- ModelStore/FileVault: drop nonisolated(unsafe) on shared, mark all instance methods nonisolated (they only read filesystem); ModelKind enum also nonisolated - AIRuntime ↔ ModelStore cross-actor call resolved by the above - LLMSession: replace deprecated Device.setDefault(device:) with task-scoped Device.withDefaultDevice(.cpu, body:); wrap both load and generate so the TaskLocal propagates through ModelContainer.perform
99 lines
4.1 KiB
Swift
99 lines
4.1 KiB
Swift
import Foundation
|
|
import MLX
|
|
import MLXLLM
|
|
import MLXLMCommon
|
|
|
|
/// 封装 MLX 语言模型的流式生成,actor 保证单线程访问。
|
|
/// 基于 mlx-swift-examples 2.29.1(commit 9bff95ca)的 API。
|
|
actor LLMSession {
|
|
let container: ModelContainer
|
|
|
|
init(container: ModelContainer) {
|
|
self.container = container
|
|
}
|
|
|
|
/// 在 simulator 把默认设备强切为 CPU(MLX 的 Metal backend 在部分 Sim 路径会 abort)。
|
|
/// 真机走 body 默认设备(GPU/ANE)。
|
|
/// 用 task-scoped `withDefaultDevice`,TaskLocal 会传递到 child Task / actor 调用。
|
|
private static func withDeviceOverride<R>(
|
|
_ body: () async throws -> R
|
|
) async rethrows -> R {
|
|
#if targetEnvironment(simulator)
|
|
return try await Device.withDefaultDevice(.cpu, body)
|
|
#else
|
|
return try await body()
|
|
#endif
|
|
}
|
|
|
|
/// 从本地目录加载模型(包含 config.json + weights + tokenizer)。
|
|
static func load(folderURL: URL) async throws -> LLMSession {
|
|
let configuration = ModelConfiguration(directory: folderURL)
|
|
let container = try await withDeviceOverride {
|
|
try await LLMModelFactory.shared.loadContainer(
|
|
configuration: configuration
|
|
)
|
|
}
|
|
return LLMSession(container: container)
|
|
}
|
|
|
|
/// 流式生成。返回的 AsyncThrowingStream 被取消时,内部 Task 也会取消。
|
|
/// - Parameters:
|
|
/// - prompt: 原始 prompt 文本(经 processor 转 LMInput)
|
|
/// - maxTokens: 最大 token 数,由 GenerateParameters 控制
|
|
func generate(prompt: String, maxTokens: Int) -> AsyncThrowingStream<TokenChunk, Error> {
|
|
AsyncThrowingStream { continuation in
|
|
let task = Task {
|
|
do {
|
|
try await Self.withDeviceOverride {
|
|
let parameters = GenerateParameters(
|
|
maxTokens: maxTokens,
|
|
temperature: Float(0.6),
|
|
topP: Float(0.9)
|
|
)
|
|
|
|
try await container.perform { (context: ModelContext) in
|
|
let userInput = UserInput(prompt: prompt)
|
|
let lmInput = try await context.processor.prepare(input: userInput)
|
|
|
|
let start = Date()
|
|
var produced = 0
|
|
|
|
for await event in try MLXLMCommon.generate(
|
|
input: lmInput,
|
|
parameters: parameters,
|
|
context: context
|
|
) {
|
|
if Task.isCancelled { break }
|
|
|
|
switch event {
|
|
case .chunk(let text):
|
|
produced += 1
|
|
let elapsed = Date().timeIntervalSince(start)
|
|
let rate = elapsed > 0 ? Double(produced) / elapsed : 0
|
|
continuation.yield(TokenChunk(text: text, decodeRate: rate))
|
|
|
|
case .info:
|
|
// 生成完成统计,是流的最后一个事件
|
|
break
|
|
|
|
case .toolCall:
|
|
// 纯文本生成不会触发,switch 穷举
|
|
break
|
|
}
|
|
}
|
|
// 注:研究笔记里曾建议尾部 MLX.GPU.synchronize() 以确保
|
|
// GPU 操作全部完成。但 AsyncStream 已经 yield 真实解码后的
|
|
// 文字,GPU 是否完全空闲不影响数据正确性。去掉此调用同时省
|
|
// 一份 transitive import MLX 的依赖,简化 SPM 链接。
|
|
}
|
|
}
|
|
continuation.finish()
|
|
} catch {
|
|
continuation.finish(throwing: error)
|
|
}
|
|
}
|
|
continuation.onTermination = { _ in task.cancel() }
|
|
}
|
|
}
|
|
}
|