9082 educate users on mat mul precision #9103
Conversation
…on-mat-mul-precision
yaoshiang
requested review from
tengyifei,
lsy323,
ManfeiBai and
zpcore
as code owners
|
PDF of doc is here |
…on-mat-mul-precision
…on-mat-mul-precision
| # | | ||
| # |  | ||
|
|
||
| # | ## Higher precision math on lower precision hardware |
There was a problem hiding this comment.
Perhaps we can just point readers to BFloat16: The secret to high peformance on Cloud TPUs instead of duplicating that content here.
There was a problem hiding this comment.
If you'd rather keep this section, let me know and I can clean it up. My preference would be to use the existing documentation.
There was a problem hiding this comment.
that doc doesn't share the math of how the 3 and 6 pass methods work, and we have had customer questions, so I think it's important we keep this description.
There was a problem hiding this comment.
OK, sounds good. Let me take another pass at it.
There was a problem hiding this comment.
Any ETA on your review Mike? Maybe I could merge it as is and add another one later to address any comments?
This reverts commit 6e3fdf9.
…on-mat-mul-precision
…hub.com/pytorch/xla into 9082-educate-users-on-mat-mul-precision
|
Added logger.warning, added issue on examples I saw of logging.warning instead of logging.get_logger.warning, added unit test to assert the warning message, and consumed the warning message in all other unit tests to avoid polluting the unit test harness stdout. |
Adds runnable tutorial to teach users about mat mul precision.
Also includes new manual build instructions for runnable tutorials.