Update __truediv__ comment in python bindings. #3586
Conversation
|
!build |
| // __div__. | ||
| // In python, __truediv__ (/) always returns a float regardless of whether | ||
| // the input arguments are float or integer. __truediv__ (/) corresponds with | ||
| // pytorch torch.true_divide(a, b). The __div__ operator is depreciated in |
There was a problem hiding this comment.
| // pytorch torch.true_divide(a, b). The __div__ operator is depreciated in | |
| // pytorch torch.true_divide(a, b). The __div__ operator is deprecated in |
| // In nvfuser, truediv function has the same semantics as python's | ||
| // operator __truediv__ (/). The div function truncates the result instead | ||
| // of promoting it to float, so it has the same semantics as the C++'s (/) | ||
| // operator. In pytorch, torch.div(a, b, rounding_mode='trunc') corresponds | ||
| // C-style integer division. |
There was a problem hiding this comment.
The first sentence here is incorrect I think. On Python 3.12, __truediv__ of ints promotes to double first and does no truncation:
[ins] In [1]: x = 3
[ins] In [2]: y = 4
[ins] In [3]: x / y
Out[3]: 0.75
[ins] In [4]: x.__truediv__(4)
Out[4]: 0.75
To get truncdiv in Python I think you need to manually truncate:
[ins] In [17]: ((-30)/4).__trunc__()
Out[17]: -7
[ins] In [18]: (-30)//4
Out[18]: -8
There was a problem hiding this comment.
How is the first sentence incorrect? It is what the comment says in csrc/ops/arith.h
// truediv: promote to float for integer division, has the same semantics as the
// python's operator /
https://github.com/NVIDIA/Fuser/blob/main/csrc/ops/arith.h#L426-L431
// div: don't promote to float, instead, truncate the result, this has the same
// semantics as the C++'s operator /
https://github.com/NVIDIA/Fuser/blob/main/csrc/ops/arith.h#L432-L437
There was a problem hiding this comment.
I updated the comment to directly mention csrc/ops/arith.h.
We are changing the dunder __truediv__ in our python API to map to c-style division.
| // python's operator __truediv__ (/). The div function in csrc/ops/arith.h | ||
| // truncates the result instead of promoting it to float. It has the same | ||
| // semantics as the C++'s (/) operator. In pytorch, | ||
| // torch.div(a, b, rounding_mode='trunc') corresponds C-style integerw |
There was a problem hiding this comment.
| // torch.div(a, b, rounding_mode='trunc') corresponds C-style integerw | |
| // torch.div(a, b, rounding_mode='trunc') corresponds C-style integer |
|
!build |
This PR updates the comments for
__truediv__operator defined in python bindings. The current comment does not reflect what the code actually does.Reference: #2837 (review)