abetlen / abetlen/llama-cpp-python

Which Python + CUDA version do you recommend for LLaVA on Windows 10 Pro 64-bit?

Open
#1,418 1 comment 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
10.6k
Forks
1.4k
PR merge metrics
PR metrics pending

Description

Hey, I would appreciate recommendation on the specific versions of Python + CUDA environment on Windows that have been tested by you guys and proven to work well with LLaVAv1.6 and LLaVAv1.6 with CUDA GPU support in the latest releases. Can you provide me with some suggestions?

I have noticed that the LLaVA projector works fine for me when using "cuBLAS", but encounters issues when using "CUDA".
I tried to run it on a Python 3.11.4 64-bit + CUDA 12.1 and then with the latest CUDA 12.4.1 environments with llama-cpp-python installed with the adequate wheels, and without wheels through CMAKE_ARGS = "-DLLAMA_CUDA=on" , but couldn't get either LLaVAv1.5 or LLaVAv1.6 projectors to work correctly on release versions above 0.2.64.

There has been a change I can't identify between the release of 0.2.64 and 0.2.65, most likely related to CUDA/cuBLAS or chat_handler, that breaks how projectors work for me.

I fought with it the whole yesterday and eventually just gave up and reverted back to 0.2.64. The project I'm working on is currently running Bakllava-1-7b or LLaVAv1.5 without issues in a Python 3.10.11 64-bit + CUDA 11.7 environment using "cuBLAS on" with release version 0.2.64. I would like to upgrade to LLaVAv.1.6, at some point in the future.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.