bhack
bhack
But in any case, if you think that the CI image is pretty clean to be distributed it is the same. I was talking about the nightly docker images cause...
If the CI and dev envs are the same yes. Is the cache populated from the main only on every merge commit or also from PR builds? Cause @huydhn just...
Ok so you are differentiating by the workflow key prefix. So what was the posioning risk you was talking about?
I am compiling a recent commit from source and I could confirm that it is a computational resources pain. Also I think that if with this we could have really...
We had some not ready dependency last time I have tried and also some usability issues (see that ticket) I've also upstream fixed https://github.com/huggingface/transformers/issues/28999
See also how many misalignment occasion we could have by design https://github.com/pytorch/pytorch/issues/126005 We need to care about fragmentations if we want to really maintain a valid cache for (community) users/developers.
Is this still on the Cold Storage bucket?
12.4.1 upstream CUDA images reintroduced cudnn layers https://github.com/pytorch/pytorch/issues/126005#issuecomment-2125777085
There are still no 12.5 upstream images: https://gitlab.com/nvidia/container-images/cuda/-/tree/master/dist?ref_type=heads
`12.5` upstream images are available: https://hub.docker.com/r/nvidia/cuda/tags But what do you want to track in this ticket? Cause if it is only about cudnn 9.1.1 check @Skylion007 comment at https://github.com/pytorch/pytorch/pull/123475#issuecomment-2142175029