Theodor Badea

preserve collectives

8

DRAFT PR: Discussion initiation, mostly addressing https://github.com/mlcommons/chakra/issues/161 Issue: some collectives are missing from the .et file, but are present in the kineto file. It all starts from the host trace,...

add collective dependencies as data_deps

5

Currently, there is no explicit dependency between collectives belonging to the same process group. Their sequential execution is deducted only from the duration of their parent compute node and their...

wrong linkage between GPU ops and their CPU parents

6

Introduced in: https://github.com/mlcommons/chakra/pull/185 Previously, linkage was done based on the Ev idx attribute of the nodes and on the timestamps. This PR introduced linkage based on External id attribute, but...