markbaseengine/Sources/MarkBase/BatchTemps.swift

import Metal

// ══════════════════════════════════════════════════════════════════
// Batch Forward Temps - Extended buffers for batch processing
// ══════════════════════════════════════════════════════════════════

extension ForwardTemps {
    
    /// Create batch-specific temporary buffers
    /// These are separate from single-token buffers to avoid interference
    public func createBatchBuffers(
        device: MTLDevice,
        batchSize: Int,
        hiddenSize: Int,
        nHeads: Int,
        headDim: Int,
        intermediateSize: Int
    ) throws -> BatchTemps {
        return try BatchTemps(
            device: device,
            batchSize: batchSize,
            hiddenSize: hiddenSize,
            nHeads: nHeads,
            headDim: headDim,
            intermediateSize: intermediateSize
        )
    }
}

/// Batch-specific temporary buffers for parallel layer processing
public struct BatchTemps {
    public let hBatch: MTLBuffer        // [batchSize, hiddenSize] - hidden state batch
    public let qBatch: MTLBuffer        // [batchSize, nHeads * headDim] - query batch
    public let nsBatch: MTLBuffer       // [batchSize, nHeads * headDim] - norm scratch batch
    public let interBatch: MTLBuffer    // [batchSize, intermediateSize] - intermediate batch
    
    public init(
        device: MTLDevice,
        batchSize: Int,
        hiddenSize: Int,
        nHeads: Int,
        headDim: Int,
        intermediateSize: Int
    ) throws {
        func buf(_ n: Int) throws -> MTLBuffer {
            guard let b = device.makeBuffer(length: n * MemoryLayout<Float>.stride,
                                            options: .storageModeShared)
            else { throw NSError(domain: "BatchTemps", code: -1, 
                                 userInfo: [NSLocalizedDescriptionKey: "Buffer creation failed"]) }
            return b
        }
        
        hBatch = try buf(batchSize * hiddenSize)
        qBatch = try buf(batchSize * nHeads * headDim)
        nsBatch = try buf(batchSize * nHeads * headDim)
        interBatch = try buf(batchSize * intermediateSize)
    }
}