maozdemir
maozdemir
@PulpCattel ; > The extra things, like downloading models, etc, could be done in follow-ups I think. I'd also prefer, in case they become more complicated, to have Python scripts...
imo at this point `load_single_document` is redundant and can be merged into `load_documents` function, which will remove the need of doing this. `os.path.join` is already case insensitive and the current...
Your CPU is not really amazing, it's not surprising to have such a performance issue. It is possible to use GPU, though. You just have to modify this for Windows...
> Maybe we should try this instead @maozdemir https://python.langchain.com/en/latest/modules/indexes/document_loaders/examples/file_directory.html?highlight=TextLoader#c-auto-detect-encodings Yes, my langchain was on 0.0.171 and apparently that feature was proposed in 0.0.174. Your suggestion works much better. @imartinez
use Q4_0 models.
I did notice the duplicate in README.md. Will correct that
@johnbrisbin > 1. Rather than scraping nvidia-smi, have you considered using pycuda? It is simple to get free memory as a plain number from the API, though admittedly the scraping...
> In your README.md section, it looks like you need a: cd llama-cpp-python just before the: $Env:CMAKE_ARGS="-DLLAMA_CUBLAS=on"; $Env:FORCE_CMAKE=1; py ./setup.py install else setup.py is not found. The build that was...
@Kaszanas Check the repo's README. https://github.com/maozdemir/privateGPT/tree/gpu