Module program 

The decode engine’s program: one step of a checkpoint as the commands its units run, in order.

A step holds its token-rows — every row of every sequence the step advances — in the token-row store, each row a vector in every Slot. A command is one unit’s job over every row of the step, in the step’s row order, finished before the next command starts: Command::Matvec streams one matrix through the array and the block-scale lanes once for all the rows; the norms, the gates and the products are the vector unit’s micro-programs; Command::Conv and Command::Delta carry a Gated-DeltaNet block’s ring and memory, and Command::CacheAppend and Command::Attend an attention block’s cache, each over every row’s own sequence. A command that carries a sequence’s state takes its rows in order, so a sequence’s rows see one another as a prompt run a token at a time does.

lower reads the step off the model’s one program — the decode step author writes, lifted into the one IR, fused by the shared table under the engine’s shim and read block by block (super::lower_graph) — and super::model::EngineModel runs exactly that list: the program is the model, so what the sequencer issues and the cycle twin prices is what the model computes. hand_lowering writes the same commands from the spec directly; it is kept only as the oracle the step-graph reading is held to.

Structs

Lowered

A program and the cut of the model’s one program it was read off.

Program

One step of a checkpoint at a knob point.

Enums

Command

One command of a step.

ProgramError

Why the engine has no program for a model: the author cannot write its step, the front end cannot lift it, or its step graph holds a block the engine has no unit for.

Slot

A vector each token-row of a step carries, one of the token-row store’s slots.

Functions

hand_lowering

The step of spec at numerics written from the spec directly: the embedding, each layer’s mixer and FFN around the residual, the final norm and the head. The oracle lower is held to, never a program the engine is given.

lower

The engine’s step of spec at numerics, read off the decode step the author writes over a cache of ctx positions: the model’s one program, lifted by the shared front end into its step graph and read block by block. The commands do not depend on ctx (the cache’s length is the layout’s); the identity records the cut read.

Type Aliases

Word

A command as the sequencer reads it: 256 bits, [7:0] the command, [15:8] its layer, seven 5-bit slots from bit 16, two 20-bit tensors from bit 51 (each its index in Program::tensors, the weight image’s order), four 16-bit counts from bit 91, a 20-bit element offset from bit 155, an 8-bit rotary table from bit 175 (its index in Program::ropes) and a 32-bit f32 constant from bit 183.