CUDA version and updated installation instructions

[sidharthrajaram]

Edits to the README regarding installation/CUDA based on discussion on these issues: #783 and #717

• Updated libraries to install for CUDA 12 (because of latest ctranslate2 support for only CUDA 12)
• Added note regarding CUDA 11 support.

[jimydavis]

pip install nvidia-cudnn-cu12 needs to be pinned to either

• ~=8.9
• ^=8.9

because in v9, libcudnn_ops_infer.so.8 seems to be replaced with libcudnn_ops.so.9 amidst other changes. I am not familiar enough with nvidia's practice on whether it should be a tilde or a caret.

For cublas, at least as of commit 91c8307, I have tested that nvidia-cublas-cu12==12.4.5.8 is fully working.

[Purfview]

Btw, @nguyendc-systran mentioned that they will keep the support for CUDA 11 for a while, maybe they dropped that idea:

OpenNMT/CTranslate2#1590 (comment)

[minhthuc2502]

We tried to support CUDA 12 and 11 but releasing 2 versions in parallel quite complicated to maintain. In the end, we decided to only support CUDA 12 but the Ctranslate2 source can always build with CUDA 11.

[Purfview]

Looks good.

[bil-ash]

May be alongside update ctranslate2 dependency to latest 4.2.0 because it supports flash attention as well as performance improvements for quantized models on CPU.

[sidharthrajaram]

May be alongside update ctranslate2 dependency to latest 4.2.0 because it supports flash attention as well as performance improvements for quantized models on CPU.

@bil-ash this PR mainly contains updates to the installation instructions due to lack of CUDA 11 support in the latest versions of ctranslate2. Upgrading the ctranslate2 dependency would be beyond the scope of this particular PR I think.

[sidharthrajaram]

Is this good to go, @Purfview ?

[Purfview]

@sidharthrajaram FYI, I'm not a maintainer of this repo.