-
Notifications
You must be signed in to change notification settings - Fork 2
/
processor.go
479 lines (416 loc) · 14.5 KB
/
processor.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
package sentencepiece
import (
"fmt"
"io"
"os"
"strconv"
"strings"
"unicode/utf8"
"github.com/eliben/go-sentencepiece/internal/model"
"github.com/eliben/go-sentencepiece/internal/prefixmatcher"
"github.com/eliben/go-sentencepiece/internal/priorityqueue"
"google.golang.org/protobuf/proto"
)
const debugEncode = false
// Processor represents a SentencePiece processor (tokenizer).
// A Processor converts input text into a sequence of tokens LLMs use, and back.
// The mapping between token IDs and the text they represent is read from the
// model proto (provided to the constructor); it's the same between all calls
// to the Encode method.
//
// The term "processor" comes from the original C++ SentencePiece library and
// its Python bindings.
type Processor struct {
model *model.ModelProto
pieces map[string]int
reserved map[string]int
// unknownID is the token identifier of the UNKNOWN piece
unknownID int
// userDefinedMatcher is a prefix matcher for symbols that are of
// "user-defined" type in the model proto.
userDefinedMatcher *prefixmatcher.PrefixMatcher
// byte2Token is a cache of byte values and the tokens they represent
byte2Token map[byte]Token
// idToByte maps IDs to byte values they represent
idToByte map[int]byte
// maxPieceLength is the maximum length of a piece in the model.
// This is used to preallocate a buffer for merging symbols.
maxPieceLength int
}
// NewProcessorFromPath creates a new Processor from a file path to the protobuf
// data.
func NewProcessorFromPath(protoFile string) (*Processor, error) {
f, err := os.Open(protoFile)
if err != nil {
return nil, fmt.Errorf("unable to read %q: %v", protoFile, err)
}
defer f.Close()
return NewProcessor(f)
}
// NewProcessor creates a new Processor from a reader with the protobuf data.
func NewProcessor(protoReader io.Reader) (*Processor, error) {
b, err := io.ReadAll(protoReader)
if err != nil {
return nil, fmt.Errorf("unable to read protobuf data: %v", err)
}
var mp model.ModelProto
err = proto.Unmarshal(b, &mp)
if err != nil {
return nil, fmt.Errorf("unable to unmarshal protobuf: %v", err)
}
tspec := mp.GetTrainerSpec()
if tspec.GetModelType() != model.TrainerSpec_BPE {
return nil, fmt.Errorf("model type %s not supported", tspec.GetModelType())
}
nspec := mp.GetNormalizerSpec()
if *nspec.AddDummyPrefix || *nspec.RemoveExtraWhitespaces {
return nil, fmt.Errorf("normalizer spec options not supported: %s", nspec)
}
userDefined := make(map[string]bool)
pieces := make(map[string]int)
reserved := make(map[string]int)
byte2Token := make(map[byte]Token)
idToByte := make(map[int]byte)
unkID := -1
maxPieceLength := 0
for i, piece := range mp.GetPieces() {
isNormalPiece := (piece.GetType() == model.ModelProto_SentencePiece_NORMAL ||
piece.GetType() == model.ModelProto_SentencePiece_USER_DEFINED ||
piece.GetType() == model.ModelProto_SentencePiece_UNUSED)
if isNormalPiece {
pieces[piece.GetPiece()] = i
maxPieceLength = max(maxPieceLength, len(piece.GetPiece()))
} else {
reserved[piece.GetPiece()] = i
}
if piece.GetType() == model.ModelProto_SentencePiece_USER_DEFINED {
userDefined[piece.GetPiece()] = true
} else if piece.GetType() == model.ModelProto_SentencePiece_UNKNOWN {
if unkID > 0 {
return nil, fmt.Errorf("unk redefined")
}
unkID = i
} else if piece.GetType() == model.ModelProto_SentencePiece_BYTE {
if !tspec.GetByteFallback() {
return nil, fmt.Errorf("byte piece %q is found although `byte_fallback=false`", piece.GetPiece())
}
bv := convertHexValue(piece.GetPiece())
if bv >= 0 && bv < 256 {
byte2Token[byte(bv)] = Token{ID: i, Text: piece.GetPiece()}
idToByte[i] = byte(bv)
}
}
}
if unkID < 0 {
return nil, fmt.Errorf("unk symbol is not defined")
}
// In case byte_fallback is specified, make sure that all 256 possible byte
// values were found.
if tspec.GetByteFallback() {
for i := 0; i < 256; i++ {
if _, found := byte2Token[byte(i)]; !found {
return nil, fmt.Errorf("byte value 0x%02X not found", i)
}
}
}
return &Processor{
model: &mp,
userDefinedMatcher: prefixmatcher.NewFromSet(userDefined),
byte2Token: byte2Token,
idToByte: idToByte,
unknownID: unkID,
pieces: pieces,
reserved: reserved,
maxPieceLength: maxPieceLength,
}, nil
}
// Encode tokenizes the input text and returns a list of Tokens.
func (proc *Processor) Encode(text string) []Token {
text = normalize(text)
// We begin by having each symbol a single Unicode character (or a
// user-defined string), and will iteratively merge them into larger and
// larger symbols until we have the final list of tokens.
// Since this list of symbols changes a lot, we represent it as a
// doubly-linked list in the symList slice. Each element in this slice has
// prev/next links to the next "live" symbol in the list; noMerge means this
// is a user-defined symbol we're not allowed to merge with neighbors.
// After the algorithm is finished, many elements in symList will be "dead"
// (unreachable by next/prev links from the first element).
// This representation is inspired by the implementation of bpe::Model
// in the SentencePiece C++ library.
type symListElem struct {
prev, next int
noMerge bool
symbol string
}
symList := make([]symListElem, 0, len(text))
for {
// Match the next symbol in text
slen, found := proc.symbolMatch(text)
// Append a list element for this symbol; note that this element will be
// at index len(symList), so prev/next are set up accordingly.
sym := symListElem{
noMerge: found,
symbol: text[:slen],
prev: len(symList) - 1,
next: len(symList) + 1,
}
symList = append(symList, sym)
// Advance the text slice to the next symbol; if no more text, we're done.
text = text[slen:]
if len(text) == 0 {
break
}
}
if len(symList) == 0 {
return nil
}
symList[len(symList)-1].next = -1
nTokens := len(symList)
debugShowSymList := func(prefix string) {
if debugEncode {
fmt.Println(prefix)
for i, elem := range symList {
fmt.Printf("[%3d]: [prev: %3v, next: %3d, noMerge: %v] %q\n", i, elem.prev, elem.next, elem.noMerge, elem.symbol)
}
}
}
debugShowSymList("initial")
// To avoid repeating work, we manage a priority queue of "merge candidates".
// Each candidate has pointers to the symList list for the left and right
// symbol in the pair, as well as the combined symbol's score.
// The priority of merging is determined by this score, with position as
// the tie-breaker (earlier pairs are preferred).
type mergeCandidate struct {
left, right int
length int
score float32
}
mergeQueue := priorityqueue.New(len(symList), func(a, b mergeCandidate) int {
if a.score > b.score || (a.score == b.score && a.left < b.left) {
return 1
}
return -1
})
// findMerged looks for x+y in the vocabulary, and returns the
// merged piece, its ID and true if found. buf is a reusable buffer used to
// merge two strings together without allocations.
buf := make([]byte, proc.maxPieceLength)
findMerged := func(x, y symListElem) (string, int, bool) {
buf = buf[:len(x.symbol)+len(y.symbol)]
copy(buf, x.symbol)
copy(buf[len(x.symbol):], y.symbol)
if id, found := proc.pieces[string(buf)]; found {
return proc.model.GetPieces()[id].GetPiece(), id, true
}
return "", 0, false
}
// suggestNewMergePair is called to potentially add a new mergeCandidate to
// mergeQueue. The candidate is added if it's valid, both its parts are
// allowed to merge, and it appears in the vocabulary.
suggestNewMergePair := func(left, right int) {
if left == -1 || right == -1 || symList[left].noMerge || symList[right].noMerge {
return
}
if mergedSymbol, id, ok := findMerged(symList[left], symList[right]); ok {
mergeQueue.Insert(mergeCandidate{
left: left,
right: right,
length: len(mergedSymbol),
score: proc.model.GetPieces()[id].GetScore(),
})
}
}
// Seed the merge queue with all pairs of symbols from symList
for i := 1; i < len(symList); i++ {
suggestNewMergePair(i-1, i)
}
// candidateIsDead indicates that a candidate is out of date: one of its
// parts was already merged with another symbol, so we don't want to consider
// it any more.
candidateIsDead := func(candidate mergeCandidate) bool {
leftSymbol := symList[candidate.left].symbol
rightSymbol := symList[candidate.right].symbol
return leftSymbol == "" || rightSymbol == "" || len(leftSymbol)+len(rightSymbol) != candidate.length
}
// Main loop
mergeQueueDead := 0
for mergeQueue.Len() > 0 {
candidate := mergeQueue.PopMax()
leftSymbol := symList[candidate.left]
rightSymbol := symList[candidate.right]
if candidateIsDead(candidate) {
mergeQueueDead--
continue
}
// If there are lots more dead merge candidates than live ones, remove the
// dead. This is a relatively expensive operation but it's performed rarely,
// and it makes the priority queue smaller - making all subsequent
// operations faster.
// The factor of 3 was determined empirically.
if mergeQueueDead*3 > mergeQueue.Len() {
mergeQueue.RemoveFunc(candidateIsDead)
mergeQueueDead = 0
}
// Do the merge:
// 1. Merge the concatenation of leftSymbol and rightSymbol into leftSymbol
mergedSymbol, _, ok := findMerged(leftSymbol, rightSymbol)
if !ok {
panic("failed to merge symbols")
}
symList[candidate.left].symbol = mergedSymbol
nTokens--
// 2. Update prev/next pointers
symList[candidate.left].next = rightSymbol.next
if rightSymbol.next >= 0 {
symList[rightSymbol.next].prev = candidate.left
}
// 3. Mark the right element in the pair as outdated (it's been merged
// into the left one).
symList[candidate.right].symbol = ""
mergeQueueDead++
// 4. Add merge suggestions for the newly merged symbol with its neighbors
suggestNewMergePair(leftSymbol.prev, candidate.left)
suggestNewMergePair(candidate.left, rightSymbol.next)
}
// Collect the final list of tokens from the remaining elements of symList.
tokens := make([]Token, 0, nTokens)
for i := 0; i >= 0; i = symList[i].next {
symbol := symList[i].symbol
id := proc.symbolToID(symbol)
if id == proc.unknownID && proc.model.GetTrainerSpec().GetByteFallback() {
// Decompose this symbol into bytes, and report each byte as a separate
// token.
for i := 0; i < len(symbol); i++ {
tokens = append(tokens, proc.byte2Token[symbol[i]])
}
} else {
tokens = append(tokens, Token{ID: id, Text: symbol})
}
}
return tokens
}
// symbolMatch finds the length of the first symbol in text. A symbol is either
// a user-defined symbol from the proto or a single rune. The second return
// value is true iff a user-defined symbol was matched.
func (proc *Processor) symbolMatch(text string) (int, bool) {
prefixLen := proc.userDefinedMatcher.FindPrefixLen(text)
if prefixLen > 0 {
return prefixLen, true
}
// Not found a user-defined prefix; get the length of next rune.
_, rlen := utf8.DecodeRuneInString(text)
return rlen, false
}
const (
symbolBOS = "<bos>"
symbolEOS = "<eos>"
symbolUNK = "<unk>"
symbolPAD = "<pad>"
)
// symbolToID finds the right ID for the given textual symbol, or returns
// proc.unknownID if the symbol is unknown.
func (proc *Processor) symbolToID(symbol string) int {
if id, found := proc.reserved[symbol]; found {
return id
}
if id, found := proc.pieces[symbol]; found {
return id
}
return proc.unknownID
}
// convertHexValue converts strings of the form "<0xXY>" to the (unsigned)
// integer value of the hexadecimal number XY. -1 is returned for bad input.
func convertHexValue(bv string) int {
bv = strings.TrimPrefix(bv, "<0x")
bv = strings.TrimSuffix(bv, ">")
n, err := strconv.ParseInt(bv, 16, 32)
if err != nil {
return -1
}
return int(n)
}
// Decode translates a list of IDs produced by [Encode] back into the string
// it represents.
func (proc *Processor) Decode(ids []int) string {
var sb strings.Builder
for i := 0; i < len(ids); {
// Find a run of IDs that represent single bytes starting at i.
nextNonByte := i
for nextNonByte < len(ids) && proc.isByteID(ids[nextNonByte]) {
nextNonByte++
}
numBytes := nextNonByte - i
// Handle a run of numBytes IDs, by decoding them into utf8 runes.
if numBytes > 0 {
buf := make([]byte, 0, numBytes)
for bi := i; bi < nextNonByte; bi++ {
buf = append(buf, proc.idToByte[ids[bi]])
}
for len(buf) > 0 {
// DecodeRune returns utf8.RuneError ('\uFFFD') for bad UTF8 encodings,
// and this is exactly what SentencePiece is supposed to emit for them.
// So we don't do any special handling for UTF8 decode errors here.
r, size := utf8.DecodeRune(buf)
sb.WriteRune(r)
buf = buf[size:]
}
}
if nextNonByte >= len(ids) {
break
}
// Here nextNonByte is the index of an ID that's not a single byte.
id := ids[nextNonByte]
if proc.isControlID(id) {
// Don't emit anything for control IDs
} else if id == proc.unknownID {
// Special "unk_surface" string for unknown IDs
sb.WriteString(proc.model.GetTrainerSpec().GetUnkSurface())
} else {
piece := proc.model.GetPieces()[id].GetPiece()
sb.WriteString(replaceSeparatorsBySpace(piece))
}
i = nextNonByte + 1
}
return sb.String()
}
// DecodeTokens is a convenience wrapper around [Decode], accepting a list of
// tokens as returned by [Encode]. It only uses the ID fields of tokens to
// decode the text.
func (proc *Processor) DecodeTokens(tokens []Token) string {
ids := make([]int, len(tokens))
for i, t := range tokens {
ids[i] = t.ID
}
return proc.Decode(ids)
}
func (proc *Processor) isByteID(id int) bool {
return proc.model.GetPieces()[id].GetType() == model.ModelProto_SentencePiece_BYTE
}
func (proc *Processor) isControlID(id int) bool {
return proc.model.GetPieces()[id].GetType() == model.ModelProto_SentencePiece_CONTROL
}
// ModelInfo stores information about the model proto loaded by the processor.
type ModelInfo struct {
VocabularySize int
BeginningOfSentenceID int
EndOfSentenceID int
UnknownID int
PadID int
}
// ModelInfo returns information about the loaded proto model file.
func (proc *Processor) ModelInfo() *ModelInfo {
getControlID := func(symbol string) int {
if id := proc.symbolToID(symbol); proc.isControlID(id) {
return id
}
return -1
}
return &ModelInfo{
VocabularySize: len(proc.model.GetPieces()),
BeginningOfSentenceID: getControlID(symbolBOS),
EndOfSentenceID: getControlID(symbolEOS),
PadID: getControlID(symbolPAD),
UnknownID: proc.unknownID,
}
}