can't use mmap because of ggml?

[aicoder2048]

llama.cpp: can't use mmap because tensors are not aligned; convert to new format to avoid this
llama_model_load_internal: format = 'ggml' (old version with low tokenizer quality and no mmap support)

[imfurkaann]

.

[dennis-gonzales]

same error

[lrodriguezcim]

same!

[nikhil-xb]

I am facing the same problem while loading the llama embeddings. After a bit of exploration, I found that there has been a new quantization format(speed improvement) introduced in the llama-cpp-python supporting repository. Refer this pull request.

Now this project privateGPT and many others are uses old quantization formats. As of now, there don't seem to be a way to convert old quantized format to new ones other than retrieving source models f16 and quantizing them to new formats.

Would recommend following this thread for more info: ggerganov/llama.cpp#1408