179 lines
4.8 KiB
Go
179 lines
4.8 KiB
Go
package controller
|
|
|
|
import (
|
|
"encoding/json"
|
|
"fmt"
|
|
"github.com/gin-gonic/gin"
|
|
"github.com/pkoukk/tiktoken-go"
|
|
"io"
|
|
"net/http"
|
|
"one-api/common"
|
|
"strconv"
|
|
"strings"
|
|
)
|
|
|
|
var stopFinishReason = "stop"
|
|
|
|
// tokenEncoderMap won't grow after initialization
|
|
var tokenEncoderMap = map[string]*tiktoken.Tiktoken{}
|
|
var defaultTokenEncoder *tiktoken.Tiktoken
|
|
|
|
func InitTokenEncoders() {
|
|
common.SysLog("initializing token encoders")
|
|
gpt35TokenEncoder, err := tiktoken.EncodingForModel("gpt-3.5-turbo")
|
|
if err != nil {
|
|
common.FatalLog(fmt.Sprintf("failed to get gpt-3.5-turbo token encoder: %s", err.Error()))
|
|
}
|
|
defaultTokenEncoder = gpt35TokenEncoder
|
|
gpt4TokenEncoder, err := tiktoken.EncodingForModel("gpt-4")
|
|
if err != nil {
|
|
common.FatalLog(fmt.Sprintf("failed to get gpt-4 token encoder: %s", err.Error()))
|
|
}
|
|
for model, _ := range common.ModelRatio {
|
|
if strings.HasPrefix(model, "gpt-3.5") {
|
|
tokenEncoderMap[model] = gpt35TokenEncoder
|
|
} else if strings.HasPrefix(model, "gpt-4") {
|
|
tokenEncoderMap[model] = gpt4TokenEncoder
|
|
} else {
|
|
tokenEncoderMap[model] = nil
|
|
}
|
|
}
|
|
common.SysLog("token encoders initialized")
|
|
}
|
|
|
|
func getTokenEncoder(model string) *tiktoken.Tiktoken {
|
|
tokenEncoder, ok := tokenEncoderMap[model]
|
|
if ok && tokenEncoder != nil {
|
|
return tokenEncoder
|
|
}
|
|
if ok {
|
|
tokenEncoder, err := tiktoken.EncodingForModel(model)
|
|
if err != nil {
|
|
common.SysError(fmt.Sprintf("failed to get token encoder for model %s: %s, using encoder for gpt-3.5-turbo", model, err.Error()))
|
|
tokenEncoder = defaultTokenEncoder
|
|
}
|
|
tokenEncoderMap[model] = tokenEncoder
|
|
return tokenEncoder
|
|
}
|
|
return defaultTokenEncoder
|
|
}
|
|
|
|
func getTokenNum(tokenEncoder *tiktoken.Tiktoken, text string) int {
|
|
if common.ApproximateTokenEnabled {
|
|
return int(float64(len(text)) * 0.38)
|
|
}
|
|
return len(tokenEncoder.Encode(text, nil, nil))
|
|
}
|
|
|
|
func countTokenMessages(messages []Message, 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
|
|
tokenNum += getTokenNum(tokenEncoder, message.Content)
|
|
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
|
|
}
|
|
|
|
func countTokenInput(input any, model string) int {
|
|
switch input.(type) {
|
|
case string:
|
|
return countTokenText(input.(string), model)
|
|
case []string:
|
|
text := ""
|
|
for _, s := range input.([]string) {
|
|
text += s
|
|
}
|
|
return countTokenText(text, model)
|
|
}
|
|
return 0
|
|
}
|
|
|
|
func countTokenText(text string, model string) int {
|
|
tokenEncoder := getTokenEncoder(model)
|
|
return getTokenNum(tokenEncoder, text)
|
|
}
|
|
|
|
func errorWrapper(err error, code string, statusCode int) *OpenAIErrorWithStatusCode {
|
|
openAIError := OpenAIError{
|
|
Message: err.Error(),
|
|
Type: "one_api_error",
|
|
Code: code,
|
|
}
|
|
return &OpenAIErrorWithStatusCode{
|
|
OpenAIError: openAIError,
|
|
StatusCode: statusCode,
|
|
}
|
|
}
|
|
|
|
func shouldDisableChannel(err *OpenAIError, statusCode int) bool {
|
|
if !common.AutomaticDisableChannelEnabled {
|
|
return false
|
|
}
|
|
if err == nil {
|
|
return false
|
|
}
|
|
if statusCode == http.StatusUnauthorized {
|
|
return true
|
|
}
|
|
if err.Type == "insufficient_quota" || err.Code == "invalid_api_key" || err.Code == "account_deactivated" {
|
|
return true
|
|
}
|
|
return false
|
|
}
|
|
|
|
func setEventStreamHeaders(c *gin.Context) {
|
|
c.Writer.Header().Set("Content-Type", "text/event-stream")
|
|
c.Writer.Header().Set("Cache-Control", "no-cache")
|
|
c.Writer.Header().Set("Connection", "keep-alive")
|
|
c.Writer.Header().Set("Transfer-Encoding", "chunked")
|
|
c.Writer.Header().Set("X-Accel-Buffering", "no")
|
|
}
|
|
|
|
func relayErrorHandler(resp *http.Response) (openAIErrorWithStatusCode *OpenAIErrorWithStatusCode) {
|
|
openAIErrorWithStatusCode = &OpenAIErrorWithStatusCode{
|
|
StatusCode: resp.StatusCode,
|
|
OpenAIError: OpenAIError{
|
|
Message: fmt.Sprintf("bad response status code %d", resp.StatusCode),
|
|
Type: "upstream_error",
|
|
Code: "bad_response_status_code",
|
|
Param: strconv.Itoa(resp.StatusCode),
|
|
},
|
|
}
|
|
responseBody, err := io.ReadAll(resp.Body)
|
|
if err != nil {
|
|
return
|
|
}
|
|
err = resp.Body.Close()
|
|
if err != nil {
|
|
return
|
|
}
|
|
var textResponse TextResponse
|
|
err = json.Unmarshal(responseBody, &textResponse)
|
|
if err != nil {
|
|
return
|
|
}
|
|
openAIErrorWithStatusCode.OpenAIError = textResponse.Error
|
|
return
|
|
}
|