Wei Wang
Wei Wang
Yes @atalman : I updated the PR description with this official link: https://developer.nvidia.com/cuda-gpus
Let's target pytorchbot merge by end of day today? Adding ciflow/binaries and ciflow/trunk as this PR seems to be ready for merge. It would be great if @malfet can take...
Yes, you can use pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu124 (we are actively developing and maintaining it, e.g. tomorrow the wheel will feature cudnn v9). As an alternative, below is...
Currently CUDA 12.4 for aarch64 is only available in nightly. p.s. cuda 12.4 support was only added after v2.3.0, so it is not possible to use cuda12.4 with 2.3.0. There...
This must have been done in a different PR. closing.
Please fix lint :) and then let's try merge it.
I realize test-infra windows AMI changes may also be needed: example https://github.com/pytorch/test-infra/pull/1523/files A more recent reference: https://github.com/pytorch/test-infra/commit/23642019138af1bb4e1c429ad476e37b9dc3ff5f
In addition to the #123649 above that Andrey mentioned (to fix the libcudnn_graph.so error), we recommend NOT updating the cudnn version of cu11.8. The build may be successful, though the...
Once https://github.com/pytorch/pytorch/pull/138732 goes in, I believe we can rebase this PR and get the signals needed to merge this PR.
Please note that this is testing assuming cuda toolkit was not installed on user side. e.g. using vanilla ubuntu:24.04 this is to test that the pytorch binary/wheel is self-sufficient without...