// Author: Simon-Pierre Boucher — contact@spboucher.ai
//
// Chip name, core count and unified memory via sysctl — feeds the memory
// estimate badge in the New Run form.
import Foundation

enum SystemInfo {
    static func sysctlString(_ name: String) -> String? {
        var size = 0
        guard sysctlbyname(name, nil, &size, nil, 0) == 0, size > 0 else { return nil }
        var buf = [CChar](repeating: 0, count: size)
        guard sysctlbyname(name, &buf, &size, nil, 0) == 0 else { return nil }
        return String(cString: buf)
    }

    static func sysctlInt(_ name: String) -> Int64? {
        var value: Int64 = 0
        var size = MemoryLayout<Int64>.size
        guard sysctlbyname(name, &value, &size, nil, 0) == 0 else { return nil }
        return value
    }

    static var chipName: String { sysctlString("machdep.cpu.brand_string") ?? "Apple Silicon" }
    static var memoryBytes: Int64 { sysctlInt("hw.memsize") ?? 0 }
    static var cpuCores: Int { Int(sysctlInt("hw.ncpu") ?? 0) }

    /// Rough training footprint: weights+grads+AdamW moments in f32 (4 copies
    /// of params × 4 bytes) + activation estimate per micro-batch.
    static func estimatedTrainingBytes(config: ForgeConfig) -> Int64 {
        let params = Int64(config.model.paramCount)
        let states = params * 4 * 4
        let actPerToken = Int64(config.model.nLayers * config.model.dModel * 24)
        let activations = actPerToken
            * Int64(config.train.batchSize * config.model.contextLength)
        return states + activations
    }
}