OVERVIEW
SOUP is a composable state-and-memory architecture.
SOUP combines configurable sequence mixers and FFN paths with its built-in Observer State Memory and SOUP Fusion blocks. The integrated 1.0.0b1 API supports uniform or per-layer widths, mixer choices, FFN choices and the shared MLBricks backend policy.
CONFIGURATION
mixer | Built-in choices includeesaandbolt. |
ffn | Built-in choices includesaffnand conventionalffn. |
width | May be uniform or configured per layer where supported. |
| Observer memory + Fusion | Built into SOUP; not public interchangeable slots in this release. |
BACKEND POLICY
Auto routing remains element-wise.
SOUP does not need to force one backend across the entire architecture. Backend-aware mixers and FFN/residual elements can resolve their supported routes independently while the model exposes a singleset_backendcontrol.
GENERATION
prepare_generation()prepares SOUP's inference-oriented recurrent plan.prefillcreates the cache anddecode_stepadvances it. Custom components that are not compatible with the optimized plan can still use the recurrent methods without callingprepare_generation().