feat(opti): muon optimizer

[beniz]

No description provided.

[JiwaniZakir]

The Muon-specific flags (--train_muon_adjust_lr_fn, --train_muon_nesterov, --train_muon_ns_steps) have no apparent validation that they're only meaningful when --train_optim muon is selected — users who accidentally set --train_muon_ns_steps 10 while using Adam will get silent no-ops with no warning. A guard in the option parsing or register_optimizer that raises or warns when Muon-specific options are set alongside a non-Muon optimizer would prevent this confusion.

The switch from self.netG_A.parameters() to self.get_named_parameters(("G_A", self.netG_A)) in b2b_model.py is the right approach for Muon, since it needs to distinguish 2D parameters (eligible for Newton-Schulz orthogonalization) from 1D parameters (biases, norms) that fall back to AdamW — but the diff is truncated so it's not possible to verify whether build_named_parameters in util/optimizer_factory.py actually implements this split correctly. If 1D parameters like LayerNorm weights are accidentally routed through the Nesterov/orthogonalization path, training will silently diverge.

The default --train_muon_ns_steps 5 is reasonable, but --train_muon_nesterov True as a flag default that is True out of the box is unusual for a flag-type option — in most argparse setups a flag implies store_true with a default of False, so documenting it as defaulting to True in both options.md and options.rst may be misleading or incorrect depending on how the option is actually registered.