OVERVIEW
Bolt as a layer or a complete language model.
Boltis the canonical public Bolt attention component andBoltAttentionis its alias.Attentionprovides the standard causal-attention surface under the same MLBricks backend policy, whileBoltModelis the ready-made language-model architecture.
PUBLIC SURFACES
Bolt/BoltAttention | Bolt sequence component. |
Attention | Standard causal attention component. |
BoltModel | Ready-made Bolt language model. |
BoltConfig | Configuration alias for the ready-made model. |
GENERATION
Explicit prefill and decode cache APIs.
Bolt exposesprefillanddecode_stepfor recurrent generation. The cache is created during prefill and passed back into subsequent decode steps.
BACKENDS
Bolt usesbackend="auto"by default. Usenativeonly when the required native route is available, orpytorchfor the reference path.