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 TurboFieldfare
import TurboFieldfareRepackCore
public enum AppVisionPackInstallationStatus: Equatable, Sendable {
case missing
case partial(String)
case complete
/// This model layout cannot host a companion pack at all — the directory is
/// not named `*.gturbo`, so there is no path to put one at.
///
/// Distinct from `.partial`, which means a pack is there and damaged and
/// offers Repair. Reporting this as `.partial` printed a raw
/// `invalidTextModelPath(...)` under "Needs repair" and offered to download
/// a pack that could never be located.
case unsupportedLayout
}
public enum AppVisionPackInstallationProbe {
public static func status(at textModelDirectory: URL) -> AppVisionPackInstallationStatus {
let textModelDirectory = textModelDirectory.standardizedFileURL
let companion: URL
do {
companion = try VisionPackLocation.companionURL(forTextModel: textModelDirectory)
} catch {
// The only failure here is a directory name that cannot carry a
// companion, which is a property of the layout rather than of any
// pack on disk.
return .unsupportedLayout
}
guard FileManager.default.fileExists(atPath: companion.path) else {
return .missing
}
do {
_ = try VisionPackVerifier.verify(
directory: companion,
textModelDirectory: textModelDirectory,
verifyWeights: false)
return .complete
} catch {
return .partial("\(error)")
}
}
}