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 AppKit
import TurboFieldfareAppCore
/// Which images a turn's inline row is built from.
///
/// One cell per attachment, always. An image whose file is gone or will not
/// decode used to be dropped, and the turn then rendered as one that never had
/// a picture — under an answer that discusses it — so the transcript stopped
/// agreeing with the context the model was actually given. A conversation can
/// lose a stored file to anything outside this app, and the reader has to be
/// able to tell "there was a picture here" from "there was not".
public enum TranscriptImageCells {
public static func images(
for attachments: [ChatImage],
load: (ChatImage) -> NSImage?,
unreadable: () -> NSImage
) -> [NSImage] {
attachments.map { load($0) ?? unreadable() }
}
}