bytepairencoding.go 6.09 KB
Newer Older
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
package model

import (
	"cmp"
	"fmt"
	"iter"
	"log/slog"
	"slices"
	"strings"

	"github.com/dlclark/regexp2"
	heap "github.com/emirpasic/gods/v2/trees/binaryheap"
	"github.com/ollama/ollama/logutil"
)

type BytePairEncoding struct {
	vocab   *Vocabulary
	regexps []*regexp2.Regexp
}

var _ TextProcessor = (*BytePairEncoding)(nil)

func NewBytePairEncoding(vocab *Vocabulary, pretokenizers ...string) BytePairEncoding {
	if len(pretokenizers) == 0 {
		// set default byte-level pretokenizer if none provided, e.g.
		// https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs#L44
		pretokenizers = []string{`'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+`}
	}

	return BytePairEncoding{
		vocab: vocab,
		regexps: slices.Collect(func(yield func(*regexp2.Regexp) bool) {
			for _, p := range pretokenizers {
				if !yield(regexp2.MustCompile(p, regexp2.RE2)) {
					return
				}
			}
		}),
	}
}

func (bpe BytePairEncoding) Vocabulary() *Vocabulary {
	return bpe.vocab
}

func (bpe BytePairEncoding) Is(id int32, special Special) bool {
	return bpe.vocab.Is(id, special)
}

func (bpe *BytePairEncoding) split(s string) iter.Seq[string] {
	parts := []string{s}
	for _, re := range bpe.regexps {
		parts = slices.Collect(func(yield func(string) bool) {
			for _, part := range parts {
				r := []rune(part)
				var offset int
				for m, _ := re.FindRunesMatch(r); m != nil; m, _ = re.FindNextMatch(m) {
					if offset-m.Index != 0 {
						if !yield(string(r[:m.Index])) {
							return
						}
					}

					if !yield(m.String()) {
						return
					}

					offset = m.Index + m.Length
				}

				if offset < len(r) {
					if !yield(string(r[offset:])) {
						return
					}
				}
			}
		})
	}

	return slices.Values(parts)
}

// fragment is a string fragment and their corresponding token IDs
type fragment struct {
	value string
	ids   []int32
}

// pair is a pair of runes and its rank
type pair struct {
	a, b  int
	rank  int
	value string
}

type merge struct {
	p, n  int
	runes []rune
}

func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
	fragments := []fragment{{value: s}}
	for _, special := range bpe.vocab.SpecialVocabulary() {
		// TODO: process special tokens concurrently
		id := bpe.vocab.Encode(special)
		for i := 0; i < len(fragments); i++ {
			frag := fragments[i]
			if len(frag.ids) > 0 {
				continue
			}

			var middle []fragment
			switch i := strings.Index(frag.value, special); {
			case i < 0:
				middle = append(middle, frag)
			case i > 0:
				middle = append(middle, fragment{value: frag.value[:i]})
				fallthrough
			default:
				middle = append(middle, fragment{value: special, ids: []int32{id}})
				if rest := frag.value[i+len(special):]; rest != "" {
					middle = append(middle, fragment{value: rest})
				}
			}

			fragments = append(fragments[:i], append(middle, fragments[i+1:]...)...)
		}
	}

	var ids []int32
	for _, frag := range fragments {
		if len(frag.ids) > 0 {
			ids = append(ids, frag.ids...)
			continue
		}

		for split := range bpe.split(frag.value) {
			// TODO: process splits concurrently
			var sb strings.Builder
			for _, b := range []byte(split) {
				r := rune(b)
				switch {
				case r == 0x00ad:
					r = 0x0143
				case r <= 0x0020:
					r = r + 0x0100
				case r >= 0x007f && r <= 0x00a0:
					r = r + 0x00a2
				}

				sb.WriteRune(r)
			}

			// short circuit if the fragment is in the vocabulary
			if id := bpe.vocab.Encode(sb.String()); id >= 0 {
				ids = append(ids, id)
				continue
			}

			runes := []rune(sb.String())
			merges := make([]merge, len(runes))
			for r := range runes {
				merges[r] = merge{
					p:     r - 1,
					n:     r + 1,
					runes: []rune{runes[r]},
				}
			}

			pairwise := func(a, b int) *pair {
				if a < 0 || b >= len(runes) {
					return nil
				}

				left, right := string(merges[a].runes), string(merges[b].runes)
				rank := bpe.vocab.Merge(left, right)
				if rank < 0 {
					return nil
				}

				return &pair{
					a:     a,
					b:     b,
					rank:  rank,
					value: left + right,
				}
			}

			pairs := heap.NewWith(func(i, j *pair) int {
				return cmp.Compare(i.rank, j.rank)
			})

			for i := range len(runes) - 1 {
				if pair := pairwise(i, i+1); pair != nil {
					pairs.Push(pair)
				}
			}

			for !pairs.Empty() {
				pair, _ := pairs.Pop()

				left, right := merges[pair.a], merges[pair.b]
				if len(left.runes) == 0 || len(right.runes) == 0 ||
					string(left.runes)+string(right.runes) != pair.value {
					continue
				}

				if id := bpe.vocab.Encode(pair.value); id < 0 {
					continue
				}

				merges[pair.a].runes = append(left.runes, right.runes...)
				merges[pair.b].runes = nil

				merges[pair.a].n = right.n
				if right.n < len(merges) {
					merges[right.n].p = pair.a
				}

				if pair := pairwise(merges[pair.a].p, pair.a); pair != nil {
					pairs.Push(pair)
				}

				if pair := pairwise(pair.a, merges[pair.a].n); pair != nil {
					pairs.Push(pair)
				}
			}

			for _, merge := range merges {
				if len(merge.runes) > 0 {
					// TODO: handle the edge case where the rune isn't in the vocabulary
					if id := bpe.vocab.Encode(string(merge.runes)); id >= 0 {
						ids = append(ids, id)
					}
				}
			}
		}
	}

	if addSpecial {
		ids = bpe.vocab.addSpecials(ids)
	}

	logutil.Trace("encoded", "string", s, "ids", ids)
	return ids, nil
}

type lazyIdsString struct {
	ids []int32
}

func (l lazyIdsString) LogValue() slog.Value {
	return slog.AnyValue(fmt.Sprint(l.ids))
}

func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
	var sb strings.Builder
	for _, id := range ids {
		for _, r := range bpe.vocab.Decode(id) {
			switch {
			case r == 0x0100:
				// this produces 0x00 aka NULL
				continue
			case r == 0x0143:
				r = 0x00ad
			case r > 0x0100 && r <= 0x0120:
				r = r - 0x0100
			case r > 0x0120 && r <= 0x0142:
				r = r - 0x00a2
			}

			// NOTE: not using WriteRune here because it writes the UTF-8
			// encoding of the rune which is _not_ what we want
			if err := sb.WriteByte(byte(r)); err != nil {
				return "", err
			}
		}
	}

	logutil.Trace("decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
	return sb.String(), nil
}