The pull request says a graph kernel node retains the same %gridid on each replay, allowing a barrier state from an earlier execution to satisfy a later wait.
Under that condition, a Mix or Norm waiter can proceed before its m-block initialization CTA has started and read split partials from the previous replay. The change tags executions with an epoch rather than %gridid. It stores the epoch and an exit count in unused space in the first barrier line, so the host-side barrier buffer and compiled extension do not change; the pull request describes the change as requiring JIT headers only.
The reported validation ran on an NVIDIA B200, replaying one graph node while alternating two inputs and comparing every replay bit-for-bit with eager results. For a 33-token case with SM contention, incorrect outputs fell from 185 of 200 runs to zero of 400; listed 512-token and 8,192-token contended cases also reached zero of 400 after the fix. The source does not establish performance impact, wider hardware coverage, or a packaged DeepGEMM release containing the change.