WIP

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

.circleci/config.yml

@@ -1,19 +1,20 @@
 version: 2.1
 setup: true
 orbs:
-  path-filtering: circleci/path-filtering@0.0.1
+  path-filtering: circleci/path-filtering@1.1.0
 
 workflows:
   version: 2.1
   generate-config:
     jobs:
       - path-filtering/filter:
+          filters:
+            tags:
+              only:
+                - /.*/
           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-chat/.* run-chat-workflow true
-            .* run-default-workflow true

.codespellrc

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

.gitignore

@@ -181,6 +181,8 @@ CMakeLists.txt.user
 gpt4all-chat/models/*
 build_*
 build-*
+cmake-build-*
+/gpt4all-chat/tests/python/config.py
 
 # IntelliJ
 .idea/

.gitmodules

@@ -1,7 +1,28 @@
-[submodule "llama.cpp-mainline"]
-	path = gpt4all-backend/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/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 = 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
+[submodule "gpt4all-chat/deps/QXlsx"]
+	path = gpt4all-chat/deps/QXlsx
+	url = https://github.com/nomic-ai/QXlsx.git
+[submodule "gpt4all-chat/deps/minja"]
+	path = gpt4all-chat/deps/minja
+	url = https://github.com/nomic-ai/minja.git
+[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/qcoro/qcoro.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/>
@@ -45,17 +38,12 @@ 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/>
 - pt\_BR translation
 
-Victor Emanuel ([@SINAPSA-IC](https://github.com/SINAPSA-IC))<br/>
-E-mail: contact@sinapsaro.ro<br/>
-Discord: `@sinapsa_ic_56124_99632`
-- ro\_RO translation
-
 不知火 Shiranui ([@supersonictw](https://github.com/supersonictw))<br/>
 E-mail: supersonic@livemail.tw<br/>
 Discord: `@supersonictw`
@@ -77,6 +65,6 @@ Discord: `@Tim453`
 - Flatpak
 
 Jack ([@wuodoo](https://github.com/wuodoo))<br/>
-E-mail: 2296103047@qq.com><br/>
+E-mail: 2296103047@qq.com<br/>
 Discord: `@mikage`
 - zh\_CN translation

README.md

@@ -1,67 +1,67 @@
 <h1 align="center">GPT4All</h1>
 
-<p align="center">GPT4All runs large language models (LLMs) privately on everyday desktops & laptops. <br> <br> No API calls or GPUs required - you can just download the application and <a href="https://docs.gpt4all.io/gpt4all_desktop/quickstart.html#quickstart">get started</a>
+<p align="center">
+  Now with support for DeepSeek R1 Distillations
+</p>
+
+<p align="center">
+  <a href="https://www.nomic.ai/gpt4all">Website</a> &bull; <a href="https://docs.gpt4all.io">Documentation</a> &bull; <a href="https://discord.gg/mGZE39AS3e">Discord</a> &bull; <a href="https://www.youtube.com/watch?v=gQcZDXRVJok">YouTube Tutorial</a>
+</p>
+
+<p align="center">
+  GPT4All runs large language models (LLMs) privately on everyday desktops & laptops.
+</p>
+<p align="center">
+  No API calls or GPUs required - you can just download the application and <a href="https://docs.gpt4all.io/gpt4all_desktop/quickstart.html#quickstart">get started</a>.
+</p>
+
+<p align="center">
+  Read about what's new in <a href="https://www.nomic.ai/blog/tag/gpt4all">our blog</a>.
+</p>
+<p align="center">
+  <a href="https://nomic.ai/gpt4all/#newsletter-form">Subscribe to the newsletter</a>
+</p>
 
 https://github.com/nomic-ai/gpt4all/assets/70534565/513a0f15-4964-4109-89e4-4f9a9011f311
 
-<p align="center">
-  <a href="https://gpt4all.io/installers/gpt4all-installer-win64.exe">
-    <img src="gpt4all-bindings/python/docs/assets/windows.png" width="80" height="80"><br>
-    Download for Windows
-  </a>
-</p>
-
-<p align="center">
-  <a href="https://gpt4all.io/installers/gpt4all-installer-darwin.dmg">
-    <img src="gpt4all-bindings/python/docs/assets/mac.png" width="85" height="100"><br>
-    Download for MacOS
-  </a>
-</p>
-
-<p align="center">
-  <a href="https://gpt4all.io/installers/gpt4all-installer-linux.run">
-    <img src="gpt4all-bindings/python/docs/assets/ubuntu.svg" width="120" height="120"><br>
-    Download for Ubuntu
-  </a>
-</p>
-
-<p align="center">
-  <a href='https://flathub.org/apps/io.gpt4all.gpt4all'>
-    <img width='240' alt='Get it on Flathub' src='https://flathub.org/api/badge?locale=en'><br>
-    Get it on Flathub (community maintained)
-  </a>
-</p>
-
-<p align="center">
-  <a href="https://gpt4all.io">Website</a> &bull; <a href="https://docs.gpt4all.io">Documentation</a> &bull; <a href="https://discord.gg/mGZE39AS3e">Discord</a>
-</p>
-<p align="center">
-  <a href="https://forms.nomic.ai/gpt4all-release-notes-signup">Subscribe to the newsletter</a>
-</p>
 <p align="center">
 GPT4All is made possible by our compute partner <a href="https://www.paperspace.com/">Paperspace</a>.
 </p>
-<p align="center">
- <a href="https://www.phorm.ai/query?projectId=755eecd3-24ad-49cc-abf4-0ab84caacf63"><img src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg" alt="phorm.ai"></a>
+
+## Download Links
+
+<p>
+  &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-win64.exe">
+    <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-darwin.dmg">
+    <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="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.
+</p>
+<p>
+  macOS requires Monterey 12.6 or newer. Best results with Apple Silicon M-series processors.
 </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))
-```
+See the full [System Requirements](gpt4all-chat/system_requirements.md) for more details.
 
+<br/>
+<br/>
+<p>
+  <a href='https://flathub.org/apps/io.gpt4all.gpt4all'>
+    <img style="height: 2em; width: auto" alt='Get it on Flathub' src='https://flathub.org/api/badge'><br/>
+    Flathub (community maintained)
+  </a>
+</p>
 
 ## Integrations
 
@@ -75,7 +75,7 @@ with model.chat_session():
     - Improved user workflow for LocalDocs
     - Expanded access to more model architectures
 - **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
+    - Mistral 7b base model, an updated model gallery on our website, several new local code models including Rift Coder v1.5
     - [Nomic Vulkan](https://blog.nomic.ai/posts/gpt4all-gpu-inference-with-vulkan) support for Q4\_0 and Q4\_1 quantizations in GGUF.
     - Offline build support for running old versions of the GPT4All Local LLM Chat Client.
 - **September 18th, 2023**: [Nomic Vulkan](https://blog.nomic.ai/posts/gpt4all-gpu-inference-with-vulkan) launches supporting local LLM inference on NVIDIA and AMD GPUs.
@@ -90,7 +90,7 @@ 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

@@ -0,0 +1,46 @@
+# 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()
+    endif()
+    target_compile_options("${target}" PRIVATE
+        # base options
+        -Wall
+        -Wextra
+        # extra options
+        -Wcast-align
+        -Wextra-semi
+        -Wformat=2
+        -Wmissing-include-dirs
+        -Wsuggest-override
+        -Wvla
+        # errors
+        -Werror=format-security
+        -Werror=init-self
+        -Werror=pointer-arith
+        -Werror=undef
+        # disabled warnings
+        -Wno-sign-compare
+        -Wno-unused-parameter
+    )
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+        target_compile_options("${target}" PRIVATE
+            -Wduplicated-branches
+            -Wduplicated-cond
+            -Wlogical-op
+            -Wno-reorder
+            -Wno-null-dereference
+        )
+    elseif (CMAKE_CXX_COMPILER_ID MATCHES "^(Apple)?Clang$")
+        target_compile_options("${target}" PRIVATE
+            -Wunreachable-code-break
+            -Wunreachable-code-return
+            -Werror=pointer-integer-compare
+            -Wno-reorder-ctor
+        )
+    endif()
+endfunction()

deps/CMakeLists.txt

@@ -0,0 +1,14 @@
+set(BUILD_SHARED_LIBS OFF)
+
+set(FMT_INSTALL OFF)
+set(FMT_MODULE ON)
+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_QML          OFF)
+set(QCORO_WITH_QTTEST       OFF)
+add_subdirectory(qcoro)

docs/gpt4all_api_server/home.md

@@ -0,0 +1,86 @@
+# GPT4All API Server
+
+GPT4All provides a local API server that allows you to run LLMs over an HTTP API. 
+
+## Key Features
+
+- **Local Execution**: Run models on your own hardware for privacy and offline use.
+- **LocalDocs Integration**: Run the API with relevant text snippets provided to your LLM from a [LocalDocs collection](../gpt4all_desktop/localdocs.md).
+- **OpenAI API Compatibility**: Use existing OpenAI-compatible clients and tools with your local models.
+
+## Activating the API Server
+
+1. Open the GPT4All Chat Desktop Application.
+2. Go to `Settings` > `Application` and scroll down to `Advanced`.
+3. Check the box for the `"Enable Local API Server"` setting.
+4. The server listens on port 4891 by default. You can choose another port number in the `"API Server Port"` setting.
+
+## Connecting to the API Server
+
+The base URL used for the API server is `http://localhost:4891/v1` (or `http://localhost:<PORT_NUM>/v1` if you are using a different port number). 
+
+The server only accepts HTTP connections (not HTTPS) and only listens on localhost (127.0.0.1) (e.g. not to the IPv6 localhost address `::1`.)
+
+## Examples
+
+!!! note "Example GPT4All API calls"
+
+    === "cURL"
+
+        ```bash
+        curl -X POST http://localhost:4891/v1/chat/completions -d '{
+        "model": "Phi-3 Mini Instruct",
+        "messages": [{"role":"user","content":"Who is Lionel Messi?"}],
+        "max_tokens": 50,
+        "temperature": 0.28
+        }'
+        ```
+
+    === "PowerShell"
+
+        ```powershell
+        Invoke-WebRequest -URI http://localhost:4891/v1/chat/completions -Method POST -ContentType application/json -Body '{
+        "model": "Phi-3 Mini Instruct",
+        "messages": [{"role":"user","content":"Who is Lionel Messi?"}],
+        "max_tokens": 50,
+        "temperature": 0.28
+        }'
+        ```
+
+## API Endpoints
+
+| Method | Path | Description |
+|--------|------|-------------|
+| GET | `/v1/models` | List available models |
+| GET | `/v1/models/<name>` | Get details of a specific model |
+| POST | `/v1/completions` | Generate text completions |
+| POST | `/v1/chat/completions` | Generate chat completions |
+
+## LocalDocs Integration
+
+You can use LocalDocs with the API server:
+
+1. Open the Chats view in the GPT4All application.
+2. Scroll to the bottom of the chat history sidebar.
+3. Select the server chat (it has a different background color).
+4. Activate LocalDocs collections in the right sidebar.
+
+(Note: LocalDocs can currently only be activated through the GPT4All UI, not via the API itself).
+
+Now, your API calls to your local LLM will have relevant references from your LocalDocs collection retrieved and placed in the input message for the LLM to respond to.
+
+The references retrieved for your API call can be accessed in the API response object at 
+
+`response["choices"][0]["references"]`
+
+The data included in the `references` are:
+
+- `text`: the actual text content from the snippet that was extracted from the reference document
+
+- `author`: the author of the reference document (if available)
+
+- `date`: the date of creation of the reference document (if available)
+
+- `page`: the page number the snippet is from (only available for PDF documents for now)
+
+- `title`: the title of the reference document (if available)

docs/gpt4all_desktop/chat_templates.md

@@ -0,0 +1,206 @@
+## What are chat templates?
+Natively, large language models only know how to complete plain text and do not know the difference between their input and their output. In order to support a chat with a person, LLMs are designed to use a template to convert the conversation to plain text using a specific format.
+
+For a given model, it is important to use an appropriate chat template, as each model is designed to work best with a specific format. The chat templates included with the built-in models should be sufficient for most purposes.
+
+There are two reasons you would want to alter the chat template:
+
+- You are sideloading a model and there is no chat template available,
+- You would like to have greater control over the input to the LLM than a system message provides.
+
+
+## What is a system message?
+A system message is a message that controls the responses from the LLM in a way that affects the entire conversation. System messages can be short, such as "Speak like a pirate.", or they can be long and contain a lot of context for the LLM to keep in mind.
+
+Not all models are designed to use a system message, so they work with some models better than others.
+
+
+## How do I customize the chat template or system message?
+To customize the chat template or system message, go to Settings > Model. Make sure to select the correct model at the top. If you clone a model, you can use a different chat template or system message from the base model, enabling you to use different settings for each conversation.
+
+These settings take effect immediately. After changing them, you can click "Redo last response" in the chat view, and the response will take the new settings into account.
+
+
+## Do I need to write a chat template?
+You typically do not need to write your own chat template. The exception is models that are not in the official model list and do not come with a chat template built-in. These will show a "Clear" option above the chat template field in the Model Settings page instead of a "Reset" option. See the section on [finding] or [creating] a chat template.
+
+[finding]: #how-do-i-find-a-chat-template
+[creating]: #advanced-how-do-chat-templates-work
+
+
+## What changed in GPT4All v3.5?
+GPT4All v3.5 overhauled the chat template system. There are three crucial differences:
+
+- The chat template now formats an entire conversation instead of a single pair of messages,
+- The chat template now uses Jinja syntax instead of `%1` and `%2` placeholders,
+- And the system message should no longer contain control tokens or trailing whitespace.
+
+If you are using any chat templates or system messages that had been added or altered from the default before upgrading to GPT4All v3.5 or newer, these will no longer work. See below for how to solve common errors you may see after upgrading.
+
+
+## Error/Warning: System message is not plain text.
+This is easy to fix. Go to the model's settings and look at the system prompt. There are three things to look for:
+
+- Control tokens such as `<|im_start|>`, `<|start_header_id|>`, or `<|system|>`
+- A prefix such as `### System` or `SYSTEM:`
+- Trailing whitespace, such as a space character or blank line.
+
+If you see any of these things, remove them. For example, this legacy system prompt:
+```
+<|start_header_id|>system<|end_header_id|>
+You are a helpful assistant.<|eot_id|>
+```
+
+Should become this:
+```
+You are a helpful assistant.
+```
+
+If you do not see anything that needs to be changed, you can dismiss the error by making a minor modification to the message and then changing it back.
+
+If you see a warning, your system message does not appear to be plain text. If you believe this warning is incorrect, it can be safely ignored. If in doubt, ask on the [Discord].
+
+[Discord]: https://discord.gg/mGZE39AS3e
+
+
+## Error: Legacy system prompt needs to be updated in Settings.
+This is the same as [above][above-1], but appears on the chat page.
+
+[above-1]: #errorwarning-system-message-is-not-plain-text
+
+
+## Error/Warning: Chat template is not in Jinja format.
+This is the result of attempting to use an old-style template (possibly from a previous version) in GPT4All 3.5+.
+
+Go to the Model Settings page and select the affected model. If you see a "Reset" button, and you have not intentionally modified the prompt template, you can click "Reset". Otherwise, this is what you can do:
+
+1. Back up your chat template by copying it safely to a text file and saving it. In the next step, it will be removed from GPT4All.
+2. Click "Reset" or "Clear".
+3. If you clicked "Clear", the chat template is now gone. Follow the steps to [find][finding] or [create][creating] a basic chat template for your model.
+4. Customize the chat template to suit your needs. For help, read the section about [creating] a chat template.
+
+
+## Error: Legacy prompt template needs to be updated in Settings.
+This is the same as [above][above-2], but appears on the chat page.
+
+[above-2]: #errorwarning-chat-template-is-not-in-jinja-format
+
+
+## The chat template has a syntax error.
+If there is a syntax error while editing the chat template, the details will be displayed in an error message above the input box. This could be because the chat template is not actually in Jinja format (see [above][above-2]).
+
+Otherwise, you have either typed something correctly, or the model comes with a template that is incompatible with GPT4All. See [the below section][creating] on creating chat templates and make sure that everything is correct. When in doubt, ask on the [Discord].
+
+
+## Error: No chat template configured.
+This may appear for models that are not from the official model list and do not include a chat template. Older versions of GPT4All picked a poor default in this case. You will get much better results if you follow the steps to [find][finding] or [create][creating] a chat template for your model.
+
+
+## Error: The chat template cannot be blank.
+If the button above the chat template on the Model Settings page says "Clear", see [above][above-3]. If you see "Reset", click that button to restore a reasonable default. Also see the section on [syntax errors][chat-syntax-error].
+
+[above-3]: #error-no-chat-template-configured
+[chat-syntax-error]: #the-chat-template-has-a-syntax-error
+
+
+## How do I find a chat template?
+When in doubt, you can always ask the [Discord] community for help. Below are the instructions to find one on your own.
+
+The authoritative source for a model's chat template is the HuggingFace repo that the original (non-GGUF) model came from. First, you should find this page. If you just have a model file, you can try a google search for the model's name. If you know the page you downloaded the GGUF model from, its README usually links to the original non-GGUF model.
+
+Once you have located the original model, there are two methods you can use to extract its chat template. Pick whichever one you are most comfortable with.
+
+### Using the CLI (all models)
+1. Install `jq` using your preferred package manager - e.g. Chocolatey (Windows), Homebrew (macOS), or apt (Ubuntu).
+2. Download `tokenizer_config.json` from the model's "Files and versions" tab.
+3. Open a command prompt in the directory which you have downloaded the model file.
+4. Run `jq -r ".chat_template" tokenizer_config.json`. This shows the chat template in a human-readable form. You can copy this and paste it into the settings page.
+5. (Optional) You can save the output to a text file like this: `jq -r ".chat_template" tokenizer_config.json >chat_template.txt`
+
+If the output is "null", the model does not provide a chat template. See the [below instructions][creating] on creating a chat template.
+
+### Python (open models)
+1. Install `transformers` using your preferred python package manager, e.g. `pip install transformers`. Make sure it is at least version v4.43.0.
+2. Copy the ID of the HuggingFace model, using the clipboard icon next to the name. For example, if the URL is `https://huggingface.co/NousResearch/Hermes-2-Pro-Llama-3-8B`, the ID is `NousResearch/Hermes-2-Pro-Llama-3-8B`.
+3. Open a python interpreter (`python`) and run the following commands. Change the model ID in the example to the one you copied.
+```
+>>> from transformers import AutoTokenizer
+>>> tokenizer = AutoTokenizer.from_pretrained('NousResearch/Hermes-2-Pro-Llama-3-8B')
+>>> print(tokenizer.get_chat_template())
+```
+You can copy the output and paste it into the settings page.
+4. (Optional) You can save the output to a text file like this:
+```
+>>> open('chat_template.txt', 'w').write(tokenizer.get_chat_template())
+```
+
+If you get a ValueError exception, this model does not provide a chat template. See the [below instructions][creating] on creating a chat template.
+
+
+### Python (gated models)
+Some models, such as Llama and Mistral, do not allow public access to their chat template. You must either use the CLI method above, or follow the following instructions to use Python:
+
+1. For these steps, you must have git and git-lfs installed.
+2. You must have a HuggingFace account and be logged in.
+3. You must already have access to the gated model. Otherwise, request access.
+4. You must have an SSH key configured for git access to HuggingFace.
+5. `git clone` the model's HuggingFace repo using the SSH clone URL. There is no need to download the entire model, which is very large. A good way to do this on Linux is:
+```console
+$ GIT_LFS_SKIP_SMUDGE=1 git clone hf.co:meta-llama/Llama-3.1-8B-Instruct.git
+$ cd Llama-3.1-8B-Instruct
+$ git lfs pull -I "tokenizer.*"
+```
+6. Follow the above instructions for open models, but replace the model ID with the path to the directory containing `tokenizer\_config.json`:
+```
+>>> tokenizer = AutoTokenizer.from_pretrained('.')
+```
+
+
+## Advanced: How do chat templates work?
+The chat template is applied to the entire conversation you see in the chat window. The template loops over the list of messages, each containing `role` and `content` fields. `role` is either `user`, `assistant`, or `system`.
+
+GPT4All also supports the special variables `bos_token`, `eos_token`, and `add_generation_prompt`. See the [HuggingFace docs] for what those do.
+
+[HuggingFace docs]: https://huggingface.co/docs/transformers/v4.46.3/en/chat_templating#special-variables
+
+
+## Advanced: How do I make a chat template?
+The best way to create a chat template is to start by using an existing one as a reference. Then, modify it to use the format documented for the given model. Its README page may explicitly give an example of its template. Or, it may mention the name of a well-known standard template, such as ChatML, Alpaca, Vicuna. GPT4All does not yet include presets for these templates, so they will have to be found in other models or taken from the community.
+
+For more information, see the very helpful [HuggingFace guide]. Some of this is not applicable, such as the information about tool calling and RAG - GPT4All implements those features differently.
+
+Some models use a prompt template that does not intuitively map to a multi-turn chat, because it is more intended for single instructions. The [FastChat] implementation of these templates is a useful reference for the correct way to extend them to multiple messages.
+
+[HuggingFace guide]: https://huggingface.co/docs/transformers/v4.46.3/en/chat_templating#advanced-template-writing-tips
+[FastChat]: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+
+
+# Advanced: What are GPT4All v1 templates?
+GPT4All supports its own template syntax, which is nonstandard but provides complete control over the way LocalDocs sources and file attachments are inserted into the conversation. These templates begin with `{# gpt4all v1 #}` and look similar to the example below.
+
+For standard templates, GPT4All combines the user message, sources, and attachments into the `content` field. For GPT4All v1 templates, this is not done, so they must be used directly in the template for those features to work correctly.
+
+```jinja
+{# gpt4all v1 #}
+{%- for message in messages %}
+    {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n' }}
+    {%- if message['role'] == 'user' %}
+        {%- for source in message['sources'] %}
+            {%- if loop.first %}
+                {{- '### Context:\n' }}
+            {%- endif %}
+            {{- 'Collection: ' + source['collection'] + '\n'   +
+                'Path: '       + source['path']       + '\n'   +
+                'Excerpt: '    + source['text']       + '\n\n' }}
+        {%- endfor %}
+    {%- endif %}
+    {%- for attachment in message['prompt_attachments'] %}
+        {{- attachment['processed_content'] + '\n\n' }}
+    {%- endfor %}
+    {{- message['content'] | trim }}
+    {{- '<|eot_id|>' }}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
+```

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/cookbook/use-local-ai-models-to-privately-chat-with-microsoft-excel.md

@@ -0,0 +1,85 @@
+# Using GPT4All to Privately Chat with your Microsoft Excel Spreadsheets
+Local and Private AI Chat with your Microsoft Excel Spreadsheets
+
+Microsoft Excel allows you to create, manage, and analyze data in spreadsheet format. By attaching your spreadsheets directly to GPT4All, you can privately chat with the AI to query and explore the data, enabling you to summarize, generate reports, and glean insights from your files—all within your conversation.
+
+<div style="position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;">
+  <iframe src="../../assets/gpt4all_xlsx_attachment.mp4" style="position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;" allowfullscreen title="YouTube Video"></iframe>
+</div>
+
+
+## Attach Microsoft Excel to your GPT4All Conversation
+
+!!! note "Attach Microsoft Excel to your GPT4All Conversation"
+
+    1. **Install GPT4All and Open **:
+
+        - Go to [nomic.ai/gpt4all](https://nomic.ai/gpt4all) to install GPT4All for your operating system.
+
+        - Navigate to the Chats view within GPT4All.
+
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Chat view -->
+                  <img width="1348" alt="Chat view" src="../../assets/chat_window.png">
+               </td>
+            </tr>
+         </table>
+
+    2. **Example Spreadsheet **:
+
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Spreadsheet view -->
+                  <img width="1348" alt="Spreadsheet view" src="../../assets/disney_spreadsheet.png">
+               </td>
+            </tr>
+         </table>
+
+    3. **Attach to GPT4All conversration**
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Attach view -->
+                  <img width="1348" alt="Attach view" src="../../assets/attach_spreadsheet.png">
+               </td>
+            </tr>
+         </table>
+
+    4. **Have GPT4All Summarize and Generate a Report**
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Attach view -->
+                  <img width="1348" alt="Attach view" src="../../assets/spreadsheet_chat.png">
+               </td>
+            </tr>
+         </table>
+
+
+## How It Works
+
+GPT4All parses your attached excel spreadsheet into Markdown, a format understandable to LLMs, and adds the markdown text to the context for your LLM chat. You can view the code that converts `.xslx` to Markdown [here](https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/src/xlsxtomd.cpp) in the GPT4All github repo.
+
+For example, the above spreadsheet titled `disney_income_stmt.xlsx` would be formatted the following way:
+
+```markdown
+## disney_income_stmt
+
+|Walt Disney Co.|||||||
+|---|---|---|---|---|---|---|
+|Consolidated Income Statement|||||||
+|||||||||
+|US$ in millions|||||||
+|12 months ended:|2023-09-30 00:00:00|2022-10-01 00:00:00|2021-10-02 00:00:00|2020-10-03 00:00:00|2019-09-28 00:00:00|2018-09-29 00:00:00|
+|Services|79562|74200|61768|59265|60542|50869|
+...
+...
+...
+```
+
+## Limitations
+
+It is important to double-check the claims LLMs make about the spreadsheets you provide. LLMs can make mistakes about the data they are presented, particularly for the LLMs with smaller parameter counts (~8B) that fit within the memory of consumer hardware.

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/quickstart.md

@@ -4,6 +4,8 @@ The GPT4All Desktop Application allows you to download and run large language mo
 
 With GPT4All, you can chat with models, turn your local files into information sources for models [(LocalDocs)](localdocs.md), or browse models available online to download onto your device.
 
+[Official Video Tutorial](https://www.youtube.com/watch?v=gQcZDXRVJok)
+
 ## Quickstart
 
 !!! note "Quickstart"

docs/gpt4all_desktop/settings.md

@@ -8,10 +8,11 @@
     | --- | --- | --- |
     | **Theme** | Color theme for the application. Options are `Light`, `Dark`, and `LegacyDark` | `Light` |
     | **Font Size** | Font size setting for text throughout the application. Options are Small, Medium, and Large | Small |
+    | **Language and Locale** | The language and locale of that language you wish to use | System Locale |
     | **Device** | Device that will run your models. Options are `Auto` (GPT4All chooses), `Metal` (Apple Silicon M1+), `CPU`, and `GPU` | `Auto` |
     | **Default Model** | Choose your preferred LLM to load by default on startup| Auto |
+    | **Suggestion Mode** | Generate suggested follow up questions at the end of responses | When chatting with LocalDocs | 
     | **Download Path** | Select a destination on your device to save downloaded models | Windows: `C:\Users\{username}\AppData\Local\nomic.ai\GPT4All`<br><br>Mac: `/Users/{username}/Library/Application Support/nomic.ai/GPT4All/`<br><br>Linux: `/home/{username}/.local/share/nomic.ai/GPT4All` |
-    
     | **Enable Datalake** | Opt-in to sharing interactions with GPT4All community (**anonymous** and **optional**) | Off |
 
 !!! note "Advanced Application Settings"
@@ -19,7 +20,7 @@
     | Setting | Description | Default Value |
     | --- | --- | --- |
     | **CPU Threads** | Number of concurrently running CPU threads (more can speed up responses) | 4 |
-    | **Save Chat Context** | Save chat context to disk to pick up exactly where a model left off. | Off |
+    | **Enable System Tray** | The application will minimize to the system tray / taskbar when the window is closed | Off |
     | **Enable Local Server** | Allow any application on your device to use GPT4All via an OpenAI-compatible GPT4All API | Off |
     | **API Server Port** | Local HTTP port for the local API server | 4891 |
 
@@ -30,8 +31,11 @@
     | Setting | Description | Default Value |
     | --- | --- | --- |
     | **Name** | Unique name of this model / character| set by model uploader |
-    | **System Prompt** | General instructions for the chats this model will be used for | set by model uploader |
-    | **Prompt Template** | Format of user <-> assistant interactions for the chats this model will be used for | set by model uploader |
+    | **Model File** | Filename (.gguf) of the model | set by model uploader |
+    | **System Message** | General instructions for the chats this model will be used for | set by model uploader |
+    | **Chat Template** | Format of user <-> assistant interactions for the chats this model will be used for | set by model uploader |
+    | **Chat Name Prompt** | Prompt used to automatically generate chat names | Describe the above conversation in seven words or less. |
+    | **Suggested FollowUp Prompt** | Prompt used to automatically generate follow up questions after a chat response | Suggest three very short factual follow-up questions that have not been answered yet or cannot be found inspired by the previous conversation and excerpts. |
 
 ### Clone
 

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

@@ -2,9 +2,9 @@
 
 ## 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 mentioned our [website](https://gpt4all.io/). If the problem persists, please share your experience on our [Discord](https://discord.com/channels/1076964370942267462).
+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).
 
 ## Bad Responses 
 
@@ -24,4 +24,4 @@ Including information in a prompt is not a guarantee that it will be used correc
 
 ### LocalDocs Issues
 
-Occasionally a model - particularly a smaller or overall weaker LLM - may not use the relevant text snippets from the files that were referenced via LocalDocs. If you are seeing this, it can help to use phrases like "in the docs" or "from the provided files" when prompting your model.
+Occasionally a model - particularly a smaller or overall weaker LLM - may not use the relevant text snippets from the files that were referenced via LocalDocs. If you are seeing this, it can help to use phrases like "in the docs" or "from the provided files" when prompting your model.

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.
+    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 ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
+    endif()
+endforeach()
+
+add_library(llmodel
+    src/dlhandle.cpp
+    src/llmodel.cpp
+    src/llmodel_c.cpp
+    src/llmodel_shared.cpp
+)
+gpt4all_add_warning_options(llmodel)
+target_sources(llmodel PUBLIC
+    FILE_SET public_headers TYPE HEADERS BASE_DIRS include
+    FILES include/gpt4all-backend/llmodel.h
+          include/gpt4all-backend/llmodel_c.h
+          include/gpt4all-backend/sysinfo.h
+)
+target_compile_definitions(llmodel PRIVATE LIB_FILE_EXT="${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-old/README.md

@@ -27,7 +27,7 @@ Unfortunately, no for three reasons:
 
 # What is being done to make them more compatible?
 
-A few things. Number one, we are maintaining compatibility with our current model zoo by way of the submodule pinning. However, we are also exploring how we can update to newer versions of llama.cpp without breaking our current models. This might involve an additional magic header check or it could possibly involve keeping the currently pinned submodule and also adding a new submodule with later changes and differienting them with namespaces or some other manner. Investigations continue.
+A few things. Number one, we are maintaining compatibility with our current model zoo by way of the submodule pinning. However, we are also exploring how we can update to newer versions of llama.cpp without breaking our current models. This might involve an additional magic header check or it could possibly involve keeping the currently pinned submodule and also adding a new submodule with later changes and differentiating them with namespaces or some other manner. Investigations continue.
 
 # What about GPU inference?
 

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

@@ -5,8 +5,10 @@
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
+#include <expected>
 #include <functional>
 #include <optional>
+#include <span>
 #include <stdexcept>
 #include <string>
 #include <string_view>
@@ -14,14 +16,19 @@
 #include <utility>
 #include <vector>
 
+class Dlhandle;
+
 using namespace std::string_literals;
 
 #define LLMODEL_MAX_PROMPT_BATCH 128
 
-class Dlhandle;
 class LLModel {
 public:
     using Token = int32_t;
+    using PromptCallback      = std::function<bool(std::span<const Token> batch, bool cached)>;
+    using ResponseCallback    = std::function<bool(Token token, std::string_view piece)>;
+    using EmbedCancelCallback = bool(unsigned *batchSizes, unsigned nBatch, const char *backend);
+    using ProgressCallback    = std::function<bool(float progress)>;
 
     class BadArchError: public std::runtime_error {
     public:
@@ -99,6 +106,7 @@ public:
         static int32_t maxContextLength(const std::string &modelPath);
         static int32_t layerCount(const std::string &modelPath);
         static bool isEmbeddingModel(const std::string &modelPath);
+        static auto chatTemplate(const char *modelPath) -> std::expected<std::string, std::string>;
         static void setImplementationsSearchPath(const std::string &path);
         static const std::string &implementationsSearchPath();
         static bool hasSupportedCPU();
@@ -122,9 +130,6 @@ public:
     };
 
     struct PromptContext {
-        std::vector<int32_t> tokens;    // current tokens in the context window
-        int32_t n_past = 0;             // number of tokens in past conversation
-        int32_t n_ctx = 0;              // number of tokens possible in context window
         int32_t n_predict = 200;
         int32_t top_k = 40;
         float   top_p = 0.9f;
@@ -133,38 +138,31 @@ 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
-        int32_t n_last_batch_tokens = 0;
+        float   contextErase = 0.5f;    // percent of context to erase if we exceed the context window
     };
 
-    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, int n_ctx, int ngl) = 0;
-    virtual bool isModelBlacklisted(const std::string &modelPath) const { (void)modelPath; return false; };
+    virtual bool isModelBlacklisted(const std::string &modelPath) const { (void)modelPath; return false; }
     virtual bool isEmbeddingModel(const std::string &modelPath) const { (void)modelPath; return false; }
     virtual bool isModelLoaded() const = 0;
     virtual size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) = 0;
-    virtual size_t stateSize() const { 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; }
+    virtual size_t stateSize() const = 0;
+    virtual size_t saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const = 0;
+    virtual size_t restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens) = 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,
-                        bool special = false,
-                        std::string *fakeReply = nullptr);
+    virtual void prompt(std::string_view        prompt,
+                        const PromptCallback   &promptCallback,
+                        const ResponseCallback &responseCallback,
+                        const PromptContext    &ctx);
 
-    using EmbedCancelCallback = bool(unsigned *batchSizes, unsigned nBatch, const char *backend);
+    virtual int32_t countPromptTokens(std::string_view prompt) const;
 
     virtual size_t embeddingSize() const {
         throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
@@ -209,14 +207,24 @@ public:
 
     void setProgressCallback(ProgressCallback callback) { m_progressCallback = callback; }
 
+    virtual int32_t contextLength() const = 0;
+    virtual auto specialTokens() -> std::unordered_map<std::string, std::string> const = 0;
+
 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) const = 0;
+    virtual std::vector<Token> tokenize(std::string_view str) const = 0;
+    virtual bool isSpecialToken(Token id) 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 int32_t contextLength() const = 0;
+    virtual void initSampler(const PromptContext &ctx) = 0;
+    virtual Token sampleToken() const = 0;
+    virtual bool evalTokens(int32_t nPast, std::span<const Token> tokens) const = 0;
+    virtual void shiftContext(const PromptContext &promptCtx, int32_t *nPast) = 0;
+    virtual int32_t inputLength() const = 0;
+    virtual int32_t computeModelInputPosition(std::span<const Token> input) const = 0;
+    virtual void setModelInputPosition(int32_t pos) = 0;
+    virtual void appendInputToken(Token tok) = 0;
+    virtual std::span<const Token> inputTokens() const = 0;
     virtual const std::vector<Token> &endTokens() const = 0;
     virtual bool shouldAddBOS() const = 0;
 
@@ -232,9 +240,11 @@ protected:
         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
-    void recalculateContext(PromptContext &promptCtx, std::function<bool(bool)> recalculate);
+    virtual auto chatTemplate(const char *modelPath) const -> std::expected<std::string, std::string>
+    {
+        (void)modelPath;
+        return std::unexpected("not implemented");
+    }
 
     const Implementation *m_implementation = nullptr;
 
@@ -247,16 +257,16 @@ protected:
         return true;
     }
 
-    bool 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);
+    // prefill context with prompt
+    auto decodePrompt(const PromptCallback &promptCallback,
+                      const PromptContext  &promptCtx,
+                      std::vector<Token>    embd_inp)
+        -> std::optional<int32_t>;
+    // generate a response
+    void generateResponse(const ResponseCallback &responseCallback,
+                          const PromptContext    &promptCtx,
+                          int32_t                 nPast);
 
-private:
     friend class LLMImplementation;
 };
 

gpt4all-backend-old/include/gpt4all-backend/llmodel_c.h

@@ -23,6 +23,11 @@ extern "C" {
  */
 typedef void *llmodel_model;
 
+/**
+ * A token.
+ */
+typedef int32_t token_t;
+
 /**
  * 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
@@ -30,21 +35,15 @@ typedef void *llmodel_model;
  * behavior.
  */
 struct llmodel_prompt_context {
-    float *logits;          // logits of current context
-    size_t logits_size;     // the size of the raw logits vector
-    int32_t *tokens;        // current tokens in the context window
-    size_t tokens_size;     // the size of the raw tokens vector
-    int32_t n_past;         // number of tokens in past conversation
-    int32_t n_ctx;          // number of tokens possible in context window
     int32_t n_predict;      // number of tokens to predict
     int32_t top_k;          // top k logits to sample from
-    float top_p;            // nucleus sampling probability threshold
-    float min_p;            // Min P sampling
-    float temp;             // temperature to adjust model's output distribution
+    float   top_p;          // nucleus sampling probability threshold
+    float   min_p;          // Min P sampling
+    float   temp;           // temperature to adjust model's output distribution
     int32_t n_batch;        // number of predictions to generate in parallel
-    float repeat_penalty;   // penalty factor for repeated tokens
+    float   repeat_penalty; // penalty factor for repeated tokens
     int32_t repeat_last_n;  // last n tokens to penalize
-    float context_erase;    // percent of context to erase if we exceed the context window
+    float   context_erase;  // percent of context to erase if we exceed the context window
 };
 
 struct llmodel_gpu_device {
@@ -63,10 +62,12 @@ typedef struct llmodel_gpu_device llmodel_gpu_device;
 
 /**
  * Callback type for prompt processing.
- * @param token_id The token id of the prompt.
+ * @param token_ids An array of token ids of the prompt.
+ * @param n_token_ids The number of tokens in the array.
+ * @param cached Whether the tokens were already in cache.
  * @return a bool indicating whether the model should keep processing.
  */
-typedef bool (*llmodel_prompt_callback)(int32_t token_id);
+typedef bool (*llmodel_prompt_callback)(const token_t *token_ids, size_t n_token_ids, bool cached);
 
 /**
  * Callback type for response.
@@ -74,14 +75,7 @@ typedef bool (*llmodel_prompt_callback)(int32_t token_id);
  * @param response The response string. NOTE: a token_id of -1 indicates the string is an error string.
  * @return a bool indicating whether the model should keep generating.
  */
-typedef bool (*llmodel_response_callback)(int32_t token_id, const char *response);
-
-/**
- * Callback type for recalculation of context.
- * @param whether the model is recalculating the context.
- * @return a bool indicating whether the model should keep generating.
- */
-typedef bool (*llmodel_recalculate_callback)(bool is_recalculating);
+typedef bool (*llmodel_response_callback)(token_t token_id, const char *response);
 
 /**
  * Embedding cancellation callback for use with llmodel_embed.
@@ -92,6 +86,8 @@ typedef bool (*llmodel_recalculate_callback)(bool is_recalculating);
  */
 typedef bool (*llmodel_emb_cancel_callback)(unsigned *batch_sizes, unsigned n_batch, const char *backend);
 
+typedef void (*llmodel_special_token_callback)(const char *name, const char *token);
+
 /**
  * Create a llmodel instance.
  * Recognises correct model type from file at model_path
@@ -150,46 +146,57 @@ bool llmodel_isModelLoaded(llmodel_model model);
  * @param model A pointer to the llmodel_model instance.
  * @return the size in bytes of the internal state of the model
  */
-uint64_t llmodel_get_state_size(llmodel_model model);
+uint64_t llmodel_state_get_size(llmodel_model model);
 
 /**
- * Saves the internal state of the model to the specified destination address.
+ * Saves the internal state of the model.
  * NOTE: This state data is specific to the type of model you have created.
  * @param model A pointer to the llmodel_model instance.
- * @param dest A pointer to the destination.
- * @return the number of bytes copied
+ * @param state Where to store the state. This must be a buffer of at least llmodel_state_get_size() bytes.
+ * @param state_size The size of the destination for the state.
+ * @param input_tokens_out Where to store the address of the token cache state. This is dynamically allocated and must
+ * be freed with llmodel_state_free_input_tokens.
+ * @param n_input_tokens Where to store the size of the token cache state.
+ * @return The number of bytes copied. On error, zero is returned, the token cache is set to NULL, and the token cache
+ * size is set to zero.
  */
-uint64_t llmodel_save_state_data(llmodel_model model, uint8_t *dest);
+uint64_t llmodel_state_get_data(llmodel_model model, uint8_t *state_out, uint64_t state_size,
+                                token_t **input_tokens_out, uint64_t *n_input_tokens);
+
+/**
+ * Frees the temporary token cache buffer created by a call to llmodel_state_get_data().
+ * @param input_tokens The token cache buffer.
+ */
+void llmodel_state_free_input_tokens(token_t *input_tokens);
 
 /**
  * Restores the internal state of the model using data from the specified address.
  * NOTE: This state data is specific to the type of model you have created.
  * @param model A pointer to the llmodel_model instance.
- * @param src A pointer to the src.
- * @return the number of bytes read
+ * @param state A pointer to the state data.
+ * @param state_size The size of the state data.
+ * @param input_tokens The token cache associated with the saved state.
+ * @param n_input_tokens The number of tokens in input_tokens.
+ * @return The number of bytes read, or zero on error.
  */
-uint64_t llmodel_restore_state_data(llmodel_model model, const uint8_t *src);
+uint64_t llmodel_state_set_data(llmodel_model model, const uint8_t *state, uint64_t state_size,
+                                const token_t *input_tokens, uint64_t n_input_tokens);
 
 /**
  * Generate a response using the model.
  * @param model A pointer to the llmodel_model instance.
  * @param prompt A string representing the input prompt.
- * @param prompt_template A string representing the input prompt template.
  * @param prompt_callback A callback function for handling the processing of prompt.
  * @param response_callback A callback function for handling the generated response.
- * @param recalculate_callback A callback function for handling recalculation requests.
- * @param special True if special tokens in the prompt should be processed, false otherwise.
- * @param fake_reply A string to insert into context as the model's reply, or NULL to generate one.
  * @param ctx A pointer to the llmodel_prompt_context structure.
+ * @param error A pointer to a string; will only be set on error.
  */
-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,
-                    bool special,
-                    const char *fake_reply);
+bool llmodel_prompt(llmodel_model               model,
+                    const char                 *prompt,
+                    llmodel_prompt_callback     prompt_callback,
+                    llmodel_response_callback   response_callback,
+                    llmodel_prompt_context     *ctx,
+                    const char                **error);
 
 /**
  * Generate an embedding using the model.
@@ -301,6 +308,10 @@ const char *llmodel_model_backend_name(llmodel_model model);
  */
 const char *llmodel_model_gpu_device_name(llmodel_model model);
 
+int32_t llmodel_count_prompt_tokens(llmodel_model model, const char *prompt, const char **error);
+
+void llmodel_model_foreach_special_token(llmodel_model model, llmodel_special_token_callback callback);
+
 #ifdef __cplusplus
 }
 #endif

gpt4all-backend-old/llama.cpp.cmake

@@ -378,19 +378,7 @@ function(include_ggml SUFFIX)
         find_package(CUDAToolkit REQUIRED)
         set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
 
-        if (NOT DEFINED GGML_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(GGML_CUDA_ARCHITECTURES "60;61;70;75") # needed for f16 CUDA intrinsics
-            else()
-                set(GGML_CUDA_ARCHITECTURES "52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
-                #set(GGML_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
-            endif()
-        endif()
-        message(STATUS "Using CUDA architectures: ${GGML_CUDA_ARCHITECTURES}")
+        # architectures are set in gpt4all-backend/CMakeLists.txt
 
         set(GGML_HEADERS_CUDA ${DIRECTORY}/ggml/include/ggml-cuda.h)
         file(GLOB   GGML_HEADERS_CUDA "${DIRECTORY}/ggml/src/ggml-cuda/*.cuh")
@@ -823,7 +811,8 @@ function(include_ggml SUFFIX)
             list(APPEND XC_FLAGS -std=${GGML_METAL_STD})
         endif()
 
-        set(GGML_METALLIB ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib)
+        set(GGML_METALLIB "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib")
+        set(GGML_METALLIB "${GGML_METALLIB}" PARENT_SCOPE)
         add_custom_command(
             OUTPUT ${GGML_METALLIB}
             COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
@@ -834,7 +823,6 @@ function(include_ggml SUFFIX)
             DEPENDS ${DIRECTORY}/ggml/src/ggml-metal.metal ${DIRECTORY}/ggml/src/ggml-common.h
             COMMENT "Compiling Metal kernels"
             )
-        set_source_files_properties(${GGML_METALLIB} DIRECTORY ${CMAKE_SOURCE_DIR} PROPERTIES GENERATED ON)
 
         add_custom_target(
             ggml-metal ALL
@@ -990,10 +978,13 @@ function(include_ggml SUFFIX)
 
     add_library(llama${SUFFIX} STATIC
                 ${DIRECTORY}/include/llama.h
+                ${DIRECTORY}/src/llama-grammar.cpp
+                ${DIRECTORY}/src/llama-sampling.cpp
+                ${DIRECTORY}/src/llama-vocab.cpp
                 ${DIRECTORY}/src/llama.cpp
-                ${DIRECTORY}/src/unicode.h
-                ${DIRECTORY}/src/unicode.cpp
                 ${DIRECTORY}/src/unicode-data.cpp
+                ${DIRECTORY}/src/unicode.cpp
+                ${DIRECTORY}/src/unicode.h
                 )
 
     target_include_directories(llama${SUFFIX} PUBLIC  ${DIRECTORY}/include ${DIRECTORY}/ggml/include)
@@ -1018,9 +1009,6 @@ function(include_ggml SUFFIX)
         C_STANDARD 11
         C_STANDARD_REQUIRED true
         )
-    if (GGML_CUDA_ARCHITECTURES)
-        set_property(TARGET ggml${SUFFIX} llama${SUFFIX} PROPERTY CUDA_ARCHITECTURES "${GGML_CUDA_ARCHITECTURES}")
-    endif()
 
     target_compile_options(ggml${SUFFIX} PRIVATE "${GGML_COMPILE_OPTS}")
     target_compile_options(llama${SUFFIX} PRIVATE "${GGML_COMPILE_OPTS}")

gpt4all-backend-old/src/llamamodel.cpp

@@ -2,6 +2,7 @@
 #include "llamamodel_impl.h"
 
 #include "llmodel.h"
+#include "utils.h"
 
 #include <ggml.h>
 #include <llama.h>
@@ -52,6 +53,8 @@ static const std::vector<const char *> KNOWN_ARCHES {
     "gpt2",
     // "gptj", -- no inference code
     "gptneox",
+    "granite",
+    "granitemoe",
     "mpt",
     "baichuan",
     "starcoder",
@@ -79,6 +82,7 @@ static const std::vector<const char *> KNOWN_ARCHES {
     "command-r",
     // "dbrx", -- 16x12B parameters
     "olmo",
+    "olmoe",
     "openelm",
     // "arctic", -- 10B+128x3.66B parameters
     "deepseek2",
@@ -103,26 +107,34 @@ static bool llama_verbose()
     return var && *var;
 }
 
-static void llama_log_callback(enum ggml_log_level level, const char *text, void *userdata)
+static void llama_log_callback(ggml_log_level level, const char *text, void *userdata, bool warn)
 {
     (void)userdata;
-    if (llama_verbose() || level <= GGML_LOG_LEVEL_ERROR) {
-        fputs(text, stderr);
-    }
-}
 
-#ifdef GGML_USE_CUDA
-static void cuda_log_callback(enum ggml_log_level level, const char *text, void *userdata)
-{
-    (void)userdata;
-    if (llama_verbose() || level <= GGML_LOG_LEVEL_WARN) {
-        fputs(text, stderr);
+    static ggml_log_level lastlevel = GGML_LOG_LEVEL_NONE;
+    if (!llama_verbose()) {
+        auto efflevel = level == GGML_LOG_LEVEL_CONT ? lastlevel : level;
+        lastlevel = efflevel;
+        switch (efflevel) {
+            case GGML_LOG_LEVEL_CONT:
+                UNREACHABLE();
+                break;
+            case GGML_LOG_LEVEL_WARN:
+                if (warn) break;
+                [[fallthrough]];
+            case GGML_LOG_LEVEL_NONE: // not used?
+            case GGML_LOG_LEVEL_INFO:
+            case GGML_LOG_LEVEL_DEBUG:
+                return; // suppress
+            case GGML_LOG_LEVEL_ERROR:
+                ;
+        }
     }
+
+    fputs(text, stderr);
 }
-#endif
 
 struct gpt_params {
-    int32_t seed          = -1;   // RNG seed
     int32_t n_keep        = 0;    // number of tokens to keep from initial prompt
 
     // sampling parameters
@@ -137,37 +149,6 @@ struct gpt_params {
     bool use_mlock         = false; // use mlock to keep model in memory
 };
 
-static int llama_sample_top_p_top_k(
-        llama_context *ctx,
-        const llama_token *last_n_tokens_data,
-        int last_n_tokens_size,
-        int top_k,
-        float top_p,
-        float min_p,
-        float temp,
-        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);
-    for (int token_id = 0; token_id < n_vocab; token_id++) {
-        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-    llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
-    // Sample 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_min_p(ctx, &candidates_p, min_p, 1);
-    llama_sample_temp(ctx, &candidates_p, temp);
-    return llama_sample_token(ctx, &candidates_p);
-}
-
 const char *get_arch_name(gguf_context *ctx_gguf)
 {
     const int kid = gguf_find_key(ctx_gguf, "general.architecture");
@@ -224,7 +205,7 @@ static int32_t get_arch_key_u32(std::string const &modelPath, std::string const
         if (keyidx != -1) {
             value = gguf_get_val_u32(ctx, keyidx);
         } else {
-            std::cerr << __func__ << ": " << key << "not found in " << modelPath << "\n";
+            std::cerr << __func__ << ": " << key << " not found in " << modelPath << "\n";
         }
     }
 
@@ -234,21 +215,27 @@ cleanup:
 }
 
 struct LLamaPrivate {
-    const std::string modelPath;
-    bool modelLoaded = false;
-    int device = -1;
-    std::string deviceName;
-    llama_model *model = nullptr;
-    llama_context *ctx = nullptr;
-    llama_model_params model_params;
-    llama_context_params ctx_params;
-    int64_t n_threads = 0;
-    std::vector<LLModel::Token> end_tokens;
-    const char *backend_name = nullptr;
+    bool                         modelLoaded  = false;
+    int                          device       = -1;
+    std::string                  deviceName;
+    int64_t                      n_threads    = 0;
+    std::vector<LLModel::Token>  end_tokens;
+    const char                  *backend_name = nullptr;
+    std::vector<LLModel::Token>  inputTokens;
+
+    llama_model          *model        = nullptr;
+    llama_context        *ctx          = nullptr;
+    llama_model_params    model_params;
+    llama_context_params  ctx_params;
+    llama_sampler        *sampler_chain;
 };
 
 LLamaModel::LLamaModel()
-    : d_ptr(new LLamaPrivate) {}
+    : d_ptr(std::make_unique<LLamaPrivate>())
+{
+    auto sparams = llama_sampler_chain_default_params();
+    d_ptr->sampler_chain = llama_sampler_chain_init(sparams);
+}
 
 // default hparams (LLaMA 7B)
 struct llama_file_hparams {
@@ -437,10 +424,9 @@ bool LLamaModel::loadModel(const std::string &modelPath, int n_ctx, int ngl)
         }
     }
 
-    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;
+    d_ptr->ctx_params.n_ctx  = n_ctx;
+    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.
@@ -506,6 +492,7 @@ LLamaModel::~LLamaModel()
         llama_free(d_ptr->ctx);
     }
     llama_free_model(d_ptr->model);
+    llama_sampler_free(d_ptr->sampler_chain);
 }
 
 bool LLamaModel::isModelLoaded() const
@@ -515,30 +502,41 @@ bool LLamaModel::isModelLoaded() const
 
 size_t LLamaModel::stateSize() const
 {
-    return llama_get_state_size(d_ptr->ctx);
+    return llama_state_get_size(d_ptr->ctx);
 }
 
-size_t LLamaModel::saveState(uint8_t *dest) const
+size_t LLamaModel::saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const
 {
-    return llama_copy_state_data(d_ptr->ctx, dest);
+    size_t bytesWritten = llama_state_get_data(d_ptr->ctx, stateOut.data(), stateOut.size());
+    if (bytesWritten)
+        inputTokensOut.assign(d_ptr->inputTokens.begin(), d_ptr->inputTokens.end());
+    return bytesWritten;
 }
 
-size_t LLamaModel::restoreState(const uint8_t *src)
+size_t LLamaModel::restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens)
 {
-    // const_cast is required, see: https://github.com/ggerganov/llama.cpp/pull/1540
-    return llama_set_state_data(d_ptr->ctx, const_cast<uint8_t*>(src));
+    size_t bytesRead = llama_state_set_data(d_ptr->ctx, state.data(), state.size());
+    if (bytesRead)
+        d_ptr->inputTokens.assign(inputTokens.begin(), inputTokens.end());
+    return bytesRead;
 }
 
-std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, const std::string &str, bool special) const
+std::vector<LLModel::Token> LLamaModel::tokenize(std::string_view str) const
 {
-    const bool wantBOS = ctx.n_past == 0 && ctx.tokens.empty();
-    const bool useBOS = wantBOS && shouldAddBOS();
     std::vector<LLModel::Token> fres(str.length() + 4);
-    auto fres_len = llama_tokenize(d_ptr->model, str.c_str(), str.length(), fres.data(), fres.size(), useBOS, special);
+    int32_t fres_len = llama_tokenize(
+        d_ptr->model, str.data(), str.length(), fres.data(), fres.size(), /*add_special*/ true, /*parse_special*/ true
+    );
     fres.resize(fres_len);
     return fres;
 }
 
+bool LLamaModel::isSpecialToken(Token id) const
+{
+    return llama_token_get_attr(d_ptr->model, id)
+        & (LLAMA_TOKEN_ATTR_CONTROL | LLAMA_TOKEN_ATTR_USER_DEFINED | LLAMA_TOKEN_ATTR_UNKNOWN);
+}
+
 std::string LLamaModel::tokenToString(Token id) const
 {
     std::vector<char> result(8, 0);
@@ -555,27 +553,66 @@ std::string LLamaModel::tokenToString(Token id) const
     return std::string(result.data(), result.size());
 }
 
-LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
+void LLamaModel::initSampler(const PromptContext &promptCtx)
 {
-    const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
-    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.min_p, promptCtx.temp,
-        promptCtx.repeat_penalty, promptCtx.n_last_batch_tokens - 1);
+    auto *model = d_ptr->model;
+    auto *chain = d_ptr->sampler_chain;
+
+    // clear sampler chain
+    for (int i = llama_sampler_chain_n(chain) - 1; i >= 0; i--) {
+        auto *smpl = llama_sampler_chain_remove(chain, i);
+        llama_sampler_free(smpl);
+    }
+
+    // build new chain
+    llama_sampler_chain_add(chain,
+        llama_sampler_init_penalties(
+            llama_n_vocab(model),
+            llama_token_eos(model),
+            llama_token_nl(model),
+            promptCtx.repeat_last_n,
+            promptCtx.repeat_penalty,
+            // TODO(jared): consider making the below configurable
+            /*penalty_freq*/    0.0f,
+            /*penalty_present*/ 0.0f,
+            /*penalize_nl*/     true,
+            /*ignore_eos*/      false
+        )
+    );
+    if (promptCtx.temp == 0.0f) {
+        llama_sampler_chain_add(chain, llama_sampler_init_greedy());
+    } else {
+        struct llama_sampler *samplers[] = {
+            llama_sampler_init_top_k(promptCtx.top_k),
+            llama_sampler_init_top_p(promptCtx.top_p, 1),
+            llama_sampler_init_min_p(promptCtx.min_p, 1),
+            llama_sampler_init_temp(promptCtx.temp),
+            llama_sampler_init_softmax(),
+            llama_sampler_init_dist(LLAMA_DEFAULT_SEED),
+        };
+        for (auto *smpl : samplers)
+            llama_sampler_chain_add(chain, smpl);
+    }
 }
 
-bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
+LLModel::Token LLamaModel::sampleToken() const
 {
-    llama_kv_cache_seq_rm(d_ptr->ctx, 0, ctx.n_past, -1);
+    return llama_sampler_sample(d_ptr->sampler_chain, d_ptr->ctx, -1);
+}
+
+bool LLamaModel::evalTokens(int32_t nPast, std::span<const Token> tokens) const
+{
+    assert(!tokens.empty());
+
+    llama_kv_cache_seq_rm(d_ptr->ctx, 0, nPast, -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.pos     [i] = nPast + i;
         batch.n_seq_id[i] = 1;
         batch.seq_id  [i][0] = 0;
         batch.logits  [i] = false;
@@ -589,11 +626,86 @@ bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &toke
     return res == 0;
 }
 
+void LLamaModel::shiftContext(const PromptContext &promptCtx, int32_t *nPast)
+{
+    // infinite text generation via context shifting
+
+    // erase up to n_ctx*contextErase tokens
+    int n_keep = shouldAddBOS();
+    int n_past = *nPast;
+    int n_discard = std::min(n_past - n_keep, int(contextLength() * promptCtx.contextErase));
+
+    assert(n_discard > 0);
+    if (n_discard <= 0)
+        return;
+
+    std::cerr << "Llama: context full, swapping: n_past = " << n_past << ", n_keep = " << n_keep
+              << ", n_discard = " << n_discard << "\n";
+
+    // erase the first n_discard tokens from the context
+    llama_kv_cache_seq_rm (d_ptr->ctx, 0, n_keep,             n_keep + n_discard);
+    llama_kv_cache_seq_add(d_ptr->ctx, 0, n_keep + n_discard, n_past,             -n_discard);
+
+    auto &inp = d_ptr->inputTokens;
+    inp.erase(inp.begin() + n_keep, inp.begin() + n_keep + n_discard);
+    *nPast = inp.size();
+}
+
 int32_t LLamaModel::contextLength() const
 {
     return llama_n_ctx(d_ptr->ctx);
 }
 
+auto LLamaModel::specialTokens() -> std::unordered_map<std::string, std::string> const
+{
+    if (!d_ptr->model)
+        throw std::logic_error("model not loaded");
+
+    std::unordered_map<std::string, std::string> tokens;
+    if (auto id = llama_token_bos(d_ptr->model); id != LLAMA_TOKEN_NULL)
+        tokens.emplace("bos_token", tokenToString(id));
+    if (auto id = llama_token_eos(d_ptr->model); id != LLAMA_TOKEN_NULL)
+        tokens.emplace("eos_token", tokenToString(id));
+    return tokens;
+}
+
+int32_t LLamaModel::inputLength() const
+{
+    return d_ptr->inputTokens.size();
+}
+
+int32_t LLamaModel::computeModelInputPosition(std::span<const Token> input) const
+{
+    // find common prefix
+    auto cacheIt = d_ptr->inputTokens.begin();
+    auto inputIt = input.begin();
+    while (cacheIt < d_ptr->inputTokens.end() && inputIt < input.end() && *cacheIt == *inputIt) {
+        ++cacheIt; ++inputIt;
+    }
+    // tell the caller to ignore the tokens between [begin, inputIt)
+    return inputIt - input.begin();
+}
+
+void LLamaModel::setModelInputPosition(int32_t pos)
+{
+    auto &inp = d_ptr->inputTokens;
+    assert(pos >= 0);
+    assert(pos <= inp.size());
+    // truncate token cache to end at the new n_past
+    if (pos < inp.size())
+        inp.resize(pos);
+}
+
+void LLamaModel::appendInputToken(Token tok)
+{
+    d_ptr->inputTokens.push_back(tok);
+}
+
+auto LLamaModel::inputTokens() const -> std::span<const Token>
+{
+    return d_ptr->inputTokens;
+}
+
 const std::vector<LLModel::Token> &LLamaModel::endTokens() const
 {
     return d_ptr->end_tokens;
@@ -601,10 +713,7 @@ const std::vector<LLModel::Token> &LLamaModel::endTokens() const
 
 bool LLamaModel::shouldAddBOS() const
 {
-    int add_bos = llama_add_bos_token(d_ptr->model);
-    if (add_bos != -1) { return add_bos; }
-    auto vocab_type = llama_vocab_type(d_ptr->model);
-    return vocab_type == LLAMA_VOCAB_TYPE_SPM || vocab_type == LLAMA_VOCAB_TYPE_WPM;
+    return llama_add_bos_token(d_ptr->model);
 }
 
 int32_t LLamaModel::maxContextLength(std::string const &modelPath) const
@@ -617,6 +726,37 @@ int32_t LLamaModel::layerCount(std::string const &modelPath) const
     return get_arch_key_u32(modelPath, "block_count");
 }
 
+// TODO(jared): reduce redundant code and operations by combining all metadata getters for unloaded
+//              models into a class that keeps the model file open
+auto LLamaModel::chatTemplate(const char *modelPath) const -> std::expected<std::string, std::string>
+{
+    auto *ctx = load_gguf(modelPath);
+    if (!ctx)
+        return std::unexpected("failed to open model file");
+
+    std::expected<std::string, std::string> result;
+    enum gguf_type ktype;
+    const int kid = gguf_find_key(ctx, "tokenizer.chat_template");
+    if (kid == -1) {
+        result = std::unexpected("key not found");
+        goto cleanup;
+    }
+
+    ktype = gguf_get_kv_type(ctx, kid);
+    if (ktype != GGUF_TYPE_STRING) {
+        result = std::unexpected(
+            "expected key type STRING (" + std::to_string(GGUF_TYPE_STRING) + "), got " + std::to_string(ktype)
+        );
+        goto cleanup;
+    }
+
+    result = gguf_get_val_str(ctx, kid);
+
+cleanup:
+    gguf_free(ctx);
+    return result;
+}
+
 #ifdef GGML_USE_VULKAN
 static const char *getVulkanVendorName(uint32_t vendorID)
 {
@@ -946,7 +1086,7 @@ void LLamaModel::embedInternal(
     const llama_token bos_token = llama_token_bos(d_ptr->model);
     const llama_token eos_token = llama_token_eos(d_ptr->model);
 
-    bool useBOS = shouldAddBOS();
+    bool useBOS = llama_add_bos_token(d_ptr->model);
     bool useEOS = llama_vocab_type(d_ptr->model) == LLAMA_VOCAB_TYPE_WPM;
 
     // no EOS, optional BOS
@@ -954,13 +1094,16 @@ void LLamaModel::embedInternal(
         if (!text.empty() && text[0] != ' ') {
             text = ' ' + text; // normalize for SPM - our fork of llama.cpp doesn't add a space prefix
         }
-        wantBOS &= useBOS;
 
         tokens.resize(text.length()+4);
-        int32_t n_tokens = llama_tokenize(d_ptr->model, text.c_str(), text.length(), tokens.data(), tokens.size(), wantBOS, false);
+        int32_t n_tokens = llama_tokenize_gpt4all(
+            d_ptr->model, text.c_str(), text.length(), tokens.data(), tokens.size(), /*add_special*/ wantBOS,
+            /*parse_special*/ false, /*insert_space*/ false
+        );
         if (n_tokens) {
             (void)eos_token;
-            assert((useEOS && wantBOS) == (eos_token != -1 && tokens[n_tokens - 1] == eos_token));
+            (void)useBOS;
+            assert((useEOS && wantBOS && useBOS) == (eos_token != -1 && tokens[n_tokens - 1] == eos_token));
             if (useEOS && wantBOS)
                 n_tokens--; // erase EOS/SEP
         }
@@ -1186,9 +1329,9 @@ DLL_EXPORT bool is_arch_supported(const char *arch)
 
 DLL_EXPORT LLModel *construct()
 {
-    llama_log_set(llama_log_callback, nullptr);
+    llama_log_set([](auto l, auto t, auto u) { llama_log_callback(l, t, u, false); }, nullptr);
 #ifdef GGML_USE_CUDA
-    ggml_backend_cuda_log_set_callback(cuda_log_callback, nullptr);
+    ggml_backend_cuda_log_set_callback([](auto l, auto t, auto u) { llama_log_callback(l, t, u, true); }, nullptr);
 #endif
     return new LLamaModel;
 }

gpt4all-backend-old/src/llamamodel_impl.h

@@ -6,10 +6,12 @@
 
 #include "llmodel.h"
 
-#include <functional>
 #include <memory>
+#include <span>
 #include <string>
+#include <string_view>
 #include <vector>
+#include <unordered_map>
 
 struct LLamaPrivate;
 struct EmbModelSpec;
@@ -27,8 +29,8 @@ public:
     bool isModelLoaded() const 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;
+    size_t saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const override;
+    size_t restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens) override;
     void setThreadCount(int32_t n_threads) override;
     int32_t threadCount() const override;
     std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired = 0) const override;
@@ -47,25 +49,36 @@ public:
     void embed(const std::vector<std::string> &texts, float *embeddings, bool isRetrieval, int dimensionality = -1,
                size_t *tokenCount = nullptr, bool doMean = true, bool atlas = false) override;
 
-private:
-    std::unique_ptr<LLamaPrivate> d_ptr;
-    bool m_supportsEmbedding = false;
-    bool m_supportsCompletion = false;
+    int32_t contextLength() const override;
+    auto specialTokens() -> std::unordered_map<std::string, std::string> const override;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) const override;
+    std::vector<Token> tokenize(std::string_view str) const override;
+    bool isSpecialToken(Token id) 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;
+    void initSampler(const PromptContext &ctx) override;
+    Token sampleToken() const override;
+    bool evalTokens(int32_t nPast, std::span<const Token> tokens) const override;
+    void shiftContext(const PromptContext &promptCtx, int32_t *nPast) override;
+    int32_t inputLength() const override;
+    int32_t computeModelInputPosition(std::span<const Token> input) const override;
+    void setModelInputPosition(int32_t pos) override;
+    void appendInputToken(Token tok) override;
+    std::span<const Token> inputTokens() 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;
+    auto chatTemplate(const char *modelPath) const -> std::expected<std::string, std::string> override;
 
     void embedInternal(const std::vector<std::string> &texts, float *embeddings, std::string prefix, int dimensionality,
                        size_t *tokenCount, bool doMean, bool atlas, EmbedCancelCallback *cancelCb,
                        const EmbModelSpec *spec);
+
+private:
+    std::unique_ptr<LLamaPrivate> d_ptr;
+    bool m_supportsEmbedding = false;
+    bool m_supportsCompletion = false;
 };
 
 #endif // LLAMAMODEL_H

gpt4all-backend-old/src/llmodel.cpp

@@ -140,9 +140,14 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
             std::string path;
             // Split the paths string by the delimiter and process each path.
             while (std::getline(ss, path, ';')) {
-                std::u8string u8_path(path.begin(), path.end());
+                fs::directory_iterator iter;
+                try {
+                    iter = fs::directory_iterator(std::u8string(path.begin(), path.end()));
+                } catch (const fs::filesystem_error &) {
+                    continue; // skip nonexistent path
+                }
                 // Iterate over all libraries
-                for (const auto &f : fs::directory_iterator(u8_path)) {
+                for (const auto &f : iter) {
                     const fs::path &p = f.path();
 
                     if (p.extension() != LIB_FILE_EXT) continue;
@@ -326,6 +331,12 @@ bool LLModel::Implementation::isEmbeddingModel(const std::string &modelPath)
     return llama && llama->isEmbeddingModel(modelPath);
 }
 
+auto LLModel::Implementation::chatTemplate(const char *modelPath) -> std::expected<std::string, std::string>
+{
+    auto *llama = constructGlobalLlama();
+    return llama ? llama->chatTemplate(modelPath) : std::unexpected("backend not available");
+}
+
 void LLModel::Implementation::setImplementationsSearchPath(const std::string& path)
 {
     s_implementations_search_path = path;

gpt4all-backend-old/src/llmodel_c.cpp

@@ -7,16 +7,20 @@
 #include <cstdlib>
 #include <cstring>
 #include <exception>
-#include <functional>
 #include <iostream>
 #include <memory>
 #include <optional>
 #include <string>
+#include <string_view>
 #include <vector>
+#include <span>
+
+namespace ranges = std::ranges;
+
+static_assert(sizeof(token_t) == sizeof(LLModel::Token));
 
 struct LLModelWrapper {
     LLModel *llModel = nullptr;
-    LLModel::PromptContext promptContext;
     ~LLModelWrapper() { delete llModel; }
 };
 
@@ -84,80 +88,80 @@ bool llmodel_isModelLoaded(llmodel_model model)
     return wrapper->llModel->isModelLoaded();
 }
 
-uint64_t llmodel_get_state_size(llmodel_model model)
+uint64_t llmodel_state_get_size(llmodel_model model)
 {
     auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->stateSize();
 }
 
-uint64_t llmodel_save_state_data(llmodel_model model, uint8_t *dest)
+uint64_t llmodel_state_get_data(llmodel_model model, uint8_t *state_out, uint64_t state_size,
+                                token_t **input_tokens_out, uint64_t *n_input_tokens)
 {
     auto *wrapper = static_cast<LLModelWrapper *>(model);
-    return wrapper->llModel->saveState(dest);
+    std::vector<LLModel::Token> inputTokens;
+    auto bytesWritten = wrapper->llModel->saveState({state_out, size_t(state_size)}, inputTokens);
+    if (bytesWritten) {
+        auto *buf = new LLModel::Token[inputTokens.size()];
+        ranges::copy(inputTokens, buf);
+        *input_tokens_out = buf;
+        *n_input_tokens = uint64_t(inputTokens.size());
+    } else {
+        *input_tokens_out = nullptr;
+        *n_input_tokens = 0;
+    }
+    return bytesWritten;
 }
 
-uint64_t llmodel_restore_state_data(llmodel_model model, const uint8_t *src)
+void llmodel_state_free_input_tokens(LLModel::Token *input_tokens)
 {
-    auto *wrapper = static_cast<LLModelWrapper *>(model);
-    return wrapper->llModel->restoreState(src);
+    delete[] input_tokens;
 }
 
-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,
-                    bool special,
-                    const char *fake_reply)
+uint64_t llmodel_state_set_data(llmodel_model model, const uint8_t *state, uint64_t state_size,
+                                const token_t *input_tokens, uint64_t n_input_tokens)
 {
     auto *wrapper = static_cast<LLModelWrapper *>(model);
+    return wrapper->llModel->restoreState({state, size_t(state_size)}, {input_tokens, size_t(n_input_tokens)});
+}
 
-    auto response_func = [response_callback](int32_t token_id, const std::string &response) {
-        return response_callback(token_id, response.c_str());
-    };
-
-    if (size_t(ctx->n_past) < wrapper->promptContext.tokens.size())
-        wrapper->promptContext.tokens.resize(ctx->n_past);
+bool llmodel_prompt(llmodel_model               model,
+                    const char                 *prompt,
+                    llmodel_prompt_callback     prompt_callback,
+                    llmodel_response_callback   response_callback,
+                    llmodel_prompt_context     *ctx,
+                    const char                **error)
+{
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
 
     // Copy the C prompt context
-    wrapper->promptContext.n_past = ctx->n_past;
-    wrapper->promptContext.n_ctx = ctx->n_ctx;
-    wrapper->promptContext.n_predict = ctx->n_predict;
-    wrapper->promptContext.top_k = ctx->top_k;
-    wrapper->promptContext.top_p = ctx->top_p;
-    wrapper->promptContext.min_p = ctx->min_p;
-    wrapper->promptContext.temp = ctx->temp;
-    wrapper->promptContext.n_batch = ctx->n_batch;
-    wrapper->promptContext.repeat_penalty = ctx->repeat_penalty;
-    wrapper->promptContext.repeat_last_n = ctx->repeat_last_n;
-    wrapper->promptContext.contextErase = ctx->context_erase;
+    LLModel::PromptContext promptContext {
+        .n_predict      = ctx->n_predict,
+        .top_k          = ctx->top_k,
+        .top_p          = ctx->top_p,
+        .min_p          = ctx->min_p,
+        .temp           = ctx->temp,
+        .n_batch        = ctx->n_batch,
+        .repeat_penalty = ctx->repeat_penalty,
+        .repeat_last_n  = ctx->repeat_last_n,
+        .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;
+    auto prompt_func = [prompt_callback](std::span<const LLModel::Token> token_ids, bool cached) {
+        return prompt_callback(token_ids.data(), token_ids.size(), cached);
+    };
+    auto response_func = [response_callback](LLModel::Token token_id, std::string_view piece) {
+        return response_callback(token_id, piece.data());
+    };
 
     // Call the C++ prompt method
-    wrapper->llModel->prompt(prompt, prompt_template, prompt_callback, response_func, recalculate_callback,
-                             wrapper->promptContext, special, fake_reply_p);
+    try {
+        wrapper->llModel->prompt(prompt, prompt_func, response_func, promptContext);
+    } catch (std::exception const &e) {
+        llmodel_set_error(error, e.what());
+        return false;
+    }
 
-    // Update the C context by giving access to the wrappers raw pointers to std::vector data
-    // which involves no copies
-    ctx->tokens = wrapper->promptContext.tokens.data();
-    ctx->tokens_size = wrapper->promptContext.tokens.size();
-
-    // Update the rest of the C prompt context
-    ctx->n_past = wrapper->promptContext.n_past;
-    ctx->n_ctx = wrapper->promptContext.n_ctx;
-    ctx->n_predict = wrapper->promptContext.n_predict;
-    ctx->top_k = wrapper->promptContext.top_k;
-    ctx->top_p = wrapper->promptContext.top_p;
-    ctx->min_p = wrapper->promptContext.min_p;
-    ctx->temp = wrapper->promptContext.temp;
-    ctx->n_batch = wrapper->promptContext.n_batch;
-    ctx->repeat_penalty = wrapper->promptContext.repeat_penalty;
-    ctx->repeat_last_n = wrapper->promptContext.repeat_last_n;
-    ctx->context_erase = wrapper->promptContext.contextErase;
+    return true;
 }
 
 float *llmodel_embed(
@@ -296,3 +300,21 @@ const char *llmodel_model_gpu_device_name(llmodel_model model)
     const auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->gpuDeviceName();
 }
+
+int32_t llmodel_count_prompt_tokens(llmodel_model model, const char *prompt, const char **error)
+{
+    auto *wrapper = static_cast<const LLModelWrapper *>(model);
+    try {
+        return wrapper->llModel->countPromptTokens(prompt);
+    } catch (const std::exception& e) {
+        llmodel_set_error(error, e.what());
+        return -1;
+    }
+}
+
+void llmodel_model_foreach_special_token(llmodel_model model, llmodel_special_token_callback callback)
+{
+    auto *wrapper = static_cast<const LLModelWrapper *>(model);
+    for (auto &[name, token] : wrapper->llModel->specialTokens())
+        callback(name.c_str(), token.c_str());
+}

gpt4all-backend-old/src/llmodel_shared.cpp

@@ -0,0 +1,298 @@
+#include "llmodel.h"
+
+#include <algorithm>
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+#include <iostream>
+#include <iterator>
+#include <optional>
+#include <ranges>
+#include <stdexcept>
+#include <string>
+#include <string_view>
+#include <vector>
+
+namespace ranges = std::ranges;
+namespace views  = std::ranges::views;
+
+void LLModel::prompt(
+    std::string_view        prompt,
+    const PromptCallback   &promptCallback,
+    const ResponseCallback &responseCallback,
+    const PromptContext    &promptCtx
+) {
+    if (!isModelLoaded())
+        throw std::invalid_argument("Attempted to prompt an unloaded model.");
+    if (!supportsCompletion())
+        throw std::invalid_argument("Not a text completion model.");
+    if (!promptCtx.n_batch)
+        throw std::invalid_argument("Batch size cannot be zero.");
+    if (!promptCtx.n_predict)
+        return; // nothing requested
+
+    auto embd_inp = tokenize(prompt);
+    if (embd_inp.empty())
+        throw std::invalid_argument("Prompt tokenized to zero tokens.");
+
+    if (auto res = decodePrompt(promptCallback, promptCtx, std::move(embd_inp)))
+        generateResponse(responseCallback, promptCtx, /*n_past*/ *res);
+}
+
+int32_t LLModel::countPromptTokens(std::string_view prompt) const
+{
+    if (!isModelLoaded())
+        throw std::invalid_argument("Attempted to tokenize with an unloaded model.");
+    return int32_t(tokenize(prompt).size());
+}
+
+auto LLModel::decodePrompt(
+    const PromptCallback &promptCallback,
+    const PromptContext  &promptCtx,
+    std::vector<Token>    embd_inp
+) -> std::optional<int32_t>
+{
+    assert(!embd_inp.empty());
+
+    int32_t nCtx = contextLength();
+    int32_t n_batch = std::min(promptCtx.n_batch, LLMODEL_MAX_PROMPT_BATCH);
+
+    // Find the greatest n_past where the beginning of embd_inp matches the end of the token cache, starting at the
+    // requested n_past.
+    // This is used to skip unnecessary work when the prompt shares a common prefix with the previous result.
+    int32_t nPast = computeModelInputPosition(embd_inp);
+
+    // always decode up to a full batch before generating, even if cached
+    nPast -= std::min(n_batch, nPast);
+
+    // TODO(jared): generalize this to find the smallest new_embd_inp.size() - nPast given the cache
+    if (!nPast && int32_t(embd_inp.size()) > nCtx) {
+        // no cache hit -> shift the input before even processing
+
+        int32_t nKeep     = shouldAddBOS();
+        auto    newLength = int32_t(nCtx * (1.f - promptCtx.contextErase));
+        int32_t nDiscard  = int32_t(embd_inp.size()) - std::max(1, std::min(nCtx, newLength));
+
+        // execute the callback even for skipped tokens. this misrepresents the position of BOS but we don't care
+        auto discardedTokens = embd_inp | views::drop(nKeep) | views::take(nDiscard);
+        if (!promptCallback(discardedTokens, true))
+            return std::nullopt;
+
+        // erase nDiscard tokens
+        embd_inp.erase(discardedTokens.begin(), discardedTokens.end());
+        assert(int32_t(embd_inp.size()) <= nCtx);
+
+        // check the cache again, just in case
+        nPast = computeModelInputPosition(embd_inp);
+        nPast -= std::min(n_batch, nPast);
+    }
+
+    setModelInputPosition(nPast);
+
+    // execute the callback even for skipped tokens
+    if (!promptCallback(embd_inp | views::take(nPast), true))
+        return std::nullopt;
+
+    // process the prompt in batches
+    for (int32_t i = nPast; i < embd_inp.size();) {
+        auto batch_end = std::min(i + n_batch, int32_t(embd_inp.size()));
+        std::span batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
+
+        // Check if the context has run out...
+        if (nPast + int32_t(batch.size()) > nCtx) {
+            shiftContext(promptCtx, &nPast);
+            assert(nPast + int32_t(batch.size()) <= nCtx);
+        }
+
+        // FIXME(Adam): We should find a way to bubble these strings to the UI level to allow for translation
+        if (!evalTokens(nPast, batch))
+            throw std::runtime_error("An internal error was encountered during prompt processing.");
+
+        for (auto &tok : batch) {
+            appendInputToken(tok);
+            nPast++;
+            if (!promptCallback({ &tok, 1 }, false))
+                return std::nullopt;
+        }
+        i = batch_end;
+    }
+
+    return nPast;
+}
+
+/*
+ * If string s overlaps with the string key such that some prefix of the key is at the end
+ * of the string, return the position in s where the first match starts. Otherwise, return
+ * std::string::npos. Examples:
+ * s = "bfo",  key = "foo" -> 1
+ * s = "fooa", key = "foo" -> npos
+ */
+static std::string::size_type stringsOverlap(const std::string &s, const std::string &key)
+{
+    if (s.empty() || key.empty())
+        throw std::invalid_argument("arguments to stringsOverlap must not be empty");
+
+    for (int start = std::max(0, int(s.size()) - int(key.size())); start < s.size(); start++) {
+        if (s.compare(start, s.size(), key, 0, s.size() - start) == 0)
+            return start;
+    }
+    return std::string::npos;
+}
+
+void LLModel::generateResponse(
+    const ResponseCallback &responseCallback,
+    const PromptContext    &promptCtx,
+    int32_t                 nPast
+) {
+    static const char *stopSequences[] {
+        "### System", "### Instruction", "### Human", "### User", "### Response", "### Assistant", "### Context",
+        "<|im_start|>", "<|im_end|>", "<|endoftext|>",
+    };
+
+    initSampler(promptCtx);
+
+    std::string cachedResponse;
+    std::vector<Token> cachedTokens;
+    int n_predicted = 0;
+
+    // Predict next tokens
+    for (bool stop = false; !stop;) {
+        // Sample next token
+        std::optional<Token> new_tok = sampleToken();
+        std::string new_piece = tokenToString(new_tok.value());
+        cachedTokens.push_back(new_tok.value());
+        cachedResponse += new_piece;
+
+        auto accept = [this, &promptCtx, &new_tok, &nPast] {
+            // Shift context if out of space
+            if (nPast >= contextLength()) {
+                shiftContext(promptCtx, &nPast);
+                assert(nPast < contextLength());
+            }
+
+            // Accept the token
+            Token tok = std::exchange(new_tok, std::nullopt).value();
+            if (!evalTokens(nPast, { &tok, 1 }))
+                throw std::runtime_error("An internal error was encountered during response generation.");
+
+            appendInputToken(tok);
+            nPast++;
+        };
+
+        // Check for EOS
+        auto lengthLimit = std::string::npos;
+        for (const auto token : endTokens()) {
+            if (new_tok == token) {
+                stop = true;
+                lengthLimit = cachedResponse.size() - new_piece.size();
+            }
+        }
+
+        if (lengthLimit != std::string::npos) {
+            // EOS matched
+        } else if (!isSpecialToken(new_tok.value())) {
+            // Check if the response contains a stop sequence
+            for (const auto &p : stopSequences) {
+                auto match = cachedResponse.find(p);
+                if (match != std::string::npos) stop = true;
+                lengthLimit = std::min(lengthLimit, match);
+                if (match == 0) break;
+            }
+
+            // Check if the response matches the start of a stop sequence
+            if (lengthLimit == std::string::npos) {
+                for (const auto &p : stopSequences) {
+                    auto match = stringsOverlap(cachedResponse, p);
+                    lengthLimit = std::min(lengthLimit, match);
+                    if (match == 0) break;
+                }
+            }
+        } else if (ranges::find(stopSequences, new_piece) < std::end(stopSequences)) {
+            // Special tokens must exactly match a stop sequence
+            stop = true;
+            lengthLimit = cachedResponse.size() - new_piece.size();
+        }
+
+        // Empty the cache, up to the length limit
+        std::string::size_type responseLength = 0;
+        while (!cachedTokens.empty()) {
+            Token tok = cachedTokens.front();
+            std::string piece = tokenToString(tok);
+
+            // Stop if the piece (or part of it) does not fit within the length limit
+            if (responseLength + (stop ? 1 : piece.size()) > lengthLimit)
+                break;
+
+            // Remove token from cache
+            assert(cachedResponse.starts_with(piece));
+            cachedTokens.erase(cachedTokens.begin(), cachedTokens.begin() + 1);
+            cachedResponse.erase(cachedResponse.begin(), cachedResponse.begin() + piece.size());
+
+            // Accept the token, if needed (not cached)
+            if (cachedTokens.empty() && new_tok)
+                accept();
+
+            // Send the token
+            if (!responseCallback(tok, piece) || ++n_predicted >= promptCtx.n_predict) {
+                stop = true;
+                break;
+            }
+
+            // FIXME(jared): we could avoid printing partial stop sequences if we didn't have to
+            // output token IDs and could cache a partial token for the next prompt call
+            responseLength += piece.size();
+        }
+        assert(cachedTokens.empty() == cachedResponse.empty());
+
+        // Accept the token, if needed (in cache)
+        if (new_tok) {
+            assert(!cachedTokens.empty() && cachedTokens.back() == new_tok);
+            if (stop) {
+                cachedTokens.pop_back();
+            } else {
+                accept();
+            }
+        }
+    }
+
+    if (inputLength() < cachedTokens.size()) {
+        /* This is theoretically possible if the longest stop sequence is greater than
+         * n_ctx * contextErase tokens. */
+        throw std::runtime_error("shifted too much context, can't go back");
+    }
+
+#ifndef NDEBUG
+    auto inp = inputTokens();
+    auto discard_start = inp.end() - cachedTokens.size();
+    assert(std::equal(discard_start, inp.end(), cachedTokens.begin()));
+#endif
+}
+
+void LLModel::embed(
+    const std::vector<std::string> &texts, float *embeddings, std::optional<std::string> prefix, int dimensionality,
+    size_t *tokenCount, bool doMean, bool atlas, EmbedCancelCallback *cancelCb
+) {
+    (void)texts;
+    (void)embeddings;
+    (void)prefix;
+    (void)dimensionality;
+    (void)tokenCount;
+    (void)doMean;
+    (void)atlas;
+    (void)cancelCb;
+    throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
+}
+
+void LLModel::embed(
+    const std::vector<std::string> &texts, float *embeddings, bool isRetrieval, int dimensionality, size_t *tokenCount,
+    bool doMean, bool atlas
+) {
+    (void)texts;
+    (void)embeddings;
+    (void)isRetrieval;
+    (void)dimensionality;
+    (void)tokenCount;
+    (void)doMean;
+    (void)atlas;
+    throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
+}

gpt4all-backend-old/src/utils.h

@@ -0,0 +1,17 @@
+#pragma once
+
+#include <cassert>
+
+#ifdef NDEBUG
+#   ifdef __has_builtin
+#       if __has_builtin(__builtin_unreachable)
+#           define UNREACHABLE() __builtin_unreachable()
+#       else
+#           define UNREACHABLE() do {} while (0)
+#       endif
+#   else
+#       define UNREACHABLE() do {} while (0)
+#   endif
+#else
+#   define UNREACHABLE() assert(!"Unreachable statement was reached")
+#endif

gpt4all-backend-test/CMakeLists.txt

@@ -0,0 +1,8 @@
+cmake_minimum_required(VERSION 3.29)
+project(gpt4all-backend-test VERSION 0.1 LANGUAGES CXX)
+
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin")
+include(../common/common.cmake)
+
+add_subdirectory(../gpt4all-backend "${CMAKE_CURRENT_BINARY_DIR}/gpt4all-backend")
+add_subdirectory(src)

gpt4all-backend-test/src/CMakeLists.txt

@@ -0,0 +1,21 @@
+set(TARGET test-backend)
+
+configure_file(config.cppm.in "${CMAKE_CURRENT_BINARY_DIR}/config.cppm")
+
+add_executable(${TARGET}
+    main.cpp
+)
+target_compile_features(${TARGET} PUBLIC cxx_std_23)
+if (CMAKE_COMPILER_IS_GNUCXX)
+    target_compile_options(${TARGET} PUBLIC -fmodules-ts)
+endif()
+target_sources(${TARGET} PRIVATE
+    FILE_SET gpt4all_backend TYPE CXX_MODULES BASE_DIRS
+        "${CMAKE_CURRENT_BINARY_DIR}"
+    FILES
+        "${CMAKE_CURRENT_BINARY_DIR}/config.cppm"
+)
+gpt4all_add_warning_options(${TARGET})
+target_link_libraries(${TARGET} PRIVATE
+    gpt4all-backend
+)

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

@@ -0,0 +1,10 @@
+module;
+
+#include <QString>
+
+export module gpt4all.test.config;
+
+using namespace Qt::Literals::StringLiterals;
+
+
+export inline QString OLLAMA_URL = u"@G4A_TEST_OLLAMA_URL@"_s;

gpt4all-backend-test/src/main.cpp

@@ -0,0 +1,14 @@
+#include <QLatin1StringView>
+
+import fmt;
+import gpt4all.backend.main;
+import gpt4all.test.config;
+
+using gpt4all::backend::LLMProvider;
+
+
+int main()
+{
+    LLMProvider provider(OLLAMA_URL);
+    fmt::print("Server version: {}", provider.getVersion());
+}

gpt4all-backend/CMakeLists.txt

@@ -1,157 +1,11 @@
-cmake_minimum_required(VERSION 3.21)  # for PROJECT_IS_TOP_LEVEL
-set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
-set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+cmake_minimum_required(VERSION 3.29)
+project(gpt4all-backend VERSION 0.1 LANGUAGES CXX)
 
-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()
+set(CMAKE_CXX_STANDARD 23)  # make sure fmt module is compiled with the same C++ version as us
+include(../common/common.cmake)
 
-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()
+find_package(Qt6 6.8 COMPONENTS Concurrent Core Network REQUIRED)
 
-# 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 20)
-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: ${IPO_ERROR}")
-else()
-    message(STATUS "Interprocedural optimization support detected")
-endif()
-
-set(DIRECTORY llama.cpp-mainline)
-include(llama.cpp.cmake)
-
-set(BUILD_VARIANTS)
-if (APPLE)
-    list(APPEND BUILD_VARIANTS metal)
-endif()
-if (LLMODEL_KOMPUTE)
-    list(APPEND BUILD_VARIANTS kompute kompute-avxonly)
-else()
-    list(PREPEND BUILD_VARIANTS cpu cpu-avxonly)
-endif()
-if (LLMODEL_VULKAN)
-    list(APPEND BUILD_VARIANTS vulkan vulkan-avxonly)
-endif()
-if (LLMODEL_CUDA)
-    if (DEFINED CMAKE_CUDA_ARCHITECTURES)
-        set(GGML_CUDA_ARCHITECTURES "${CMAKE_CUDA_ARCHITECTURES}")
-    endif()
-
-    include(CheckLanguage)
-    check_language(CUDA)
-    if (NOT CMAKE_CUDA_COMPILER)
-        message(WARNING "CUDA Toolkit not found. To build without CUDA, use -DLLMODEL_CUDA=OFF.")
-    endif()
-    enable_language(CUDA)
-    list(APPEND BUILD_VARIANTS cuda cuda-avxonly)
-endif()
-if (LLMODEL_ROCM)
-    enable_language(HIP)
-    list(APPEND BUILD_VARIANTS rocm rocm-avxonly)
-endif()
-
-set(CMAKE_VERBOSE_MAKEFILE ON)
-
-# 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})
-
-    # 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
-        llamamodel.cpp llmodel_shared.cpp)
-    target_compile_definitions(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
-        LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
-    prepare_target(llamamodel-mainline llama-mainline)
-
-    if (NOT PROJECT_IS_TOP_LEVEL AND BUILD_VARIANT STREQUAL cuda)
-        set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
-    endif()
-endforeach()
-
-add_library(llmodel
-    llmodel.h llmodel.cpp llmodel_shared.cpp
-    llmodel_c.h llmodel_c.cpp
-    dlhandle.cpp
-)
-target_compile_definitions(llmodel PRIVATE LIB_FILE_EXT="${CMAKE_SHARED_LIBRARY_SUFFIX}")
-
-set_target_properties(llmodel PROPERTIES
-                              VERSION ${PROJECT_VERSION}
-                              SOVERSION ${PROJECT_VERSION_MAJOR})
-
-set(COMPONENT_NAME_MAIN ${PROJECT_NAME})
-set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install)
+set(FMT_MODULE ON)
+add_subdirectory(../deps "${CMAKE_CURRENT_BINARY_DIR}/common_deps")
+add_subdirectory(src)

gpt4all-backend/llmodel_shared.cpp

@@ -1,312 +0,0 @@
-#include "llmodel.h"
-
-#include <algorithm>
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <functional>
-#include <iostream>
-#include <optional>
-#include <regex>
-#include <stdexcept>
-#include <string>
-#include <unordered_set>
-#include <vector>
-
-// 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)
-{
-    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);
-        assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
-        if (!evalTokens(promptCtx, batch)) {
-            std::cerr << "LLModel ERROR: Failed to process prompt\n";
-            goto stop_generating;
-        }
-        promptCtx.n_past += batch.size();
-        if (!recalculate(true))
-            goto stop_generating;
-        i = batch_end;
-    }
-    assert(promptCtx.n_past == int32_t(promptCtx.tokens.size()));
-
-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,
-                     bool special,
-                     std::string *fakeReply)
-{
-    if (!isModelLoaded()) {
-        std::cerr << implementation().modelType() << " ERROR: prompt won't work with an unloaded model!\n";
-        return;
-    }
-
-    if (!supportsCompletion()) {
-        std::string errorMessage = "ERROR: this model does not support text completion or chat!";
-        responseCallback(-1, errorMessage);
-        std::cerr << implementation().modelType() << " " << errorMessage << "\n";
-        return;
-    }
-
-    // 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
-    if (!decodePrompt(promptCallback, responseCallback, recalculateCallback, promptCtx, embd_inp))
-        return; // error
-
-    // decode the assistant's reply, either generated or spoofed
-    if (fakeReply == nullptr) {
-        generateResponse(responseCallback, recalculateCallback, promptCtx);
-    } else {
-        embd_inp = tokenize(promptCtx, *fakeReply, false);
-        if (!decodePrompt(promptCallback, responseCallback, recalculateCallback, promptCtx, embd_inp))
-            return; // error
-    }
-
-    // decode the rest of the prompt template
-    // template: end of assistant prompt
-    std::string asstSuffix;
-    if (placeholders.size() >= 2) {
-        size_t start = placeholders[1].position() + placeholders[1].length();
-        asstSuffix = promptTemplate.substr(start);
-    } else {
-        asstSuffix = "\n\n"; // default to a blank link, good for e.g. Alpaca
-    }
-    if (!asstSuffix.empty()) {
-        embd_inp = tokenize(promptCtx, asstSuffix, true);
-        decodePrompt(promptCallback, responseCallback, recalculateCallback, promptCtx, embd_inp);
-    }
-}
-
-// returns false on error
-bool 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();
-
-    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() <<
-            " tokens and the context window is " << promptCtx.n_ctx << "!\n";
-        return false;
-    }
-
-    promptCtx.n_predict = std::min(promptCtx.n_predict, promptCtx.n_ctx - (int) embd_inp.size());
-    promptCtx.n_past = std::min(promptCtx.n_past, promptCtx.n_ctx);
-    promptCtx.n_batch = std::min(promptCtx.n_batch, LLMODEL_MAX_PROMPT_BATCH);
-
-    // process the prompt in batches
-    size_t i = 0;
-    while (i < embd_inp.size()) {
-        size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
-        std::vector<Token> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
-
-        // Check if the context has run out...
-        if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
-            recalculateContext(promptCtx, recalculateCallback);
-            assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
-        }
-
-        if (!evalTokens(promptCtx, batch)) {
-            std::cerr << implementation().modelType() << " ERROR: Failed to process prompt\n";
-            return false;
-        }
-
-        size_t tokens = batch_end - i;
-        for (size_t t = 0; t < tokens; ++t) {
-            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 false;
-        }
-        i = batch_end;
-    }
-
-    return true;
-}
-
-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
-        = { "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant", "### Context" };
-
-    // predict next tokens
-    for (int i = 0; i < promptCtx.n_predict; i++) {
-
-        // sample next token
-        auto id = sampleToken(promptCtx);
-
-        // Check if the context has run out...
-        if (promptCtx.n_past + 1 > promptCtx.n_ctx) {
-            recalculateContext(promptCtx, recalculateCallback);
-            assert(promptCtx.n_past + 1 <= promptCtx.n_ctx);
-        }
-
-        if (!evalTokens(promptCtx, { id })) {
-            std::cerr << implementation().modelType() << " ERROR: Failed to predict next token\n";
-            return;
-        }
-
-        // display text
-        for (const auto token : endTokens()) {
-            if (id == token) return;
-        }
-
-        const std::string str = tokenToString(id);
-
-        // Check if the provided str is part of our reverse prompts
-        bool foundPartialReversePrompt = false;
-        const std::string completed = cachedResponse + std::string(str);
-        if (reversePrompts.find(completed) != reversePrompts.end())
-            return;
-
-        // Check if it partially matches our reverse prompts and if so, cache
-        for (const auto& s : reversePrompts) {
-            if (s.compare(0, completed.size(), completed) == 0) {
-                foundPartialReversePrompt = true;
-                cachedResponse = completed;
-                break;
-            }
-        }
-
-        // Regardless the token gets added to our cache
-        cachedTokens.push_back(id);
-
-        // Continue if we have found a partial match
-        if (foundPartialReversePrompt)
-            continue;
-
-        // Empty the cache
-        for (auto t : cachedTokens) {
-            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;
-        }
-        cachedTokens.clear();
-    }
-}
-
-void LLModel::embed(
-    const std::vector<std::string> &texts, float *embeddings, std::optional<std::string> prefix, int dimensionality,
-    size_t *tokenCount, bool doMean, bool atlas, EmbedCancelCallback *cancelCb
-) {
-    (void)texts;
-    (void)embeddings;
-    (void)prefix;
-    (void)dimensionality;
-    (void)tokenCount;
-    (void)doMean;
-    (void)atlas;
-    (void)cancelCb;
-    throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
-}
-
-void LLModel::embed(
-    const std::vector<std::string> &texts, float *embeddings, bool isRetrieval, int dimensionality, size_t *tokenCount,
-    bool doMean, bool atlas
-) {
-    (void)texts;
-    (void)embeddings;
-    (void)isRetrieval;
-    (void)dimensionality;
-    (void)tokenCount;
-    (void)doMean;
-    (void)atlas;
-    throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
-}

gpt4all-backend/llmodel_shared.h

@@ -1,49 +0,0 @@
-#pragma once
-
-#include <ggml.h>
-
-#include <cstddef>
-#include <cstdint>
-#include <vector>
-
-struct llm_buffer {
-    uint8_t * addr = NULL;
-    size_t size = 0;
-
-    void resize(size_t size) {
-        delete[] addr;
-        addr = new uint8_t[size];
-        this->size = size;
-    }
-
-    ~llm_buffer() {
-        delete[] addr;
-    }
-};
-
-struct llm_kv_cache {
-    struct ggml_tensor * k;
-    struct ggml_tensor * v;
-
-    struct ggml_context * ctx = NULL;
-
-    llm_buffer buf;
-
-    int n; // number of tokens currently in the cache
-
-    ~llm_kv_cache() {
-        if (ctx) {
-            ggml_free(ctx);
-        }
-    }
-};
-
-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) {
-        buf.resize(plan.work_size);
-        plan.work_data = buf.addr;
-    }
-    ggml_graph_compute(graph, &plan);
-}

gpt4all-backend/src/CMakeLists.txt

@@ -0,0 +1,20 @@
+set(TARGET gpt4all-backend)
+
+add_library(${TARGET} STATIC
+    main.cpp
+)
+target_compile_features(${TARGET} PUBLIC cxx_std_23)
+if (CMAKE_COMPILER_IS_GNUCXX)
+    target_compile_options(${TARGET} PUBLIC -fmodules-ts)
+endif()
+target_sources(${TARGET} PUBLIC
+    FILE_SET gpt4all_backend TYPE CXX_MODULES FILES
+    main.cppm
+)
+gpt4all_add_warning_options(${TARGET})
+target_link_libraries(${TARGET} PUBLIC
+    QCoro6::Coro Qt6::Core Qt6::Network
+)
+target_link_libraries(${TARGET} PRIVATE
+    QCoro6::Network fmt::fmt
+)

gpt4all-backend/src/main.cpp

@@ -0,0 +1,33 @@
+module;
+
+#include <expected>
+#include <memory>
+
+#include <QFuture>
+#include <QNetworkAccessManager>
+#include <QNetworkReply>
+#include <QNetworkRequest>
+#include <QByteArray>
+#include <QString>
+
+#include <QCoro/QCoroNetworkReply>
+
+using namespace Qt::Literals::StringLiterals;
+
+module gpt4all.backend.main;
+
+
+namespace gpt4all::backend {
+
+auto LLMProvider::getVersion() const -> QCoro::Task<DataOrNetErr<QString>>
+{
+    QNetworkAccessManager nam;
+    std::unique_ptr<QNetworkReply> reply(co_await nam.get(QNetworkRequest(m_baseUrl.resolved(u"/api/version"_s))));
+    if (auto err = reply->error())
+        co_return std::unexpected(err);
+
+    // TODO(jared): parse JSON here instead of just returning the data
+    co_return QString::fromUtf8(reply->readAll());
+}
+
+} // namespace gpt4all::backend

gpt4all-backend/src/main.cppm

@@ -0,0 +1,38 @@
+module;
+
+#include <expected>
+
+#include <QNetworkReply>
+#include <QUrl>
+
+#include <QCoro/QCoroTask>
+
+class QString;
+template <typename T> class QFuture;
+
+export module gpt4all.backend.main;
+
+
+export namespace gpt4all::backend {
+
+template <typename T>
+using DataOrNetErr = std::expected<T, QNetworkReply::NetworkError>;
+
+
+class LLMProvider {
+public:
+    LLMProvider(QUrl baseUrl)
+        : m_baseUrl(baseUrl)
+        {}
+
+    const QUrl &baseUrl() const { return m_baseUrl; }
+    void getBaseUrl(QUrl value) { m_baseUrl = std::move(value); }
+
+    /// Retrieve the Ollama version, e.g. "0.5.1"
+    auto getVersion() const -> QCoro::Task<DataOrNetErr<QString>>;
+
+private:
+    QUrl m_baseUrl;
+};
+
+} // namespace gpt4all::backend