-
Notifications
You must be signed in to change notification settings - Fork 699
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
fix build of TensorFlow 2.5+ on aarch64 #17101
Conversation
The patch is actually wrong and may yield wrong results. Use the upstream patch from TF 2.10
Use updated patch from easybuilders#17101
Use updated patch from easybuilders#17101
Use updated patch from easybuilders#17101
Woops, this is blocking CI for the other TF PRs as the ECs updated here are not updated in the other PRs but all ECs get parsed and checksum-checked for some reason? |
I'll look into that, but it'll take a while (lots of missing dependencies there currently)... |
What do you mean by "blocking CI"? Which other PRs are you referring to? |
Test report by @boegel |
@Flamefire Not sure how useful that test report is, since it's all failures (but at least all dependencies are in place now...) |
One hit an old bug in TensorFlow which is fixed in newer versions for the others the failure is not contained in the log. I can see if I can backport the patch for that old failure. |
Does this help (full log for |
Some have build failures:
The first is similar to one I see in 2.11 on PPC but hard to tell what's wrong. The other 2 build errors are not in the log. And in the one you attached it seemed to succeed and then fail in the test validation -.- The others are failing tests which look like real bugs. Maybe try to build the ones that build without this PR and if they also fail I'd say this PR is good. From all I can tell the change is correct: https://developer.arm.com/architectures/instruction-sets/intrinsics/#q=vdotq_lane_s32
--> 3rd argument must be an s8 Edit: I found a patch for the AARCH64 build: tensorflow/tensorflow@4933ada |
4a3b335
to
4b1f3a7
Compare
Test report by @boegel |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Test report by @boegel |
@boegel I'd like to get this in to fix the CI failures on other TF PRs using the new checksum and/or patch version. It would be good to know if the current ECs/patch also fail on aarch64. If the failures look similar then the issue is not introduced here. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lgtm
Although problems remains on Thanks a lot @Flamefire! |
Going in, thanks @Flamefire! |
Use updated patch from easybuilders#17101
The patch is actually wrong and may yield wrong results. Use the upstream patch from TF 2.10
I misunderstood the comment when creating the original patch which I noticed when checking the patches for TF 2.11 where the issue was fixed in the TensorFlow repo.
@boegel Can you test this on your ARM machine and include this for the next release please?