why sm_61 was dropped from the nightly build?

Just switched to cu111 nightly build after realizing cu11.0 build wasn’t getting updated anymore just to discover that

GeForce GTX 1070 Ti with CUDA capability sm_61 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_70 sm_80 sm_86.
If you want to use the GeForce GTX 1070 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

What can be done to restore the support for sm_61 in nightly? It makes no sense that you’d drop support for not very old hardware and require users to build from scratch which is far from easy and takes an insane amount of resources and most of the time fails. I know it costs a bit more to build support for another arch, but it is still much cheaper and simpler when the user doesn’t have to do it when you multiply this by the number of users (except I suppose there aren’t many nightly users, so I’m probably a minority).

I use the nightly build because I am working on integrating new pytorch features, like pipeline, and I didn’t know that having an older card is suddenly going to be a problem. 🙁

Also the part:

If you want to use the GeForce GTX 1070 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

was not thought through, as there are no instructions on how to use GeForce GTX 1070 Ti GPU with PyTorch on that page. It probably need to say that:

You have to build from source to use older cards and you will find instructions on how to 
build from source at https://pytorch.org/get-started/locally/

Thanks.

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser @anjali411 @seemethere @malfet @walterddr @ngimel @aocsa

1 possible answer(s) on “why sm_61 was dropped from the nightly build?