xla
9668dd45 - Made literal to tensor conversion use CopyTensors, which is case or relayout is considerably faster.

Commit

7 years ago

Made literal to tensor conversion use CopyTensors, which is case or relayout is considerably faster. With BF16 is even faster as we do not go through the BF16 -> F32 conversion pass.

References

#138 - Finishing up handling the 3way nature that XLA tensors can have (XLA data, IR Node, or ATEN tensor)

Author

dlibenzi

Committer

dlibenzi

Parents

70ef4e84

xla 9668dd45 - Made literal to tensor conversion use CopyTensors, which is case or relayout is considerably faster.

xla
9668dd45 - Made literal to tensor conversion use CopyTensors, which is case or relayout is considerably faster.