DeepSpeed

[muon] Per-head Muon for linear-attention layers: read head geometry from the owning module

#8436

Merged