import Accelerate import Foundation /// Scores Chinese text: natural-log probability of `text` following `context`. public protocol LanguageModel: AnyObject { func logProb(_ text: String, context: String) -> Double } /// GPT-2 forward pass (CPU, Accelerate) for the character-level CKIP gpt2-tiny-chinese /// model (GPL-3.0), stored by training/lm/export_lm.py as float16 ("AZLM" format). /// Weights are expanded to float32 in memory (~46 MB for gpt2-tiny). public final class GPT2LanguageModel: LanguageModel { public enum LoadError: Error { case badFormat(String) } let nLayer: Int, nHead: Int, d: Int, vocabSize: Int let eps: Float private var weights: [Float] = [] private var offsets: [String: Int] = [:] private var vocab: [Character: Int32] = [:] static let cls: Int32 = 101, unk: Int32 = 100 /// Longest input scored (context + text); longer contexts are cut from the left. public var maxTokens = 64 public init(modelURL: URL, vocabURL: URL) throws { let data = try Data(contentsOf: modelURL, options: .mappedIfSafe) guard data.count > 8, data.prefix(4) == Data("AZLM".utf8) else { throw LoadError.badFormat("magic") } let headerLength = Int(data[4]) | Int(data[5]) << 8 | Int(data[6]) << 16 | Int(data[7]) << 24 guard let header = try JSONSerialization.jsonObject(with: data.subdata(in: 8..<(8 + headerLength))) as? [String: Any], let nLayer = header["n_layer"] as? Int, let nHead = header["n_head"] as? Int, let d = header["n_embd"] as? Int, let vocabSize = header["vocab"] as? Int, let eps = header["eps"] as? Double, let tensors = header["tensors"] as? [[Any]] else { throw LoadError.badFormat("header") } self.nLayer = nLayer; self.nHead = nHead; self.d = d; self.vocabSize = vocabSize; self.eps = Float(eps) for t in tensors { guard let name = t[0] as? String, let offset = t[2] as? Int else { throw LoadError.badFormat("tensor") } offsets[name] = offset } let start = 8 + headerLength let count = (data.count - start) / 2 weights = [Float](repeating: 0, count: count) try data.withUnsafeBytes { (raw: UnsafeRawBufferPointer) in guard let base = raw.baseAddress?.advanced(by: start) else { throw LoadError.badFormat("data") } try weights.withUnsafeMutableBufferPointer { out in var src = vImage_Buffer(data: UnsafeMutableRawPointer(mutating: base), height: 1, width: vImagePixelCount(count), rowBytes: count * 2) var dst = vImage_Buffer(data: out.baseAddress, height: 1, width: vImagePixelCount(count), rowBytes: count * 4) guard vImageConvert_Planar16FtoPlanarF(&src, &dst, 0) == kvImageNoError else { throw LoadError.badFormat("float16") } } } var index: Int32 = 0 for line in try String(contentsOf: vocabURL, encoding: .utf8).split(separator: "\n", omittingEmptySubsequences: false) { if line.count == 1, let c = line.first, vocab[c] == nil { vocab[c] = index } index += 1 } for name in ["wte.weight", "wpe.weight", "ln_f.weight"] where offsets[name] == nil { throw LoadError.badFormat("missing \(name)") } } func tokens(_ text: String) -> [Int32] { text.map { vocab[$0] ?? GPT2LanguageModel.unk } } public func logProb(_ text: String, context: String) -> Double { let target = tokens(text) guard !target.isEmpty else { return 0 } var ctx = tokens(context) if ctx.count + target.count + 1 > maxTokens { ctx = Array(ctx.suffix(max(0, maxTokens - target.count - 1))) } let input = [GPT2LanguageModel.cls] + ctx + target.dropLast() let h = hidden(input) // [T, d] return weights.withUnsafeBufferPointer { w in logSoftmaxScore(h, rows: input.count, from: ctx.count, targets: target, w: w) } } // MARK: - Forward pass private func w(_ name: String) -> Int { offsets[name]! } /// Final hidden states [T, d] (row-major). func hidden(_ ids: [Int32]) -> [Float] { let T = ids.count, d = self.d var x = [Float](repeating: 0, count: T * d) weights.withUnsafeBufferPointer { W in let wte = w("wte.weight"), wpe = w("wpe.weight") for t in 0..) -> [Float] { let d = self.d var out = [Float](repeating: 0, count: x.count) let g = W.baseAddress! + w(name + ".weight"), b = W.baseAddress! + w(name + ".bias") for r in 0..) -> [Float] { var out = [Float](repeating: 0, count: rows * outDim) vDSP_mmul(x, 1, W.baseAddress! + w(name + ".weight"), 1, &out, 1, vDSP_Length(rows), vDSP_Length(outDim), vDSP_Length(inDim)) let bias = W.baseAddress! + w(name + ".bias") out.withUnsafeMutableBufferPointer { op in for r in 0.. [Float] { let d = self.d, hd = d / nHead let scale = 1 / Float(hd).squareRoot() var out = [Float](repeating: 0, count: T * d) var scores = [Float](repeating: 0, count: T) qkv.withUnsafeBufferPointer { q in out.withUnsafeMutableBufferPointer { o in for head in 0..) -> Double { let d = self.d, V = vocabSize let rowsUsed = T - start // hᵀ for the rows we need: [d, rowsUsed] var hT = [Float](repeating: 0, count: d * rowsUsed) h.withUnsafeBufferPointer { hp in vDSP_mtrans(hp.baseAddress! + start * d, 1, &hT, 1, vDSP_Length(d), vDSP_Length(rowsUsed)) } // logitsᵀ [V, rowsUsed] = wte [V, d] · hᵀ [d, rowsUsed] var logitsT = [Float](repeating: 0, count: V * rowsUsed) vDSP_mmul(W.baseAddress! + w("wte.weight"), 1, hT, 1, &logitsT, 1, vDSP_Length(V), vDSP_Length(rowsUsed), vDSP_Length(d)) var logits = [Float](repeating: 0, count: V * rowsUsed) // [rowsUsed, V] vDSP_mtrans(logitsT, 1, &logits, 1, vDSP_Length(rowsUsed), vDSP_Length(V)) var total = 0.0 var expBuf = [Float](repeating: 0, count: V) logits.withUnsafeMutableBufferPointer { lp in expBuf.withUnsafeMutableBufferPointer { eb in let e = eb.baseAddress! for i in 0..