MarkBase Admin
ac75faa0cc
CI / build-and-test (push) Has been cancelled
- E4B-MarkBase model (42 layers, 4.4GB) loaded successfully - All Phase 1-6 tests passed (model loading, forward pass, vision/audio towers, token generation, performance) - All stress tests passed (5/5 in 127.6s) - Concurrent inference - Memory stress (67.5 tok/s, 0 NaN) - Continuous generation - Batch processing - Long-running stability - Swift Metal inference engine with multimodal support
70 lines
2.7 KiB
Swift
70 lines
2.7 KiB
Swift
import Foundation
|
|
import Metal
|
|
|
|
// ─────────────────────────────────────────────────────────────
|
|
// Float16 Support for MarkBase
|
|
// ─────────────────────────────────────────────────────────────
|
|
|
|
/// Float16 data type for memory-efficient computation
|
|
public typealias Float16 = Swift.Float16
|
|
|
|
/// Float16 buffer utilities
|
|
public enum Float16Utils {
|
|
/// Convert Float32 array to Float16
|
|
public static func toFloat16(_ values: [Float]) -> [Float16] {
|
|
return values.map { Float16($0) }
|
|
}
|
|
|
|
/// Convert Float16 array to Float32
|
|
public static func toFloat32(_ values: [Float16]) -> [Float] {
|
|
return values.map { Float($0) }
|
|
}
|
|
|
|
/// Create MTLBuffer from Float16 array
|
|
public static func makeBuffer(device: MTLDevice, values: [Float16]) -> MTLBuffer? {
|
|
return device.makeBuffer(
|
|
bytes: values,
|
|
length: values.count * MemoryLayout<Float16>.stride,
|
|
options: .storageModeShared
|
|
)
|
|
}
|
|
|
|
/// Read Float16 from MTLBuffer
|
|
public static func readFloat16(from buffer: MTLBuffer, count: Int) -> [Float16] {
|
|
let ptr = buffer.contents().assumingMemoryBound(to: Float16.self)
|
|
return Array(UnsafeBufferPointer(start: ptr, count: count))
|
|
}
|
|
|
|
/// Convert Float32 MTLBuffer to Float16
|
|
public static func convertBuffer(
|
|
from buffer: MTLBuffer,
|
|
device: MTLDevice,
|
|
count: Int
|
|
) -> MTLBuffer? {
|
|
let float32Ptr = buffer.contents().assumingMemoryBound(to: Float.self)
|
|
let float32Values = Array<Float>(UnsafeBufferPointer(start: float32Ptr, count: count))
|
|
let float16Values = toFloat16(float32Values)
|
|
return makeBuffer(device: device, values: float16Values)
|
|
}
|
|
}
|
|
|
|
/// Float16 quantization for model weights
|
|
public struct Float16Quantizer {
|
|
/// Quantize Float32 weights to Float16
|
|
public static func quantize(weights: [Float]) -> [Float16] {
|
|
return Float16Utils.toFloat16(weights)
|
|
}
|
|
|
|
/// Dequantize Float16 weights to Float32
|
|
public static func dequantize(weights: [Float16]) -> [Float] {
|
|
return Float16Utils.toFloat32(weights)
|
|
}
|
|
|
|
/// Calculate memory savings
|
|
public static func memorySavings(float32Count: Int, float16Count: Int) -> Double {
|
|
let float32Size = float32Count * MemoryLayout<Float>.stride
|
|
let float16Size = float16Count * MemoryLayout<Float16>.stride
|
|
return 1.0 - (Double(float16Size) / Double(float32Size))
|
|
}
|
|
}
|