sparse MoE plus muon acceleration

hardware portability and vendor lock-in

Model-side optimizations are co-designed with a specific domestic accelerator stack, so portability must be evaluated at the kernel and optimizer level rather than at the API level.

1 source videos