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
Hi, I'm a beginner and trying to understand how device topology is determined when the in_strategy tuple does not match the actual number of available devices.
For example, if I have 8 devices and use the following sharding configuration:
ops.MatMul().shard(((2, 2), (2, 1)))
What would the resulting device topology mapping look like?
Would mat_1 be mapped to something like device[-1][2][2], and mat_2 to device[-1][2]?
Also, will automatic broadcast happen in this case?