Create Highlevel Bindings for FP8 Datatype #788
Conversation
|
Automatic reviews are disabled for this repository. |
| from numba.cuda.extending import register_jitable | ||
|
|
||
|
|
||
| @register_jitable |
There was a problem hiding this comment.
The use of @register_jittable is recommended by code agents. Is this a good choice in today's Numba?
There was a problem hiding this comment.
It only makes sense if you want to call the function as a pure Python function.
There was a problem hiding this comment.
(So register_jitable just makes no sense in the context of Numba-CUDA)
There was a problem hiding this comment.
This PR shows that register_jittable may enable jitting function that's not directly callable in pure python. As shown by
@register_jitable
def bfloat16_to_e8m0(x, saturate, rounding):
return _cvt_bfloat16raw_to_e8m0(
_bfloat16_as_bfloat16_raw(x), saturate, rounding
)Where _bfloat16_as_bfloat16_raw is written as numba intrinsics and may not be called with these arguments as-is.
Proposing updating the docstring of register_jittable and keeping the function.
|
/ok to test 894b79d |
2 participants
This PR adds high level exposure of fp8 data type to Numba-CUDA.
Supported features include:
Supported FP8 variants (element wise):
Supported packed FP8 variants:
This PR also adds tests for packed type bindings introduced in #686.
closes #200