server: improve correctness of request parsing and responses (#2929)

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

.circleci/continue_config.yml

@@ -317,9 +317,9 @@ jobs:
             wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
-              bison build-essential ccache cuda-compiler-11-8 flex gperf libcublas-dev-11-8 libfontconfig1 libfreetype6
+              bison build-essential ccache cuda-compiler-11-8 flex g++-12 gperf libcublas-dev-11-8 libfontconfig1
-              libgl1-mesa-dev libmysqlclient21 libnvidia-compute-550-server libodbc2 libpq5 libwayland-dev libx11-6
+              libfreetype6 libgl1-mesa-dev libmysqlclient21 libnvidia-compute-550-server libodbc2 libpq5 libwayland-dev
-              libx11-xcb1 libxcb-cursor0 libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-keysyms1 libxcb-randr0
+              libx11-6 libx11-xcb1 libxcb-cursor0 libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-keysyms1 libxcb-randr0
               libxcb-render-util0 libxcb-shape0 libxcb-shm0 libxcb-sync1 libxcb-util1 libxcb-xfixes0 libxcb-xinerama0
               libxcb-xkb1 libxcb1 libxext6 libxfixes3 libxi6 libxkbcommon-x11-0 libxkbcommon0 libxrender1 patchelf
               python3 vulkan-sdk
@@ -352,6 +352,8 @@ jobs:
             ~/Qt/Tools/CMake/bin/cmake \
               -S ../gpt4all-chat -B . \
               -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_C_COMPILER=gcc-12 \
+              -DCMAKE_CXX_COMPILER=g++-12 \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache \
@@ -391,9 +393,9 @@ jobs:
             wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
-              bison build-essential ccache cuda-compiler-11-8 flex gperf libcublas-dev-11-8 libfontconfig1 libfreetype6
+              bison build-essential ccache cuda-compiler-11-8 flex g++-12 gperf libcublas-dev-11-8 libfontconfig1
-              libgl1-mesa-dev libmysqlclient21 libnvidia-compute-550-server libodbc2 libpq5 libwayland-dev libx11-6
+              libfreetype6 libgl1-mesa-dev libmysqlclient21 libnvidia-compute-550-server libodbc2 libpq5 libwayland-dev
-              libx11-xcb1 libxcb-cursor0 libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-keysyms1 libxcb-randr0
+              libx11-6 libx11-xcb1 libxcb-cursor0 libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-keysyms1 libxcb-randr0
               libxcb-render-util0 libxcb-shape0 libxcb-shm0 libxcb-sync1 libxcb-util1 libxcb-xfixes0 libxcb-xinerama0
               libxcb-xkb1 libxcb1 libxext6 libxfixes3 libxi6 libxkbcommon-x11-0 libxkbcommon0 libxrender1 patchelf
               python3 vulkan-sdk
@@ -426,6 +428,8 @@ jobs:
             ~/Qt/Tools/CMake/bin/cmake \
               -S ../gpt4all-chat -B . \
               -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_C_COMPILER=gcc-12 \
+              -DCMAKE_CXX_COMPILER=g++-12 \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache \
@@ -447,7 +451,7 @@ jobs:
 
   build-offline-chat-installer-windows:
     machine:
-      image: 'windows-server-2019-vs2019:2022.08.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -538,7 +542,7 @@ jobs:
 
   sign-offline-chat-installer-windows:
     machine:
-      image: 'windows-server-2019-vs2019:2022.08.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -568,7 +572,7 @@ jobs:
 
   build-online-chat-installer-windows:
     machine:
-      image: 'windows-server-2019-vs2019:2022.08.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -666,7 +670,7 @@ jobs:
 
   sign-online-chat-installer-windows:
     machine:
-      image: 'windows-server-2019-vs2019:2022.08.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -720,9 +724,9 @@ jobs:
             wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
-              bison build-essential ccache cuda-compiler-11-8 flex gperf libcublas-dev-11-8 libfontconfig1 libfreetype6
+              bison build-essential ccache cuda-compiler-11-8 flex g++-12 gperf libcublas-dev-11-8 libfontconfig1
-              libgl1-mesa-dev libmysqlclient21 libnvidia-compute-550-server libodbc2 libpq5 libwayland-dev libx11-6
+              libfreetype6 libgl1-mesa-dev libmysqlclient21 libnvidia-compute-550-server libodbc2 libpq5 libwayland-dev
-              libx11-xcb1 libxcb-cursor0 libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-keysyms1 libxcb-randr0
+              libx11-6 libx11-xcb1 libxcb-cursor0 libxcb-glx0 libxcb-icccm4 libxcb-image0 libxcb-keysyms1 libxcb-randr0
               libxcb-render-util0 libxcb-shape0 libxcb-shm0 libxcb-sync1 libxcb-util1 libxcb-xfixes0 libxcb-xinerama0
               libxcb-xkb1 libxcb1 libxext6 libxfixes3 libxi6 libxkbcommon-x11-0 libxkbcommon0 libxrender1 python3
               vulkan-sdk
@@ -744,6 +748,8 @@ jobs:
             ~/Qt/Tools/CMake/bin/cmake \
               -S gpt4all-chat -B build \
               -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_C_COMPILER=gcc-12 \
+              -DCMAKE_CXX_COMPILER=g++-12 \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache \
@@ -758,7 +764,7 @@ jobs:
 
   build-gpt4all-chat-windows:
     machine:
-      image: 'windows-server-2019-vs2019:2022.08.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -864,8 +870,8 @@ jobs:
           paths:
             - ../.ccache
 
-  build-ts-docs: 
+  build-ts-docs:
-    docker: 
+    docker:
       - image: cimg/base:stable
     steps:
       - checkout
@@ -887,7 +893,7 @@ jobs:
     docker:
       - image: circleci/python:3.8
     steps:
-      - checkout 
+      - checkout
       - run:
           name: Install dependencies
           command: |
@@ -928,7 +934,8 @@ jobs:
             wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
-              build-essential ccache cmake cuda-compiler-11-8 libcublas-dev-11-8 libnvidia-compute-550-server vulkan-sdk
+              build-essential ccache cmake cuda-compiler-11-8 g++-12 libcublas-dev-11-8 libnvidia-compute-550-server
+              vulkan-sdk
             )
             sudo apt-get update
             sudo apt-get install -y "${packages[@]}"
@@ -942,6 +949,8 @@ jobs:
             cd gpt4all-backend
             cmake -B build \
               -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_C_COMPILER=gcc-12 \
+              -DCMAKE_CXX_COMPILER=g++-12 \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache \
@@ -1014,7 +1023,7 @@ jobs:
 
   build-py-windows:
     machine:
-      image: 'windows-server-2019-vs2019:2022.08.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -1118,11 +1127,12 @@ jobs:
           name: Install 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 wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
             wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
-              build-essential ccache cmake cuda-compiler-11-8 libcublas-dev-11-8 libnvidia-compute-550-server vulkan-sdk
+              build-essential ccache cmake cuda-compiler-11-8 g++-12 libcublas-dev-11-8 libnvidia-compute-550-server
+              vulkan-sdk
             )
             sudo apt-get update
             sudo apt-get install -y "${packages[@]}"
@@ -1135,6 +1145,9 @@ jobs:
             mkdir -p runtimes/build
             cd runtimes/build
             cmake ../.. \
+              -DCMAKE_BUILD_TYPE=Release \
+              -DCMAKE_C_COMPILER=gcc-12 \
+              -DCMAKE_C_COMPILER=g++-12 \
               -DCMAKE_BUILD_TYPE=Release \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
@@ -1204,7 +1217,7 @@ jobs:
 
   build-bindings-backend-windows:
     machine:
-      image: 'windows-server-2022-gui:2023.03.1'
+      image: windows-server-2022-gui:current
       resource_class: windows.large
       shell: powershell.exe -ExecutionPolicy Bypass
     steps:
@@ -1230,7 +1243,7 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            choco install -y ccache cmake ninja --installargs 'ADD_CMAKE_TO_PATH=System'  
+            choco install -y ccache cmake ninja --installargs 'ADD_CMAKE_TO_PATH=System'
       - run:
           name: Build Libraries
           command: |
@@ -1263,8 +1276,8 @@ jobs:
           paths:
             - runtimes/win-x64_msvc/*.dll
 
-  build-nodejs-linux: 
+  build-nodejs-linux:
-    docker: 
+    docker:
       - image: cimg/base:stable
     steps:
       - checkout
@@ -1280,10 +1293,10 @@ jobs:
           pkg-manager: yarn
           override-ci-command: yarn install
       - run:
-          command: | 
+          command: |
             cd gpt4all-bindings/typescript
             yarn prebuildify -t 18.16.0 --napi
-      - run: 
+      - run:
           command: |
             mkdir -p gpt4all-backend/prebuilds/linux-x64
             mkdir -p gpt4all-backend/runtimes/linux-x64
@@ -1292,10 +1305,10 @@ jobs:
       - persist_to_workspace:
           root: gpt4all-backend
           paths:
-            - prebuilds/linux-x64/*.node 
+            - prebuilds/linux-x64/*.node
             - runtimes/linux-x64/*-*.so
 
-  build-nodejs-macos: 
+  build-nodejs-macos:
     macos:
       xcode: 15.4.0
     steps:
@@ -1312,12 +1325,12 @@ jobs:
           pkg-manager: yarn
           override-ci-command: yarn install
       - run:
-          command: | 
+          command: |
             cd gpt4all-bindings/typescript
             yarn prebuildify -t 18.16.0 --napi
-      - run: 
+      - run:
           name: "Persisting all necessary things to workspace"
-          command: |  
+          command: |
             mkdir -p gpt4all-backend/prebuilds/darwin-x64
             mkdir -p gpt4all-backend/runtimes/darwin
             cp /tmp/gpt4all-backend/runtimes/osx-x64/*-*.* gpt4all-backend/runtimes/darwin
@@ -1328,7 +1341,7 @@ jobs:
             - prebuilds/darwin-x64/*.node
             - runtimes/darwin/*-*.*
 
-  build-nodejs-windows: 
+  build-nodejs-windows:
     executor:
       name: win/default
       size: large
@@ -1342,29 +1355,29 @@ jobs:
          command: wget https://nodejs.org/dist/v18.16.0/node-v18.16.0-x86.msi -P C:\Users\circleci\Downloads\
          shell: cmd.exe
       - run: MsiExec.exe /i C:\Users\circleci\Downloads\node-v18.16.0-x86.msi /qn
-      - run: 
+      - run:
           command: |
               Start-Process powershell -verb runAs -Args "-start GeneralProfile"
               nvm install 18.16.0
               nvm use 18.16.0
-      - run: node --version 
+      - run: node --version
       - run: corepack enable
-      - run:           
+      - run:
           command: |
             npm install -g yarn
             cd gpt4all-bindings/typescript
             yarn install
       - run:
-          command: | 
+          command: |
             cd gpt4all-bindings/typescript
-            yarn prebuildify -t 18.16.0 --napi 
+            yarn prebuildify -t 18.16.0 --napi
-      - run: 
+      - run:
           command: |
             mkdir -p gpt4all-backend/prebuilds/win32-x64
             mkdir -p gpt4all-backend/runtimes/win32-x64
             cp /tmp/gpt4all-backend/runtimes/win-x64_msvc/*-*.dll gpt4all-backend/runtimes/win32-x64
             cp gpt4all-bindings/typescript/prebuilds/win32-x64/*.node gpt4all-backend/prebuilds/win32-x64
-        
+
       - persist_to_workspace:
           root: gpt4all-backend
           paths:
@@ -1372,7 +1385,7 @@ jobs:
             - runtimes/win32-x64/*-*.dll
 
   prepare-npm-pkg:
-    docker: 
+    docker:
       - image: cimg/base:stable
     steps:
       - attach_workspace:
@@ -1383,19 +1396,19 @@ jobs:
           node-version: "18.16"
       - run: node --version
       - run: corepack enable
-      - run: 
+      - run:
           command: |
             cd gpt4all-bindings/typescript
             # excluding llmodel. nodejs bindings dont need llmodel.dll
             mkdir -p runtimes/win32-x64/native
             mkdir -p prebuilds/win32-x64/
-            cp /tmp/gpt4all-backend/runtimes/win-x64_msvc/*-*.dll runtimes/win32-x64/native/ 
+            cp /tmp/gpt4all-backend/runtimes/win-x64_msvc/*-*.dll runtimes/win32-x64/native/
-            cp /tmp/gpt4all-backend/prebuilds/win32-x64/*.node prebuilds/win32-x64/    
+            cp /tmp/gpt4all-backend/prebuilds/win32-x64/*.node prebuilds/win32-x64/
 
-            mkdir -p runtimes/linux-x64/native 
+            mkdir -p runtimes/linux-x64/native
             mkdir -p prebuilds/linux-x64/
-            cp /tmp/gpt4all-backend/runtimes/linux-x64/*-*.so runtimes/linux-x64/native/ 
+            cp /tmp/gpt4all-backend/runtimes/linux-x64/*-*.so runtimes/linux-x64/native/
-            cp /tmp/gpt4all-backend/prebuilds/linux-x64/*.node prebuilds/linux-x64/    
+            cp /tmp/gpt4all-backend/prebuilds/linux-x64/*.node prebuilds/linux-x64/
 
             # darwin has univeral runtime libraries
             mkdir -p runtimes/darwin/native
@@ -1403,22 +1416,22 @@ jobs:
 
             cp /tmp/gpt4all-backend/runtimes/darwin/*-*.* runtimes/darwin/native/
 
-            cp /tmp/gpt4all-backend/prebuilds/darwin-x64/*.node prebuilds/darwin-x64/    
+            cp /tmp/gpt4all-backend/prebuilds/darwin-x64/*.node prebuilds/darwin-x64/
-            
+
             # Fallback build if user is not on above prebuilds
             mv -f binding.ci.gyp binding.gyp
 
             mkdir gpt4all-backend
             cd ../../gpt4all-backend
             mv llmodel.h llmodel.cpp llmodel_c.cpp llmodel_c.h sysinfo.h dlhandle.h ../gpt4all-bindings/typescript/gpt4all-backend/
-            
+
       # Test install
       - node/install-packages:
           app-dir: gpt4all-bindings/typescript
           pkg-manager: yarn
           override-ci-command: yarn install
-      - run: 
+      - run:
-          command: | 
+          command: |
             cd gpt4all-bindings/typescript
             yarn run test
       - run:
@@ -1552,7 +1565,7 @@ workflows:
             - build-py-linux
             - build-py-macos
   build-bindings:
-    when: 
+    when:
       or:
        - << pipeline.parameters.run-all-workflows >>
        - << pipeline.parameters.run-python-workflow >>
@@ -1585,8 +1598,8 @@ workflows:
           requires:
             - hold
 
-      # NodeJs Jobs 
+      # NodeJs Jobs
-      - prepare-npm-pkg: 
+      - prepare-npm-pkg:
           filters:
             branches:
               only:

.gitmodules

@@ -8,3 +8,6 @@
 [submodule "gpt4all-chat/deps/SingleApplication"]
 	path = gpt4all-chat/deps/SingleApplication
 	url = https://github.com/nomic-ai/SingleApplication.git
+[submodule "gpt4all-chat/deps/fmt"]
+	path = gpt4all-chat/deps/fmt
+	url = https://github.com/fmtlib/fmt.git

gpt4all-backend/CMakeLists.txt

@@ -33,7 +33,7 @@ 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)
 
-set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD 23)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
 set(BUILD_SHARED_LIBS ON)

gpt4all-backend/deps/llama.cpp-mainline

@@ -1 +1 @@
-Subproject commit 443665aec4721ecf57df8162e7e093a0cd674a76
+Subproject commit ced74fbad4b258507f3ec06e77eec9445583511a

gpt4all-backend/include/gpt4all-backend/llmodel.h

@@ -162,7 +162,7 @@ public:
                         bool allowContextShift,
                         PromptContext &ctx,
                         bool special = false,
-                        std::string *fakeReply = nullptr);
+                        std::optional<std::string_view> fakeReply = {});
 
     using EmbedCancelCallback = bool(unsigned *batchSizes, unsigned nBatch, const char *backend);
 
@@ -212,7 +212,7 @@ public:
 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 &ctx, const std::string &str, bool special = false) = 0;
+    virtual std::vector<Token> tokenize(PromptContext &ctx, std::string_view str, bool special = false) = 0;
     virtual bool isSpecialToken(Token id) const = 0;
     virtual std::string tokenToString(Token id) const = 0;
     virtual Token sampleToken(PromptContext &ctx) const = 0;
@@ -249,7 +249,8 @@ protected:
                       std::function<bool(int32_t, const std::string&)> responseCallback,
                       bool allowContextShift,
                       PromptContext &promptCtx,
-                      std::vector<Token> embd_inp);
+                      std::vector<Token> embd_inp,
+                      bool isResponse = false);
     void generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
                           bool allowContextShift,
                           PromptContext &promptCtx);

gpt4all-backend/src/llamamodel.cpp

@@ -536,13 +536,13 @@ 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, bool special)
+std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, std::string_view str, bool special)
 {
     bool atStart = m_tokenize_last_token == -1;
     bool insertSpace = atStart || isSpecialToken(m_tokenize_last_token);
     std::vector<LLModel::Token> fres(str.length() + 4);
     int32_t fres_len = llama_tokenize_gpt4all(
-        d_ptr->model, str.c_str(), str.length(), fres.data(), fres.size(), /*add_special*/ atStart,
+        d_ptr->model, str.data(), str.length(), fres.data(), fres.size(), /*add_special*/ atStart,
         /*parse_special*/ special, /*insert_space*/ insertSpace
     );
     fres.resize(fres_len);

gpt4all-backend/src/llamamodel_impl.h

@@ -8,6 +8,7 @@
 
 #include <memory>
 #include <string>
+#include <string_view>
 #include <vector>
 
 struct LLamaPrivate;
@@ -52,7 +53,7 @@ private:
     bool m_supportsCompletion = false;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) override;
+    std::vector<Token> tokenize(PromptContext &ctx, std::string_view str, bool special) override;
     bool isSpecialToken(Token id) const override;
     std::string tokenToString(Token id) const override;
     Token sampleToken(PromptContext &ctx) const override;

gpt4all-backend/src/llmodel_c.cpp

@@ -12,6 +12,7 @@
 #include <memory>
 #include <optional>
 #include <string>
+#include <string_view>
 #include <vector>
 
 struct LLModelWrapper {
@@ -130,13 +131,10 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
     wrapper->promptContext.repeat_last_n = ctx->repeat_last_n;
     wrapper->promptContext.contextErase = ctx->context_erase;
 
-    std::string fake_reply_str;
-    if (fake_reply) { fake_reply_str = fake_reply; }
-    auto *fake_reply_p = fake_reply ? &fake_reply_str : nullptr;
-
     // Call the C++ prompt method
     wrapper->llModel->prompt(prompt, prompt_template, prompt_callback, response_func, allow_context_shift,
-                             wrapper->promptContext, special, fake_reply_p);
+                             wrapper->promptContext, special,
+                             fake_reply ? std::make_optional<std::string_view>(fake_reply) : std::nullopt);
 
     // Update the C context by giving access to the wrappers raw pointers to std::vector data
     // which involves no copies

gpt4all-backend/src/llmodel_shared.cpp

@@ -11,6 +11,7 @@
 #include <sstream>
 #include <stdexcept>
 #include <string>
+#include <string_view>
 #include <vector>
 
 namespace ranges = std::ranges;
@@ -45,7 +46,7 @@ void LLModel::prompt(const std::string &prompt,
                      bool allowContextShift,
                      PromptContext &promptCtx,
                      bool special,
-                     std::string *fakeReply)
+                     std::optional<std::string_view> fakeReply)
 {
     if (!isModelLoaded()) {
         std::cerr << implementation().modelType() << " ERROR: prompt won't work with an unloaded model!\n";
@@ -129,11 +130,11 @@ void LLModel::prompt(const std::string &prompt,
         return; // error
 
     // decode the assistant's reply, either generated or spoofed
-    if (fakeReply == nullptr) {
+    if (!fakeReply) {
         generateResponse(responseCallback, allowContextShift, promptCtx);
     } else {
         embd_inp = tokenize(promptCtx, *fakeReply, false);
-        if (!decodePrompt(promptCallback, responseCallback, allowContextShift, promptCtx, embd_inp))
+        if (!decodePrompt(promptCallback, responseCallback, allowContextShift, promptCtx, embd_inp, true))
             return; // error
     }
 
@@ -157,7 +158,8 @@ bool LLModel::decodePrompt(std::function<bool(int32_t)> promptCallback,
                            std::function<bool(int32_t, const std::string&)> responseCallback,
                            bool allowContextShift,
                            PromptContext &promptCtx,
-                           std::vector<Token> embd_inp) {
+                           std::vector<Token> embd_inp,
+                           bool isResponse) {
     if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
         responseCallback(-1, "ERROR: The prompt size exceeds the context window size and cannot be processed.");
         std::cerr << implementation().modelType() << " ERROR: The prompt is " << embd_inp.size() <<
@@ -196,7 +198,9 @@ bool LLModel::decodePrompt(std::function<bool(int32_t)> promptCallback,
         for (size_t t = 0; t < tokens; ++t) {
             promptCtx.tokens.push_back(batch.at(t));
             promptCtx.n_past += 1;
-            if (!promptCallback(batch.at(t)))
+            Token tok = batch.at(t);
+            bool res = isResponse ? responseCallback(tok, tokenToString(tok)) : promptCallback(tok);
+            if (!res)
                 return false;
         }
         i = batch_end;

gpt4all-chat/CHANGELOG.md

@@ -26,6 +26,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - Fix a typo in Model Settings (by [@3Simplex](https://github.com/3Simplex) in [#2916](https://github.com/nomic-ai/gpt4all/pull/2916))
 - Fix the antenna icon tooltip when using the local server ([#2922](https://github.com/nomic-ai/gpt4all/pull/2922))
 - Fix a few issues with locating files and handling errors when loading remote models on startup ([#2875](https://github.com/nomic-ai/gpt4all/pull/2875))
+- Significantly improve API server request parsing and response correctness ([#2929](https://github.com/nomic-ai/gpt4all/pull/2929))
 
 ## [3.2.1] - 2024-08-13
 

gpt4all-chat/CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.16)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
-set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD 23)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 if(APPLE)
@@ -64,6 +64,12 @@ message(STATUS "Qt 6 root directory: ${Qt6_ROOT_DIR}")
 
 set (CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
+set(FMT_INSTALL OFF)
+set(BUILD_SHARED_LIBS_SAVED "${BUILD_SHARED_LIBS}")
+set(BUILD_SHARED_LIBS OFF)
+add_subdirectory(deps/fmt)
+set(BUILD_SHARED_LIBS "${BUILD_SHARED_LIBS_SAVED}")
+
 add_subdirectory(../gpt4all-backend llmodel)
 
 set(CHAT_EXE_RESOURCES)
@@ -240,7 +246,7 @@ else()
     PRIVATE Qt6::Quick Qt6::Svg Qt6::HttpServer Qt6::Sql Qt6::Pdf)
 endif()
 target_link_libraries(chat
-    PRIVATE llmodel SingleApplication)
+    PRIVATE llmodel SingleApplication fmt::fmt)
 
 
 # -- install --

gpt4all-chat/deps/fmt

@@ -0,0 +1 @@
+Subproject commit 0c9fce2ffefecfdce794e1859584e25877b7b592

gpt4all-chat/src/chat.cpp

@@ -239,16 +239,17 @@ void Chat::newPromptResponsePair(const QString &prompt)
     resetResponseState();
     m_chatModel->updateCurrentResponse(m_chatModel->count() - 1, false);
     m_chatModel->appendPrompt("Prompt: ", prompt);
-    m_chatModel->appendResponse("Response: ", prompt);
+    m_chatModel->appendResponse("Response: ", QString());
     emit resetResponseRequested();
 }
 
+// the server needs to block until response is reset, so it calls resetResponse on its own m_llmThread
 void Chat::serverNewPromptResponsePair(const QString &prompt)
 {
     resetResponseState();
     m_chatModel->updateCurrentResponse(m_chatModel->count() - 1, false);
     m_chatModel->appendPrompt("Prompt: ", prompt);
-    m_chatModel->appendResponse("Response: ", prompt);
+    m_chatModel->appendResponse("Response: ", QString());
 }
 
 bool Chat::restoringFromText() const

gpt4all-chat/src/chatapi.cpp

@@ -93,7 +93,7 @@ void ChatAPI::prompt(const std::string &prompt,
                      bool allowContextShift,
                      PromptContext &promptCtx,
                      bool special,
-                     std::string *fakeReply) {
+                     std::optional<std::string_view> fakeReply) {
 
     Q_UNUSED(promptCallback);
     Q_UNUSED(allowContextShift);
@@ -121,7 +121,7 @@ void ChatAPI::prompt(const std::string &prompt,
     if (fakeReply) {
         promptCtx.n_past += 1;
         m_context.append(formattedPrompt);
-        m_context.append(QString::fromStdString(*fakeReply));
+        m_context.append(QString::fromUtf8(fakeReply->data(), fakeReply->size()));
         return;
     }
 

gpt4all-chat/src/chatapi.h

@@ -12,9 +12,10 @@
 
 #include <cstddef>
 #include <cstdint>
-#include <stdexcept>
 #include <functional>
+#include <stdexcept>
 #include <string>
+#include <string_view>
 #include <vector>
 
 class QNetworkAccessManager;
@@ -72,7 +73,7 @@ public:
                 bool allowContextShift,
                 PromptContext &ctx,
                 bool special,
-                std::string *fakeReply) override;
+                std::optional<std::string_view> fakeReply) override;
 
     void setThreadCount(int32_t n_threads) override;
     int32_t threadCount() const override;
@@ -97,7 +98,7 @@ protected:
     // them as they are only called from the default implementation of 'prompt' which we override and
     // completely replace
 
-    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) override
+    std::vector<Token> tokenize(PromptContext &ctx, std::string_view str, bool special) override
     {
         (void)ctx;
         (void)str;

gpt4all-chat/src/chatllm.cpp

@@ -626,16 +626,16 @@ void ChatLLM::regenerateResponse()
     m_ctx.tokens.erase(m_ctx.tokens.end() - m_promptResponseTokens, m_ctx.tokens.end());
     m_promptResponseTokens = 0;
     m_promptTokens = 0;
-    m_response = std::string();
+    m_response = m_trimmedResponse = std::string();
-    emit responseChanged(QString::fromStdString(m_response));
+    emit responseChanged(QString::fromStdString(m_trimmedResponse));
 }
 
 void ChatLLM::resetResponse()
 {
     m_promptTokens = 0;
     m_promptResponseTokens = 0;
-    m_response = std::string();
+    m_response = m_trimmedResponse = std::string();
-    emit responseChanged(QString::fromStdString(m_response));
+    emit responseChanged(QString::fromStdString(m_trimmedResponse));
 }
 
 void ChatLLM::resetContext()
@@ -645,9 +645,12 @@ void ChatLLM::resetContext()
     m_ctx = LLModel::PromptContext();
 }
 
-QString ChatLLM::response() const
+QString ChatLLM::response(bool trim) const
 {
-    return QString::fromStdString(remove_leading_whitespace(m_response));
+    std::string resp = m_response;
+    if (trim)
+        resp = remove_leading_whitespace(resp);
+    return QString::fromStdString(resp);
 }
 
 ModelInfo ChatLLM::modelInfo() const
@@ -705,7 +708,8 @@ bool ChatLLM::handleResponse(int32_t token, const std::string &response)
     // check for error
     if (token < 0) {
         m_response.append(response);
-        emit responseChanged(QString::fromStdString(remove_leading_whitespace(m_response)));
+        m_trimmedResponse = remove_leading_whitespace(m_response);
+        emit responseChanged(QString::fromStdString(m_trimmedResponse));
         return false;
     }
 
@@ -715,7 +719,8 @@ bool ChatLLM::handleResponse(int32_t token, const std::string &response)
     m_timer->inc();
     Q_ASSERT(!response.empty());
     m_response.append(response);
-    emit responseChanged(QString::fromStdString(remove_leading_whitespace(m_response)));
+    m_trimmedResponse = remove_leading_whitespace(m_response);
+    emit responseChanged(QString::fromStdString(m_trimmedResponse));
     return !m_stopGenerating;
 }
 
@@ -741,7 +746,7 @@ bool ChatLLM::prompt(const QList<QString> &collectionList, const QString &prompt
 
 bool ChatLLM::promptInternal(const QList<QString> &collectionList, const QString &prompt, const QString &promptTemplate,
     int32_t n_predict, int32_t top_k, float top_p, float min_p, float temp, int32_t n_batch, float repeat_penalty,
-    int32_t repeat_penalty_tokens)
+    int32_t repeat_penalty_tokens, std::optional<QString> fakeReply)
 {
     if (!isModelLoaded())
         return false;
@@ -751,7 +756,7 @@ bool ChatLLM::promptInternal(const QList<QString> &collectionList, const QString
 
     QList<ResultInfo> databaseResults;
     const int retrievalSize = MySettings::globalInstance()->localDocsRetrievalSize();
-    if (!collectionList.isEmpty()) {
+    if (!fakeReply && !collectionList.isEmpty()) {
         emit requestRetrieveFromDB(collectionList, prompt, retrievalSize, &databaseResults); // blocks
         emit databaseResultsChanged(databaseResults);
     }
@@ -797,7 +802,8 @@ bool ChatLLM::promptInternal(const QList<QString> &collectionList, const QString
         m_ctx.n_predict = old_n_predict; // now we are ready for a response
     }
     m_llModelInfo.model->prompt(prompt.toStdString(), promptTemplate.toStdString(), promptFunc, responseFunc,
-                                /*allowContextShift*/ true, m_ctx);
+                                /*allowContextShift*/ true, m_ctx, false,
+                                fakeReply.transform(std::mem_fn(&QString::toStdString)));
 #if defined(DEBUG)
     printf("\n");
     fflush(stdout);
@@ -805,9 +811,9 @@ bool ChatLLM::promptInternal(const QList<QString> &collectionList, const QString
     m_timer->stop();
     qint64 elapsed = totalTime.elapsed();
     std::string trimmed = trim_whitespace(m_response);
-    if (trimmed != m_response) {
+    if (trimmed != m_trimmedResponse) {
-        m_response = trimmed;
+        m_trimmedResponse = trimmed;
-        emit responseChanged(QString::fromStdString(m_response));
+        emit responseChanged(QString::fromStdString(m_trimmedResponse));
     }
 
     SuggestionMode mode = MySettings::globalInstance()->suggestionMode();
@@ -1078,6 +1084,7 @@ bool ChatLLM::deserialize(QDataStream &stream, int version, bool deserializeKV,
     QString response;
     stream >> response;
     m_response = response.toStdString();
+    m_trimmedResponse = trim_whitespace(m_response);
     QString nameResponse;
     stream >> nameResponse;
     m_nameResponse = nameResponse.toStdString();
@@ -1306,10 +1313,9 @@ void ChatLLM::processRestoreStateFromText()
 
         auto &response = *it++;
         Q_ASSERT(response.first != "Prompt: ");
-        auto responseText = response.second.toStdString();
 
         m_llModelInfo.model->prompt(prompt.second.toStdString(), promptTemplate.toStdString(), promptFunc, nullptr,
-                                    /*allowContextShift*/ true, m_ctx, false, &responseText);
+                                    /*allowContextShift*/ true, m_ctx, false, response.second.toUtf8().constData());
     }
 
     if (!m_stopGenerating) {

gpt4all-chat/src/chatllm.h

@@ -116,7 +116,7 @@ public:
     void setForceUnloadModel(bool b) { m_forceUnloadModel = b; }
     void setMarkedForDeletion(bool b) { m_markedForDeletion = b; }
 
-    QString response() const;
+    QString response(bool trim = true) const;
 
     ModelInfo modelInfo() const;
     void setModelInfo(const ModelInfo &info);
@@ -198,7 +198,7 @@ Q_SIGNALS:
 protected:
     bool promptInternal(const QList<QString> &collectionList, const QString &prompt, const QString &promptTemplate,
         int32_t n_predict, int32_t top_k, float top_p, float min_p, float temp, int32_t n_batch, float repeat_penalty,
-        int32_t repeat_penalty_tokens);
+        int32_t repeat_penalty_tokens, std::optional<QString> fakeReply = {});
     bool handlePrompt(int32_t token);
     bool handleResponse(int32_t token, const std::string &response);
     bool handleNamePrompt(int32_t token);
@@ -221,6 +221,7 @@ private:
     bool loadNewModel(const ModelInfo &modelInfo, QVariantMap &modelLoadProps);
 
     std::string m_response;
+    std::string m_trimmedResponse;
     std::string m_nameResponse;
     QString m_questionResponse;
     LLModelInfo m_llModelInfo;

gpt4all-chat/src/localdocsmodel.h

@@ -20,24 +20,25 @@ class LocalDocsCollectionsModel : public QSortFilterProxyModel
     Q_OBJECT
     Q_PROPERTY(int count READ count NOTIFY countChanged)
     Q_PROPERTY(int updatingCount READ updatingCount NOTIFY updatingCountChanged)
+
 public:
     explicit LocalDocsCollectionsModel(QObject *parent);
+    int count() const { return rowCount(); }
+    int updatingCount() const;
 
 public Q_SLOTS:
-    int count() const { return rowCount(); }
     void setCollections(const QList<QString> &collections);
-    int updatingCount() const;
 
 Q_SIGNALS:
     void countChanged();
     void updatingCountChanged();
 
-private Q_SLOT:
-    void maybeTriggerUpdatingCountChanged();
-
 protected:
     bool filterAcceptsRow(int sourceRow, const QModelIndex &sourceParent) const override;
 
+private Q_SLOTS:
+    void maybeTriggerUpdatingCountChanged();
+
 private:
     QList<QString> m_collections;
     int m_updatingCount = 0;

gpt4all-chat/src/modellist.h

@@ -18,10 +18,12 @@
 #include <QVector>
 #include <Qt>
 #include <QtGlobal>
-#include <QtQml>
+
+#include <utility>
 
 using namespace Qt::Literals::StringLiterals;
 
+
 struct ModelInfo {
     Q_GADGET
     Q_PROPERTY(QString id READ id WRITE setId)
@@ -523,7 +525,7 @@ private:
 
 protected:
     explicit ModelList();
-    ~ModelList() { for (auto *model: m_models) { delete model; } }
+    ~ModelList() override { for (auto *model: std::as_const(m_models)) { delete model; } }
     friend class MyModelList;
 };
 

gpt4all-chat/src/mysettings.h

@@ -8,6 +8,7 @@
 #include <QSettings>
 #include <QString>
 #include <QStringList>
+#include <QTranslator>
 #include <QVector>
 
 #include <cstdint>

gpt4all-chat/src/server.h

@@ -4,22 +4,29 @@
 #include "chatllm.h"
 #include "database.h"
 
-#include <QHttpServerRequest>
+#include <QHttpServer>
 #include <QHttpServerResponse>
-#include <QObject>
+#include <QJsonObject>
 #include <QList>
+#include <QObject>
 #include <QString>
 
+#include <memory>
+#include <optional>
+#include <utility>
+
 class Chat;
-class QHttpServer;
+class ChatRequest;
+class CompletionRequest;
+
 
 class Server : public ChatLLM
 {
     Q_OBJECT
 
 public:
-    Server(Chat *parent);
+    explicit Server(Chat *chat);
-    virtual ~Server();
+    ~Server() override = default;
 
 public Q_SLOTS:
     void start();
@@ -27,14 +34,17 @@ public Q_SLOTS:
 Q_SIGNALS:
     void requestServerNewPromptResponsePair(const QString &prompt);
 
+private:
+    auto handleCompletionRequest(const CompletionRequest &request) -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>;
+    auto handleChatRequest(const ChatRequest &request) -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>;
+
 private Q_SLOTS:
-    QHttpServerResponse handleCompletionRequest(const QHttpServerRequest &request, bool isChat);
     void handleDatabaseResultsChanged(const QList<ResultInfo> &results) { m_databaseResults = results; }
     void handleCollectionListChanged(const QList<QString> &collectionList) { m_collections = collectionList; }
 
 private:
     Chat *m_chat;
-    QHttpServer *m_server;
+    std::unique_ptr<QHttpServer> m_server;
     QList<ResultInfo> m_databaseResults;
     QList<QString> m_collections;
 };