diff --git a/common/model-ratio.go b/common/model-ratio.go index 681f0ae7..74c74a90 100644 --- a/common/model-ratio.go +++ b/common/model-ratio.go @@ -6,6 +6,29 @@ import ( "time" ) +var DalleSizeRatios = map[string]map[string]float64{ + "dall-e-2": { + "256x256": 1, + "512x512": 1.125, + "1024x1024": 1.25, + }, + "dall-e-3": { + "1024x1024": 1, + "1024x1792": 2, + "1792x1024": 2, + }, +} + +var DalleGenerationImageAmounts = map[string][2]int{ + "dall-e-2": {1, 10}, + "dall-e-3": {1, 1}, // OpenAI allows n=1 currently. +} + +var DalleImagePromptLengthLimitations = map[string]int{ + "dall-e-2": 1000, + "dall-e-3": 4000, +} + // ModelRatio // https://platform.openai.com/docs/models/model-endpoint-compatibility // https://cloud.baidu.com/doc/WENXINWORKSHOP/s/Blfmc9dlf @@ -36,7 +59,11 @@ var ModelRatio = map[string]float64{ "text-davinci-003": 10, "text-davinci-edit-001": 10, "code-davinci-edit-001": 10, - "whisper-1": 15, // $0.006 / minute -> $0.006 / 150 words -> $0.006 / 200 tokens -> $0.03 / 1k tokens + "whisper-1": 15, // $0.006 / minute -> $0.006 / 150 words -> $0.006 / 200 tokens -> $0.03 / 1k tokens + "tts-1": 7.5, // $0.015 / 1K characters + "tts-1-1106": 7.5, + "tts-1-hd": 15, // $0.030 / 1K characters + "tts-1-hd-1106": 15, "davinci": 10, "curie": 10, "babbage": 10, @@ -45,7 +72,8 @@ var ModelRatio = map[string]float64{ "text-search-ada-doc-001": 10, "text-moderation-stable": 0.1, "text-moderation-latest": 0.1, - "dall-e": 8, + "dall-e-2": 8, // $0.016 - $0.020 / image + "dall-e-3": 20, // $0.040 - $0.120 / image "claude-instant-1": 0.815, // $1.63 / 1M tokens "claude-2": 5.51, // $11.02 / 1M tokens "ERNIE-Bot": 0.8572, // ￥0.012 / 1k tokens diff --git a/controller/channel-test.go b/controller/channel-test.go index c5d4d26c..1b0b745a 100644 --- a/controller/channel-test.go +++ b/controller/channel-test.go @@ -5,6 +5,7 @@ import ( "encoding/json" "errors" "fmt" + "io" "net/http" "one-api/common" "one-api/model" @@ -45,8 +46,8 @@ func testChannel(channel *model.Channel, request ChatRequest) (err error, openai if channel.Type == common.ChannelTypeAzure { requestURL = getFullRequestURL(channel.GetBaseURL(), fmt.Sprintf("/openai/deployments/%s/chat/completions?api-version=2023-03-15-preview", request.Model), channel.Type) } else { - if channel.GetBaseURL() != "" { - requestURL = channel.GetBaseURL() + if baseURL := channel.GetBaseURL(); len(baseURL) > 0 { + requestURL = baseURL } requestURL = getFullRequestURL(requestURL, "/v1/chat/completions", channel.Type) @@ -71,10 +72,14 @@ func testChannel(channel *model.Channel, request ChatRequest) (err error, openai } defer resp.Body.Close() var response TextResponse - err = json.NewDecoder(resp.Body).Decode(&response) + body, err := io.ReadAll(resp.Body) if err != nil { return err, nil } + err = json.Unmarshal(body, &response) + if err != nil { + return fmt.Errorf("Error: %s\nResp body: %s", err, body), nil + } if response.Usage.CompletionTokens == 0 { return errors.New(fmt.Sprintf("type %s, code %v, message %s", response.Error.Type, response.Error.Code, response.Error.Message)), &response.Error } diff --git a/controller/model.go b/controller/model.go index 7bd9d097..59ea22e8 100644 --- a/controller/model.go +++ b/controller/model.go @@ -55,12 +55,21 @@ func init() { // https://platform.openai.com/docs/models/model-endpoint-compatibility openAIModels = []OpenAIModels{ { - Id: "dall-e", + Id: "dall-e-2", Object: "model", Created: 1677649963, OwnedBy: "openai", Permission: permission, - Root: "dall-e", + Root: "dall-e-2", + Parent: nil, + }, + { + Id: "dall-e-3", + Object: "model", + Created: 1677649963, + OwnedBy: "openai", + Permission: permission, + Root: "dall-e-3", Parent: nil, }, { @@ -72,6 +81,42 @@ func init() { Root: "whisper-1", Parent: nil, }, + { + Id: "tts-1", + Object: "model", + Created: 1677649963, + OwnedBy: "openai", + Permission: permission, + Root: "tts-1", + Parent: nil, + }, + { + Id: "tts-1-1106", + Object: "model", + Created: 1677649963, + OwnedBy: "openai", + Permission: permission, + Root: "tts-1-1106", + Parent: nil, + }, + { + Id: "tts-1-hd", + Object: "model", + Created: 1677649963, + OwnedBy: "openai", + Permission: permission, + Root: "tts-1-hd", + Parent: nil, + }, + { + Id: "tts-1-hd-1106", + Object: "model", + Created: 1677649963, + OwnedBy: "openai", + Permission: permission, + Root: "tts-1-hd-1106", + Parent: nil, + }, { Id: "gpt-3.5-turbo", Object: "model", diff --git a/controller/relay-audio.go b/controller/relay-audio.go index 53833108..01267fbf 100644 --- a/controller/relay-audio.go +++ b/controller/relay-audio.go @@ -5,7 +5,6 @@ import ( "context" "encoding/json" "errors" - "fmt" "github.com/gin-gonic/gin" "io" "net/http" @@ -21,6 +20,22 @@ func relayAudioHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode channelId := c.GetInt("channel_id") userId := c.GetInt("id") group := c.GetString("group") + tokenName := c.GetString("token_name") + + var ttsRequest TextToSpeechRequest + if relayMode == RelayModeAudioSpeech { + // Read JSON + err := common.UnmarshalBodyReusable(c, &ttsRequest) + // Check if JSON is valid + if err != nil { + return errorWrapper(err, "invalid_json", http.StatusBadRequest) + } + audioModel = ttsRequest.Model + // Check if text is too long 4096 + if len(ttsRequest.Input) > 4096 { + return errorWrapper(errors.New("input is too long (over 4096 characters)"), "text_too_long", http.StatusBadRequest) + } + } preConsumedTokens := common.PreConsumedQuota modelRatio := common.GetModelRatio(audioModel) @@ -31,22 +46,32 @@ func relayAudioHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode if err != nil { return errorWrapper(err, "get_user_quota_failed", http.StatusInternalServerError) } - if userQuota-preConsumedQuota < 0 { - return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden) - } - err = model.CacheDecreaseUserQuota(userId, preConsumedQuota) - if err != nil { - return errorWrapper(err, "decrease_user_quota_failed", http.StatusInternalServerError) - } - if userQuota > 100*preConsumedQuota { - // in this case, we do not pre-consume quota - // because the user has enough quota - preConsumedQuota = 0 - } - if preConsumedQuota > 0 { - err := model.PreConsumeTokenQuota(tokenId, preConsumedQuota) + + quota := 0 + // Check if user quota is enough + if relayMode == RelayModeAudioSpeech { + quota = int(float64(len(ttsRequest.Input)) * modelRatio * groupRatio) + if quota > userQuota { + return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden) + } + } else { + if userQuota-preConsumedQuota < 0 { + return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden) + } + err = model.CacheDecreaseUserQuota(userId, preConsumedQuota) if err != nil { - return errorWrapper(err, "pre_consume_token_quota_failed", http.StatusForbidden) + return errorWrapper(err, "decrease_user_quota_failed", http.StatusInternalServerError) + } + if userQuota > 100*preConsumedQuota { + // in this case, we do not pre-consume quota + // because the user has enough quota + preConsumedQuota = 0 + } + if preConsumedQuota > 0 { + err := model.PreConsumeTokenQuota(tokenId, preConsumedQuota) + if err != nil { + return errorWrapper(err, "pre_consume_token_quota_failed", http.StatusForbidden) + } } } @@ -93,47 +118,32 @@ func relayAudioHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode if err != nil { return errorWrapper(err, "close_request_body_failed", http.StatusInternalServerError) } - var audioResponse AudioResponse - defer func(ctx context.Context) { - go func() { - quota := countTokenText(audioResponse.Text, audioModel) + if relayMode == RelayModeAudioSpeech { + defer func(ctx context.Context) { + go postConsumeQuota(ctx, tokenId, quota, userId, channelId, modelRatio, groupRatio, audioModel, tokenName) + }(c.Request.Context()) + } else { + responseBody, err := io.ReadAll(resp.Body) + if err != nil { + return errorWrapper(err, "read_response_body_failed", http.StatusInternalServerError) + } + err = resp.Body.Close() + if err != nil { + return errorWrapper(err, "close_response_body_failed", http.StatusInternalServerError) + } + var whisperResponse WhisperResponse + err = json.Unmarshal(responseBody, &whisperResponse) + if err != nil { + return errorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError) + } + defer func(ctx context.Context) { + quota := countTokenText(whisperResponse.Text, audioModel) quotaDelta := quota - preConsumedQuota - err := model.PostConsumeTokenQuota(tokenId, quotaDelta) - if err != nil { - common.SysError("error consuming token remain quota: " + err.Error()) - } - err = model.CacheUpdateUserQuota(userId) - if err != nil { - common.SysError("error update user quota cache: " + err.Error()) - } - if quota != 0 { - tokenName := c.GetString("token_name") - logContent := fmt.Sprintf("模型倍率 %.2f，分组倍率 %.2f", modelRatio, groupRatio) - model.RecordConsumeLog(ctx, userId, channelId, 0, 0, audioModel, tokenName, quota, logContent) - model.UpdateUserUsedQuotaAndRequestCount(userId, quota) - channelId := c.GetInt("channel_id") - model.UpdateChannelUsedQuota(channelId, quota) - } - }() - }(c.Request.Context()) - - responseBody, err := io.ReadAll(resp.Body) - - if err != nil { - return errorWrapper(err, "read_response_body_failed", http.StatusInternalServerError) + go postConsumeQuota(ctx, tokenId, quotaDelta, userId, channelId, modelRatio, groupRatio, audioModel, tokenName) + }(c.Request.Context()) + resp.Body = io.NopCloser(bytes.NewBuffer(responseBody)) } - err = resp.Body.Close() - if err != nil { - return errorWrapper(err, "close_response_body_failed", http.StatusInternalServerError) - } - err = json.Unmarshal(responseBody, &audioResponse) - if err != nil { - return errorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError) - } - - resp.Body = io.NopCloser(bytes.NewBuffer(responseBody)) - for k, v := range resp.Header { c.Writer.Header().Set(k, v[0]) } diff --git a/controller/relay-image.go b/controller/relay-image.go index ccd52dce..1d1b71ba 100644 --- a/controller/relay-image.go +++ b/controller/relay-image.go @@ -6,15 +6,28 @@ import ( "encoding/json" "errors" "fmt" - "github.com/gin-gonic/gin" "io" "net/http" "one-api/common" "one-api/model" + + "github.com/gin-gonic/gin" ) +func isWithinRange(element string, value int) bool { + if _, ok := common.DalleGenerationImageAmounts[element]; !ok { + return false + } + + min := common.DalleGenerationImageAmounts[element][0] + max := common.DalleGenerationImageAmounts[element][1] + + return value >= min && value <= max +} + func relayImageHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode { - imageModel := "dall-e" + imageModel := "dall-e-2" + imageSize := "1024x1024" tokenId := c.GetInt("token_id") channelType := c.GetInt("channel") @@ -31,19 +44,44 @@ func relayImageHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode } } + // Size validation + if imageRequest.Size != "" { + imageSize = imageRequest.Size + } + + // Model validation + if imageRequest.Model != "" { + imageModel = imageRequest.Model + } + + imageCostRatio, hasValidSize := common.DalleSizeRatios[imageModel][imageSize] + + // Check if model is supported + if hasValidSize { + if imageRequest.Quality == "hd" && imageModel == "dall-e-3" { + if imageSize == "1024x1024" { + imageCostRatio *= 2 + } else { + imageCostRatio *= 1.5 + } + } + } else { + return errorWrapper(errors.New("size not supported for this image model"), "size_not_supported", http.StatusBadRequest) + } + // Prompt validation if imageRequest.Prompt == "" { - return errorWrapper(errors.New("prompt is required"), "required_field_missing", http.StatusBadRequest) + return errorWrapper(errors.New("prompt is required"), "prompt_missing", http.StatusBadRequest) } - // Not "256x256", "512x512", or "1024x1024" - if imageRequest.Size != "" && imageRequest.Size != "256x256" && imageRequest.Size != "512x512" && imageRequest.Size != "1024x1024" { - return errorWrapper(errors.New("size must be one of 256x256, 512x512, or 1024x1024"), "invalid_field_value", http.StatusBadRequest) + // Check prompt length + if len(imageRequest.Prompt) > common.DalleImagePromptLengthLimitations[imageModel] { + return errorWrapper(errors.New("prompt is too long"), "prompt_too_long", http.StatusBadRequest) } - // N should between 1 and 10 - if imageRequest.N != 0 && (imageRequest.N < 1 || imageRequest.N > 10) { - return errorWrapper(errors.New("n must be between 1 and 10"), "invalid_field_value", http.StatusBadRequest) + // Number of generated images validation + if isWithinRange(imageModel, imageRequest.N) == false { + return errorWrapper(errors.New("invalid value of n"), "n_not_within_range", http.StatusBadRequest) } // map model name @@ -82,16 +120,7 @@ func relayImageHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode ratio := modelRatio * groupRatio userQuota, err := model.CacheGetUserQuota(userId) - sizeRatio := 1.0 - // Size - if imageRequest.Size == "256x256" { - sizeRatio = 1 - } else if imageRequest.Size == "512x512" { - sizeRatio = 1.125 - } else if imageRequest.Size == "1024x1024" { - sizeRatio = 1.25 - } - quota := int(ratio*sizeRatio*1000) * imageRequest.N + quota := int(ratio*imageCostRatio*1000) * imageRequest.N if consumeQuota && userQuota-quota < 0 { return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden) diff --git a/controller/relay-text.go b/controller/relay-text.go index 4806229e..018c8d8a 100644 --- a/controller/relay-text.go +++ b/controller/relay-text.go @@ -369,6 +369,8 @@ func relayTextHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode { } case APITypeTencent: req.Header.Set("Authorization", apiKey) + case APITypePaLM: + // do not set Authorization header default: req.Header.Set("Authorization", "Bearer "+apiKey) } diff --git a/controller/relay-utils.go b/controller/relay-utils.go index e888efa7..e2b77a97 100644 --- a/controller/relay-utils.go +++ b/controller/relay-utils.go @@ -1,11 +1,13 @@ package controller import ( + "context" "encoding/json" "fmt" "io" "net/http" "one-api/common" + "one-api/model" "strconv" "strings" @@ -192,3 +194,20 @@ func getFullRequestURL(baseURL string, requestURL string, channelType int) strin return fullRequestURL } + +func postConsumeQuota(ctx context.Context, tokenId int, quota int, userId int, channelId int, modelRatio float64, groupRatio float64, modelName string, tokenName string) { + err := model.PostConsumeTokenQuota(tokenId, quota) + if err != nil { + common.SysError("error consuming token remain quota: " + err.Error()) + } + err = model.CacheUpdateUserQuota(userId) + if err != nil { + common.SysError("error update user quota cache: " + err.Error()) + } + if quota != 0 { + logContent := fmt.Sprintf("模型倍率 %.2f，分组倍率 %.2f", modelRatio, groupRatio) + model.RecordConsumeLog(ctx, userId, channelId, 0, 0, modelName, tokenName, quota, logContent) + model.UpdateUserUsedQuotaAndRequestCount(userId, quota) + model.UpdateChannelUsedQuota(channelId, quota) + } +} diff --git a/controller/relay.go b/controller/relay.go index 1926110e..863267b4 100644 --- a/controller/relay.go +++ b/controller/relay.go @@ -24,7 +24,9 @@ const ( RelayModeModerations RelayModeImagesGenerations RelayModeEdits - RelayModeAudio + RelayModeAudioSpeech + RelayModeAudioTranscription + RelayModeAudioTranslation ) // https://platform.openai.com/docs/api-reference/chat @@ -77,16 +79,30 @@ type TextRequest struct { //Stream bool `json:"stream"` } +// ImageRequest docs: https://platform.openai.com/docs/api-reference/images/create type ImageRequest struct { - Prompt string `json:"prompt"` - N int `json:"n"` - Size string `json:"size"` + Model string `json:"model"` + Prompt string `json:"prompt" binding:"required"` + N int `json:"n"` + Size string `json:"size"` + Quality string `json:"quality"` + ResponseFormat string `json:"response_format"` + Style string `json:"style"` + User string `json:"user"` } -type AudioResponse struct { +type WhisperResponse struct { Text string `json:"text,omitempty"` } +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"` @@ -183,14 +199,22 @@ func Relay(c *gin.Context) { relayMode = RelayModeImagesGenerations } else if strings.HasPrefix(c.Request.URL.Path, "/v1/edits") { relayMode = RelayModeEdits - } else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio") { - relayMode = RelayModeAudio + } else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/speech") { + relayMode = RelayModeAudioSpeech + } else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/transcription") { + relayMode = RelayModeAudioTranscription + } else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/translation") { + relayMode = RelayModeAudioTranslation } var err *OpenAIErrorWithStatusCode switch relayMode { case RelayModeImagesGenerations: err = relayImageHelper(c, relayMode) - case RelayModeAudio: + case RelayModeAudioSpeech: + fallthrough + case RelayModeAudioTranslation: + fallthrough + case RelayModeAudioTranscription: err = relayAudioHelper(c, relayMode) default: err = relayTextHelper(c, relayMode) diff --git a/middleware/distributor.go b/middleware/distributor.go index d80945fc..c4ddc3a0 100644 --- a/middleware/distributor.go +++ b/middleware/distributor.go @@ -40,10 +40,7 @@ func Distribute() func(c *gin.Context) { } else { // Select a channel for the user var modelRequest ModelRequest - var err error - if !strings.HasPrefix(c.Request.URL.Path, "/v1/audio") { - err = common.UnmarshalBodyReusable(c, &modelRequest) - } + err := common.UnmarshalBodyReusable(c, &modelRequest) if err != nil { abortWithMessage(c, http.StatusBadRequest, "无效的请求") return @@ -60,10 +57,10 @@ func Distribute() func(c *gin.Context) { } if strings.HasPrefix(c.Request.URL.Path, "/v1/images/generations") { if modelRequest.Model == "" { - modelRequest.Model = "dall-e" + modelRequest.Model = "dall-e-2" } } - if strings.HasPrefix(c.Request.URL.Path, "/v1/audio") { + if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/transcriptions") || strings.HasPrefix(c.Request.URL.Path, "/v1/audio/translations") { if modelRequest.Model == "" { modelRequest.Model = "whisper-1" } diff --git a/router/relay-router.go b/router/relay-router.go index e84f02db..912f4989 100644 --- a/router/relay-router.go +++ b/router/relay-router.go @@ -29,6 +29,7 @@ func SetRelayRouter(router *gin.Engine) { relayV1Router.POST("/engines/:model/embeddings", controller.Relay) relayV1Router.POST("/audio/transcriptions", controller.Relay) relayV1Router.POST("/audio/translations", controller.Relay) + relayV1Router.POST("/audio/speech", controller.Relay) relayV1Router.GET("/files", controller.RelayNotImplemented) relayV1Router.POST("/files", controller.RelayNotImplemented) relayV1Router.DELETE("/files/:id", controller.RelayNotImplemented)