Muon creator Keller Jordan requested logs to validate the change
@kellerjordan0 The right thing is to give a heads up to the library owner to see if there is any alternative options to use. So I filed a bug here: https://github.com/facebookresearch/optimizers/issues/265
@_arohan_ Any update on this?
@_arohan_ Are you planning on releasing the reproducible logfile that was automatically generated by your run?
@kellerjordan0 The right thing is to give a heads up to the library owner to see if there is any alternative options to use. So I filed a bug here: https://github.com/facebookresearch/optimizers/issues/265
@kellerjordan0 Yes I am planning to. Why would have a different impression.
Looks like the library already supports a PseudoInverse when matrices are rank deficient. Let me see if it reproduces, so we should get a clean log.
@_arohan_ Are you planning on releasing the reproducible logfile that was automatically generated by your run?
Muon creator Keller Jordan requested logs to validate the change
@kellerjordan0 The right thing is to give a heads up to the library owner to see if there is any alternative options to use. So I filed a bug here: https://github.com/facebookresearch/optimizers/issues/265
@_arohan_ Any update on this?