spb/prisme Public MIT

1//2//  InferenceQueue.swift3//  Prisme4//5//  Author: Simon-Pierre Boucher <contact@spboucher.ai>6//78import Foundation910/// Priority of an inference request. User gestures preempt page enrichment,11/// which preempts background work (CLAUDE.md §8).12enum InferencePriority: Int, Comparable, Sendable {13    case background = 014    case activePage = 115    case userGesture = 21617    static func < (lhs: Self, rhs: Self) -> Bool { lhs.rawValue < rhs.rawValue }18}1920/// Caps in-flight model requests and wakes waiters by priority. The Neural21/// Engine serializes inference anyway — an unbounded fan-out only hides the22/// real queue and wastes energy.23actor InferenceQueue {24    private let maxConcurrent: Int25    private var inFlight = 026    private var waiters: [(priority: InferencePriority, order: Int, continuation: CheckedContinuation<Void, Never>)] = []27    private var counter = 02829    init(maxConcurrent: Int) {30        self.maxConcurrent = maxConcurrent31    }3233    func run<T: Sendable>(34        _ priority: InferencePriority,35        operation: @Sendable () async throws -> T36    ) async rethrows -> T {37        await acquire(priority)38        defer { release() }39        return try await operation()40    }4142    private func acquire(_ priority: InferencePriority) async {43        if inFlight < maxConcurrent {44            inFlight += 145            return46        }47        counter += 148        let order = counter49        await withCheckedContinuation { continuation in50            waiters.append((priority, order, continuation))51        }52    }5354    private func release() {55        if let index = waiters.indices.max(by: { lhs, rhs in56            (waiters[lhs].priority, -waiters[lhs].order) < (waiters[rhs].priority, -waiters[rhs].order)57        }) {58            let next = waiters.remove(at: index)59            next.continuation.resume()60        } else {61            inFlight -= 162        }63    }64}65