Piro Inference Architecture

durableWeights = LoadWeights()

adaptationState = InitializeAdaptationState(durableWeights)

contextState = InitializeContextState()

x = Embed(PiroInput)

for each observedChunk inChunkText(x.text):

runtimeWeights = BindAdaptationState(durableWeights, adaptationState)

prediction = PredictNextToken(observedChunk, contextState,runtimeWeights)

adaptationState = AdaptationStep(adaptationState, observedChunk,prediction)

contextState = AdvanceContextState(observedChunk, contextState)

for step in range(maxNewTokens):

logits = OutputHead(contextState, adaptationState)

token = Argmax(logits)

Emit(token)

if token == eosToken:

break

contextState = AdvanceContextState(token, contextState)

durableWeights = ConsolidateWeights(durableWeights, adaptationState)

SaveWeights(durableWeights)