WIP: provider types are recongized properly in qml

2025 is too soon to use C++ features from 2020 without running into bugs
in every build tool that touches the project.

.circleci/config.yml

@@ -17,6 +17,4 @@ workflows:
           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-chat/.* run-chat-workflow true

.codespellrc

@@ -1,3 +1,3 @@
 [codespell]
-ignore-words-list = blong, afterall, assistent, crasher, requestor
+ignore-words-list = blong, afterall, assistent, crasher, requestor, nam
 skip = ./.git,./gpt4all-chat/translations,*.pdf,*.svg,*.lock

.github/ISSUE_TEMPLATE/bindings-bug.md

@@ -1,35 +0,0 @@
----
-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. -->

.gitmodules

@@ -1,16 +1,16 @@
-[submodule "llama.cpp-mainline"]
-	path = gpt4all-backend/deps/llama.cpp-mainline
+[submodule "gpt4all-backend-old/deps/llama.cpp-mainline"]
+	path = gpt4all-backend-old/deps/llama.cpp-mainline
 	url = https://github.com/nomic-ai/llama.cpp.git
-	branch = master
+    branch = master
 [submodule "gpt4all-chat/usearch"]
 	path = gpt4all-chat/deps/usearch
 	url = https://github.com/nomic-ai/usearch.git
 [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
+[submodule "deps/fmt"]
+	path = deps/fmt
+	url = https://github.com/nomic-ai/fmt.git
 [submodule "gpt4all-chat/deps/DuckX"]
 	path = gpt4all-chat/deps/DuckX
 	url = https://github.com/nomic-ai/DuckX.git
@@ -23,3 +23,12 @@
 [submodule "gpt4all-chat/deps/json"]
 	path = gpt4all-chat/deps/json
 	url = https://github.com/nlohmann/json.git
+[submodule "gpt4all-backend/deps/qcoro"]
+	path = deps/qcoro
+	url = https://github.com/nomic-ai/qcoro.git
+[submodule "gpt4all-backend/deps/date"]
+	path = gpt4all-backend/deps/date
+	url = https://github.com/HowardHinnant/date.git
+[submodule "gpt4all-chat/deps/generator"]
+	path = gpt4all-chat/deps/generator
+	url = https://github.com/TartanLlama/generator.git

MAINTAINERS.md

@@ -29,13 +29,6 @@ Jared Van Bortel ([@cebtenzzre](https://github.com/cebtenzzre))<br/>
 E-mail: jared@nomic.ai<br/>
 Discord: `@cebtenzzre`
 - gpt4all-backend
-- Python binding
-- Python CLI app
-
-Jacob Nguyen ([@jacoobes](https://github.com/jacoobes))<br/>
-Discord: `@jacoobes`<br/>
-E-mail: `jacoobes@sern.dev`
-- TypeScript binding
 
 Dominik ([@cosmic-snow](https://github.com/cosmic-snow))<br/>
 E-mail: cosmic-snow@mailfence.com<br/>
@@ -45,7 +38,7 @@ Discord: `@cosmic__snow`
 Max Cembalest ([@mcembalest](https://github.com/mcembalest))<br/>
 E-mail: max@nomic.ai<br/>
 Discord: `@maxcembalest.`
-- Official documentation (gpt4all-bindings/python/docs -> https://docs.gpt4all.io/)
+- Official documentation (docs -> https://docs.gpt4all.io/)
 
 Thiago Ramos ([@thiagojramos](https://github.com/thiagojramos))<br/>
 E-mail: thiagojramos@outlook.com<br/>

README.md

@@ -32,24 +32,35 @@ GPT4All is made possible by our compute partner <a href="https://www.paperspace.
 
 <p>
   &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-win64.exe">
-    <img src="gpt4all-bindings/python/docs/assets/windows.png" style="height: 1em; width: auto" /> Windows Installer
+    <img src="docs/assets/windows.png" style="height: 1em; width: auto" /> Windows Installer
+  </a> &mdash;
+</p>
+<p>
+  &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-win64-arm.exe">
+    <img src="gpt4all-bindings/python/docs/assets/windows.png" style="height: 1em; width: auto" /> Windows ARM Installer
   </a> &mdash;
 </p>
 <p>
   &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-darwin.dmg">
-    <img src="gpt4all-bindings/python/docs/assets/mac.png" style="height: 1em; width: auto" /> macOS Installer
+    <img src="docs/assets/mac.png" style="height: 1em; width: auto" /> macOS Installer
   </a> &mdash;
 </p>
 <p>
   &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-linux.run">
-    <img src="gpt4all-bindings/python/docs/assets/ubuntu.svg" style="height: 1em; width: auto" /> Ubuntu Installer
+    <img src="docs/assets/ubuntu.svg" style="height: 1em; width: auto" /> Ubuntu Installer
   </a> &mdash;
 </p>
 <p>
-  Windows and Linux require Intel Core i3 2nd Gen / AMD Bulldozer, or better. x86-64 only, no ARM.
+  The Windows and Linux builds require Intel Core i3 2nd Gen / AMD Bulldozer, or better.
 </p>
 <p>
-  macOS requires Monterey 12.6 or newer. Best results with Apple Silicon M-series processors.
+  The Windows ARM build supports Qualcomm Snapdragon and Microsoft SQ1/SQ2 processors.
+</p>
+<p>
+  The Linux build is x86-64 only (no ARM).
+</p>
+<p>
+  The macOS build requires Monterey 12.6 or newer. Best results with Apple Silicon M-series processors.
 </p>
 
 See the full [System Requirements](gpt4all-chat/system_requirements.md) for more details.
@@ -63,24 +74,6 @@ See the full [System Requirements](gpt4all-chat/system_requirements.md) for more
   </a>
 </p>
 
-## Install GPT4All Python
-
-`gpt4all` gives you access to LLMs with our Python client around [`llama.cpp`](https://github.com/ggerganov/llama.cpp) implementations. 
-
-Nomic contributes to open source software like [`llama.cpp`](https://github.com/ggerganov/llama.cpp) to make LLMs accessible and efficient **for all**.
-
-```bash
-pip install gpt4all
-```
-
-```python
-from gpt4all import GPT4All
-model = GPT4All("Meta-Llama-3-8B-Instruct.Q4_0.gguf") # downloads / loads a 4.66GB LLM
-with model.chat_session():
-    print(model.generate("How can I run LLMs efficiently on my laptop?", max_tokens=1024))
-```
-
-
 ## Integrations
 
 :parrot::link: [Langchain](https://python.langchain.com/v0.2/docs/integrations/providers/gpt4all/)
@@ -108,7 +101,7 @@ Please see CONTRIBUTING.md and follow the issues, bug reports, and PR markdown t
 
 Check project discord, with project owners, or through existing issues/PRs to avoid duplicate work.
 Please make sure to tag all of the above with relevant project identifiers or your contribution could potentially get lost.
-Example tags: `backend`, `bindings`, `python-bindings`, `documentation`, etc.
+Example tags: `backend`, `documentation`, etc.
 
 ## Citation
 

common/common.cmake

@@ -1,3 +1,8 @@
+# enable color diagnostics with ninja
+if (CMAKE_GENERATOR STREQUAL Ninja AND CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+    add_compile_options(-fdiagnostics-color=always)
+endif()
+
 function(gpt4all_add_warning_options target)
     if (MSVC)
         return()

deps/CMakeLists.txt

@@ -0,0 +1,22 @@
+set(BUILD_SHARED_LIBS OFF)
+
+set(FMT_INSTALL OFF)
+add_subdirectory(fmt)
+
+set(BUILD_TESTING           OFF)
+set(QCORO_BUILD_EXAMPLES    OFF)
+set(QCORO_WITH_QTDBUS       OFF)
+set(QCORO_WITH_QTWEBSOCKETS OFF)
+set(QCORO_WITH_QTQUICK      OFF)
+set(QCORO_WITH_QTTEST       OFF)
+add_subdirectory(qcoro)
+
+set(GPT4ALL_BOOST_TAG 1.87.0)
+FetchContent_Declare(
+    boost
+    URL "https://github.com/boostorg/boost/releases/download/boost-${GPT4ALL_BOOST_TAG}/boost-${GPT4ALL_BOOST_TAG}-cmake.tar.xz"
+    URL_HASH "SHA256=7da75f171837577a52bbf217e17f8ea576c7c246e4594d617bfde7fafd408be5"
+)
+
+set(BOOST_INCLUDE_LIBRARIES json describe system)
+FetchContent_MakeAvailable(boost)

docs/gpt4all_desktop/cookbook/use-local-ai-models-to-privately-chat-with-Obsidian.md

@@ -46,7 +46,7 @@ Obsidian for Desktop is a powerful management and note-taking software designed
             <tr>
                <td>
                   <!-- Screenshot of adding collection in LocalDocs -->
-                  <img width="1348" alt="Screenshot of adding collection" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/124ef867a9d9afd9e14d3858cd77bce858f79773/gpt4all-bindings/python/docs/assets/obsidian_adding_collection.png">
+                  <img width="1348" alt="Screenshot of adding collection" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/main/docs/assets/obsidian_adding_collection.png">
                </td>
             </tr>
          </table>
@@ -65,7 +65,7 @@ Obsidian for Desktop is a powerful management and note-taking software designed
             <tr>
                <td>
                   <!-- Screenshot of accessing LocalDocs in chats -->
-                  <img width="1447" alt="Accessing LocalDocs in chats" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/124ef867a9d9afd9e14d3858cd77bce858f79773/gpt4all-bindings/python/docs/assets/obsidian_docs.png">
+                  <img width="1447" alt="Accessing LocalDocs in chats" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/main/docs/assets/obsidian_docs.png">
                </td>
             </tr>
          </table>
@@ -76,7 +76,7 @@ Obsidian for Desktop is a powerful management and note-taking software designed
             <tr>
                <td>
                   <!-- Screenshot of interacting sources -->
-                  <img width="662" alt="osbsidian user interaction" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/124ef867a9d9afd9e14d3858cd77bce858f79773/gpt4all-bindings/python/docs/assets/osbsidian_user_interaction.png">
+                  <img width="662" alt="osbsidian user interaction" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/main/docs/assets/osbsidian_user_interaction.png">
                </td>
             </tr>
          </table>
@@ -84,7 +84,7 @@ Obsidian for Desktop is a powerful management and note-taking software designed
             <tr>
                <td>
                   <!-- Screenshot of viewing sources -->
-                  <img width="662" alt="osbsidian GPT4ALL response" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/124ef867a9d9afd9e14d3858cd77bce858f79773/gpt4all-bindings/python/docs/assets/obsidian_response.png">
+                  <img width="662" alt="osbsidian GPT4ALL response" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/main/docs/assets/obsidian_response.png">
                </td>
             </tr>
          </table>
@@ -96,7 +96,7 @@ Obsidian for Desktop is a powerful management and note-taking software designed
             <tr>
                <td>
                   <!-- Referenced Files  -->
-                  <img width="643" alt="Referenced Files" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/124ef867a9d9afd9e14d3858cd77bce858f79773/gpt4all-bindings/python/docs/assets/obsidian_sources.png">
+                  <img width="643" alt="Referenced Files" src="https://raw.githubusercontent.com/nomic-ai/gpt4all/main/docs/assets/obsidian_sources.png">
                </td>
             </tr>
          </table>
@@ -104,6 +104,3 @@ Obsidian for Desktop is a powerful management and note-taking software designed
 ## How It Works
 
 Obsidian for Desktop syncs your Obsidian notes to your computer, while LocalDocs integrates these files into your LLM chats using embedding models. These models find semantically similar snippets from your files to enhance the context of your interactions.
-
-To learn more about embedding models and explore further, refer to the [Nomic Python SDK documentation](https://docs.nomic.ai/atlas/capabilities/embeddings).
-

docs/gpt4all_desktop/localdocs.md

@@ -44,5 +44,3 @@ LocalDocs brings the information you have from files on-device into your LLM cha
 ## How It Works
 
 A LocalDocs collection uses Nomic AI's free and fast on-device embedding models to index your folder into text snippets that each get an **embedding vector**. These vectors allow us to find snippets from your files that are semantically similar to the questions and prompts you enter in your chats. We then include those semantically similar snippets in the prompt to the LLM.
-
-To try the embedding models yourself, we recommend using the [Nomic Python SDK](https://docs.nomic.ai/atlas/capabilities/embeddings)

docs/gpt4all_desktop/settings.md

@@ -49,7 +49,6 @@ You can **clone** an existing model, which allows you to save a configuration of
     |----------------------------|------------------------------------------|-----------|
     | **Context Length**         | Maximum length of input sequence in tokens        | 2048      |
     | **Max Length**             | Maximum length of response in tokens     | 4096      |
-    | **Prompt Batch Size**      | Token batch size for parallel processing | 128      |
     | **Temperature**            | Lower temperature gives more likely generations | 0.7       |
     | **Top P**                  | Prevents choosing highly unlikely tokens  | 0.4       |
     | **Top K**                  | Size of selection pool for tokens         | 40        |

docs/gpt4all_help/faq.md

@@ -6,32 +6,16 @@
 
 We support models with a `llama.cpp` implementation which have been uploaded to [HuggingFace](https://huggingface.co/).
 
-### Which embedding models are supported?
-
-We support SBert and Nomic Embed Text v1 & v1.5.
-
 ## Software
 
 ### What software do I need?
 
 All you need is to [install GPT4all](../index.md) onto you Windows, Mac, or Linux computer.
 
-### Which SDK languages are supported?
-
-Our SDK is in Python for usability, but these are light bindings around [`llama.cpp`](https://github.com/ggerganov/llama.cpp) implementations that we contribute to for efficiency and accessibility on everyday computers.
-
 ### Is there an API?
 
 Yes, you can run your model in server-mode with our [OpenAI-compatible API](https://platform.openai.com/docs/api-reference/completions), which you can configure in [settings](../gpt4all_desktop/settings.md#application-settings)
 
-### Can I monitor a GPT4All deployment?
-
-Yes, GPT4All [integrates](../gpt4all_python/monitoring.md) with [OpenLIT](https://github.com/openlit/openlit) so you can deploy LLMs with user interactions and hardware usage automatically monitored for full observability.
-
-### Is there a command line interface (CLI)?
-
-[Yes](https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings/cli), we have a lightweight use of the Python client as a CLI. We welcome further contributions!
-
 ## Hardware
 
 ### What hardware do I need?

docs/gpt4all_help/troubleshooting.md

@@ -2,7 +2,7 @@
 
 ## Error Loading Models
 
-It is possible you are trying to load a model from HuggingFace whose weights are not compatible with our [backend](https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-bindings).
+It is possible you are trying to load a model from HuggingFace whose weights are not compatible with our [backend](https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-backend).
 
 Try downloading one of the officially supported models listed on the main models page in the application. If the problem persists, please share your experience on our [Discord](https://discord.com/channels/1076964370942267462).
 

docs/index.md

@@ -12,17 +12,3 @@ No API calls or GPUs required - you can just download the application and [get s
         [Download for Mac](https://gpt4all.io/installers/gpt4all-installer-darwin.dmg)     
         [Download for Linux](https://gpt4all.io/installers/gpt4all-installer-linux.run)
     </div>
-
-!!! note "Python SDK"
-    Use GPT4All in Python to program with LLMs implemented with the [`llama.cpp`](https://github.com/ggerganov/llama.cpp) backend and [Nomic's C backend](https://github.com/nomic-ai/gpt4all/tree/main/gpt4all-backend). Nomic contributes to open source software like [`llama.cpp`](https://github.com/ggerganov/llama.cpp) to make LLMs accessible and efficient **for all**.
-
-    ```bash
-    pip install gpt4all
-    ```
-
-    ```python
-    from gpt4all import GPT4All
-    model = GPT4All("Meta-Llama-3-8B-Instruct.Q4_0.gguf") # downloads / loads a 4.66GB LLM
-    with model.chat_session():
-        print(model.generate("How can I run LLMs efficiently on my laptop?", max_tokens=1024))
-    ```

gpt4all-backend-old/CMakeLists.txt

@@ -0,0 +1,189 @@
+cmake_minimum_required(VERSION 3.23)  # for FILE_SET
+
+include(../common/common.cmake)
+
+set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (APPLE)
+    option(BUILD_UNIVERSAL "Build a Universal binary on macOS" ON)
+else()
+    option(LLMODEL_KOMPUTE "llmodel: use Kompute"              ON)
+    option(LLMODEL_VULKAN  "llmodel: use Vulkan"               OFF)
+    option(LLMODEL_CUDA    "llmodel: use CUDA"                 ON)
+    option(LLMODEL_ROCM    "llmodel: use ROCm"                 OFF)
+endif()
+
+if (APPLE)
+  if (BUILD_UNIVERSAL)
+    # Build a Universal binary on macOS
+    # This requires that the found Qt library is compiled as Universal binaries.
+    set(CMAKE_OSX_ARCHITECTURES "arm64;x86_64" CACHE STRING "" FORCE)
+  else()
+    # Build for the host architecture on macOS
+    if (NOT CMAKE_OSX_ARCHITECTURES)
+      set(CMAKE_OSX_ARCHITECTURES "${CMAKE_HOST_SYSTEM_PROCESSOR}" CACHE STRING "" FORCE)
+    endif()
+  endif()
+endif()
+
+# Include the binary directory for the generated header file
+include_directories("${CMAKE_CURRENT_BINARY_DIR}")
+
+set(LLMODEL_VERSION_MAJOR 0)
+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)
+
+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)
+
+# Check for IPO support
+include(CheckIPOSupported)
+check_ipo_supported(RESULT IPO_SUPPORTED OUTPUT IPO_ERROR)
+if (NOT IPO_SUPPORTED)
+    message(WARNING "Interprocedural optimization is not supported by your toolchain! This will lead to bigger file sizes and worse performance: ${IPO_ERROR}")
+else()
+    message(STATUS "Interprocedural optimization support detected")
+endif()
+
+set(DIRECTORY deps/llama.cpp-mainline)
+include(llama.cpp.cmake)
+
+set(BUILD_VARIANTS)
+if (APPLE)
+    list(APPEND BUILD_VARIANTS metal)
+endif()
+if (LLMODEL_KOMPUTE)
+    list(APPEND BUILD_VARIANTS kompute kompute-avxonly)
+else()
+    list(PREPEND BUILD_VARIANTS cpu cpu-avxonly)
+endif()
+if (LLMODEL_VULKAN)
+    list(APPEND BUILD_VARIANTS vulkan vulkan-avxonly)
+endif()
+if (LLMODEL_CUDA)
+    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+    # Defaults must be set before enable_language(CUDA).
+    # Keep this in sync with the arch list in ggml/src/CMakeLists.txt (plus 5.0 for non-F16 branch).
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # 52 == lowest CUDA 12 standard
+        # 60 == f16 CUDA intrinsics
+        # 61 == integer CUDA intrinsics
+        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75") # needed for f16 CUDA intrinsics
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "50;52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
+            #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    include(CheckLanguage)
+    check_language(CUDA)
+    if (NOT CMAKE_CUDA_COMPILER)
+        message(WARNING "CUDA Toolkit not found. To build without CUDA, use -DLLMODEL_CUDA=OFF.")
+    endif()
+    enable_language(CUDA)
+    list(APPEND BUILD_VARIANTS cuda cuda-avxonly)
+endif()
+if (LLMODEL_ROCM)
+    enable_language(HIP)
+    list(APPEND BUILD_VARIANTS rocm rocm-avxonly)
+endif()
+
+# Go through each build variant
+foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
+    # Determine flags
+    if (BUILD_VARIANT MATCHES avxonly)
+        set(GPT4ALL_ALLOW_NON_AVX OFF)
+    else()
+        set(GPT4ALL_ALLOW_NON_AVX ON)
+    endif()
+    set(GGML_AVX2 ${GPT4ALL_ALLOW_NON_AVX})
+    set(GGML_F16C ${GPT4ALL_ALLOW_NON_AVX})
+    set(GGML_FMA  ${GPT4ALL_ALLOW_NON_AVX})
+
+    set(GGML_METAL   OFF)
+    set(GGML_KOMPUTE OFF)
+    set(GGML_VULKAN  OFF)
+    set(GGML_CUDA    OFF)
+    set(GGML_ROCM    OFF)
+    if (BUILD_VARIANT MATCHES metal)
+        set(GGML_METAL   ON)
+    elseif (BUILD_VARIANT MATCHES kompute)
+        set(GGML_KOMPUTE ON)
+    elseif (BUILD_VARIANT MATCHES vulkan)
+        set(GGML_VULKAN  ON)
+    elseif (BUILD_VARIANT MATCHES cuda)
+        set(GGML_CUDA    ON)
+    elseif (BUILD_VARIANT MATCHES rocm)
+        set(GGML_HIPBLAS ON)
+    endif()
+
+    # Include GGML
+    include_ggml(-mainline-${BUILD_VARIANT})
+
+    if (BUILD_VARIANT MATCHES metal)
+        set(GGML_METALLIB "${GGML_METALLIB}" PARENT_SCOPE)
+    endif()
+
+    # Function for preparing individual implementations
+    function(prepare_target TARGET_NAME BASE_LIB)
+        set(TARGET_NAME ${TARGET_NAME}-${BUILD_VARIANT})
+        message(STATUS "Configuring model implementation target ${TARGET_NAME}")
+        # Link to ggml/llama
+        target_link_libraries(${TARGET_NAME}
+            PRIVATE ${BASE_LIB}-${BUILD_VARIANT})
+        # Let it know about its build variant
+        target_compile_definitions(${TARGET_NAME}
+            PRIVATE GGML_BUILD_VARIANT="${BUILD_VARIANT}")
+        # Enable IPO if possible
+# FIXME: Doesn't work with msvc reliably. See https://github.com/nomic-ai/gpt4all/issues/841
+#        set_property(TARGET ${TARGET_NAME}
+#                     PROPERTY INTERPROCEDURAL_OPTIMIZATION ${IPO_SUPPORTED})
+    endfunction()
+
+    # Add each individual implementations
+    add_library(llamamodel-mainline-${BUILD_VARIANT} SHARED
+        src/llamamodel.cpp src/llmodel_shared.cpp)
+    gpt4all_add_warning_options(llamamodel-mainline-${BUILD_VARIANT})
+    target_compile_definitions(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
+        LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
+    target_include_directories(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
+        src include/gpt4all-backend
+    )
+    prepare_target(llamamodel-mainline llama-mainline)
+
+    if (NOT PROJECT_IS_TOP_LEVEL AND BUILD_VARIANT STREQUAL cuda)
+        set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
+    endif()
+endforeach()
+
+add_library(llmodel
+    src/dlhandle.cpp
+    src/llmodel.cpp
+    src/llmodel_c.cpp
+    src/llmodel_shared.cpp
+)
+gpt4all_add_warning_options(llmodel)
+target_sources(llmodel PUBLIC
+    FILE_SET public_headers TYPE HEADERS BASE_DIRS include
+    FILES include/gpt4all-backend/llmodel.h
+          include/gpt4all-backend/llmodel_c.h
+          include/gpt4all-backend/sysinfo.h
+)
+target_compile_definitions(llmodel PRIVATE LIB_FILE_EXT="${CMAKE_SHARED_LIBRARY_SUFFIX}")
+target_include_directories(llmodel PRIVATE src include/gpt4all-backend)
+
+set_target_properties(llmodel PROPERTIES
+                              VERSION ${PROJECT_VERSION}
+                              SOVERSION ${PROJECT_VERSION_MAJOR})
+
+set(COMPONENT_NAME_MAIN ${PROJECT_NAME})
+set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install)

gpt4all-backend-test/CMakeLists.txt

@@ -0,0 +1,10 @@
+cmake_minimum_required(VERSION 3.28...3.31)
+project(gpt4all-backend-test VERSION 0.1 LANGUAGES CXX)
+
+set(G4A_TEST_OLLAMA_URL "http://localhost:11434/" CACHE STRING "The base URL of the Ollama server to use.")
+
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin")
+include(../common/common.cmake)
+
+add_subdirectory(../gpt4all-backend gpt4all-backend)
+add_subdirectory(src)

gpt4all-backend-test/src/CMakeLists.txt

@@ -0,0 +1,15 @@
+set(TARGET test-backend)
+
+configure_file(config.h.in "${CMAKE_CURRENT_BINARY_DIR}/include/config.h")
+
+add_executable(${TARGET}
+    main.cpp
+)
+target_compile_features(${TARGET} PUBLIC cxx_std_23)
+target_include_directories(${TARGET} PRIVATE
+    "${CMAKE_CURRENT_BINARY_DIR}/include"
+)
+gpt4all_add_warning_options(${TARGET})
+target_link_libraries(${TARGET} PRIVATE
+    gpt4all-backend
+)

gpt4all-backend-test/src/config.h.in

@@ -0,0 +1,6 @@
+#pragma once
+
+#include <QString>
+
+
+inline const QString OLLAMA_URL = QStringLiteral("@G4A_TEST_OLLAMA_URL@");

gpt4all-backend-test/src/main.cpp

@@ -0,0 +1,70 @@
+#include "config.h"
+
+#include "pretty.h"
+
+#include <QCoro/QCoroTask> // IWYU pragma: keep
+#include <boost/json.hpp>
+#include <fmt/base.h>
+#include <gpt4all-backend/formatters.h> // IWYU pragma: keep
+#include <gpt4all-backend/ollama-client.h>
+
+#include <QCoreApplication>
+#include <QTimer>
+#include <QString>
+#include <QUrl>
+
+#include <coroutine>
+#include <expected>
+#include <variant>
+
+namespace json = boost::json;
+using namespace Qt::Literals::StringLiterals;
+using gpt4all::backend::OllamaClient;
+
+
+template <typename T>
+static std::string to_json(const T &value)
+{ return pretty_print(json::value_from(value)); }
+
+static void run()
+{
+    fmt::print("Connecting to server at {}\n", OLLAMA_URL);
+    OllamaClient provider(OLLAMA_URL);
+
+    auto versionResp = QCoro::waitFor(provider.version());
+    if (versionResp) {
+        fmt::print("Version response: {}\n", to_json(*versionResp));
+    } else {
+        fmt::print("Error retrieving version: {}\n", versionResp.error().errorString);
+        return QCoreApplication::exit(1);
+    }
+
+    auto modelsResponse = QCoro::waitFor(provider.list());
+    if (modelsResponse) {
+        fmt::print("Available models:\n");
+        for (const auto & model : modelsResponse->models)
+            fmt::print("{}\n", model.model);
+        if (!modelsResponse->models.empty())
+            fmt::print("First model: {}\n", to_json(modelsResponse->models.front()));
+    } else {
+        fmt::print("Error retrieving available models: {}\n", modelsResponse.error().errorString);
+        return QCoreApplication::exit(1);
+    }
+
+    auto showResponse = QCoro::waitFor(provider.show({ .model = "DeepSeek-R1-Distill-Llama-70B-Q4_K_S" }));
+    if (showResponse) {
+        fmt::print("Show response: {}\n", to_json(*showResponse));
+    } else {
+        fmt::print("Error retrieving model info: {}\n", showResponse.error().errorString);
+        return QCoreApplication::exit(1);
+    }
+
+    QCoreApplication::exit(0);
+}
+
+int main(int argc, char *argv[])
+{
+    QCoreApplication app(argc, argv);
+    QTimer::singleShot(0, &run);
+    return app.exec();
+}

gpt4all-backend-test/src/pretty.h

@@ -0,0 +1,95 @@
+#pragma once
+
+#include <boost/json.hpp>
+
+#include <sstream>
+#include <string>
+
+
+inline void pretty_print( std::ostream& os, boost::json::value const& jv, std::string* indent = nullptr )
+{
+    std::string indent_;
+    if(! indent)
+        indent = &indent_;
+    switch(jv.kind())
+    {
+    case boost::json::kind::object:
+    {
+        os << "{\n";
+        indent->append(4, ' ');
+        auto const& obj = jv.get_object();
+        if(! obj.empty())
+        {
+            auto it = obj.begin();
+            for(;;)
+            {
+                os << *indent << boost::json::serialize(it->key()) << ": ";
+                pretty_print(os, it->value(), indent);
+                if(++it == obj.end())
+                    break;
+                os << ",\n";
+            }
+        }
+        os << "\n";
+        indent->resize(indent->size() - 4);
+        os << *indent << "}";
+        break;
+    }
+
+    case boost::json::kind::array:
+    {
+        os << "[\n";
+        indent->append(4, ' ');
+        auto const& arr = jv.get_array();
+        if(! arr.empty())
+        {
+            auto it = arr.begin();
+            for(;;)
+            {
+                os << *indent;
+                pretty_print( os, *it, indent);
+                if(++it == arr.end())
+                    break;
+                os << ",\n";
+            }
+        }
+        os << "\n";
+        indent->resize(indent->size() - 4);
+        os << *indent << "]";
+        break;
+    }
+
+    case boost::json::kind::string:
+    {
+        os << boost::json::serialize(jv.get_string());
+        break;
+    }
+
+    case boost::json::kind::uint64:
+    case boost::json::kind::int64:
+    case boost::json::kind::double_:
+        os << jv;
+        break;
+
+    case boost::json::kind::bool_:
+        if(jv.get_bool())
+            os << "true";
+        else
+            os << "false";
+        break;
+
+    case boost::json::kind::null:
+        os << "null";
+        break;
+    }
+
+    if(indent->empty())
+        os << "\n";
+}
+
+inline std::string pretty_print( boost::json::value const& jv, std::string* indent = nullptr )
+{
+    std::ostringstream ss;
+    pretty_print(ss, jv, indent);
+    return ss.str();
+}

gpt4all-backend/CMakeLists.txt

@@ -1,189 +1,22 @@
-cmake_minimum_required(VERSION 3.23)  # for FILE_SET
+cmake_minimum_required(VERSION 3.28...3.31)
+project(gpt4all-backend VERSION 0.1 LANGUAGES CXX)
 
+set(CMAKE_CXX_STANDARD 23)  # make sure fmt is compiled with the same C++ version as us
 include(../common/common.cmake)
 
-set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
-set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+find_package(Qt6 6.8 COMPONENTS Concurrent Core Network REQUIRED)
 
-if (APPLE)
-    option(BUILD_UNIVERSAL "Build a Universal binary on macOS" ON)
-else()
-    option(LLMODEL_KOMPUTE "llmodel: use Kompute"              ON)
-    option(LLMODEL_VULKAN  "llmodel: use Vulkan"               OFF)
-    option(LLMODEL_CUDA    "llmodel: use CUDA"                 ON)
-    option(LLMODEL_ROCM    "llmodel: use ROCm"                 OFF)
-endif()
+add_subdirectory(../deps common_deps)
+add_subdirectory(deps)
+add_subdirectory(src)
 
-if (APPLE)
-  if (BUILD_UNIVERSAL)
-    # Build a Universal binary on macOS
-    # This requires that the found Qt library is compiled as Universal binaries.
-    set(CMAKE_OSX_ARCHITECTURES "arm64;x86_64" CACHE STRING "" FORCE)
-  else()
-    # Build for the host architecture on macOS
-    if (NOT CMAKE_OSX_ARCHITECTURES)
-      set(CMAKE_OSX_ARCHITECTURES "${CMAKE_HOST_SYSTEM_PROCESSOR}" CACHE STRING "" FORCE)
-    endif()
-  endif()
-endif()
-
-# Include the binary directory for the generated header file
-include_directories("${CMAKE_CURRENT_BINARY_DIR}")
-
-set(LLMODEL_VERSION_MAJOR 0)
-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)
-
-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)
-
-# Check for IPO support
-include(CheckIPOSupported)
-check_ipo_supported(RESULT IPO_SUPPORTED OUTPUT IPO_ERROR)
-if (NOT IPO_SUPPORTED)
-    message(WARNING "Interprocedural optimization is not supported by your toolchain! This will lead to bigger file sizes and worse performance: ${IPO_ERROR}")
-else()
-    message(STATUS "Interprocedural optimization support detected")
-endif()
-
-set(DIRECTORY deps/llama.cpp-mainline)
-include(llama.cpp.cmake)
-
-set(BUILD_VARIANTS)
-if (APPLE)
-    list(APPEND BUILD_VARIANTS metal)
-endif()
-if (LLMODEL_KOMPUTE)
-    list(APPEND BUILD_VARIANTS kompute kompute-avxonly)
-else()
-    list(PREPEND BUILD_VARIANTS cpu cpu-avxonly)
-endif()
-if (LLMODEL_VULKAN)
-    list(APPEND BUILD_VARIANTS vulkan vulkan-avxonly)
-endif()
-if (LLMODEL_CUDA)
-    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
-
-    # Defaults must be set before enable_language(CUDA).
-    # Keep this in sync with the arch list in ggml/src/CMakeLists.txt.
-    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        # 52 == lowest CUDA 12 standard
-        # 60 == f16 CUDA intrinsics
-        # 61 == integer CUDA intrinsics
-        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75") # needed for f16 CUDA intrinsics
-        else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
-            #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
-        endif()
-    endif()
-    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
-
-    include(CheckLanguage)
-    check_language(CUDA)
-    if (NOT CMAKE_CUDA_COMPILER)
-        message(WARNING "CUDA Toolkit not found. To build without CUDA, use -DLLMODEL_CUDA=OFF.")
-    endif()
-    enable_language(CUDA)
-    list(APPEND BUILD_VARIANTS cuda cuda-avxonly)
-endif()
-if (LLMODEL_ROCM)
-    enable_language(HIP)
-    list(APPEND BUILD_VARIANTS rocm rocm-avxonly)
-endif()
-
-# Go through each build variant
-foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
-    # Determine flags
-    if (BUILD_VARIANT MATCHES avxonly)
-        set(GPT4ALL_ALLOW_NON_AVX OFF)
-    else()
-        set(GPT4ALL_ALLOW_NON_AVX ON)
-    endif()
-    set(GGML_AVX2 ${GPT4ALL_ALLOW_NON_AVX})
-    set(GGML_F16C ${GPT4ALL_ALLOW_NON_AVX})
-    set(GGML_FMA  ${GPT4ALL_ALLOW_NON_AVX})
-
-    set(GGML_METAL   OFF)
-    set(GGML_KOMPUTE OFF)
-    set(GGML_VULKAN  OFF)
-    set(GGML_CUDA    OFF)
-    set(GGML_ROCM    OFF)
-    if (BUILD_VARIANT MATCHES metal)
-        set(GGML_METAL   ON)
-    elseif (BUILD_VARIANT MATCHES kompute)
-        set(GGML_KOMPUTE ON)
-    elseif (BUILD_VARIANT MATCHES vulkan)
-        set(GGML_VULKAN  ON)
-    elseif (BUILD_VARIANT MATCHES cuda)
-        set(GGML_CUDA    ON)
-    elseif (BUILD_VARIANT MATCHES rocm)
-        set(GGML_HIPBLAS ON)
-    endif()
-
-    # Include GGML
-    include_ggml(-mainline-${BUILD_VARIANT})
-
-    if (BUILD_VARIANT MATCHES metal)
-        set(GGML_METALLIB "${GGML_METALLIB}" PARENT_SCOPE)
-    endif()
-
-    # Function for preparing individual implementations
-    function(prepare_target TARGET_NAME BASE_LIB)
-        set(TARGET_NAME ${TARGET_NAME}-${BUILD_VARIANT})
-        message(STATUS "Configuring model implementation target ${TARGET_NAME}")
-        # Link to ggml/llama
-        target_link_libraries(${TARGET_NAME}
-            PRIVATE ${BASE_LIB}-${BUILD_VARIANT})
-        # Let it know about its build variant
-        target_compile_definitions(${TARGET_NAME}
-            PRIVATE GGML_BUILD_VARIANT="${BUILD_VARIANT}")
-        # Enable IPO if possible
-# FIXME: Doesn't work with msvc reliably. See https://github.com/nomic-ai/gpt4all/issues/841
-#        set_property(TARGET ${TARGET_NAME}
-#                     PROPERTY INTERPROCEDURAL_OPTIMIZATION ${IPO_SUPPORTED})
-    endfunction()
-
-    # Add each individual implementations
-    add_library(llamamodel-mainline-${BUILD_VARIANT} SHARED
-        src/llamamodel.cpp src/llmodel_shared.cpp)
-    gpt4all_add_warning_options(llamamodel-mainline-${BUILD_VARIANT})
-    target_compile_definitions(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
-        LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-    target_include_directories(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
-        src include/gpt4all-backend
-    )
-    prepare_target(llamamodel-mainline llama-mainline)
-
-    if (NOT PROJECT_IS_TOP_LEVEL AND BUILD_VARIANT STREQUAL cuda)
-        set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
-    endif()
-endforeach()
-
-add_library(llmodel
-    src/dlhandle.cpp
-    src/llmodel.cpp
-    src/llmodel_c.cpp
-    src/llmodel_shared.cpp
+target_sources(gpt4all-backend PUBLIC
+    FILE_SET public_headers TYPE HEADERS BASE_DIRS include FILES
+        include/gpt4all-backend/formatters.h
+        include/gpt4all-backend/generation-params.h
+        include/gpt4all-backend/json-helpers.h
+        include/gpt4all-backend/ollama-client.h
+        include/gpt4all-backend/ollama-model.h
+        include/gpt4all-backend/ollama-types.h
+        include/gpt4all-backend/rest.h
 )
-gpt4all_add_warning_options(llmodel)
-target_sources(llmodel PUBLIC
-    FILE_SET public_headers TYPE HEADERS BASE_DIRS include
-    FILES include/gpt4all-backend/llmodel.h
-          include/gpt4all-backend/llmodel_c.h
-          include/gpt4all-backend/sysinfo.h
-)
-target_compile_definitions(llmodel PRIVATE LIB_FILE_EXT="${CMAKE_SHARED_LIBRARY_SUFFIX}")
-target_include_directories(llmodel PRIVATE src include/gpt4all-backend)
-
-set_target_properties(llmodel PROPERTIES
-                              VERSION ${PROJECT_VERSION}
-                              SOVERSION ${PROJECT_VERSION_MAJOR})
-
-set(COMPONENT_NAME_MAIN ${PROJECT_NAME})
-set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install)

gpt4all-backend/deps/CMakeLists.txt

@@ -0,0 +1,5 @@
+include(FetchContent)
+
+set(BUILD_SHARED_LIBS OFF)
+
+add_subdirectory(date)

gpt4all-backend/include/gpt4all-backend/formatters.h

@@ -0,0 +1,33 @@
+#pragma once
+
+#include <fmt/base.h>
+#include <fmt/format.h>
+
+#include <QByteArray>
+#include <QString>
+#include <QStringView>
+#include <QUtf8StringView>
+#include <QVariant>
+
+#include <string_view>
+
+
+// fmtlib formatters for QString and QVariant
+
+#define MAKE_FORMATTER(type, conversion)                                         \
+    template <>                                                                  \
+    struct fmt::formatter<type, char> : fmt::formatter<std::string_view, char> { \
+        template <typename FmtContext>                                           \
+        FmtContext::iterator format(const type &value, FmtContext &ctx) const    \
+        {                                                                        \
+            auto valueUtf8 = (conversion);                                       \
+            std::string_view view(valueUtf8.cbegin(), valueUtf8.cend());         \
+            return formatter<std::string_view, char>::format(view, ctx);         \
+        }                                                                        \
+    }
+
+MAKE_FORMATTER(QLatin1StringView, value                    );
+MAKE_FORMATTER(QString,           value.toUtf8()           );
+MAKE_FORMATTER(QStringView,       value.toUtf8()           );
+MAKE_FORMATTER(QUtf8StringView,   value                    );
+MAKE_FORMATTER(QVariant,          value.toString().toUtf8());

gpt4all-backend/include/gpt4all-backend/generation-params.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include <QtTypes>
+
+
+namespace gpt4all::backend {
+
+
+struct GenerationParams {
+    uint  n_predict;
+    float temperature;
+    float top_p;
+    // int32_t top_k = 40;
+    // float   min_p = 0.0f;
+    // int32_t n_batch = 9;
+    // float   repeat_penalty = 1.10f;
+    // int32_t repeat_last_n = 64;     // last n tokens to penalize
+    // float   contextErase = 0.5f;    // percent of context to erase if we exceed the context window
+};
+
+
+} // namespace gpt4all::backend

gpt4all-backend/include/gpt4all-backend/json-helpers.h

@@ -0,0 +1,15 @@
+#pragma once
+
+class QString;
+namespace boost::json {
+    class value;
+    struct value_from_tag;
+    template <typename T> struct value_to_tag;
+}
+
+
+/// Allows QString to be serialized to JSON.
+void tag_invoke(const boost::json::value_from_tag &, boost::json::value &value, const QString &qstr);
+
+/// Allows JSON strings to be deserialized as QString.
+QString tag_invoke(const boost::json::value_to_tag<QString> &, const boost::json::value &value);