I am using CUDA 12.5 now (for another software). Plan to downgrade the CUDA to 12.4 (for that software and PyTorch). Like to know what is the approximate timeframe for CUDA 12.4 to be compatible with Pytorch 2.3? Weeks, months?
Thanks.
P/S I guess CUDA 12.5 will be much later, a year later?
Your locally installed CUDA toolkit won’t be used if you install the PyTorch binaries since they ship with their own CUDA runtime dependencies.
If you want to use a specific CUDA version you can build PyTorch from source using it, as it’s compatible with the latest CUDA releases.
Honestly, I do not fully understand it due to my limited knowledge. Let me recap what I did.
Upgraded from Davinci Resolve Studio18 to DRS 19 a few months ago. Updated the CUDA to the latest one 12.5 a few weeks ago.
Uninstalled and reinstalled Pycharm, Python and Pytorch a few weeks ago. Why? Got many errors (think due to my own making, not knowing what I was configuring). Before the reinstallation, I got Pytorch to access my GPU Cuda with the correct Pytorch and Cuda versions.
Now no access between Pytorch 2.3 and Cuda 12.5. I think Pytorch 2.3 currently does not support Cuda 12.5. Support for Cuda 12.4 in Pytorch 2.3 is coming.
Please help withe the following:
“Your locally installed CUDA toolkit won’t be used if you install the PyTorch binaries since they ship with their own CUDA runtime dependencies.”
*** Meaning Pytorch and DRS 19 will use their respective installed Cuda versions? Use PIP to install specific Cuda version for Pytorch?
“If you want to use a specific CUDA version you can build PyTorch from source using it, as it’s compatible with the latest CUDA releases.”
*** What is meant by “build Pytorch from source”? Saw that a few times on the web.
*** “as it’s compatible with the latest CUDA releases.” Possible to install any CUDA version for Pytorch 2.3?
PyTorch 2.3 does support CUDA 12.5 if you build from source. The binaries (pip wheels and condo binaries) whip with CUDA 11.8 or 12.1.
The current nightly binaries ship additionally with CUDA 12.4.
However, as already mentioned your locally installed CUDA toolkit won’t be used and the shipped CUDA dependencies will be used.
I don’t know what DRS does, but PyTorch binaries install all needed CUDA runtime dependencies during the install process and will use these.