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 TurboFieldfareFormat
enum GTurboLayoutValidator {
static func validate(path: String,
plan: RepackPlan,
audit: RepackAudit? = nil) throws {
let data = try Posix.readBoundedData(path, maximumBytes: 16 * 1024 * 1024)
let layout: GTurboPackedExpertsLayoutV1
do { layout = try GTurboPackedExpertsLayoutCodec.decode(data) }
catch {
throw RepackError.configurationInvalid(
detail: "layout.json validation failed: \(error)")
}
var validatedLogicalExperts = 0
for layer in layout.layers {
guard let planLayer = plan.layers.first(where: { $0.layerIndex == layer.layer }) else {
throw RepackError.configurationInvalid(detail: "layout.json validation failed: malformed layer")
}
guard layer.experts.count == planLayer.expertsPerLayer,
layout.expertStride == planLayer.expertStride else {
throw RepackError.configurationInvalid(detail:
"layout.json validation failed: plan mismatch in layer \(layer.layer)")
}
validatedLogicalExperts += layer.experts.count
}
guard layout.layers.count == plan.layers.count else {
throw RepackError.configurationInvalid(
detail: "layout.json validation failed: layer count mismatch")
}
audit?.packedExpertLayoutAuditLogicalIDCount = validatedLogicalExperts
audit?.packedExpertLayoutOffsetValidationPassed = true
}
}