[Feature] Support old MPT GGUF conversions with duplicated output tensor
dlippold opened this issue · comments
Dietmar Lippold commented
Bug Report
The fine-tuned MPT model from https://huggingface.co/maddes8cht/mosaicml-mpt-7b-instruct-gguf/ in quantization Q4_1 was usabel in release 2.7.2 but not longer in 2.7.3 and later. In particular it is currently not usable.
When I try to load the model file I get the following error message:
Could not load model due to invalid model file for mosaicml-mpt-7b-instruct-Q4_1.gguf
The reason of the problem may have to do with #2006
Steps to Reproduce
- Download the model file from the specified URL
- Start GPT4all
- Choose the downloaded model file
Expected Behavior
The model file should be loaded.
Your Environment
- GPT4All version: 2.7.2, 2.7.3, 2.7.5
- Operating System: Ubuntu Linux 22.04.
- Chat model used (if applicable): see above
Jared Van Bortel commented
I fixed this upstream in ggerganov/llama.cpp#6139 which should make it into the next release of GPT4All (already included in #2310).