//
//  GenerationParams.swift
//  Zyquo MLX
//
//  Author: Simon-Pierre Boucher
//  Mail: contact@spboucher.ai
//

import Foundation
import MLXLMCommon

/// App-level generation parameters (Codable for persistence, presets, and
/// per-run notes). Mapped 1:1 onto `MLXLMCommon.GenerateParameters`
/// (verified fields — docs/MLX-RESEARCH.md §4.3).
struct GenerationParams: Codable, Hashable, Sendable {
    var temperature: Float = 0.6
    var topP: Float = 1.0
    var topK: Int = 0
    var minP: Float = 0.0
    var maxTokens: Int? = 2048
    var repetitionPenalty: Float?
    var repetitionContextSize: Int = 20
    var seed: UInt64?
    /// KV-cache quantization bits (nil = off).
    var kvBits: Int?
    var kvGroupSize: Int = 64
    /// Cap on KV cache length (RotatingKVCache when set).
    var maxKVSize: Int?

    var asMLX: GenerateParameters {
        GenerateParameters(
            maxTokens: maxTokens,
            maxKVSize: maxKVSize,
            kvBits: kvBits,
            kvGroupSize: kvGroupSize,
            temperature: temperature,
            topP: topP,
            topK: topK,
            minP: minP,
            repetitionPenalty: repetitionPenalty,
            repetitionContextSize: repetitionContextSize,
            seed: seed
        )
    }
}

/// Final statistics for one generation (from `GenerateCompletionInfo`).
struct InferenceStats: Codable, Hashable, Sendable {
    var promptTokens: Int
    var generatedTokens: Int
    /// Time to first token ≈ prompt processing time.
    var ttft: TimeInterval
    var generateTime: TimeInterval
    var stopReason: String

    var tokensPerSecond: Double {
        generateTime > 0 ? Double(generatedTokens) / generateTime : 0
    }
}