Module tp 

K3 device-binding adapter for the shared tensor-parallel storage plan. Runtime loading partitions packed and dense weights before GPU allocation.

Head counts on config are already per-rank. Full sizes = local * tp.

Functions

supports_tp

K3 does tensor-parallel. Callers must not fail-fast supports_tp false.

tensor_plan

(kind, full_out, full_in) for one checkpoint key.