Module model 

The decode engine’s model: a hybrid Gated-DeltaNet checkpoint (qwen35, Qwen3.5 and Qwen3.8) run over a batch of sequences in the engine’s arithmetic, every number from super::numerics.

Each sequence carries its own state: per Gated-DeltaNet block its conv ring and its memory, per attention block its key and value caches, held in the formats the knobs name (EngineNumerics). A step takes any rows of any sequences; a sequence’s rows run in order through its own state, and every row’s arithmetic is its own, so a sequence’s numbers never depend on what else rides the batch.

The matrices run as the array runs them: a block-format row’s exact sums against its activations quantized as the knob names, per 32 elements or per the matrix’s own block (numerics::dot), an f32 matrix’s dot in the vector units’ lanes.

A step is the checkpoint’s program at the model’s knob point (super::program::lower: the model’s one program, read off its step graph), its commands run in order over the step’s token-rows: the model is the interpreter of the program the engine’s sequencer issues.

Structs

EngineModel

A hybrid Gated-DeltaNet checkpoint in the engine’s arithmetic.

EngineNumerics

The knobs of the engine’s arithmetic.

Replay

What a lazily flushed head does at a token whose log, the token’s own entry appended, holds count tokens (Replay::at).

Enums

CacheFormat

How an attention block’s keys and values are stored.

EngineError

Why the engine’s model cannot run a checkpoint.

StateFormat

How a sequence’s recurrent memory and conv ring are stored between tokens.

Constants

CACHE_BUDGET

The default bytes of tensor data the model keeps in memory.

CODED_BUDGET

The default bytes of decoded rows the model keeps: a coded weight takes four bytes.

FLUSH_TOKENS

Tokens a lazily flushed memory carries in its log between flushes.

TRAILING_TOKENS

Tokens of each window a trailing flush replays: half of FLUSH_TOKENS, so its log, two windows at its longest, is as long as a flush’s every FLUSH_TOKENS.

Functions

argmax

The greedy token: the index of the largest logit (super::numerics::argmax).

runs

Whether the engine runs spec at numerics, or what of it the engine does not run yet, by name.