flashinfer.comm.init_tp_moe_finalize_rmsnorm

flashinfer.comm.init_tp_moe_finalize_rmsnorm(group=None, backend: str = 'auto') → bool

Collectively initialize the process-wide TP multicast workspace.

group is the exact tensor-parallel group. Its size must exceed one and evenly divide K_EXPERTS. The default process group is used when it is omitted. backend selects the kernels the group is prepared for: "cute" the CuTe DSL kernels of this module, "cake" the Cake kernels of flashinfer.cake_tp_moe_finalize_rmsnorm_backend, "auto" both (a call then runs the Cake kernels where a Cake program serves its token count on this architecture and world size, the CuTe kernels otherwise). Availability is reduced across the group, so every rank returns the same verdict: whether a prepared backend serves the group.