            llmodel.llmodel_model_destroy(self.model)
 
-    def memory_needed(self, model_path: str) -> int:
-        model_path_enc = model_path.encode("utf-8")
-        self.model = llmodel.llmodel_model_create(model_path_enc)
-
-        if self.model is not None:
-            return llmodel.llmodel_required_mem(self.model, model_path_enc)
-        else:
-            raise ValueError("Unable to instantiate model")
-
-    def list_gpu(self, model_path: str) -> list:
-        """
-        Lists available GPU devices that satisfy the model's memory requirements.
-
-        Parameters
-        ----------
-        model_path : str
-            Path to the model.
-
-        Returns
-        -------
-        list
-            A list of LLModelGPUDevice structures representing 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)
+    def _list_gpu(self, mem_required: int) -> list[LLModelGPUDevice]:
         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(self.model, 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 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):
+        mem_required = llmodel.llmodel_required_mem(self.model, self.model_path, self.n_ctx, self.ngl)
 
-            # 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)]
+        if llmodel.llmodel_gpu_init_gpu_device_by_string(self.model, mem_required, device.encode()):
+            return
 
-            # Identify GPUs that are unavailable due to insufficient memory or features
-            unavailable_gpus = set(all_gpus) - set(available_gpus)
+        # Retrieve all GPUs without considering memory requirements.
+        num_devices = ctypes.c_int32(0)
+        all_devices_ptr = llmodel.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 = [d.name.decode() for d in all_devices_ptr[:num_devices.value]]
 
-            # 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)
+        # Retrieve GPUs that meet the memory requirements using list_gpu
+        available_gpus = [device.name.decode() for device in self._list_gpu(mem_required)]
 
-    def load_model(self, model_path: str) -> bool:
+        # Identify GPUs that are unavailable due to insufficient memory or features
+        unavailable_gpus = set(all_gpus).difference(available_gpus)
+
+        # 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)
+
+    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 = llmodel.llmodel_model_create(model_path_enc)
-
-        if self.model is not None:
-            llmodel.llmodel_loadModel(self.model, model_path_enc)
-        else:
-            raise ValueError("Unable to instantiate model")
-
-        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 not llmodel.llmodel_isModelLoaded(self.model):
@@ -297,7 +249,7 @@ 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,
@@ -311,8 +263,11 @@ class LLModel:
                 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
@@ -328,7 +283,7 @@ class LLModel:
             raise ValueError("Text must not be None or empty")
 
         embedding_size = ctypes.c_size_t()
-        c_text = ctypes.c_char_p(text.encode('utf-8'))
+        c_text = ctypes.c_char_p(text.encode())
         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)]
         llmodel.llmodel_free_embedding(embedding_ptr)
@@ -337,6 +292,7 @@ class LLModel:
     def prompt_model(
         self,
         prompt: str,
+        prompt_template: str,
         callback: ResponseCallbackType,
         n_predict: int = 4096,
         top_k: int = 40,
@@ -347,6 +303,7 @@ class LLModel:
         repeat_last_n: int = 10,
         context_erase: float = 0.75,
         reset_context: bool = False,
+        special: bool = False,
     ):
         """
         Generate response from model from a prompt.
@@ -373,9 +330,6 @@ class LLModel:
             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,
@@ -390,21 +344,20 @@ 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,
         )
 
 
     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()
-
-        output_queue: Queue = Queue()
+        output_queue: Queue[str | Sentinel] = Queue()
 
         # Put response tokens into an output queue
         def _generator_callback_wrapper(callback: ResponseCallbackType) -> ResponseCallbackType:
@@ -419,15 +372,15 @@ 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()
@@ -435,7 +388,7 @@ class LLModel:
         # Generator
         while True:
             response = output_queue.get()
-            if response is TERMINATING_SYMBOL:
+            if isinstance(response, Sentinel):
                 break
             yield response
 
@@ -458,7 +411,7 @@ 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()
 
@@ -467,7 +420,7 @@ 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

@@ -1,16 +1,23 @@
 """
 Python only API for running all GPT4All models.
 """
+from __future__ import annotations
+
 import os
+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
 
 import requests
+from requests.exceptions import ChunkedEncodingError
 from tqdm import tqdm
+from urllib3.exceptions import IncompleteRead, ProtocolError
 
-from . import pyllmodel
+from . import _pyllmodel
 
 # TODO: move to config
 DEFAULT_MODEL_DIRECTORY = os.path.join(str(Path.home()), ".cache", "gpt4all").replace("\\", "\\\\")
@@ -29,17 +36,14 @@ class Embed4All:
     Python class that handles embeddings for GPT4All.
     """
 
-    def __init__(
-        self,
-        n_threads: Optional[int] = None,
-    ):
+    def __init__(self, model_name: Optional[str] = None, n_threads: Optional[int] = None, **kwargs):
         """
         Constructor
 
         Args:
             n_threads: number of CPU threads used by GPT4All. Default is None, then the number of threads are determined automatically.
         """
-        self.gpt4all = GPT4All(model_name='ggml-all-MiniLM-L6-v2-f16.bin', n_threads=n_threads)
+        self.gpt4all = GPT4All(model_name or 'all-MiniLM-L6-v2-f16.gguf', n_threads=n_threads, **kwargs)
 
     def embed(self, text: str) -> List[float]:
         """
@@ -62,17 +66,20 @@ class GPT4All:
     def __init__(
         self,
         model_name: str,
-        model_path: Optional[str] = None,
+        model_path: Optional[Union[str, os.PathLike[str]]] = None,
         model_type: Optional[str] = None,
         allow_download: bool = True,
         n_threads: Optional[int] = None,
         device: Optional[str] = "cpu",
+        n_ctx: int = 2048,
+        ngl: int = 100,
+        verbose: bool = False,
     ):
         """
         Constructor
 
         Args:
-            model_name: Name of GPT4All or custom model. Including ".bin" file extension is optional but encouraged.
+            model_name: Name of GPT4All or custom model. Including ".gguf" file extension is optional but encouraged.
             model_path: Path to directory containing model file or, if file does not exist, where to download model.
                 Default is None, in which case models will be stored in `~/.cache/gpt4all/`.
             model_type: Model architecture. This argument currently does not have any functionality and is just used as
@@ -87,15 +94,17 @@ class GPT4All:
                 Default is "cpu".
 
                 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()
         # Retrieve model and download if allowed
-        self.config: ConfigType = self.retrieve_model(model_name, model_path=model_path, allow_download=allow_download)
-        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.config: ConfigType = self.retrieve_model(model_name, model_path=model_path, allow_download=allow_download, verbose=verbose)
+        self.model = _pyllmodel.LLModel(self.config["path"], n_ctx, ngl)
+        if device is not None and device != "cpu":
+            self.model.init_gpu(device)
+        self.model.load_model()
         # Set n_threads
         if n_threads is not None:
             self.model.set_thread_count(n_threads)
@@ -107,19 +116,22 @@ class GPT4All:
     @staticmethod
     def list_models() -> List[ConfigType]:
         """
-        Fetch model list from https://gpt4all.io/models/models.json.
+        Fetch model list from https://gpt4all.io/models/models2.json.
 
         Returns:
             Model list in JSON format.
         """
-        return requests.get("https://gpt4all.io/models/models.json").json()
+        resp = requests.get("https://gpt4all.io/models/models2.json")
+        if resp.status_code != 200:
+            raise ValueError(f'Request failed: HTTP {resp.status_code} {resp.reason}')
+        return resp.json()
 
     @staticmethod
     def retrieve_model(
         model_name: str,
-        model_path: Optional[str] = None,
+        model_path: Optional[Union[str, os.PathLike[str]]] = None,
         allow_download: bool = True,
-        verbose: bool = True,
+        verbose: bool = False,
     ) -> ConfigType:
         """
         Find model file, and if it doesn't exist, download the model.
@@ -135,7 +147,7 @@ class GPT4All:
             Model config.
         """
 
-        model_filename = append_bin_suffix_if_missing(model_name)
+        model_filename = append_extension_if_missing(model_name)
 
         # get the config for the model
         config: ConfigType = DEFAULT_MODEL_CONFIG
@@ -162,7 +174,7 @@ class GPT4All:
                 )
             model_path = DEFAULT_MODEL_DIRECTORY
         else:
-            model_path = model_path.replace("\\", "\\\\")
+            model_path = str(model_path).replace("\\", "\\\\")
 
         if not os.path.exists(model_path):
             raise ValueError(f"Invalid model directory: {model_path}")
@@ -172,7 +184,7 @@ class GPT4All:
             config.pop("url", None)
             config["path"] = model_dest
             if verbose:
-                print("Found model file at ", model_dest)
+                print("Found model file at", model_dest, file=sys.stderr)
 
         # If model file does not exist, download
         elif allow_download:
@@ -187,7 +199,7 @@ class GPT4All:
     @staticmethod
     def download_model(
         model_filename: str,
-        model_path: str,
+        model_path: Union[str, os.PathLike[str]],
         verbose: bool = True,
         url: Optional[str] = None,
     ) -> str:
@@ -195,7 +207,7 @@ class GPT4All:
         Download model from https://gpt4all.io.
 
         Args:
-            model_filename: Filename of model (with .bin extension).
+            model_filename: Filename of model (with .gguf extension).
             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)
@@ -207,38 +219,67 @@ class GPT4All:
         def get_download_url(model_filename):
             if url:
                 return url
-            return f"https://gpt4all.io/models/{model_filename}"
+            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)
 
-        response = requests.get(download_url, stream=True)
+        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)
+            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')
+            return response
+
+        response = make_request()
+
         total_size_in_bytes = int(response.headers.get("content-length", 0))
         block_size = 2**20  # 1 MB
 
-        with tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True) as progress_bar:
+        with open(download_path, "wb") as file, \
+                tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True) as progress_bar:
             try:
-                with open(download_path, "wb") as file:
-                    for data in response.iter_content(block_size):
-                        progress_bar.update(len(data))
-                        file.write(data)
+                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:
-                if os.path.exists(download_path):
-                    if verbose:
-                        print("Cleaning up the interrupted download...")
+                if verbose:
+                    print("Cleaning up the interrupted download...", file=sys.stderr)
+                try:
                     os.remove(download_path)
+                except OSError:
+                    pass
                 raise
 
-        # Validate download was successful
-        if total_size_in_bytes != 0 and progress_bar.n != total_size_in_bytes:
-            raise RuntimeError("An error occurred during download. Downloaded file may not work.")
-
-        # Sleep for a little bit so OS can remove file lock
-        time.sleep(2)
+        if os.name == 'nt':
+            time.sleep(2)  # Sleep for a little bit so Windows can remove file lock
 
         if verbose:
-            print("Model downloaded at: ", download_path)
+            print("Model downloaded at:", download_path, file=sys.stderr)
         return download_path
 
     def generate(
@@ -253,7 +294,7 @@ class GPT4All:
         n_batch: int = 8,
         n_predict: Optional[int] = None,
         streaming: bool = False,
-        callback: pyllmodel.ResponseCallbackType = pyllmodel.empty_response_callback,
+        callback: _pyllmodel.ResponseCallbackType = _pyllmodel.empty_response_callback,
     ) -> Union[str, Iterable[str]]:
         """
         Generate outputs from any GPT4All model.
@@ -275,6 +316,10 @@ class GPT4All:
             Either the entire completion or a generator that yields the completion token by token.
         """
 
+        if re.search(r"%1(?![0-9])", self._current_prompt_template):
+            raise ValueError("Prompt template containing a literal '%1' is not supported. For a prompt "
+                             "placeholder, please use '{0}' instead.")
+
         # Preparing the model request
         generate_kwargs: Dict[str, Any] = dict(
             temp=temp,
@@ -288,16 +333,29 @@ class GPT4All:
 
         if self._is_chat_session_activated:
             # check if there is only one message, i.e. system prompt:
-            generate_kwargs["reset_context"] = len(self.current_chat_session) == 1
+            reset = len(self.current_chat_session) == 1
+            generate_kwargs["reset_context"] = reset
             self.current_chat_session.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 "",
-            )
+            if self._format_chat_prompt_template.__func__ is GPT4All._format_chat_prompt_template:
+                if reset:
+                    # ingest system prompt
+                    self.model.prompt_model(self.current_chat_session[0]["content"], "%1",
+                                            n_batch=n_batch, n_predict=0, special=True)
+                prompt_template = self._current_prompt_template.format("%1")
+            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.current_chat_session[-1:],
+                    self.current_chat_session[0]["content"] if reset else "",
+                )
+                prompt_template = "%1"
         else:
+            prompt_template = "%1"
             generate_kwargs["reset_context"] = True
 
         # Prepare the callback, process the model response
@@ -311,10 +369,9 @@ class GPT4All:
             output_collector = self.current_chat_session
 
         def _callback_wrapper(
-            callback: pyllmodel.ResponseCallbackType,
+            callback: _pyllmodel.ResponseCallbackType,
             output_collector: List[MessageType],
-        ) -> pyllmodel.ResponseCallbackType:
-
+        ) -> _pyllmodel.ResponseCallbackType:
             def _callback(token_id: int, response: str) -> bool:
                 nonlocal callback, output_collector
 
@@ -327,14 +384,16 @@ 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,
         )
 
@@ -385,24 +444,6 @@ 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:
@@ -422,7 +463,7 @@ def empty_chat_session(system_prompt: str = "") -> List[MessageType]:
     return [{"role": "system", "content": system_prompt}]
 
 
-def append_bin_suffix_if_missing(model_name):
-    if not model_name.endswith(".bin"):
-        model_name += ".bin"
+def append_extension_if_missing(model_name):
+    if not model_name.endswith((".bin", ".gguf")):
+        model_name += ".gguf"
     return model_name

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

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import time
 from io import StringIO