forked from MartialBE/one-hub
-
Notifications
You must be signed in to change notification settings - Fork 0
/
token.go
257 lines (234 loc) · 7.61 KB
/
token.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
package common
import (
"errors"
"fmt"
"math"
"one-api/common/config"
"one-api/common/logger"
"strings"
"one-api/common/image"
"one-api/types"
"github.com/MartialBE/tiktoken-go"
"github.com/spf13/viper"
)
var tokenEncoderMap = map[string]*tiktoken.Tiktoken{}
var gpt35TokenEncoder *tiktoken.Tiktoken
var gpt4TokenEncoder *tiktoken.Tiktoken
var gpt4oTokenEncoder *tiktoken.Tiktoken
func InitTokenEncoders() {
if viper.GetBool("disable_token_encoders") {
config.DisableTokenEncoders = true
logger.SysLog("token encoders disabled")
return
}
logger.SysLog("initializing token encoders")
var err error
gpt35TokenEncoder, err = tiktoken.EncodingForModel("gpt-3.5-turbo")
if err != nil {
logger.FatalLog(fmt.Sprintf("failed to get gpt-3.5-turbo token encoder: %s", err.Error()))
}
gpt4TokenEncoder, err = tiktoken.EncodingForModel("gpt-4")
if err != nil {
logger.FatalLog(fmt.Sprintf("failed to get gpt-4 token encoder: %s", err.Error()))
}
gpt4oTokenEncoder, err = tiktoken.EncodingForModel("gpt-4o")
if err != nil {
logger.FatalLog(fmt.Sprintf("failed to get gpt-4o token encoder: %s", err.Error()))
}
logger.SysLog("token encoders initialized")
}
func getTokenEncoder(model string) *tiktoken.Tiktoken {
if config.DisableTokenEncoders {
return nil
}
tokenEncoder, ok := tokenEncoderMap[model]
if ok {
return tokenEncoder
}
if strings.HasPrefix(model, "gpt-3.5") {
tokenEncoder = gpt35TokenEncoder
} else if strings.HasPrefix(model, "gpt-4o") {
tokenEncoder = gpt4oTokenEncoder
} else if strings.HasPrefix(model, "gpt-4") {
tokenEncoder = gpt4TokenEncoder
} else {
var err error
tokenEncoder, err = tiktoken.EncodingForModel(model)
if err != nil {
logger.SysError(fmt.Sprintf("failed to get token encoder for model %s: %s, using encoder for gpt-3.5-turbo", model, err.Error()))
tokenEncoder = gpt35TokenEncoder
}
}
tokenEncoderMap[model] = tokenEncoder
return tokenEncoder
}
func getTokenNum(tokenEncoder *tiktoken.Tiktoken, text string) int {
if config.DisableTokenEncoders || config.ApproximateTokenEnabled {
return int(float64(len(text)) * 0.38)
}
return len(tokenEncoder.Encode(text, nil, nil))
}
func CountTokenMessages(messages []types.ChatCompletionMessage, model string) int {
tokenEncoder := getTokenEncoder(model)
// Reference:
// https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
// https://github.com/pkoukk/tiktoken-go/issues/6
//
// Every message follows <|start|>{role/name}\n{content}<|end|>\n
var tokensPerMessage int
var tokensPerName int
if model == "gpt-3.5-turbo-0301" {
tokensPerMessage = 4
tokensPerName = -1 // If there's a name, the role is omitted
} else {
tokensPerMessage = 3
tokensPerName = 1
}
tokenNum := 0
for _, message := range messages {
tokenNum += tokensPerMessage
switch v := message.Content.(type) {
case string:
tokenNum += getTokenNum(tokenEncoder, v)
case []any:
for _, it := range v {
m := it.(map[string]any)
switch m["type"] {
case "text":
tokenNum += getTokenNum(tokenEncoder, m["text"].(string))
case "image_url":
imageUrl, ok := m["image_url"].(map[string]any)
if ok {
url := imageUrl["url"].(string)
detail := ""
if imageUrl["detail"] != nil {
detail = imageUrl["detail"].(string)
}
imageTokens, err := countImageTokens(url, detail)
if err != nil {
//Due to the excessive length of the error information, only extract and record the most critical part.
logger.SysError("error counting image tokens: " + err.Error())
} else {
tokenNum += imageTokens
}
}
}
}
}
tokenNum += getTokenNum(tokenEncoder, message.Role)
if message.Name != nil {
tokenNum += tokensPerName
tokenNum += getTokenNum(tokenEncoder, *message.Name)
}
}
tokenNum += 3 // Every reply is primed with <|start|>assistant<|message|>
return tokenNum
}
const (
lowDetailCost = 85
highDetailCostPerTile = 170
additionalCost = 85
)
// https://platform.openai.com/docs/guides/vision/calculating-costs
// https://github.com/openai/openai-cookbook/blob/05e3f9be4c7a2ae7ecf029a7c32065b024730ebe/examples/How_to_count_tokens_with_tiktoken.ipynb
func countImageTokens(url string, detail string) (_ int, err error) {
// var fetchSize = true
var width, height int
// Reference: https://platform.openai.com/docs/guides/vision/low-or-high-fidelity-image-understanding
// detail == "auto" is undocumented on how it works, it just said the model will use the auto setting which will look at the image input size and decide if it should use the low or high setting.
// According to the official guide, "low" disable the high-res model,
// and only receive low-res 512px x 512px version of the image, indicating
// that image is treated as low-res when size is smaller than 512px x 512px,
// then we can assume that image size larger than 512px x 512px is treated
// as high-res. Then we have the following logic:
// if detail == "" || detail == "auto" {
// width, height, err = image.GetImageSize(url)
// if err != nil {
// return 0, err
// }
// fetchSize = false
// // not sure if this is correct
// if width > 512 || height > 512 {
// detail = "high"
// } else {
// detail = "low"
// }
// }
// However, in my test, it seems to be always the same as "high".
// The following image, which is 125x50, is still treated as high-res, taken
// 255 tokens in the response of non-stream chat completion api.
// https://upload.wikimedia.org/wikipedia/commons/1/10/18_Infantry_Division_Messina.jpg
if detail == "" || detail == "auto" {
// assume by test, not sure if this is correct
detail = "high"
}
switch detail {
case "low":
return lowDetailCost, nil
case "high":
width, height, err = image.GetImageSize(url)
if err != nil {
return 0, err
}
if width > 2048 || height > 2048 { // max(width, height) > 2048
ratio := float64(2048) / math.Max(float64(width), float64(height))
width = int(float64(width) * ratio)
height = int(float64(height) * ratio)
}
if width > 768 && height > 768 { // min(width, height) > 768
ratio := float64(768) / math.Min(float64(width), float64(height))
width = int(float64(width) * ratio)
height = int(float64(height) * ratio)
}
numSquares := int(math.Ceil(float64(width)/512) * math.Ceil(float64(height)/512))
result := numSquares*highDetailCostPerTile + additionalCost
return result, nil
default:
return 0, errors.New("invalid detail option")
}
}
func CountTokenInput(input any, model string) int {
switch v := input.(type) {
case string:
return CountTokenText(v, model)
case []string:
text := ""
for _, s := range v {
text += s
}
return CountTokenText(text, model)
}
return 0
}
func CountTokenText(text string, model string) int {
tokenEncoder := getTokenEncoder(model)
return getTokenNum(tokenEncoder, text)
}
func CountTokenImage(input interface{}) (int, error) {
switch v := input.(type) {
case types.ImageRequest:
// 处理 ImageRequest
return calculateToken(v.Model, v.Size, v.N, v.Quality)
case types.ImageEditRequest:
// 处理 ImageEditsRequest
return calculateToken(v.Model, v.Size, v.N, "")
default:
return 0, errors.New("unsupported type")
}
}
func calculateToken(model string, size string, n int, quality string) (int, error) {
imageCostRatio, hasValidSize := DalleSizeRatios[model][size]
if hasValidSize {
if quality == "hd" && model == "dall-e-3" {
if size == "1024x1024" {
imageCostRatio *= 2
} else {
imageCostRatio *= 1.5
}
}
} else {
imageCostRatio = 1
// return 0, errors.New("size not supported for this image model")
}
return int(imageCostRatio*1000) * n, nil
}