perf: initialize all token encoder when starting (close #459, close $460)

* 兼容讯飞v2接口

* Revert "兼容讯飞v2接口"

This reverts commit 21f05d1294.

* fix: fix implementation

---------

Co-authored-by: JustSong <songquanpeng@foxmail.com>
Co-authored-by: JustSong <39998050+songquanpeng@users.noreply.github.com>

README.ja.md

@@ -52,7 +52,7 @@ _✨ 標準的な OpenAI API フォーマットを通じてすべての LLM に
 
 > **警告**: この README は ChatGPT によって翻訳されています。翻訳ミスを発見した場合は遠慮なく PR を投稿してください。
 
-> **警告**： 英語版の Docker イメージは `justsong/one-api-ja` です。
+> **警告**： 英語版の Docker イメージは `justsong/one-api-en` です。
 
 > **注**: Docker からプルされた最新のイメージは、`alpha` リリースかもしれません。安定性が必要な場合は、手動でバージョンを指定してください。
 
@@ -89,7 +89,7 @@ _✨ 標準的な OpenAI API フォーマットを通じてすべての LLM に
 
 ## デプロイメント
 ### Docker デプロイメント
-デプロイコマンド: `docker run --name one-api -d --restart always -p 3000:3000 -e TZ=Asia/Shanghai -v /home/ubuntu/data/one-api:/data justsong/one-api-ja`。
+デプロイコマンド: `docker run --name one-api -d --restart always -p 3000:3000 -e TZ=Asia/Shanghai -v /home/ubuntu/data/one-api:/data justsong/one-api-en`。
 
 コマンドを更新する: `docker run --rm -v /var/run/docker.sock:/var/run/docker.sock containrr/watchtower -cR`。
 

README.md

@@ -51,11 +51,13 @@ _✨ 通过标准的 OpenAI API 格式访问所有的大模型，开箱即用
   <a href="https://iamazing.cn/page/reward">赞赏支持</a>
 </p>
 
-> **Note**：本项目为开源项目，使用者必须在遵循 OpenAI 的[使用条款](https://openai.com/policies/terms-of-use)以及**法律法规**的情况下使用，不得用于非法用途。
+> **Note**
+> 本项目为开源项目，使用者必须在遵循 OpenAI 的[使用条款](https://openai.com/policies/terms-of-use)以及**法律法规**的情况下使用，不得用于非法用途。
+> 
+> 根据[《生成式人工智能服务管理暂行办法》](http://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm)的要求，请勿对中国地区公众提供一切未经备案的生成式人工智能服务。
 
-> **Note**：使用 Docker 拉取的最新镜像可能是 `alpha` 版本，如果追求稳定性请手动指定版本。
+> **Warning**
-
+> 使用 Docker 拉取的最新镜像可能是 `alpha` 版本，如果追求稳定性请手动指定版本。
-> **Warning**：从 `v0.3` 版本升级到 `v0.4` 版本需要手动迁移数据库，请手动执行[数据库迁移脚本](./bin/migration_v0.3-v0.4.sql)。
 
 ## 功能
 1. 支持多种大模型：

common/model-ratio.go

@@ -1,6 +1,9 @@
 package common
 
-import "encoding/json"
+import (
+	"encoding/json"
+	"strings"
+)
 
 // ModelRatio
 // https://platform.openai.com/docs/models/model-endpoint-compatibility
@@ -38,8 +41,8 @@ var ModelRatio = map[string]float64{
 	"text-moderation-stable":  0.1,
 	"text-moderation-latest":  0.1,
 	"dall-e":                  8,
-	"claude-instant-1":        0.75,
+	"claude-instant-1":        0.815,  // $1.63 / 1M tokens
-	"claude-2":                30,
+	"claude-2":                5.51,   // $11.02 / 1M tokens
 	"ERNIE-Bot":               0.8572, // ￥0.012 / 1k tokens
 	"ERNIE-Bot-turbo":         0.5715, // ￥0.008 / 1k tokens
 	"Embedding-V1":            0.1429, // ￥0.002 / 1k tokens
@@ -73,3 +76,19 @@ func GetModelRatio(name string) float64 {
 	}
 	return ratio
 }
+
+func GetCompletionRatio(name string) float64 {
+	if strings.HasPrefix(name, "gpt-3.5") {
+		return 1.333333
+	}
+	if strings.HasPrefix(name, "gpt-4") {
+		return 2
+	}
+	if strings.HasPrefix(name, "claude-instant-1") {
+		return 3.38
+	}
+	if strings.HasPrefix(name, "claude-2") {
+		return 2.965517
+	}
+	return 1
+}

controller/channel-test.go

@@ -174,7 +174,7 @@ func testAllChannels(notify bool) error {
 				err = errors.New(fmt.Sprintf("响应时间 %.2fs 超过阈值 %.2fs", float64(milliseconds)/1000.0, float64(disableThreshold)/1000.0))
 				disableChannel(channel.Id, channel.Name, err.Error())
 			}
-			if shouldDisableChannel(openaiErr) {
+			if shouldDisableChannel(openaiErr, -1) {
 				disableChannel(channel.Id, channel.Name, err.Error())
 			}
 			channel.UpdateResponseTime(milliseconds)

controller/relay-ali.go

@@ -177,9 +177,11 @@ func aliStreamHandler(c *gin.Context, resp *http.Response) (*OpenAIErrorWithStat
 				common.SysError("error unmarshalling stream response: " + err.Error())
 				return true
 			}
-			usage.PromptTokens += aliResponse.Usage.InputTokens
+			if aliResponse.Usage.OutputTokens != 0 {
-			usage.CompletionTokens += aliResponse.Usage.OutputTokens
+				usage.PromptTokens = aliResponse.Usage.InputTokens
-			usage.TotalTokens += aliResponse.Usage.InputTokens + aliResponse.Usage.OutputTokens
+				usage.CompletionTokens = aliResponse.Usage.OutputTokens
+				usage.TotalTokens = aliResponse.Usage.InputTokens + aliResponse.Usage.OutputTokens
+			}
 			response := streamResponseAli2OpenAI(&aliResponse)
 			response.Choices[0].Delta.Content = strings.TrimPrefix(response.Choices[0].Delta.Content, lastResponseText)
 			lastResponseText = aliResponse.Output.Text

controller/relay-baidu.go

@@ -215,9 +215,11 @@ func baiduStreamHandler(c *gin.Context, resp *http.Response) (*OpenAIErrorWithSt
 				common.SysError("error unmarshalling stream response: " + err.Error())
 				return true
 			}
-			usage.PromptTokens += baiduResponse.Usage.PromptTokens
+			if baiduResponse.Usage.TotalTokens != 0 {
-			usage.CompletionTokens += baiduResponse.Usage.CompletionTokens
+				usage.TotalTokens = baiduResponse.Usage.TotalTokens
-			usage.TotalTokens += baiduResponse.Usage.TotalTokens
+				usage.PromptTokens = baiduResponse.Usage.PromptTokens
+				usage.CompletionTokens = baiduResponse.Usage.TotalTokens - baiduResponse.Usage.PromptTokens
+			}
 			response := streamResponseBaidu2OpenAI(&baiduResponse)
 			jsonResponse, err := json.Marshal(response)
 			if err != nil {

controller/relay-text.go

@@ -315,6 +315,10 @@ func relayTextHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode {
 			return errorWrapper(err, "close_request_body_failed", http.StatusInternalServerError)
 		}
 		isStream = isStream || strings.HasPrefix(resp.Header.Get("Content-Type"), "text/event-stream")
+
+		if resp.StatusCode != http.StatusOK {
+			return relayErrorHandler(resp)
+		}
 	}
 
 	var textResponse TextResponse
@@ -326,14 +330,7 @@ func relayTextHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode {
 		go func() {
 			if consumeQuota {
 				quota := 0
-				completionRatio := 1.0
+				completionRatio := common.GetCompletionRatio(textRequest.Model)
-				if strings.HasPrefix(textRequest.Model, "gpt-3.5") {
-					completionRatio = 1.333333
-				}
-				if strings.HasPrefix(textRequest.Model, "gpt-4") {
-					completionRatio = 2
-				}
-
 				promptTokens = textResponse.Usage.PromptTokens
 				completionTokens = textResponse.Usage.CompletionTokens
 

controller/relay-utils.go

@@ -1,16 +1,38 @@
 package controller
 
 import (
+	"encoding/json"
 	"fmt"
 	"github.com/gin-gonic/gin"
 	"github.com/pkoukk/tiktoken-go"
+	"io"
+	"net/http"
 	"one-api/common"
+	"strconv"
 )
 
 var stopFinishReason = "stop"
 
 var tokenEncoderMap = map[string]*tiktoken.Tiktoken{}
 
+func InitTokenEncoders() {
+	common.SysLog("initializing token encoders")
+	fallbackTokenEncoder, err := tiktoken.EncodingForModel("gpt-3.5-turbo")
+	if err != nil {
+		common.FatalLog(fmt.Sprintf("failed to get fallback token encoder: %s", err.Error()))
+	}
+	for model, _ := range common.ModelRatio {
+		tokenEncoder, err := tiktoken.EncodingForModel(model)
+		if err != nil {
+			common.SysError(fmt.Sprintf("using fallback encoder for model %s", model))
+			tokenEncoderMap[model] = fallbackTokenEncoder
+			continue
+		}
+		tokenEncoderMap[model] = tokenEncoder
+	}
+	common.SysLog("token encoders initialized")
+}
+
 func getTokenEncoder(model string) *tiktoken.Tiktoken {
 	if tokenEncoder, ok := tokenEncoderMap[model]; ok {
 		return tokenEncoder
@@ -95,13 +117,16 @@ func errorWrapper(err error, code string, statusCode int) *OpenAIErrorWithStatus
 	}
 }
 
-func shouldDisableChannel(err *OpenAIError) bool {
+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
 	}
@@ -115,3 +140,30 @@ func setEventStreamHeaders(c *gin.Context) {
 	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:    "one_api_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
+}

controller/relay-xunfei.go

@@ -75,7 +75,7 @@ type XunfeiChatResponse struct {
 	} `json:"payload"`
 }
 
-func requestOpenAI2Xunfei(request GeneralOpenAIRequest, xunfeiAppId string) *XunfeiChatRequest {
+func requestOpenAI2Xunfei(request GeneralOpenAIRequest, xunfeiAppId string, domain string) *XunfeiChatRequest {
 	messages := make([]XunfeiMessage, 0, len(request.Messages))
 	for _, message := range request.Messages {
 		if message.Role == "system" {
@@ -96,7 +96,7 @@ func requestOpenAI2Xunfei(request GeneralOpenAIRequest, xunfeiAppId string) *Xun
 	}
 	xunfeiRequest := XunfeiChatRequest{}
 	xunfeiRequest.Header.AppId = xunfeiAppId
-	xunfeiRequest.Parameter.Chat.Domain = "general"
+	xunfeiRequest.Parameter.Chat.Domain = domain
 	xunfeiRequest.Parameter.Chat.Temperature = request.Temperature
 	xunfeiRequest.Parameter.Chat.TopK = request.N
 	xunfeiRequest.Parameter.Chat.MaxTokens = request.MaxTokens
@@ -178,15 +178,28 @@ func buildXunfeiAuthUrl(hostUrl string, apiKey, apiSecret string) string {
 
 func xunfeiStreamHandler(c *gin.Context, textRequest GeneralOpenAIRequest, appId string, apiSecret string, apiKey string) (*OpenAIErrorWithStatusCode, *Usage) {
 	var usage Usage
+	query := c.Request.URL.Query()
+	apiVersion := query.Get("api-version")
+	if apiVersion == "" {
+		apiVersion = c.GetString("api_version")
+	}
+	if apiVersion == "" {
+		apiVersion = "v1.1"
+		common.SysLog("api_version not found, use default: " + apiVersion)
+	}
+	domain := "general"
+	if apiVersion == "v2.1" {
+		domain = "generalv2"
+	}
+	hostUrl := fmt.Sprintf("wss://spark-api.xf-yun.com/%s/chat", apiVersion)
 	d := websocket.Dialer{
 		HandshakeTimeout: 5 * time.Second,
 	}
-	hostUrl := "wss://aichat.xf-yun.com/v1/chat"
 	conn, resp, err := d.Dial(buildXunfeiAuthUrl(hostUrl, apiKey, apiSecret), nil)
 	if err != nil || resp.StatusCode != 101 {
 		return errorWrapper(err, "dial_failed", http.StatusInternalServerError), nil
 	}
-	data := requestOpenAI2Xunfei(textRequest, appId)
+	data := requestOpenAI2Xunfei(textRequest, appId, domain)
 	err = conn.WriteJSON(data)
 	if err != nil {
 		return errorWrapper(err, "write_json_failed", http.StatusInternalServerError), nil

controller/relay.go

@@ -185,7 +185,7 @@ func Relay(c *gin.Context) {
 		channelId := c.GetInt("channel_id")
 		common.SysError(fmt.Sprintf("relay error (channel #%d): %s", channelId, err.Message))
 		// https://platform.openai.com/docs/guides/error-codes/api-errors
-		if shouldDisableChannel(&err.OpenAIError) {
+		if shouldDisableChannel(&err.OpenAIError, err.StatusCode) {
 			channelId := c.GetInt("channel_id")
 			channelName := c.GetString("channel_name")
 			disableChannel(channelId, channelName, err.Message)

i18n/en.json

@@ -520,5 +520,8 @@
   "代理": "Proxy",
   "此项可选，用于通过代理站来进行 API 调用，请输入代理站地址，格式为：https://domain.com": "This is optional, used to make API calls through the proxy site, please enter the proxy site address, the format is: https://domain.com",
   "取消密码登录将导致所有未绑定其他登录方式的用户（包括管理员）无法通过密码登录，确认取消？": "Canceling password login will cause all users (including administrators) who have not bound other login methods to be unable to log in via password, confirm cancel?",
-  "按照如下格式输入：": "Enter in the following format:"
+  "按照如下格式输入：": "Enter in the following format:",
+  "模型版本": "Model version",
+  "请输入星火大模型版本，注意是接口地址中的版本号，例如：v2.1": "Please enter the version of the Starfire model, note that it is the version number in the interface address, for example: v2.1",
+  "点击查看": "click to view"
 }

main.go

@@ -77,6 +77,7 @@ func main() {
 		}
 		go controller.AutomaticallyTestChannels(frequency)
 	}
+	controller.InitTokenEncoders()
 
 	// Initialize HTTP server
 	server := gin.Default()

middleware/distributor.go

@@ -107,7 +107,7 @@ func Distribute() func(c *gin.Context) {
 		c.Set("model_mapping", channel.ModelMapping)
 		c.Request.Header.Set("Authorization", fmt.Sprintf("Bearer %s", channel.Key))
 		c.Set("base_url", channel.BaseURL)
-		if channel.Type == common.ChannelTypeAzure {
+		if channel.Type == common.ChannelTypeAzure || channel.Type == common.ChannelTypeXunfei {
 			c.Set("api_version", channel.Other)
 		}
 		c.Next()

web/src/components/LogsTable.js

@@ -43,6 +43,7 @@ function renderType(type) {
 
 const LogsTable = () => {
   const [logs, setLogs] = useState([]);
+  const [showStat, setShowStat] = useState(false);
   const [loading, setLoading] = useState(true);
   const [activePage, setActivePage] = useState(1);
   const [searchKeyword, setSearchKeyword] = useState('');
@@ -92,6 +93,17 @@ const LogsTable = () => {
     }
   };
 
+  const handleEyeClick = async () => {
+    if (!showStat) {
+      if (isAdminUser) {
+        await getLogStat();
+      } else {
+        await getLogSelfStat();
+      }
+    }
+    setShowStat(!showStat);
+  };
+
   const loadLogs = async (startIdx) => {
     let url = '';
     let localStartTimestamp = Date.parse(start_timestamp) / 1000;
@@ -129,13 +141,8 @@ const LogsTable = () => {
 
   const refresh = async () => {
     setLoading(true);
-    setActivePage(1)
+    setActivePage(1);
     await loadLogs(0);
-    if (isAdminUser) {
-      getLogStat().then();
-    } else {
-      getLogSelfStat().then();
-    }
   };
 
   useEffect(() => {
@@ -190,7 +197,12 @@ const LogsTable = () => {
   return (
     <>
       <Segment>
-        <Header as='h3'>使用明细（总消耗额度：{renderQuota(stat.quota)}）</Header>
+        <Header as='h3'>
+          使用明细（总消耗额度：
+          {showStat && renderQuota(stat.quota)}
+          {!showStat && <span onClick={handleEyeClick} style={{ cursor: 'pointer', color: 'gray' }}>点击查看</span>}
+          ）
+        </Header>
         <Form>
           <Form.Group>
             {

web/src/pages/Channel/EditChannel.js

@@ -163,6 +163,9 @@ const EditChannel = () => {
     if (localInputs.type === 3 && localInputs.other === '') {
       localInputs.other = '2023-06-01-preview';
     }
+    if (localInputs.type === 18 && localInputs.other === '') {
+      localInputs.other = 'v2.1';
+    }
     if (localInputs.model_mapping === '') {
       localInputs.model_mapping = '{}';
     }
@@ -275,6 +278,20 @@ const EditChannel = () => {
               options={groupOptions}
             />
           </Form.Field>
+          {
+            inputs.type === 18 && (
+              <Form.Field>
+                <Form.Input
+                  label='模型版本'
+                  name='other'
+                  placeholder={'请输入星火大模型版本，注意是接口地址中的版本号，例如：v2.1'}
+                  onChange={handleInputChange}
+                  value={inputs.other}
+                  autoComplete='new-password'
+                />
+              </Form.Field>
+            )
+          }
           <Form.Field>
             <Form.Dropdown
               label='模型'