Module tp

Module tp 

Source
Expand description

2026-09-25: The Kimi K3 tensor-parallel plan (metrale_model_weights::kimi_k3_host::tp::tensor_plan) in TpShardKind terms.

Head counts on config are per rank; the returned sizes are full, local heads times tp_world_size.

Owner: model-arch, Kimi K3. Invariants: none beyond the types.

Functionsยง

supports_tp
2026-09-25: Always true: the Kimi K3 loader binds tensor-parallel shards.
tensor_plan
2026-09-25: (kind, full_out, full_in) for one checkpoint key. A key the plan does not name is Replicated.