[!WARNING]
Sampling:
Mistral-Nemo-12B is very sensitive to the temperature sampler, try values near 0.3 at first or else you will get some weird results. This is mentioned by MistralAI at their Transformers section.
Flash-Attention seems to have seem weird effects with the model as well, however there is no confirmation on this.
[!IMPORTANT]
IMPORTANT:
Updated version of the model available here: mistral-doryV2-12b-gguf
How to Use:
As of July 20, 2024, llama.cpp does not support Mistral-Nemo. However, you can still use it by building from iamlemec's branch mistral-nemo at llama.cpp GitHub repository.
These quants were quantized off iamlemec's branch using commit 6515e78
Recommended to use cmake -B build -DGGML_CUDA=ON if you're using a CUDA compatible GPU.
If the build takes too long use cmake -B build --config Release -j 4, which uses 4 threads to build. Adjust the number to the amount of physical cores on your CPU.
Set -b to batch size
Set -ub to physical batch size
Set -c to context size
Set -ngl to amount of layers to load onto GPU
Change the path to where the model is actually stored.
If you need more clarification on parameters check out the llama.cpp Server Docs
