Update cudnn semantic version in README to avoid performance issue #1239
Conversation
I suggest changing the version spec of `nvidia-cublas-cu12` and `nvidia-cudnn-cu12`. Using existing one: ```bash pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.* ``` on 2025-02-07 gives ``` nvidia-cublas-cu12==12.8.3.14 nvidia-cudnn-cu12==9.7.1.26 ``` which results in cuda running 30x slower than cpu on my Ubuntu 22.04, with a 6 seconds wav audio. After I downgrade them to: ``` nvidia-cublas-cu12==12.4.5.8 nvidia-cudnn-cu12==9.1.0.70 ``` which is what `torch==2.6.0` currently depends on, everything works fine. Then, I go through a thorough grid search over recent cublas and cudnn versions, and find that all cudnn after 9.6 (inclusive) exhibit the aforementioned performance issue. I didn't study the lower bound, though. Hence, I make it compatible with current semantic version, i.e. `>=9`). That's why I make this change.
|
Hello, we just suggest the major version for reference, since this project does not depend on the PYPI cudnn installation we cannot pin the version to anything except the major version |
|
Hello. Thank you so much for your reply! IMHO it would still be better to make
explicit in README, so that users will know they may need to adjust the cublas and/or cudnn version accordingly, depending on their own dev environments. It could be somewhat confusing to simply state |
|
unfortunately, tuning the cudnn version is not an option for most users, especially that it's closely tied with with other packages that might exist in the same environment such as pytorch which is only compatible with a single cudnn version. I know CUDA dependencies are a headache to manage that's why they are out of the scope of the requirements even though they are required, simply because there's no one solution fits all here so we leave it up to the users to decide |
|
I got your point. Thank you so much! |
I suggest changing the version spec of
nvidia-cublas-cu12andnvidia-cudnn-cu12. Using existing one:pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*on 2025-02-07 gives
which results in cuda running 30x slower than cpu on my Ubuntu 22.04, with a 6 seconds wav audio.
After I downgrade them to:
which is what
torch==2.6.0currently depends on, everything works fine. Then, I go through a thorough grid search over recent cublas and cudnn versions, and find that all cudnn after 9.6 (inclusive) exhibit the aforementioned performance issue.I didn't study the lower bound, though. Hence, I make it compatible with current semantic version, i.e.
>=9).That's why I make this change.
Thank you so much!