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 Synchronization
public final class GenerationTranscriptMailbox: Sendable {
public struct Snapshot: Sendable {
public let pendingText: String
public let completeText: String
}
private struct State: Sendable {
var pending = ""
var complete = ""
}
private let state = Mutex(State())
public init() {}
public func append(_ text: String) {
guard !text.isEmpty else { return }
state.withLock {
$0.pending += text
$0.complete += text
}
}
public func drain() -> Snapshot {
state.withLock {
let snapshot = Snapshot(pendingText: $0.pending, completeText: $0.complete)
$0.pending = ""
return snapshot
}
}
public var completeText: String {
state.withLock { $0.complete }
}
public func reset() {
state.withLock { $0 = State() }
}
}