-
-
Notifications
You must be signed in to change notification settings - Fork 15.3k
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
python.pkgs.tensorflow: 1.14.0 -> 1.15.0 #71282
Conversation
I have decided to keep the default to no-cuda for now, unless there are any objections. |
This should also fix the tensorflow build, which was broken by #69252. |
Unfortunately the build still doesn't quite work with bazel 1.0. Upstream is tracking this here: tensorflow/tensorflow#33240 |
913b48f
to
2fdfb1a
Compare
Upstream is working on a fix in tensorflow/tensorflow#33415. |
Did a quick pass at applying this patch. Unfortunately, I'm unable to build this with upstream's fix applied. Will continue looking into it some later day. Adding to here FYI: on that nixpkgs fork the build fails with:
|
Ignore my prior comment. Used an out-of-date version of this pull request. Re-testing. |
Re-tested using master + this patch, with adjustments to the deps SHA. On our build farm, both cuda and non-cuda fail to compile with similar link errors:
Unfortunate that the tensorflow build is a monolith. :( Nothing for nixpkgs to resolve afaict. Kinda surprising tho. Pain to test/build. |
@coreyoconnor This seems to stem from Relevant upstream commit: tensorflow/tensorflow@f92a0e6 Edit... or maybe not. I have a hard time telling which code is from which project here... Actually, it seems to be that they actually upgraded protobuf to 3.8.0 in v1.15.0, but that commit above is not present in 1.15.0 |
2fdfb1a
to
2583a43
Compare
Thanks to @scentini in tensorflow/tensorflow#33415, it builds now! Things left to do:
If anybody else wants to do some testing, now is the time. |
7352080
to
62108eb
Compare
… 1.14.0 -> 1.15.0
The tensorflow build was broken by the bazel 1.0 update in 73eb01b.
62108eb
to
4fb7831
Compare
this issue is resolved with timokau changes. Unfortunately, I missed the latest patch and ran into the gast error. XD Testing again. |
Yeah, sorry about that. I made some mistake during fixup before push. Should work now. |
@timokau no problem! Subsequent build failed on this
I'm building using cudatoolkit 10. Which is not the default in all-packages. I wonder if this workaround is specific to cudatoolkit 9? |
Success! Compiled without issue with cudatoolkit_10 (once I removed that workaround). So I presume yes: that is specific to cudatoolkit_9. |
Thanks for testing! I don't quite understand what you're saying though. So the workaround is not necessary for a newer cudatoolkit? That's good, but why does it fail? If the workaround is not necessary, it still shouldn't do any harm right? |
Now needed at runtime with the python2 build.
0c31f7c
to
5162862
Compare
5162862
to
8e382a7
Compare
Merging now, since this has been broken on master for so long. Any fixups can be made in followup PRs. For example, python2 is still broken. It's probably not too difficult to fix, but I don't plan to invest more time in this for now. |
Correct: the work around is not needed for newer cuda. The workaround
actually instructs nix to copy that specific path to the store. Which does
not exist on newer cuda toolkit. Thus it will fail due to not finding the
path.
(On mobile)
…On Wed, Nov 13, 2019, 12:41 AM Timo Kaufmann ***@***.***> wrote:
Thanks for testing! I don't quite understand what you're saying though. So
the workaround is not necessary for a newer cudatoolkit? That's good, but
why does it fail? If the workaround is not necessary, it still shouldn't do
any harm right?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#71282?email_source=notifications&email_token=AAAIMDPASHOEV2BJNL4YQJ3QTO4TDA5CNFSM4JBYKW7KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOED5KQTQ#issuecomment-553297998>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAAIMDICQVDEYWLWQ2FTU3LQTO4TDANCNFSM4JBYKW7A>
.
|
Oh: the default cudtoolkit is 9. Which this does compile fine for. Fixing
for cuda 10 can wait :) definitely not a merge blocker
…On Wed, Nov 13, 2019, 8:11 AM Timo Kaufmann ***@***.***> wrote:
Merging now, since this has been broken on master for so long. Any fixups
can be made in followup PRs. For example, python2 is still broken. It's
probably not too difficult to fix, but I don't plan to invest more time in
this for now.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#71282?email_source=notifications&email_token=AAAIMDN2JQ4VBWBDTZL7JJLQTQRKTA5CNFSM4JBYKW7KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOED6VF6Y#issuecomment-553472763>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAAIMDOM6T7JHJPYZFZ5BBDQTQRKTANCNFSM4JBYKW7A>
.
|
Oh, so the path doesn't exist. Its weird that the workaround is necessary at all. The best solution would probably be to fix the cudatoolkit 9 build to make the workaround unnecessary for both versions. Do feel free to create a PR to fix this either in the tensorflow build or in the cudatoolkit build. If you do, ping me and I'll review (to the best of my ability, since I don't have a nvida card) :) |
will do! Thanks again for figuring this out :) |
Motivation for this change
Continuation of #70910.
fixes #71578.
Things done
sandbox
innix.conf
on non-NixOS)nix-shell -p nix-review --run "nix-review wip"
./result/bin/
)nix path-info -S
before and after)