In BROADCAST mode (used by Mesh-TensorFlow), assume that all cores return...
In BROADCAST mode (used by Mesh-TensorFlow), assume that all cores return identical copies of the full output. Return the output from the first core instead of concatenating the outputs on axis 0. Since we have broken the assumption that the input is batch-split on axis 0, it makes no sense to assume that the output will emerge batch-split on axis 0. In the future, we may want to provide more explicit options about how to combine the outfeeds. PiperOrigin-RevId: 225869483
Loading
Please sign in to comment