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
I have a program with 2 separate process each using different slices of a TPU. I would like to create each slice such that the layout is as good as it can be for both slices.
Lets say on a v3-8 I want 1 slice of 6 and one of 2. Should it be done like this:
I assume option 2 is better because create_device_mesh should create a layout where items close in the list are physically close together, but at the same time the docs for create device mesh say it returns a mesh shape "with good collective performance" so option 1 might be better for the bigger slice, but the smaller slice may be stuck with 2 core which have bad collective performance? Or does it just not make too much difference?
If it helps this question is specifically for Sebulba where a separate slice of the TPU is used for the actor and the learner.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
Hi 👋
I have a program with 2 separate process each using different slices of a TPU. I would like to create each slice such that the layout is as good as it can be for both slices.
Lets say on a
v3-8
I want 1 slice of 6 and one of 2. Should it be done like this:Or should I create it like this:
I assume option 2 is better because
create_device_mesh
should create a layout where items close in the list are physically close together, but at the same time the docs for create device mesh say it returns a mesh shape "with good collective performance" so option 1 might be better for the bigger slice, but the smaller slice may be stuck with 2 core which have bad collective performance? Or does it just not make too much difference?If it helps this question is specifically for Sebulba where a separate slice of the TPU is used for the actor and the learner.
Beta Was this translation helpful? Give feedback.
All reactions