Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook
Swift apple-silicongemmagemma4gemma4-26b-a4bgpgpullmllm-inferencelocal-aimacosmetalon-device-aion-device-llmswift
import Foundation
import TurboFieldfareDecodeProtocol
extension DecodeConversationGate.Rejection {
/// The wording the service sends back. Names what was asked for and what is
/// actually open, because "rejected" alone is not a diagnosis.
var message: String {
switch self {
case .staleConversation(let requested, let open):
let openText = open.map(\.uuidString) ?? "none"
return "turn belongs to conversation \(requested.uuidString), "
+ "and the open conversation is \(openText)"
case .outOfOrderTurn(let requested, let committed):
let requestedText = requested.map(String.init) ?? "unset"
return "turn index \(requestedText) does not follow the "
+ "\(committed) turns committed to this conversation"
case .oneShotDuringConversation(let open):
return "a conversation is open (\(open.uuidString)); "
+ "a one-shot turn cannot run on it"
}
}
/// The terminal event a refused turn comes back as.
///
/// `.lineageLost`, not `.failed`. A refusal means the turn was composed
/// against a conversation the KV is no longer holding, so retrying it is
/// refused identically forever — and `.failed` reaches the app as a
/// retryable `.unknown`, which offers exactly that retry. `.lineageLost`
/// is the kind that says the lineage is over and only a new chat clears it.
///
func terminalEvent(generationID: UUID,
conversationEpoch: UUID?) -> DecodeServiceEvent {
DecodeServiceEvent(
kind: .lineageLost,
generationID: generationID,
error: message,
conversationEpoch: conversationEpoch)
}
}