tree fa55da321cd1b15ff60e2a2758626d207037a5e0
parent 04c243a5e47d536f7680614bb73408e51f43b093
author Christian Sigg <csigg@google.com> 1627460975 -0700
committer TensorFlower Gardener <gardener@tensorflow.org> 1627461257 -0700

Don't mark XLA GPU targets as compatible_with cloud.

XLA GPU depends on CUDA libraries which are not compatible, and without them the targets don't do anything useful (even if they build fine). This also avoids having to mark TFRT GPU targets (that XLA GPU wants to depend on) as compatible_with cloud.

Explicitly mark a few targets as compatible_with cloud, because the XLA CPU backend depends on them.

PiperOrigin-RevId: 387295973
Change-Id: Ide59fd4fae831a45029a712e4a500a3843acab0f