[Feature] Upgrade llama.cpp to support Phi-3-mini-128k-instruct and IBM Granite

[dlippold]

Feature Request

It would be nice if llama.cpp could be upgraded because there are new models supported in the recent versions, e.g.:

• IBM Granite (https://github.com/ibm-granite/granite-code-models), see https://github.com/ggerganov/llama.cpp/pull/7481, integrated into llama.cpp on May 28. Model is available here: https://huggingface.co/ibm-granite/granite-8b-code-instruct-GGUF

• Phi-3-medium-128k, see https://github.com/ggerganov/llama.cpp/pull/7225, seems to be integraed end of May. Model is available here: https://huggingface.co/bartowski/Phi-3-medium-128k-instruct-GGUF

[manyoso]

This is being worked on and we'll have a new release out soon that will have latest llama.cpp

[ThiloteE]

llama.cpp has been upgraded, hence will close this issue.

• I am raising a PR for the phi-3 model.
• The IBM Granite models are not tested and may require custom quantizations. If you do the necessary testing and provide quants that are provent to work with GPT4All, I am sure PRs are welcome. Benchmarks of that model family are looking good. Seems to be better than llama-3-8b at coding.