mirror of
https://github.com/mii443/AzooKeyKanaKanjiConverter.git
synced 2025-08-22 15:05:26 +00:00
425 lines
18 KiB
Swift
425 lines
18 KiB
Swift
import llama
|
|
import SwiftUtils
|
|
import HeapModule
|
|
import Algorithms
|
|
import Foundation
|
|
|
|
struct FixedSizeHeap<Element: Comparable> {
|
|
private var size: Int
|
|
private var heap: Heap<Element>
|
|
|
|
init(size: Int) {
|
|
self.size = size
|
|
self.heap = []
|
|
}
|
|
|
|
mutating func removeMax() {
|
|
self.heap.removeMax()
|
|
}
|
|
|
|
mutating func removeMin() {
|
|
self.heap.removeMin()
|
|
}
|
|
|
|
@discardableResult
|
|
mutating func insertIfPossible(_ element: Element) -> Bool {
|
|
if self.heap.count < self.size {
|
|
self.heap.insert(element)
|
|
return true
|
|
} else if let min = self.heap.min, element > min {
|
|
self.heap.replaceMin(with: element)
|
|
return true
|
|
} else {
|
|
return false
|
|
}
|
|
}
|
|
|
|
var unordered: [Element] {
|
|
self.heap.unordered
|
|
}
|
|
|
|
var max: Element? {
|
|
self.heap.max
|
|
}
|
|
|
|
var min: Element? {
|
|
self.heap.min
|
|
}
|
|
}
|
|
|
|
enum ZenzError: LocalizedError {
|
|
case couldNotLoadModel(path: String)
|
|
case couldNotLoadContext
|
|
|
|
var errorDescription: String? {
|
|
switch self {
|
|
case .couldNotLoadContext: "failed to load context"
|
|
case .couldNotLoadModel(path: let path): "could not load model weight at \(path)"
|
|
}
|
|
}
|
|
}
|
|
|
|
class ZenzContext {
|
|
private var model: OpaquePointer
|
|
private var context: OpaquePointer
|
|
private var prevInput: [llama_token] = []
|
|
|
|
private let n_len: Int32 = 512
|
|
|
|
init(model: OpaquePointer, context: OpaquePointer) {
|
|
self.model = model
|
|
self.context = context
|
|
}
|
|
|
|
deinit {
|
|
llama_free(context)
|
|
llama_free_model(model)
|
|
llama_backend_free()
|
|
}
|
|
|
|
private static var ctx_params: llama_context_params {
|
|
let n_threads = max(1, min(8, ProcessInfo.processInfo.processorCount - 2))
|
|
debug("Using \(n_threads) threads")
|
|
var ctx_params = llama_context_default_params()
|
|
ctx_params.seed = 1234
|
|
ctx_params.n_ctx = 512
|
|
ctx_params.n_threads = UInt32(n_threads)
|
|
ctx_params.n_threads_batch = UInt32(n_threads)
|
|
ctx_params.n_batch = 512
|
|
return ctx_params
|
|
}
|
|
|
|
static func createContext(path: String) throws -> ZenzContext {
|
|
llama_backend_init()
|
|
var model_params = llama_model_default_params()
|
|
model_params.use_mmap = true
|
|
let model = llama_load_model_from_file(path, model_params)
|
|
guard let model else {
|
|
debug("Could not load model at \(path)")
|
|
throw ZenzError.couldNotLoadModel(path: path)
|
|
}
|
|
|
|
let context = llama_new_context_with_model(model, ctx_params)
|
|
guard let context else {
|
|
debug("Could not load context!")
|
|
throw ZenzError.couldNotLoadContext
|
|
}
|
|
|
|
return ZenzContext(model: model, context: context)
|
|
}
|
|
|
|
func reset_context() throws {
|
|
llama_free(self.context)
|
|
let context = llama_new_context_with_model(self.model, Self.ctx_params)
|
|
guard let context else {
|
|
debug("Could not load context!")
|
|
throw ZenzError.couldNotLoadContext
|
|
}
|
|
self.context = context
|
|
}
|
|
|
|
private func get_logits(tokens: [llama_token], logits_start_index: Int = 0) -> UnsafeMutablePointer<Float>? {
|
|
// manage kv_cache
|
|
do {
|
|
let commonTokens = self.prevInput.commonPrefix(with: tokens)
|
|
llama_kv_cache_seq_rm(context, 0, llama_pos(commonTokens.count), -1)
|
|
}
|
|
var batch = llama_batch_init(512, 0, 1)
|
|
let n_ctx = llama_n_ctx(context)
|
|
let n_kv_req = tokens.count + (Int(n_len) - tokens.count)
|
|
if n_kv_req > n_ctx {
|
|
debug("error: n_kv_req > n_ctx, the required KV cache size is not big enough")
|
|
}
|
|
for i in tokens.indices {
|
|
llama_batch_add(&batch, tokens[i], Int32(i), [0], logits: logits_start_index <= i)
|
|
}
|
|
// 評価
|
|
if llama_decode(context, batch) != 0 {
|
|
debug("llama_decode() failed")
|
|
return nil
|
|
}
|
|
return llama_get_logits(context)
|
|
}
|
|
|
|
func evaluate(text: String, ignorePrompt: String = "") -> Float {
|
|
let tokens_list = self.tokenize(text: text, add_bos: true, add_eos: true)
|
|
guard let logits = self.get_logits(tokens: tokens_list) else {
|
|
debug("logits unavailable")
|
|
return .nan
|
|
}
|
|
let tokenizedPromptCount = ignorePrompt.isEmpty ? 1 : tokenize(text: ignorePrompt, add_bos: true, add_eos: false).count
|
|
let n_vocab = llama_n_vocab(model)
|
|
|
|
var sum: Float = 0
|
|
// 最初のプロンプト部分は無視する
|
|
for (i, token_id) in tokens_list.indexed().dropFirst(tokenizedPromptCount) {
|
|
// FIXME: there can be more efficient implementations, poossibly using Accelerate or other frameworks.
|
|
var log_prob: Float = 0
|
|
for index in ((i - 1) * Int(n_vocab)) ..< (i * Int(n_vocab)) {
|
|
log_prob += exp(logits[index])
|
|
}
|
|
log_prob = log(log_prob)
|
|
log_prob = logits[Int((i - 1) * Int(n_vocab) + Int(token_id))] - log_prob
|
|
sum += log_prob
|
|
}
|
|
return sum
|
|
}
|
|
|
|
enum CandidateEvaluationResult: Sendable, Equatable, Hashable {
|
|
case error
|
|
case pass(score: Float, alternativeConstraints: [AlternativeConstraint])
|
|
case fixRequired(prefixConstraint: [UInt8])
|
|
case wholeResult(String)
|
|
|
|
struct AlternativeConstraint: Sendable, Equatable, Hashable {
|
|
var probabilityRatio: Float
|
|
var prefixConstraint: [UInt8]
|
|
}
|
|
}
|
|
|
|
func getLearningPriority(data: DicdataElement) -> Float {
|
|
// 文字数の長い候補ほど優先的に適用されるようにする
|
|
// 積極的な複合語化の効果を期待
|
|
return if 1 <= data.ruby.count && data.ruby.count <= 4 {
|
|
Float(data.ruby.count + 2)
|
|
} else if 5 <= data.ruby.count && data.ruby.count <= 15 {
|
|
Float(data.ruby.count * 2)
|
|
} else {
|
|
30
|
|
}
|
|
}
|
|
|
|
func predict_next_character(leftSideContext: String, count: Int) -> [(character: Character, value: Float)] {
|
|
struct NextCharacterCandidate: Comparable {
|
|
static func < (lhs: NextCharacterCandidate, rhs: NextCharacterCandidate) -> Bool {
|
|
lhs.value < rhs.value
|
|
}
|
|
var character: Character
|
|
var value: Float
|
|
}
|
|
|
|
// 文末を目指して生成するためのプロンプト
|
|
// \u{EE01}を停止トークンとみなせる
|
|
let prompt_tokens = self.tokenize(text: "\u{EE00}。\u{EE02}\(leftSideContext)", add_bos: false)
|
|
let startOffset = prompt_tokens.count - 1
|
|
|
|
guard let logits = self.get_logits(tokens: prompt_tokens, logits_start_index: startOffset) else {
|
|
print("logits unavailable")
|
|
return []
|
|
}
|
|
|
|
let n_vocab = llama_n_vocab(model)
|
|
var exp_sum: Float = 0
|
|
let startIndex = (prompt_tokens.count - 1 - startOffset) * Int(n_vocab)
|
|
let endIndex = (prompt_tokens.count - startOffset) * Int(n_vocab)
|
|
|
|
// Min-Heapを使用してn-bestを計算
|
|
var minHeap: FixedSizeHeap<NextCharacterCandidate> = .init(size: count)
|
|
let token_to_penalty_weight: [llama_token: Float] = prompt_tokens.indexed().reduce(into: [:]) { dict, item in
|
|
let (index, token) = item
|
|
// 現在位置から遠いほど減衰させる
|
|
dict[token, default: 0] += 2 / Float(prompt_tokens.count - index)
|
|
}
|
|
|
|
for index in startIndex..<endIndex {
|
|
let token = llama_token(index - startIndex)
|
|
let repeat_penalty = Float(1.0 + token_to_penalty_weight[token, default: 0])
|
|
let v = exp(logits[index] / repeat_penalty)
|
|
exp_sum += v
|
|
|
|
let tokenPieceData = Data((token_to_piece(token: token)).map(UInt8.init))
|
|
let character: Character
|
|
if let validCharacter = String(data: tokenPieceData, encoding: .utf8), let c = validCharacter.first {
|
|
character = c
|
|
} else {
|
|
continue
|
|
}
|
|
minHeap.insertIfPossible(NextCharacterCandidate(character: character, value: v))
|
|
}
|
|
|
|
// Heapからソートして結果を取り出す
|
|
return minHeap.unordered.sorted { $0.value > $1.value }.map { ($0.character, $0.value / exp_sum) }
|
|
}
|
|
|
|
func evaluate_candidate(input: String, candidate: Candidate, versionDependentConfig: ConvertRequestOptions.ZenzaiVersionDependentMode) -> CandidateEvaluationResult {
|
|
print("Evaluate", candidate)
|
|
// For zenz-v1 model, \u{EE00} is a token used for 'start query', and \u{EE01} is a token used for 'start answer'
|
|
// We assume \u{EE01}\(candidate) is always splitted into \u{EE01}_\(candidate) by zenz-v1 tokenizer
|
|
let prompt: String
|
|
if case .v2(let mode) = versionDependentConfig {
|
|
if let leftSideContext = mode.leftSideContext, !leftSideContext.isEmpty {
|
|
let lsContext = leftSideContext.suffix(40)
|
|
if let profile = mode.profile, !profile.isEmpty {
|
|
let pf = profile.suffix(25)
|
|
prompt = "\u{EE00}\(input)\u{EE02}プロフィール：\(pf)・発言：\(lsContext)\u{EE01}"
|
|
} else {
|
|
prompt = "\u{EE00}\(input)\u{EE02}\(lsContext)\u{EE01}"
|
|
}
|
|
} else if let profile = mode.profile, !profile.isEmpty {
|
|
let pf = profile.suffix(25)
|
|
prompt = "\u{EE00}\(input)\u{EE02}プロフィール：\(pf)・発言：\u{EE01}"
|
|
} else {
|
|
prompt = "\u{EE00}\(input)\u{EE01}"
|
|
}
|
|
} else {
|
|
prompt = "\u{EE00}\(input)\u{EE01}"
|
|
}
|
|
// Therefore, tokens = prompt_tokens + candidate_tokens is an appropriate operation.
|
|
let prompt_tokens = self.tokenize(text: prompt, add_bos: true, add_eos: false)
|
|
let candidate_tokens = self.tokenize(text: candidate.text, add_bos: false, add_eos: false)
|
|
let tokens = prompt_tokens + candidate_tokens
|
|
let startOffset = prompt_tokens.count - 1
|
|
let pos_max = llama_kv_cache_seq_pos_max(self.context, 0)
|
|
print("pos max:", pos_max)
|
|
guard let logits = self.get_logits(tokens: tokens, logits_start_index: startOffset) else {
|
|
debug("logits unavailable")
|
|
return .error
|
|
}
|
|
let n_vocab = llama_n_vocab(model)
|
|
let is_learned_token: [(isLearned: Bool, priority: Float)] = Array(repeating: (false, 0), count: prompt_tokens.count) + candidate.data.flatMap {
|
|
// priorityは文字数にする→文字数が長いほど優先される
|
|
Array(repeating: ($0.metadata.contains(.isLearned), getLearningPriority(data: $0)), count: self.tokenize(text: $0.word, add_bos: false).count)
|
|
}
|
|
|
|
var score: Float = 0
|
|
|
|
struct AlternativeHighProbToken: Comparable {
|
|
static func < (lhs: AlternativeHighProbToken, rhs: AlternativeHighProbToken) -> Bool {
|
|
lhs.probabilityRatioToMaxProb < rhs.probabilityRatioToMaxProb
|
|
}
|
|
|
|
var token: llama_token
|
|
var constraint: [UInt8]
|
|
// 最大probabilityに対しての割合
|
|
var probabilityRatioToMaxProb: Float
|
|
}
|
|
|
|
var altTokens = FixedSizeHeap<AlternativeHighProbToken>(size: 5)
|
|
for (i, token_id) in tokens.indexed().dropFirst(prompt_tokens.count) {
|
|
// それぞれのトークンが、一つ前の予測において最も確率の高いトークンであるかをチェックする
|
|
// softmaxはmaxなので、単にlogitsの中で最も大きいものを選べば良い
|
|
// 一方実用的にはlog_probも得ておきたい。このため、ここでは明示的にsoftmaxも計算している
|
|
struct TokenAndExpLogit: Comparable {
|
|
static func < (lhs: TokenAndExpLogit, rhs: TokenAndExpLogit) -> Bool {
|
|
lhs.expLogit < rhs.expLogit
|
|
}
|
|
|
|
var token: llama_token
|
|
var expLogit: Float
|
|
}
|
|
var exp_sum: Float = 0
|
|
let startIndex = (i - 1 - startOffset) * Int(n_vocab)
|
|
let endIndex = (i - startOffset) * Int(n_vocab)
|
|
var tokenHeap = FixedSizeHeap<TokenAndExpLogit>(size: 3)
|
|
for index in startIndex ..< endIndex {
|
|
let v = exp(logits[index])
|
|
exp_sum += v
|
|
tokenHeap.insertIfPossible(TokenAndExpLogit(token: llama_token(index - startIndex), expLogit: v))
|
|
}
|
|
guard let maxItem = tokenHeap.max else {
|
|
print("Max Item could not be found for unknown reason")
|
|
return .error
|
|
}
|
|
// ここで最も良い候補であったかをチェックする
|
|
if maxItem.token != token_id {
|
|
if maxItem.token == llama_token_eos(model) {
|
|
var cchars = tokens[..<i].reduce(into: []) {
|
|
$0.append(contentsOf: token_to_piece(token: $1))
|
|
}
|
|
// adding "\0"
|
|
cchars.append(0)
|
|
let string = String(cString: cchars)
|
|
// 要求するべき制約を記述する
|
|
let wholeResult = String(string.dropFirst(prompt.count))
|
|
return .wholeResult(wholeResult)
|
|
} else {
|
|
let actual_exp: Float = exp(logits[startIndex + Int(token_id)])
|
|
// 学習されたトークンであり、なおかつactual_expのある程度大きければ、学習されたトークンを優先する
|
|
let preferLearnedToken = is_learned_token[i].isLearned && actual_exp * is_learned_token[i].priority > maxItem.expLogit
|
|
if !preferLearnedToken {
|
|
// adding "\0"
|
|
let cchars = tokens[..<i].reduce(into: []) {
|
|
$0.append(contentsOf: token_to_piece(token: $1))
|
|
} + token_to_piece(token: maxItem.token)
|
|
return .fixRequired(prefixConstraint: cchars.dropFirst(prompt.utf8.count).map(UInt8.init))
|
|
}
|
|
}
|
|
} else {
|
|
tokenHeap.removeMax()
|
|
let prefix = tokens[..<i].reduce(into: []) {
|
|
$0.append(contentsOf: token_to_piece(token: $1))
|
|
}.dropFirst(prompt.utf8.count)
|
|
|
|
for item in tokenHeap.unordered {
|
|
altTokens.insertIfPossible(
|
|
AlternativeHighProbToken(
|
|
token: item.token,
|
|
constraint: prefix.map(UInt8.init) + token_to_piece(token: item.token).map(UInt8.init),
|
|
probabilityRatioToMaxProb: item.expLogit / maxItem.expLogit
|
|
)
|
|
)
|
|
}
|
|
}
|
|
score += log(maxItem.expLogit) - log(exp_sum)
|
|
}
|
|
for item in altTokens.unordered.sorted(by: >) {
|
|
print("Item")
|
|
print(" Constraint", String(cString: item.constraint + [0]))
|
|
print(" Probability Gain", item.probabilityRatioToMaxProb)
|
|
}
|
|
return .pass(score: score, alternativeConstraints: altTokens.unordered.sorted(by: >).map {.init(probabilityRatio: $0.probabilityRatioToMaxProb, prefixConstraint: $0.constraint)})
|
|
}
|
|
|
|
private func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama_pos, _ seq_ids: [llama_seq_id], logits: Bool) {
|
|
batch.token [Int(batch.n_tokens)] = id
|
|
batch.pos [Int(batch.n_tokens)] = pos
|
|
batch.n_seq_id[Int(batch.n_tokens)] = Int32(seq_ids.count)
|
|
for i in 0..<seq_ids.count {
|
|
batch.seq_id[Int(batch.n_tokens)]![Int(i)] = seq_ids[i]
|
|
}
|
|
batch.logits [Int(batch.n_tokens)] = logits ? 1 : 0
|
|
batch.n_tokens += 1
|
|
}
|
|
|
|
private func tokenize(text: String, add_bos: Bool, add_eos: Bool = false) -> [llama_token] {
|
|
let utf8Count = text.utf8.count
|
|
let n_tokens = utf8Count + (add_bos ? 1 : 0)
|
|
let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
|
|
let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
|
|
var swiftTokens: [llama_token] = if tokenCount < 0 {
|
|
[llama_token_bos(model)]
|
|
} else {
|
|
(0..<tokenCount).map {tokens[Int($0)]}
|
|
}
|
|
tokens.deallocate()
|
|
if add_eos {
|
|
swiftTokens.append(llama_token_eos(model))
|
|
}
|
|
return swiftTokens
|
|
}
|
|
|
|
/// - note: The result does not contain null-terminator
|
|
private func token_to_piece(token: llama_token) -> [CChar] {
|
|
let result = UnsafeMutablePointer<Int8>.allocate(capacity: 8)
|
|
result.initialize(repeating: Int8(0), count: 8)
|
|
defer {
|
|
result.deallocate()
|
|
}
|
|
let nTokens = llama_token_to_piece(model, token, result, 8, false)
|
|
|
|
if nTokens < 0 {
|
|
let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
|
|
newResult.initialize(repeating: Int8(0), count: Int(-nTokens))
|
|
defer {
|
|
newResult.deallocate()
|
|
}
|
|
let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, false)
|
|
let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
|
|
return Array(bufferPointer)
|
|
} else {
|
|
let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nTokens))
|
|
return Array(bufferPointer)
|
|
}
|
|
}
|
|
}
|