refactor: refactor relay part (#935)

This commit is contained in:
JustSong
2024-01-14 19:21:03 +08:00
committed by GitHub
parent f2c51a494c
commit eed9f5fdf0
37 changed files with 1527 additions and 1437 deletions

View File

@@ -0,0 +1,6 @@
package openai
const (
ContentTypeText = "text"
ContentTypeImageURL = "image_url"
)

View File

@@ -0,0 +1,144 @@
package openai
import (
"bufio"
"bytes"
"encoding/json"
"github.com/gin-gonic/gin"
"io"
"net/http"
"one-api/common"
"one-api/relay/constant"
"strings"
)
func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*ErrorWithStatusCode, string) {
responseText := ""
scanner := bufio.NewScanner(resp.Body)
scanner.Split(func(data []byte, atEOF bool) (advance int, token []byte, err error) {
if atEOF && len(data) == 0 {
return 0, nil, nil
}
if i := strings.Index(string(data), "\n"); i >= 0 {
return i + 1, data[0:i], nil
}
if atEOF {
return len(data), data, nil
}
return 0, nil, nil
})
dataChan := make(chan string)
stopChan := make(chan bool)
go func() {
for scanner.Scan() {
data := scanner.Text()
if len(data) < 6 { // ignore blank line or wrong format
continue
}
if data[:6] != "data: " && data[:6] != "[DONE]" {
continue
}
dataChan <- data
data = data[6:]
if !strings.HasPrefix(data, "[DONE]") {
switch relayMode {
case constant.RelayModeChatCompletions:
var streamResponse ChatCompletionsStreamResponse
err := json.Unmarshal([]byte(data), &streamResponse)
if err != nil {
common.SysError("error unmarshalling stream response: " + err.Error())
continue // just ignore the error
}
for _, choice := range streamResponse.Choices {
responseText += choice.Delta.Content
}
case constant.RelayModeCompletions:
var streamResponse CompletionsStreamResponse
err := json.Unmarshal([]byte(data), &streamResponse)
if err != nil {
common.SysError("error unmarshalling stream response: " + err.Error())
continue
}
for _, choice := range streamResponse.Choices {
responseText += choice.Text
}
}
}
}
stopChan <- true
}()
common.SetEventStreamHeaders(c)
c.Stream(func(w io.Writer) bool {
select {
case data := <-dataChan:
if strings.HasPrefix(data, "data: [DONE]") {
data = data[:12]
}
// some implementations may add \r at the end of data
data = strings.TrimSuffix(data, "\r")
c.Render(-1, common.CustomEvent{Data: data})
return true
case <-stopChan:
return false
}
})
err := resp.Body.Close()
if err != nil {
return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), ""
}
return nil, responseText
}
func Handler(c *gin.Context, resp *http.Response, promptTokens int, model string) (*ErrorWithStatusCode, *Usage) {
var textResponse SlimTextResponse
responseBody, err := io.ReadAll(resp.Body)
if err != nil {
return ErrorWrapper(err, "read_response_body_failed", http.StatusInternalServerError), nil
}
err = resp.Body.Close()
if err != nil {
return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
}
err = json.Unmarshal(responseBody, &textResponse)
if err != nil {
return ErrorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError), nil
}
if textResponse.Error.Type != "" {
return &ErrorWithStatusCode{
Error: textResponse.Error,
StatusCode: resp.StatusCode,
}, nil
}
// Reset response body
resp.Body = io.NopCloser(bytes.NewBuffer(responseBody))
// We shouldn't set the header before we parse the response body, because the parse part may fail.
// And then we will have to send an error response, but in this case, the header has already been set.
// So the HTTPClient will be confused by the response.
// For example, Postman will report error, and we cannot check the response at all.
for k, v := range resp.Header {
c.Writer.Header().Set(k, v[0])
}
c.Writer.WriteHeader(resp.StatusCode)
_, err = io.Copy(c.Writer, resp.Body)
if err != nil {
return ErrorWrapper(err, "copy_response_body_failed", http.StatusInternalServerError), nil
}
err = resp.Body.Close()
if err != nil {
return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
}
if textResponse.Usage.TotalTokens == 0 {
completionTokens := 0
for _, choice := range textResponse.Choices {
completionTokens += CountTokenText(choice.Message.StringContent(), model)
}
textResponse.Usage = Usage{
PromptTokens: promptTokens,
CompletionTokens: completionTokens,
TotalTokens: promptTokens + completionTokens,
}
}
return nil, &textResponse.Usage
}

View File

@@ -0,0 +1,283 @@
package openai
type Message struct {
Role string `json:"role"`
Content any `json:"content"`
Name *string `json:"name,omitempty"`
}
type ImageURL struct {
Url string `json:"url,omitempty"`
Detail string `json:"detail,omitempty"`
}
type TextContent struct {
Type string `json:"type,omitempty"`
Text string `json:"text,omitempty"`
}
type ImageContent struct {
Type string `json:"type,omitempty"`
ImageURL *ImageURL `json:"image_url,omitempty"`
}
type OpenAIMessageContent struct {
Type string `json:"type,omitempty"`
Text string `json:"text"`
ImageURL *ImageURL `json:"image_url,omitempty"`
}
func (m Message) IsStringContent() bool {
_, ok := m.Content.(string)
return ok
}
func (m Message) StringContent() string {
content, ok := m.Content.(string)
if ok {
return content
}
contentList, ok := m.Content.([]any)
if ok {
var contentStr string
for _, contentItem := range contentList {
contentMap, ok := contentItem.(map[string]any)
if !ok {
continue
}
if contentMap["type"] == ContentTypeText {
if subStr, ok := contentMap["text"].(string); ok {
contentStr += subStr
}
}
}
return contentStr
}
return ""
}
func (m Message) ParseContent() []OpenAIMessageContent {
var contentList []OpenAIMessageContent
content, ok := m.Content.(string)
if ok {
contentList = append(contentList, OpenAIMessageContent{
Type: ContentTypeText,
Text: content,
})
return contentList
}
anyList, ok := m.Content.([]any)
if ok {
for _, contentItem := range anyList {
contentMap, ok := contentItem.(map[string]any)
if !ok {
continue
}
switch contentMap["type"] {
case ContentTypeText:
if subStr, ok := contentMap["text"].(string); ok {
contentList = append(contentList, OpenAIMessageContent{
Type: ContentTypeText,
Text: subStr,
})
}
case ContentTypeImageURL:
if subObj, ok := contentMap["image_url"].(map[string]any); ok {
contentList = append(contentList, OpenAIMessageContent{
Type: ContentTypeImageURL,
ImageURL: &ImageURL{
Url: subObj["url"].(string),
},
})
}
}
}
return contentList
}
return nil
}
type ResponseFormat struct {
Type string `json:"type,omitempty"`
}
type GeneralOpenAIRequest struct {
Model string `json:"model,omitempty"`
Messages []Message `json:"messages,omitempty"`
Prompt any `json:"prompt,omitempty"`
Stream bool `json:"stream,omitempty"`
MaxTokens int `json:"max_tokens,omitempty"`
Temperature float64 `json:"temperature,omitempty"`
TopP float64 `json:"top_p,omitempty"`
N int `json:"n,omitempty"`
Input any `json:"input,omitempty"`
Instruction string `json:"instruction,omitempty"`
Size string `json:"size,omitempty"`
Functions any `json:"functions,omitempty"`
FrequencyPenalty float64 `json:"frequency_penalty,omitempty"`
PresencePenalty float64 `json:"presence_penalty,omitempty"`
ResponseFormat *ResponseFormat `json:"response_format,omitempty"`
Seed float64 `json:"seed,omitempty"`
Tools any `json:"tools,omitempty"`
ToolChoice any `json:"tool_choice,omitempty"`
User string `json:"user,omitempty"`
}
func (r GeneralOpenAIRequest) ParseInput() []string {
if r.Input == nil {
return nil
}
var input []string
switch r.Input.(type) {
case string:
input = []string{r.Input.(string)}
case []any:
input = make([]string, 0, len(r.Input.([]any)))
for _, item := range r.Input.([]any) {
if str, ok := item.(string); ok {
input = append(input, str)
}
}
}
return input
}
type ChatRequest struct {
Model string `json:"model"`
Messages []Message `json:"messages"`
MaxTokens int `json:"max_tokens"`
}
type TextRequest struct {
Model string `json:"model"`
Messages []Message `json:"messages"`
Prompt string `json:"prompt"`
MaxTokens int `json:"max_tokens"`
//Stream bool `json:"stream"`
}
// ImageRequest docs: https://platform.openai.com/docs/api-reference/images/create
type ImageRequest struct {
Model string `json:"model"`
Prompt string `json:"prompt" binding:"required"`
N int `json:"n,omitempty"`
Size string `json:"size,omitempty"`
Quality string `json:"quality,omitempty"`
ResponseFormat string `json:"response_format,omitempty"`
Style string `json:"style,omitempty"`
User string `json:"user,omitempty"`
}
type WhisperJSONResponse struct {
Text string `json:"text,omitempty"`
}
type WhisperVerboseJSONResponse struct {
Task string `json:"task,omitempty"`
Language string `json:"language,omitempty"`
Duration float64 `json:"duration,omitempty"`
Text string `json:"text,omitempty"`
Segments []Segment `json:"segments,omitempty"`
}
type Segment struct {
Id int `json:"id"`
Seek int `json:"seek"`
Start float64 `json:"start"`
End float64 `json:"end"`
Text string `json:"text"`
Tokens []int `json:"tokens"`
Temperature float64 `json:"temperature"`
AvgLogprob float64 `json:"avg_logprob"`
CompressionRatio float64 `json:"compression_ratio"`
NoSpeechProb float64 `json:"no_speech_prob"`
}
type TextToSpeechRequest struct {
Model string `json:"model" binding:"required"`
Input string `json:"input" binding:"required"`
Voice string `json:"voice" binding:"required"`
Speed float64 `json:"speed"`
ResponseFormat string `json:"response_format"`
}
type Usage struct {
PromptTokens int `json:"prompt_tokens"`
CompletionTokens int `json:"completion_tokens"`
TotalTokens int `json:"total_tokens"`
}
type Error struct {
Message string `json:"message"`
Type string `json:"type"`
Param string `json:"param"`
Code any `json:"code"`
}
type ErrorWithStatusCode struct {
Error
StatusCode int `json:"status_code"`
}
type SlimTextResponse struct {
Choices []TextResponseChoice `json:"choices"`
Usage `json:"usage"`
Error Error `json:"error"`
}
type TextResponseChoice struct {
Index int `json:"index"`
Message `json:"message"`
FinishReason string `json:"finish_reason"`
}
type TextResponse struct {
Id string `json:"id"`
Model string `json:"model,omitempty"`
Object string `json:"object"`
Created int64 `json:"created"`
Choices []TextResponseChoice `json:"choices"`
Usage `json:"usage"`
}
type EmbeddingResponseItem struct {
Object string `json:"object"`
Index int `json:"index"`
Embedding []float64 `json:"embedding"`
}
type EmbeddingResponse struct {
Object string `json:"object"`
Data []EmbeddingResponseItem `json:"data"`
Model string `json:"model"`
Usage `json:"usage"`
}
type ImageResponse struct {
Created int `json:"created"`
Data []struct {
Url string `json:"url"`
}
}
type ChatCompletionsStreamResponseChoice struct {
Delta struct {
Content string `json:"content"`
} `json:"delta"`
FinishReason *string `json:"finish_reason,omitempty"`
}
type ChatCompletionsStreamResponse struct {
Id string `json:"id"`
Object string `json:"object"`
Created int64 `json:"created"`
Model string `json:"model"`
Choices []ChatCompletionsStreamResponseChoice `json:"choices"`
}
type CompletionsStreamResponse struct {
Choices []struct {
Text string `json:"text"`
FinishReason string `json:"finish_reason"`
} `json:"choices"`
}

View File

@@ -0,0 +1,205 @@
package openai
import (
"errors"
"fmt"
"github.com/pkoukk/tiktoken-go"
"math"
"one-api/common"
"one-api/common/image"
"strings"
)
// 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
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 {
common.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":
if fetchSize {
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)
}

View File

@@ -0,0 +1,13 @@
package openai
func ErrorWrapper(err error, code string, statusCode int) *ErrorWithStatusCode {
Error := Error{
Message: err.Error(),
Type: "one_api_error",
Code: code,
}
return &ErrorWithStatusCode{
Error: Error,
StatusCode: statusCode,
}
}