What models are supported by the GPT4All ecosystem?
Currently, there are six different model architectures that are supported:
- GPT-J - Based off of the GPT-J architecture with examples found here
- LLaMA - Based off of the LLaMA architecture with examples found here
- MPT - Based off of Mosaic ML's MPT architecture with examples found here
- Replit - Based off of Replit Inc.'s Replit architecture with examples found here
- Falcon - Based off of TII's Falcon architecture with examples found here
- StarCoder - Based off of BigCode's StarCoder architecture with examples found here
Why so many different architectures? What differentiates them?
One of the major differences is license. Currently, the LLaMA based models are subject to a non-commercial license, whereas the GPTJ and MPT base models allow commercial usage. However, its successor Llama 2 is commercially licensable, too. In the early advent of the recent explosion of activity in open source local models, the LLaMA models have generally been seen as performing better, but that is changing quickly. Every week - even every day! - new models are released with some of the GPTJ and MPT models competitive in performance/quality with LLaMA. What's more, there are some very nice architectural innovations with the MPT models that could lead to new performance/quality gains.
How does GPT4All make these models available for CPU inference?
By leveraging the ggml library written by Georgi Gerganov and a growing community of developers. There are currently multiple different versions of this library. The original GitHub repo can be found here, but the developer of the library has also created a LLaMA based version here. Currently, this backend is using the latter as a submodule.
Does that mean GPT4All is compatible with all llama.cpp models and vice versa?
The upstream llama.cpp project has introduced several compatibility breaking quantization methods recently. This is a breaking change that renders all previous models (including the ones that GPT4All uses) inoperative with newer versions of llama.cpp since that change.
Fortunately, we have engineered a submoduling system allowing us to dynamically load different versions of the underlying library so that GPT4All just works.
What are the system requirements?
Your CPU needs to support AVX or AVX2 instructions and you need enough RAM to load a model into memory.
What about GPU inference?
In newer versions of llama.cpp, there has been some added support for NVIDIA GPU's for inference. We're investigating how to incorporate this into our downloadable installers.
Ok, so bottom line... how do I make my model on Hugging Face compatible with GPT4All ecosystem right now?
- Check to make sure the Hugging Face model is available in one of our three supported architectures
- If it is, then you can use the conversion script inside of our pinned llama.cpp submodule for GPTJ and LLaMA based models
- Or if your model is an MPT model you can use the conversion script located directly in this backend directory under the scripts subdirectory
There's a problem with the download
Some bindings can download a model, if allowed to do so. For example, in Python or TypeScript if
allowDownload=true (default), a model is automatically downloaded into
.cache/gpt4all/ in the user's home folder,
unless it already exists.
In case of connection issues or errors during the download, you might want to manually verify the model file's MD5 checksum by comparing it with the one listed in models2.json.
As an alternative to the basic downloader built into the bindings, you can choose to download from the https://gpt4all.io/ website instead. Scroll down to 'Model Explorer' and pick your preferred model.
I need the chat GUI and bindings to behave the same
The chat GUI and bindings are based on the same backend. You can make them behave the same way by following these steps:
First of all, ensure that all parameters in the chat GUI settings match those passed to the generating API, e.g.:
To make comparing the output easier, set Temperature in both to 0 for now. This will make the output deterministic.
Next you'll have to compare the templates, adjusting them as necessary, based on how you're using the bindings.
- Specifically, in Python:
- With simple
generate()calls, the input has to be surrounded with system and prompt templates.
- When using a chat session, it depends on whether the bindings are allowed to download models2.json. If yes,
and in the chat GUI the default templates are used, it'll be handled automatically. If no, use
chat_session()template parameters to customize them.
- With simple
- Specifically, in Python:
Once you're done, remember to reset Temperature to its previous value in both chat GUI and your custom code.