kangkang/康康/AI/VLSession.swift

import Foundation
import MLX
import MLXVLM
import MLXLMCommon

/// 封装 MLX VL 模型(Qwen3-VL)的图像 → 文本推理。
/// 与 LLMSession 同款 actor 隔离,串行化由上游 AIRuntime 统一保证。
actor VLSession {
    let container: ModelContainer

    init(container: ModelContainer) {
        self.container = container
    }

    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 VL 模型(包含 config.json + weights + tokenizer + processor)。
    static func load(folderURL: URL) async throws -> VLSession {
        let configuration = ModelConfiguration(directory: folderURL)
        let container = try await withDeviceOverride {
            try await VLMModelFactory.shared.loadContainer(
                configuration: configuration
            )
        }
        return VLSession(container: container)
    }

    /// 一次性生成(等收完所有 token 再返回完整字符串)。
    /// VL 用于结构化 JSON 抽取,不需要流式 — 也避免半成品 JSON 抖动 UI。
    /// - Parameters:
    ///   - imageURLs: 本地 file:// URL,从 FileVault 拿
    ///   - prompt: 文本指令(VLPrompts.reportExtraction)
    ///   - maxTokens: 默认 512(JSON 体量 ≈ 200-400)
    func analyze(imageURLs: [URL],
                 prompt: String,
                 maxTokens: Int = 512) async throws -> String {
        try await Self.withDeviceOverride {
            try await container.perform { (context: ModelContext) in
                let images = imageURLs.map { UserInput.Image.url($0) }
                let userInput = UserInput(prompt: prompt, images: images)
                let lmInput = try await context.processor.prepare(input: userInput)

                let parameters = GenerateParameters(
                    maxTokens: maxTokens,
                    temperature: Float(0.2),   // JSON 要稳,温度低
                    topP: Float(0.9)
                )

                var collected = ""
                for await event in try MLXLMCommon.generate(
                    input: lmInput,
                    parameters: parameters,
                    context: context
                ) {
                    if Task.isCancelled { break }
                    if case .chunk(let text) = event {
                        collected.append(text)
                    }
                }
                return collected
            }
        }
    }
}