Reddit user questions OKLS step efficiency versus Muon
A Reddit post asks why OKLS, described as a zero-staleness implementation of KL-Shampoo, may be outperformed by Muon in step efficiency. The submitter claims that SOAP is more step-efficient than Shampoo, KL-Shampoo is more step-efficient than SOAP, and OKLS only surpasses Muon in parameter efficiency during very long training runs. The supplied post does not include benchmark measurements, implementation details, or responses that resolve these comparisons.
SOAP is more step efficeint than Shampoo, and KL-Shampoo is more step efficient than SOAP.
Then how is OKLS getting outperformed by Muon in terms of step efficiency where it's a zero-staleness implementation of KL-Shampoo?
It only gets ahead of Muon in terms of parameter efficiency in very long training runs.