Pytorch 1.12.1 - and Disable kineto due to cupti conflict

[conda-forge-linter]

Hi! This is the friendly automated conda-forge-webservice.

I've rerendered the recipe as instructed in #135.

Here's a checklist to do before merging.

• Bump the build number if needed.

[conda-forge-linter]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe) and found it was in an excellent condition.

[hmaarrfk]

This isn't having the desired effect, the 10.2 build shows:

2022-08-01T03:53:43.6127771Z    INFO (pytorch,lib/python3.10/site-packages/torch/lib/libtorch_cpu.so): Needed DSO lib/libcupti.so.10.2 found in conda-forge::cudatoolkit-10.2.89-h713d32c_10

[ngam]

Did you forget to set it to "OFF" here?

[hmaarrfk]

thanks

[ngam]

Suggestion: limit the builds to cuda112 and only use one cuda arch (e.g. 8.0) for testing here so that the ci finishes within six hours

[hmaarrfk]

Suggestion: limit the builds to cuda112 and only use one cuda arch (e.g. 8.0) for testing here so that the ci finishes within six hours

Seems like that would complicate the process even more. I would rather do one thing at once. address the cupti bugs instead of revamping the build system at the same time.

[ngam]

Suggestion: limit the builds to cuda112 and only use one cuda arch (e.g. 8.0) for testing here so that the ci finishes within six hours

Seems like that would complicate the process even more. I would rather do one thing at once. address the cupti bugs instead of revamping the build system at the same time.

Fine with me. My suggestions was merely to get this to finish within six hours for diagnostics only within this PR, not a permanent change. So once we see if disabling kineto actually does the trick, we can go back to build all the cuda arches as before. This would remove the need to build this locally (unless you got the info below from the CI?)

2022-08-01T03:53:43.6127771Z    INFO (pytorch,lib/python3.10/site-packages/torch/lib/libtorch_cpu.so): Needed DSO lib/libcupti.so.10.2 found in conda-forge::cudatoolkit-10.2.89-h713d32c_10

[hmaarrfk]

The 10.2 builds finish within the allotted time (sometimes) so you can check those. That is what I did in what you quoted.

[ngam]

10.2 builds finish within the allotted time (sometimes)

Ah, didn't know that. Okay then, we can check that in the meanwhile!

[hmaarrfk]

Well kineto is properly disabled, and cupti is nowhere to be found in the logs for now. I think we are good. I'm going to build cuda locally.

[ngam]

Let's try to change this into 1.12.1?

[hmaarrfk]

I guess i had the builds ready, but I don't see a sense in having too much uploaded.

[ngam]

I guess i had the builds ready, but I don't see a sense in having too much uploaded.

If you do, just upload them! Sorry I only wrote that thinking you haven't built them already!

Upload them and we can deal with the 1.12.1 stuff later

[ngam]

Also, I am not sure if this is their final 1.12.1 tag... pytorch/pytorch@v1.12.0...v1.12.1

They didn't announce it under releases, but it has already appeared on pypi, etc.

[hmaarrfk]

I'm pretty swamped right now with the day job.i think i can get this in by next week

[hmaarrfk]

Also.

pytorch/pytorch#81680 (comment)

[ngam]

Take your time :)

[hmaarrfk]

GPU build logs
log_files.zip

[hmaarrfk]

GPU builds are being uploaded now.