Report on issues

Open Issues Summary

• Backend and GUI: Freeze/crash when running the project with the Vulcan backend (#1572), rebase required for upstream of Vulcan backend (#1571), stuck instances in Windows, causing application failures (#1560, #1552), potential issues with multi GPU offloading (#1562), problem with older Intel CPUs (#1540), Windows installer infinite loop issue (#1378), issues with CUDA support (#1304).
• Models: Issues with model validation and changes in models.json (#1569), VRAM not freed when switching models (#1550), chat UI not recognizing new models on OS X (#1440), models not detected on NVIDIA GPUs older than Turing (#1425), issues with AVX512 instructions of Xeon Phi KNL 7200 Series Processors (#1305).
• Plugins: Enhancements requested for LocalDocs plugin (#1570), issues with LocalDocs plugin not working (#1362), filesystem types to include in LocalDoc features (#1306), OCR for PDF files in Local Docs requested (#1255).
• API: HTTP API server returning empty results (#1466), API incessantly reloading models (#1274), API memory usage doubled (#1274).
• Bindings: C# bindings failing to load models (#1534), issue with C# NuGet package (#1520), issues with TypeScript package (#1497), request for Java bindings change (#1360).
• Features and Enhancement Requests: Request for multimodal model support (#1568), Pydantic integration (#1565), feature requests for Vulkan acceleration (#1551), support for partial GPU offloading (#1562), request for support of older Intel CPUs (#1540), deletion of old chat without recalculating context (#1545), enhancement requests for better GUI (#1307), automatic increment of "New Chat" if same name exists (#1475).
• Errors and Crashes: Errors in allocating memory (#1543), chat crashes with a network error (#1549), crash due to memory corruption issues (#1374), application crashes on startup if model dir is in a Windows network drive (#1364).
• Others: Issues with language support (#1305, #1321, #1393), request for different model download mechanism (#1548), ChatGPT plugin functionality (#1391), Freewilly1 and Freewilly2 support (#1275), plugin functionality request (#1483), API response contains invalid date (#1353).

Recently closed issues indicate a progressive trend towards addressing and fixing backend and GUI related problems of the software project, although some new ones have surfaced.

Report on pull requests

Open Pull Requests:

• #1417: Added chatgpt style plugin functionality to Python bindings for GPT4All with minor modifications to existing codebase.
• #1390: Vulkan support added for TypeScript bindings and gguf support.
• #1379: Prompt context added to completion response in TypeScript bindings, allows the display of context in user interfaces.
• #1369: Updated the gpt4all API's docker container for faster and smaller operation. Stopped bundling models inside images and used external volumes.
• #1366: Added functionality to list all available GPT4All models by updating Settings.
• #1361: Plans to unify paths for backend, chat and bindings. PR lacks details.
• #1309: PR adds quantized Redmond Puffin 13B to the list of recommended models.
• #1293: Improved Python tests and code through addition of pytest markers and model folder customisation.
• #1285: Fixed a bug that prevented multi-epoch checkpoint loading.
• #1232: Added a feature to reverse prompts in python bindings.
• #1229: An example pipeline featuring GPT4All, MongoDB, ElasticSearch and Docker has been proposed.
• #1126: Enhancements proposed for the LLModel class. Makes PrintStream an alternative to System.out.
• #1077: Resolves the 'beginning of sentence' (BOS) error in text completion requests. Associated with #996.
• #1067: Introduced CodeQL workflow for better security in this open-source software project.
• #1054: User experience and documentation improvements within CLI suggested.
• #1023: Proposal of Dart binding implementation for GPT4All.
• #984: Implementation of cake frosting to standardize build pipeline in Csharp.
• #896: Some test suites are proposed for testing llmodel C and C++ interfaces.
• #860: Support for OpenAI chat models are suggested for GPT4All Python API.
• #765: New tokenizer implementation for MPT and GPT-J proposed.
• #746: CUDA and OpenCL support added in this PR, but need testing for CUDA.

Recently Closed Pull Requests:

• #1564: The default model was changed to Mistral Instruct.
• #1563: The addition of a tokenizer merges the mpt-7b-chat model.
• #1561: A clear note on the new GGUF format was added to the readme.
• #1559: The default extension was changed to .gguf in the python bindings.
• #1548: A syntax error in main.qml was fixed in the chat functionality.
• #1546: Changes implemented to avoid loading libraries during static initialization.
• #1538: Default model URLs updated.
• #1532: Extraneous and shorter text added to the accessibility name and description fields.
• #1508: Experimentation was conducted with shaders for matmat and matvec operations.
• #1493: Added functionality to restore state from text.
• #1485: Support was added for the EM German Mistral Model.
• #1478: Clearer CPU fallback messages were implemented in the chat model.
• #1462: In the Python bindings, support for the Path in GPT4All.init was added.
• #1444: Git clone SSH URI was replaced with HTTPS URL.
• #1434: Fixed the MSVC compilation issues on windows with Vulkan bindings.
• #1412: Improved Java API signatures to maintain backward compatibility.
• #1410: Non-constructive change. PR should be ignored.
• #1398: Reduced dynamic linker dependencies installed with Vulkan.
• #1387: Nonconstructive PR, likely a testing oversight.
• #1383: Change implemented related to the Vulkan backend.
• #1375: Configuration update for the Python binding.
• #1368: Integration added for Weaviate.
• #1348: Error messages were added to backend in case of model creation failure.
• #1341: Typescript tests have been fixed to be compatible with Windows.
• #1333: The gpt4all_faq.md was updated.
• #1322: Font size changer was added and current updates initiated.
• #1315: An idea to a wishlist was added.
• #1280: An adjustment between light mode and dark mode was proposed.
• #1264: Fixes were implemented for batched GPU inference.
• #1244: Addition of an error message when llmodel_model_create fails.
• #1231: Reverse prompts functionality was added in python bindings.
• #1219: The default TopP setting was updated to 0.4.
• #1190: Changes in model location procedure to find models in model directories were corrected.
• #1183: The gpt4all_faq.md document was updated.
• #1118: Wayland support was enabled only for Linux builds.
• #1114: The bug that preserved chat history between sessions was fixed in the Python bindings.
• #1081: Fixes were implemented to make eval buf larger in the model Falcon.
• #1076: The Falcon 7B model was added.
• #1070: An update was added for the codacy.yml file.
• #1053: Java bindings have been updated for the support of gpt4all version 2.4.8.
• #1012: The README of Python bindings was updated with troubleshooting information.
• #1001: Functionality improved to support loading of files if 'ggml' is found anywhere in the file name.
• #999: Readme files got updated with the addition of emojis.
• #984: Standardized the build pipeline using implemented cake frosting.
• #963: The llama.cpp mainline version was bumped to latest.
• #961: Setup.py version was bumped.
• #955: Improvement in text quality with the help of a tokenizer in MPT and GPT-J models.
• #948: Compatibility of llmodel_model_create function was fixed.
• #932: A bug in train.py file resulting in wrong checkpoint resuming was fixed.
• #867: The decode function in Python was updated to include emojis.
• #800: The issue with chatllm not cleaning up properly was resolved.
• #796: Fixed an issue where dynamic path creation disabled the launch of models
• #785: Fixed deduplication of qml for buttons and confirmation of all buttons having accessible properties.
• #769: A markdown file, which is used for documentation was misaligned, this PR fixed that.
• #763: A library loader was added for .NET Bindings. It made updates in the bindings to support the newest changes.
• #743: The use of outdated, ambiguous term "OSX" was replaced with the current term "macOS".
• #742: The creation of duplicate threads for loading and unloading models in a poorly managed memory setup was resolved.
• #691: The PR added significant features, including a simple CLI chat repl – a read-eval-print-loop for chat functions.
• #669: Fonts used in qml were corrected to a monospace variant, which solved #1026.
• #626: Issued with punctuations and capitalizations were corrected.
• #612: An error in the implementation of llmodel_model_create function was fixed.
• #607: Server mode was created with the addition of custom Ports.
• #596: Documentations were adjusted to accommodate updated API key.
• #593: Python binding prompt formatting was improved.
• #582: Windows user instructions were updated, making it easier for them to install the required packages and setup.
• #580: Fixed bugs in train.py responsible for causing wrong readings on the train tracker during longer runs.
• #566: Certain characters were removed that were causing issues with the fetch data script.
• #547: Count of max tokens to generate and the corresponding format for prompts were decided.
• #533: A checkpoint was used to improve the training process for datasets.
• #530: The selection of model was automated if only one was found.
• #522: In the PR, an attempt was made to allow to disable the verbosity feature.

Please note that the PR numbers (#) are included for easy lookup. The briefs provided do not indicate the final resolution state of the PRs. It is recommended to analyze each PR for complete information, particularly those needing corrective actions.

Report on README

Project Summary

GPT4All is an open-source project that allows the execution of powerful and customized language models on consumer-grade CPUs and virtually any GPU. The models are plug-and-play files maintained by Nomic AI and are intended to be easily trainable and deployable by individuals and enterprises.

Key Points

• Starting from v2.5.0, GPT4All only supports models in GGUF format (.gguf extension). Previous models with .bin extension will not function.
• GPT4All has Python, Typescript, Golang, C#, and Java bindings to interface with the application.
• GPT4All models can be utilized via a desktop chat client.
• GPT4All features an active issue tracker, indicating its ongoing development and support.

Uncertainties/Todos

• The readme specifies that the project will only run if the CPU supports AVX or AVX2 instructions, but there is no clarification on how to check or handle this.
• It is unclear whether or how users can contribute to the development and training of the models.

Anomalies

• No anomalies were detected.