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

import Foundation

/// Training hyperparameters. Field names/defaults mirror mlx-lm's
/// `CONFIG_DEFAULTS` exactly (docs/TRAINING-RESEARCH.md §1.2); serialized to
/// the YAML config the Python bridge consumes (`lora_parameters` and
/// `lr_schedule` are config-only upstream).
struct HyperParams: Codable, Hashable, Sendable {
    // LoRA-specific (config-only upstream)
    var rank: Int = 8
    /// MLX uses a single scale factor instead of alpha/rank (default 20.0).
    var scale: Double = 20.0
    var dropout: Double = 0.0
    /// Target modules; nil = adapt all linear layers in targeted blocks.
    var keys: [String]?

    // Core loop
    var numLayers: Int = 16 // -1 = all layers
    var batchSize: Int = 4
    var iterations: Int = 1000
    var learningRate: Double = 1e-5
    var maxSeqLength: Int = 2048
    var seed: Int = 0

    // Optimizer: adam | adamw | muon | sgd | adafactor
    var optimizer: String = "adam"
    var gradCheckpoint: Bool = false
    var gradAccumulationSteps: Int = 1
    var maskPrompt: Bool = false

    // Cadences
    var stepsPerReport: Int = 10
    var stepsPerEval: Int = 200
    var saveEvery: Int = 100
    var valBatches: Int = 25
}