Daniel Galvez comments

Results 85 comments of


                                            Daniel Galvez

Tensor.nonzero fails on GPU for tensors containing more than INT_MAX elements

If the reason for removing third_party/cub was that it was considered stable and unchanging, then this could be a good reasons to bring back third_party/cub as third_party/cccl then.

Draft: Fix the "cast ping pong" problem when we run AMP inference.

#9198 supersedes this.

Truncating keys to length 100 sometimes causes duplication

Solution is to set longFileMode= LONGFILE_POSIX as described here: http://commons.apache.org/proper/commons-compress/tar.html

Implement text normalization

https://github.com/rhasspy/gruut this may be a good alternative. Multiple languages no C++ dependency to install.

Output dataset in WebDataset format

Okay, it looks like Apache Commons Compress library supports tar file format, so I would prefer to use that, going the java route. This actually seems pretty reasonable to do....

Things to do before neurips

Poster + 3 minute talk due: Oct 18th Camera-ready paper due: November 6th Neurips (dataset release): Early December

Use model-cast-to-bfloat16 rather than AMP-to-bfloat16 for inference.

FYI, the PR description has been updated to describe the benefit from using this mode of execution instead of AMP

Use model-cast-to-bfloat16 rather than AMP-to-bfloat16 for inference.

This is ready for another round of review. Unfortunately the black + isort change made a lot of noise. But basically I removed a lot of cruft. This PR basically...

Use model-cast-to-bfloat16 rather than AMP-to-bfloat16 for inference.

@nithinraok @pzelasko @titu1994 I think this change is good to go at this point.

[cond] support torch built in function as subgraph

thank you!