A high-performance inference engine for AI models
import Foundation
struct CachedModel: Identifiable, Encodable {
let name: String
let path: String
let bytes: Int64
var id: String { path }
}
enum ModelStore {
private enum Failure: Error {
case notCached(String)
}
static func list() -> [CachedModel] {
let caches = URL.cachesDirectory.appending(path: ".cache")
var models: [CachedModel] = []
for url in contents(caches.appending(path: "huggingface/hub"))
where url.lastPathComponent.hasPrefix("models--") {
let repository = url.lastPathComponent.trimmingPrefix("models--")
models.append(model(String(repository.replacing("--", with: "/")), url))
}
for vendor in contents(caches.appending(path: "mirai/models")) {
for url in contents(vendor) {
models.append(model(url.lastPathComponent, url))
}
}
for url in contents(URL.documentsDirectory.appending(path: "models")) {
models.append(model("models/\(url.lastPathComponent)", url))
}
return models
}
static func delete(path: String) throws {
guard list().contains(where: { $0.path == path }) else {
throw Failure.notCached(path)
}
try FileManager.default.removeItem(atPath: path)
}
private static func model(_ name: String, _ url: URL) -> CachedModel {
var bytes: Int64 = 0
let files = FileManager.default.enumerator(
at: url, includingPropertiesForKeys: [.totalFileAllocatedSizeKey])
while let file = files?.nextObject() as? URL {
bytes += Int64(
(try? file.resourceValues(forKeys: [.totalFileAllocatedSizeKey])
.totalFileAllocatedSize) ?? 0)
}
return CachedModel(name: name, path: url.path, bytes: bytes)
}
private static func contents(_ url: URL) -> [URL] {
let urls = try? FileManager.default.contentsOfDirectory(
at: url, includingPropertiesForKeys: nil, options: .skipsHiddenFiles)
return (urls ?? []).sorted { $0.lastPathComponent < $1.lastPathComponent }
}
}