@maruaudio/dsp/neural/nam-a2-spec declares the standard Neural Amp Modeler A2 topology without any inference state or model content, so importers, packers and validators can check a model without loading the inference code. It fixes 23 layers with their kernel sizes (6, with 15 at layers 15 and 16) and dilations (two runs of 1, 3, 7, 17, 41, 101, 239, then 1, 13, then a third run), a 16-tap head, a LeakyReLU slope of 0.01, a 6,347-sample prewarm (the receptive field) and a 50,832-sample history ceiling.

namA2WeightCount(3) // 1871, Lite
namA2WeightCount(8) // 12146, Full

Lifecycle

None. The constants are immutable; namA2WeightCount(channels) is 179·c² + 86·c + 2 (156 convolution taps and 23 residual matrices, three bias or mix-in vectors per layer, plus the input, head and trailing scale terms), and throws for any width other than 3 or 8. Call it at import or preparation time.

When to use it

Checking a NAM file's architecture and weight count before packing or installing it, in code that must stay light: @maruaudio/nam-runtime's packer and module tests import it. @maru/processors-nam-a2 re-exports it.

Limits

• Standard A2 only, Lite or Full; it names no other NAM architecture.
• It describes the topology; validating finite weights and prewarm output belongs to @maru/processors-nam-a2.