I'm testing this locally, but it's still building.

I could copy the tensorflow bazelrc from the remote repo, and allow us to use our own for whatever we need (and prevent this kind of out-of-sync issue from happening again). It would make it much harder to edit though (you'd need to use a patch). Thoughts?

This is working on my machine, so we should be good to merge. tensorflow.bazelrc is now the bazelrc from the tensorflow repo, and .bazelrc can be whatever we want (I have readonly build cache in there for now, setting it to write in the CI will take precedence).

cc @karllessard @Craigacp

Would be good to figure out where the QuantizedConcatV2 on my machine is coming from first though.

The CI timing out since it can't reuse the cache.

The CI timing out since it can't reuse the cache.

Yes, so let's merge it, anyway we know that the current build is not working properly on GPU devices, thanks @rnett

It's timing out on master too, I think you have to manually re-run it a couple times to build the cache.

Finally got the snapshots to build, I think it took 4 passes this time... you can try it out now