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
|
package main
import (
"bufio"
"database/sql"
"fmt"
"io"
"log"
"os"
_ "github.com/duckdb/duckdb-go/v2"
"github.com/jonasknobloch/mbpe"
"go.jknobloc.com/x/dict"
"go.jknobloc.com/x/research/entropy"
"go.jknobloc.com/x/tokenizer/bpe"
"go.jknobloc.com/x/shelf"
)
type NoPreTok struct{}
func (p *NoPreTok) PreTokenize(s string) []string {
return []string{s}
}
func main() {
o := shelf.Abs("results/runs-entropy/lesci_minipile_test_256_seed_default/gpt2_256_m000_minipile_seed/lesci.db")
var db *sql.DB
if database, err := initDatabase(o); err != nil {
log.Fatal(err)
} else {
db = database
}
defer db.Close()
db.SetMaxOpenConns(1)
var file *os.File
if f, err := os.Open(shelf.Abs("results/knobloch/minipile/dict.txt")); err != nil {
log.Fatal(err)
} else {
file = f
}
tok := tokenizerMBPE()
var dict *dict.Dict[*entropy.Entry]
if d, err := loadDict(file, tok); err != nil {
log.Fatal(err)
} else {
dict = d
}
if err := entropy.Run(db, dict); err != nil {
log.Fatal(err)
}
}
func initDatabase(dsn string) (*sql.DB, error) {
var db *sql.DB
if database, err := sql.Open("duckdb", dsn); err != nil {
return nil, err
} else {
db = database
}
if err := db.Ping(); err != nil {
_ = db.Close()
return nil, err
}
return db, nil
}
func loadDict(r io.Reader, tok *mbpe.Tokenizer) (*dict.Dict[*entropy.Entry], error) {
d := dict.NewDict[*entropy.Entry]()
// TODO we could count non empty lines to pre alloc dict
scanner := bufio.NewScanner(r)
for scanner.Scan() {
line := scanner.Text()
if err := scanner.Err(); err != nil {
return nil, err
}
var s string
var n int
if _, err := fmt.Sscanf(line, "%s %d", &s, &n); err != nil {
return nil, err
}
// TODO see cmd/main.go we currently need mbpe tok ??
if n < 100 {
continue
}
ids := tok.Tokenize(s)
if len(ids) < 4 {
continue
}
b := to[int, uint8](ids) // TODO that conversion is fucking ass
// TODO vs decode each buffer suffix and use s as map key
d.Set(string(b), &entropy.Entry{
Encoded: s, // TODO encoded or decoded
TokenIDs: b,
LogProbs: make([]float32, len(ids)),
N: 0,
})
}
return d, nil
}
func to[a bpe.Integer, b bpe.Integer](s []a) []b {
r := make([]b, len(s))
// TODO handle overflow
for i, v := range s {
r[i] = b(v)
}
return r
}
func tokenizerMBPE() *mbpe.Tokenizer {
var tok *bpe.Tokenizer
v := shelf.Abs("tokenizers/minipile/tokenizer_gpt2_256_m000_minipile/vocab.json")
m := shelf.Abs("tokenizers/minipile/tokenizer_gpt2_256_m000_minipile/merges.txt")
if t, err := bpe.NewTokenizerFromFiles(v, m, bpe.DefaultConfig()); err != nil {
log.Fatal(err)
} else {
tok = t
}
foo := bpe.MBPE(tok)
foo.SetPreTokenizer(&NoPreTok{})
return foo
}
|
