Module shim 

What the decode engine runs as one unit: its answer to the shared optimizer.

The engine reads a step block by block (super::lower_graph), and the merges it makes are the shared table’s classes: • the projections of one activation whose matrices are one checkpoint tensor — a Gated-DeltaNet block’s query, key and value as row ranges following one another from the tensor’s first row, an attention block’s queries and output gate as the halves of each head — as one matvec (merged_qkv); • a convolution and the SiLU after it, the conv unit’s one pass over a Gated-DeltaNet block’s channels (conv_bias_silu); • a Gated-DeltaNet block’s write strength and forget gate, the sigmoid and the decay its delta rule reads, as the gates unit (gdn_gating); • an attention’s output gate, its result times the sigmoid of the gate, as the sigmoid-gate unit (glu_act_mul). A norm, a rotation, a cache append and an add each run as a command of their own. An op the engine has no unit for is refused as having none, for what super::capability names when it refuses a whole checkpoint for it.

Structs

EngineShim

The engine’s shim for one model at one knob point: its spec says which checkpoint tensor and view each constant reads, its numerics the contract its build keeps — which change the contract, never what the engine runs whole.

Constants

NO_UNIT

How the shim begins every refusal of an op the engine has no unit for.