You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Where the postfix FN stands for "only nan values and no infinite values" and the postfix UZ stands for "no negative zero". 1
Also for traces we use a shorthand for all the types like f32 for float32 however in this case, using the same notation as the one we currently use doesn't give enough information about the type f8 could be any of E4M3 or E5M2 variants.
Thunder should recognize all FP8 variants that exist in PyTorch. As for printing them, I suggest using f8 with the full suffix, for example, f8_e5m2fnuz.
馃殌 Feature
Currently supported dtypes are listed in this conversion dict:
lightning-thunder/thunder/core/dtypes.py
Lines 504 to 535 in 17024fa
PyTorch has recently added support for FP8E4M3 and FP8E5M2 dtypes and we need to add them to Thunder for native FP8 mixed precision training.
The first step is to be able to compile a function that does only dtype conversion.
cc @crcrpar
The text was updated successfully, but these errors were encountered: