[XLA] Support strength reducing bfloat16 dot products
There is no reason to limit ourselves to float32 dot product operations, we simply convert to and from float32 around the reduction to simulate the precision change. PiperOrigin-RevId: 213014410
Loading
Please sign in to comment