DOC Add note in Array API doc regarding support for devices without float 64 support #28034
Conversation
glemaitre
changed the title
glemaitre
left a comment
glemaitre
left a comment
There was a problem hiding this comment.
LGTM apart from those nitpicks.
| a transfer to CPU. | ||
|
|
||
| Minimizing the usage of float64 upcasting in scikit-learn is an open improvement | ||
| direction, to maybe yield better performance from devices that do not support it, |
There was a problem hiding this comment.
| direction, to maybe yield better performance from devices that do not support it, | |
| direction, to yield better performance from devices that do not support it, |
There was a problem hiding this comment.
Oh actually, I wanted to remove the "maybe"
There was a problem hiding this comment.
I can't reconstruct what I meant to say. Reading my comment now I am confused.com about what I was trying to say (it seems like a comment written while distracted) :-/
Let's ignore it?
Co-authored-by: Guillaume Lemaitre <guillaume@probabl.ai>
|
I've applied all the suggestions. I think they are uncontroversial typo fixes. The one I didn't apply is because I don't think it makes sense. @glemaitre maybe you can take another look and then merge? |
|
I'm still +1 until that this actually the policy that we use. It seems that it was questioned in this PR: #27904 (comment) I would delay a merge until we settle on the matter. |
|
A section about float64 was added to |
4 participants
What does this implement/fix? Explain your changes.
Adds a note in the array api documentation that documents scikit-learn policy regarding support of devices that do not support float64 precision operations. (basically stating that it favors consistency with CPU behavior at the cost of data transfers to CPU, over remaining on the device at the cost of capping compute to highest supported precision)
Discuted before with @betatim and @ogrisel , in particular during review of #27904