Module build 

Binding one GLM MLP site for a rank: TP slicing of the dense/shared halves, EP selection of the routed experts.

Takes load closures rather than a WeightStore, for the same reason crate::layers::glm5next_dsa::build does: the slicing is then testable without a checkpoint, and the loader wiring stays one call site.

🔴 The two axes are different, and mixing them is silent

• TP splits the width of the dense FFN and the shared expert. gate_proj/up_proj are [inter, hidden] and split by ROW; down_proj is [hidden, inter] and splits by COLUMN. Slicing down_proj by row instead gives a well-formed [hidden/tp, inter] tensor and a plausible, wrong output.
• EP splits the set of routed experts. An expert is never cut — it is owned whole. The router stays replicated so every rank selects the same ids.

Functions

build_dense_mlp

TP-slice and upload one BF16 SwiGLU MLP — a dense layer, or a routed layer’s shared expert.

build_moe

Bind one routed MoE site for this rank: replicated router, TP-sharded shared expert, and exactly the local_experts routed experts this EP rank owns.

Type Aliases

ExpertFn

One routed expert, by GLOBAL id.

LoadFn

A BF16/F32 tensor as host f32, by layer-relative name.