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
|
package llmc
import (
"fmt"
"os"
"path/filepath"
"runtime"
"sync"
"go.jknobloc.com/x/dataset"
"go.jknobloc.com/x/llm"
)
func TokenizeAll(reader dataset.Reader, tok llm.Tokenizer, eot int) <-chan []uint32 {
out := make(chan []uint32, 256)
go func() {
defer close(out)
sem := make(chan struct{}, max(1, runtime.NumCPU()-1))
var wg sync.WaitGroup
for _, text := range reader.Texts() {
sem <- struct{}{}
wg.Add(1)
go func(t string) {
defer func() { <-sem; wg.Done() }()
tokens := tok.Tokenize(t)
tokensU32 := make([]uint32, len(tokens)+1)
tokensU32[0] = uint32(eot)
for i, id := range tokens {
tokensU32[i+1] = uint32(id)
}
out <- tokensU32
}(text)
}
wg.Wait()
}()
return out
}
func WriteShards(name, data string, shardSize int, docs <-chan []uint32) error {
if err := os.MkdirAll(name, os.ModePerm); err != nil {
return err
}
buffer := make([]uint32, 0, shardSize)
shardIdx := 0
flush := func() error {
split := "train"
if shardIdx == 0 {
split = "val"
}
path := filepath.Join(name, fmt.Sprintf("%s_%s_%06d.bin", data, split, shardIdx))
d := DataFile[uint32]{
Model: GPT2,
Tokens: buffer,
}
if n, err := Serialize(&d, path); err != nil {
return err
} else {
fmt.Printf("wrote %s (%d tokens)\n", filepath.Base(path), n)
}
buffer = buffer[:0]
shardIdx++
return nil
}
for tokens := range docs {
for len(tokens) > 0 {
space := shardSize - len(buffer)
if space >= len(tokens) {
buffer = append(buffer, tokens...)
break
}
buffer = append(buffer, tokens[:space]...)
tokens = tokens[space:] // carry remainder
if err := flush(); err != nil {
return err
}
}
}
if len(buffer) > 0 {
return flush()
}
return nil
}
|
