OVERVIEW
ESA in MLBricks Kit.
Entangled State Attention (ESA) is available as both a reusable sequence layer and a ready-made language-model architecture. In MLBricks Kit 1.0.0b1 it follows the sharedauto | native | pytorchbackend policy and supports recurrent prefill/decode workflows for generation.
CHOOSE THE SURFACE
| Use | Public surface |
|---|---|
| Drop-in sequence component | ESA/esa |
| Ready-made language model | ESAModel+ESAModelConfig |
| Training | mlbricks.Trainerormlbricks.train |
| Save / load / compile / quantize | Package-level MLBricks lifecycle helpers |
BACKENDS
One public backend policy across the kit.
backend="auto"is the default. It lets MLBricks qualify supported routes and use the appropriate implementation for the element. Usenativewhen you require a supported native route, orpytorchwhen you want the reference path explicitly.
from mlbricks import ESA, backend_report
layer = ESA(embd=384, head=6, backend="auto")
print(backend_report(layer))GENERATION
Forward for full sequences, recurrent state for generation.
Use ordinaryforwardfor training and full-sequence work. For recurrent inference,prefillcreates the state anddecode_stepadvances it one step at a time.ESAModelalso exposes its higher-level generation helpers.