add sort to get_compute_capabilities #527
Conversation
|
@TimDettmers Please help review this Pull Request. Thank you. |
|
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread. |
|
Thank you for catching this! I think a question is if we should select the highest or lowest compute capability. The lowest would give the most broad support for multiple GPU. However, it is common that one uses an old GPU for monitors and a new GPU for deep learning work. As such, one would want to select the highest compute capability. @Titus-von-Koeller @younesbelkada any thoughts? |
|
I will merge #703 instead as it is more comprehensive, but lets chat a bit about which how we should sort the compute capability |
|
closed this PR as it's included in another one |
I agree, highest is a good choice. We'll see if this is contrary to people's needs or expectations. If that is the case, we can add some logic to make it configurable. Thanks a lot @Ronnoc for your contribution! |
The "sort compute capabilities" fix from bitsandbytes-foundation#703 (bitsandbytes-foundation#527) would actually do nothing due to this.
* Diagnostics: streamline debug printing code * CUDA setup: Remove unused `backup_paths` * CUDA setup: DRY OS detection * CUDA setup: Streamline `manual_override()` * CUDA setup: Use comment instead of string literal, simplify * CUDA setup: remove duplicate sort The "sort compute capabilities" fix from #703 (#527) would actually do nothing due to this. * CUDA setup: make version number replacement logic more obvious
In
get_compute_capability, the last item of ccs (ccs[-1]) will be selected as the maximum compute capability.But
ccsare not sorted inget_compute_capabilitiesThis PR adds sorting to
get_compute_capabilities