Model and training API

Maturity: verified

Signal preparation

Symbol	Contract
prepare_iq_input	normalize and patch one semantic batch, then prepare masked physical metadata
IQModelInput	aligned normalized I/Q, patches, masks, metadata, sample IDs, and traces
FixedPatchingConfig	selected fixed width within a supported-width set

Model and objective

Symbol	Contract
CausalPatchModel	abstract causal model over prepared metadata and I/Q patches
CausalPatchModelOutput	(B, M+N, D) states, metadata boundary, and aligned token mask
TinyCausalIQModel	compact PyTorch transformer used by the bounded path
NextPatchPredictionTask	predict patch i+1 from causal patch state i and calculate masked MSE

Optimizer step and catalog

Symbol	Contract
run_npp_optimizer_step	one guarded parameter update; caller owns lifecycle
PretrainCatalog	map readable component keys to validators and builders
PretrainProgram	constructed task, optimizer, input-prep config, and selected identities
PRETRAIN_CATALOG	shipped Tiny/NPP/AdamW registrations

These functions and objects do not own iteration, application artifacts, provider resources, or job lifecycle.

See Also

• Models and objectives

• Bounded configuration contract