Open mo-tenstorrent opened 2 weeks ago
This can affect device data aggregation. Only doing per device might not be enough #8900
Bumped down to P2 because with #14961 fixed, the breaking issue will be gone. We need further testing to make sure there are not other bugs.
With sub-device mesh, the rule that all worker cores are part of a single op does not apply any more.
This will break post processing of the runs with sub core mesh as it asserts if it sees workers cores of the same device are part of two different ops during the op intervals.
With opID now shared with the device, we can loosen that restriction at least on sub-device mesh runs