Merge remote-tracking branch 'origin/upstream/main'

relay/adaptor/aws/adaptor.go

@@ -4,14 +4,13 @@ import (
 	"io"
 	"net/http"
 
+	"github.com/Laisky/errors/v2"
 	"github.com/aws/aws-sdk-go-v2/aws"
 	"github.com/aws/aws-sdk-go-v2/credentials"
 	"github.com/aws/aws-sdk-go-v2/service/bedrockruntime"
 	"github.com/gin-gonic/gin"
-	"github.com/pkg/errors"
-	"github.com/songquanpeng/one-api/common/ctxkey"
 	"github.com/songquanpeng/one-api/relay/adaptor"
-	"github.com/songquanpeng/one-api/relay/adaptor/anthropic"
+	"github.com/songquanpeng/one-api/relay/adaptor/aws/utils"
 	"github.com/songquanpeng/one-api/relay/meta"
 	"github.com/songquanpeng/one-api/relay/model"
 )
@@ -19,18 +18,52 @@ import (
 var _ adaptor.Adaptor = new(Adaptor)
 
 type Adaptor struct {
-	meta      *meta.Meta
-	awsClient *bedrockruntime.Client
+	awsAdapter utils.AwsAdapter
+
+	Meta      *meta.Meta
+	AwsClient *bedrockruntime.Client
 }
 
 func (a *Adaptor) Init(meta *meta.Meta) {
-	a.meta = meta
-	a.awsClient = bedrockruntime.New(bedrockruntime.Options{
+	a.Meta = meta
+	a.AwsClient = bedrockruntime.New(bedrockruntime.Options{
 		Region:      meta.Config.Region,
 		Credentials: aws.NewCredentialsCache(credentials.NewStaticCredentialsProvider(meta.Config.AK, meta.Config.SK, "")),
 	})
 }
 
+func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+
+	adaptor := GetAdaptor(request.Model)
+	if adaptor == nil {
+		return nil, errors.New("adaptor not found")
+	}
+
+	a.awsAdapter = adaptor
+	return adaptor.ConvertRequest(c, relayMode, request)
+}
+
+func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode) {
+	if a.awsAdapter == nil {
+		return nil, utils.WrapErr(errors.New("awsAdapter is nil"))
+	}
+	return a.awsAdapter.DoResponse(c, a.AwsClient, meta)
+}
+
+func (a *Adaptor) GetModelList() (models []string) {
+	for model := range adaptors {
+		models = append(models, model)
+	}
+	return
+}
+
+func (a *Adaptor) GetChannelName() string {
+	return "aws"
+}
+
 func (a *Adaptor) GetRequestURL(meta *meta.Meta) (string, error) {
 	return "", nil
 }
@@ -39,17 +72,6 @@ func (a *Adaptor) SetupRequestHeader(c *gin.Context, req *http.Request, meta *me
 	return nil
 }
 
-func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error) {
-	if request == nil {
-		return nil, errors.New("request is nil")
-	}
-
-	claudeReq := anthropic.ConvertRequest(*request)
-	c.Set(ctxkey.RequestModel, request.Model)
-	c.Set(ctxkey.ConvertedRequest, claudeReq)
-	return claudeReq, nil
-}
-
 func (a *Adaptor) ConvertImageRequest(request *model.ImageRequest) (any, error) {
 	if request == nil {
 		return nil, errors.New("request is nil")
@@ -60,23 +82,3 @@ func (a *Adaptor) ConvertImageRequest(request *model.ImageRequest) (any, error)
 func (a *Adaptor) DoRequest(c *gin.Context, meta *meta.Meta, requestBody io.Reader) (*http.Response, error) {
 	return nil, nil
 }
-
-func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode) {
-	if meta.IsStream {
-		err, usage = StreamHandler(c, a.awsClient)
-	} else {
-		err, usage = Handler(c, a.awsClient, meta.ActualModelName)
-	}
-	return
-}
-
-func (a *Adaptor) GetModelList() (models []string) {
-	for n := range awsModelIDMap {
-		models = append(models, n)
-	}
-	return
-}
-
-func (a *Adaptor) GetChannelName() string {
-	return "aws"
-}

relay/adaptor/aws/claude/adapter.go

@@ -0,0 +1,37 @@
+package aws
+
+import (
+	"github.com/aws/aws-sdk-go-v2/service/bedrockruntime"
+	"github.com/gin-gonic/gin"
+	"github.com/pkg/errors"
+	"github.com/songquanpeng/one-api/common/ctxkey"
+	"github.com/songquanpeng/one-api/relay/adaptor/anthropic"
+	"github.com/songquanpeng/one-api/relay/adaptor/aws/utils"
+	"github.com/songquanpeng/one-api/relay/meta"
+	"github.com/songquanpeng/one-api/relay/model"
+)
+
+var _ utils.AwsAdapter = new(Adaptor)
+
+type Adaptor struct {
+}
+
+func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+
+	claudeReq := anthropic.ConvertRequest(*request)
+	c.Set(ctxkey.RequestModel, request.Model)
+	c.Set(ctxkey.ConvertedRequest, claudeReq)
+	return claudeReq, nil
+}
+
+func (a *Adaptor) DoResponse(c *gin.Context, awsCli *bedrockruntime.Client, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode) {
+	if meta.IsStream {
+		err, usage = StreamHandler(c, awsCli)
+	} else {
+		err, usage = Handler(c, awsCli, meta.ActualModelName)
+	}
+	return
+}

relay/adaptor/aws/claude/main.go

@@ -19,21 +19,13 @@ import (
 	"github.com/songquanpeng/one-api/common/helper"
 	"github.com/songquanpeng/one-api/common/logger"
 	"github.com/songquanpeng/one-api/relay/adaptor/anthropic"
+	"github.com/songquanpeng/one-api/relay/adaptor/aws/utils"
 	"github.com/songquanpeng/one-api/relay/adaptor/openai"
 	relaymodel "github.com/songquanpeng/one-api/relay/model"
 )
 
-func wrapErr(err error) *relaymodel.ErrorWithStatusCode {
-	return &relaymodel.ErrorWithStatusCode{
-		StatusCode: http.StatusInternalServerError,
-		Error: relaymodel.Error{
-			Message: fmt.Sprintf("%s", err.Error()),
-		},
-	}
-}
-
 // https://docs.aws.amazon.com/bedrock/latest/userguide/model-ids.html
-var awsModelIDMap = map[string]string{
+var AwsModelIDMap = map[string]string{
 	"claude-instant-1.2":         "anthropic.claude-instant-v1",
 	"claude-2.0":                 "anthropic.claude-v2",
 	"claude-2.1":                 "anthropic.claude-v2:1",
@@ -44,7 +36,7 @@ var awsModelIDMap = map[string]string{
 }
 
 func awsModelID(requestModel string) (string, error) {
-	if awsModelID, ok := awsModelIDMap[requestModel]; ok {
+	if awsModelID, ok := AwsModelIDMap[requestModel]; ok {
 		return awsModelID, nil
 	}
 
@@ -54,7 +46,7 @@ func awsModelID(requestModel string) (string, error) {
 func Handler(c *gin.Context, awsCli *bedrockruntime.Client, modelName string) (*relaymodel.ErrorWithStatusCode, *relaymodel.Usage) {
 	awsModelId, err := awsModelID(c.GetString(ctxkey.RequestModel))
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "awsModelID")), nil
+		return utils.WrapErr(errors.Wrap(err, "awsModelID")), nil
 	}
 
 	awsReq := &bedrockruntime.InvokeModelInput{
@@ -65,30 +57,30 @@ func Handler(c *gin.Context, awsCli *bedrockruntime.Client, modelName string) (*
 
 	claudeReq_, ok := c.Get(ctxkey.ConvertedRequest)
 	if !ok {
-		return wrapErr(errors.New("request not found")), nil
+		return utils.WrapErr(errors.New("request not found")), nil
 	}
 	claudeReq := claudeReq_.(*anthropic.Request)
 	awsClaudeReq := &Request{
 		AnthropicVersion: "bedrock-2023-05-31",
 	}
 	if err = copier.Copy(awsClaudeReq, claudeReq); err != nil {
-		return wrapErr(errors.Wrap(err, "copy request")), nil
+		return utils.WrapErr(errors.Wrap(err, "copy request")), nil
 	}
 
 	awsReq.Body, err = json.Marshal(awsClaudeReq)
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "marshal request")), nil
+		return utils.WrapErr(errors.Wrap(err, "marshal request")), nil
 	}
 
 	awsResp, err := awsCli.InvokeModel(c.Request.Context(), awsReq)
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "InvokeModel")), nil
+		return utils.WrapErr(errors.Wrap(err, "InvokeModel")), nil
 	}
 
 	claudeResponse := new(anthropic.Response)
 	err = json.Unmarshal(awsResp.Body, claudeResponse)
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "unmarshal response")), nil
+		return utils.WrapErr(errors.Wrap(err, "unmarshal response")), nil
 	}
 
 	openaiResp := anthropic.ResponseClaude2OpenAI(claudeResponse)
@@ -108,7 +100,7 @@ func StreamHandler(c *gin.Context, awsCli *bedrockruntime.Client) (*relaymodel.E
 	createdTime := helper.GetTimestamp()
 	awsModelId, err := awsModelID(c.GetString(ctxkey.RequestModel))
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "awsModelID")), nil
+		return utils.WrapErr(errors.Wrap(err, "awsModelID")), nil
 	}
 
 	awsReq := &bedrockruntime.InvokeModelWithResponseStreamInput{
@@ -119,7 +111,7 @@ func StreamHandler(c *gin.Context, awsCli *bedrockruntime.Client) (*relaymodel.E
 
 	claudeReq_, ok := c.Get(ctxkey.ConvertedRequest)
 	if !ok {
-		return wrapErr(errors.New("request not found")), nil
+		return utils.WrapErr(errors.New("request not found")), nil
 	}
 	claudeReq := claudeReq_.(*anthropic.Request)
 
@@ -127,16 +119,16 @@ func StreamHandler(c *gin.Context, awsCli *bedrockruntime.Client) (*relaymodel.E
 		AnthropicVersion: "bedrock-2023-05-31",
 	}
 	if err = copier.Copy(awsClaudeReq, claudeReq); err != nil {
-		return wrapErr(errors.Wrap(err, "copy request")), nil
+		return utils.WrapErr(errors.Wrap(err, "copy request")), nil
 	}
 	awsReq.Body, err = json.Marshal(awsClaudeReq)
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "marshal request")), nil
+		return utils.WrapErr(errors.Wrap(err, "marshal request")), nil
 	}
 
 	awsResp, err := awsCli.InvokeModelWithResponseStream(c.Request.Context(), awsReq)
 	if err != nil {
-		return wrapErr(errors.Wrap(err, "InvokeModelWithResponseStream")), nil
+		return utils.WrapErr(errors.Wrap(err, "InvokeModelWithResponseStream")), nil
 	}
 	stream := awsResp.GetStream()
 	defer stream.Close()

relay/adaptor/aws/llama3/adapter.go

@@ -0,0 +1,37 @@
+package aws
+
+import (
+	"github.com/aws/aws-sdk-go-v2/service/bedrockruntime"
+	"github.com/songquanpeng/one-api/common/ctxkey"
+
+	"github.com/gin-gonic/gin"
+	"github.com/pkg/errors"
+	"github.com/songquanpeng/one-api/relay/adaptor/aws/utils"
+	"github.com/songquanpeng/one-api/relay/meta"
+	"github.com/songquanpeng/one-api/relay/model"
+)
+
+var _ utils.AwsAdapter = new(Adaptor)
+
+type Adaptor struct {
+}
+
+func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+
+	llamaReq := ConvertRequest(*request)
+	c.Set(ctxkey.RequestModel, request.Model)
+	c.Set(ctxkey.ConvertedRequest, llamaReq)
+	return llamaReq, nil
+}
+
+func (a *Adaptor) DoResponse(c *gin.Context, awsCli *bedrockruntime.Client, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode) {
+	if meta.IsStream {
+		err, usage = StreamHandler(c, awsCli)
+	} else {
+		err, usage = Handler(c, awsCli, meta.ActualModelName)
+	}
+	return
+}

relay/adaptor/aws/llama3/main.go

@@ -0,0 +1,231 @@
+// Package aws provides the AWS adaptor for the relay service.
+package aws
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"text/template"
+
+	"github.com/songquanpeng/one-api/common/ctxkey"
+	"github.com/songquanpeng/one-api/common/random"
+
+	"github.com/aws/aws-sdk-go-v2/aws"
+	"github.com/aws/aws-sdk-go-v2/service/bedrockruntime"
+	"github.com/aws/aws-sdk-go-v2/service/bedrockruntime/types"
+	"github.com/gin-gonic/gin"
+	"github.com/pkg/errors"
+	"github.com/songquanpeng/one-api/common"
+	"github.com/songquanpeng/one-api/common/helper"
+	"github.com/songquanpeng/one-api/common/logger"
+	"github.com/songquanpeng/one-api/relay/adaptor/aws/utils"
+	"github.com/songquanpeng/one-api/relay/adaptor/openai"
+	relaymodel "github.com/songquanpeng/one-api/relay/model"
+)
+
+// Only support llama-3-8b and llama-3-70b instruction models
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-ids.html
+var AwsModelIDMap = map[string]string{
+	"llama3-8b-8192":  "meta.llama3-8b-instruct-v1:0",
+	"llama3-70b-8192": "meta.llama3-70b-instruct-v1:0",
+}
+
+func awsModelID(requestModel string) (string, error) {
+	if awsModelID, ok := AwsModelIDMap[requestModel]; ok {
+		return awsModelID, nil
+	}
+
+	return "", errors.Errorf("model %s not found", requestModel)
+}
+
+// promptTemplate with range
+const promptTemplate = `<|begin_of_text|>{{range .Messages}}<|start_header_id|>{{.Role}}<|end_header_id|>{{.StringContent}}<|eot_id|>{{end}}<|start_header_id|>assistant<|end_header_id|>
+`
+
+var promptTpl = template.Must(template.New("llama3-chat").Parse(promptTemplate))
+
+func RenderPrompt(messages []relaymodel.Message) string {
+	var buf bytes.Buffer
+	err := promptTpl.Execute(&buf, struct{ Messages []relaymodel.Message }{messages})
+	if err != nil {
+		logger.SysError("error rendering prompt messages: " + err.Error())
+	}
+	return buf.String()
+}
+
+func ConvertRequest(textRequest relaymodel.GeneralOpenAIRequest) *Request {
+	llamaRequest := Request{
+		MaxGenLen:   textRequest.MaxTokens,
+		Temperature: textRequest.Temperature,
+		TopP:        textRequest.TopP,
+	}
+	if llamaRequest.MaxGenLen == 0 {
+		llamaRequest.MaxGenLen = 2048
+	}
+	prompt := RenderPrompt(textRequest.Messages)
+	llamaRequest.Prompt = prompt
+	return &llamaRequest
+}
+
+func Handler(c *gin.Context, awsCli *bedrockruntime.Client, modelName string) (*relaymodel.ErrorWithStatusCode, *relaymodel.Usage) {
+	awsModelId, err := awsModelID(c.GetString(ctxkey.RequestModel))
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "awsModelID")), nil
+	}
+
+	awsReq := &bedrockruntime.InvokeModelInput{
+		ModelId:     aws.String(awsModelId),
+		Accept:      aws.String("application/json"),
+		ContentType: aws.String("application/json"),
+	}
+
+	llamaReq, ok := c.Get(ctxkey.ConvertedRequest)
+	if !ok {
+		return utils.WrapErr(errors.New("request not found")), nil
+	}
+
+	awsReq.Body, err = json.Marshal(llamaReq)
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "marshal request")), nil
+	}
+
+	awsResp, err := awsCli.InvokeModel(c.Request.Context(), awsReq)
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "InvokeModel")), nil
+	}
+
+	var llamaResponse Response
+	err = json.Unmarshal(awsResp.Body, &llamaResponse)
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "unmarshal response")), nil
+	}
+
+	openaiResp := ResponseLlama2OpenAI(&llamaResponse)
+	openaiResp.Model = modelName
+	usage := relaymodel.Usage{
+		PromptTokens:     llamaResponse.PromptTokenCount,
+		CompletionTokens: llamaResponse.GenerationTokenCount,
+		TotalTokens:      llamaResponse.PromptTokenCount + llamaResponse.GenerationTokenCount,
+	}
+	openaiResp.Usage = usage
+
+	c.JSON(http.StatusOK, openaiResp)
+	return nil, &usage
+}
+
+func ResponseLlama2OpenAI(llamaResponse *Response) *openai.TextResponse {
+	var responseText string
+	if len(llamaResponse.Generation) > 0 {
+		responseText = llamaResponse.Generation
+	}
+	choice := openai.TextResponseChoice{
+		Index: 0,
+		Message: relaymodel.Message{
+			Role:    "assistant",
+			Content: responseText,
+			Name:    nil,
+		},
+		FinishReason: llamaResponse.StopReason,
+	}
+	fullTextResponse := openai.TextResponse{
+		Id:      fmt.Sprintf("chatcmpl-%s", random.GetUUID()),
+		Object:  "chat.completion",
+		Created: helper.GetTimestamp(),
+		Choices: []openai.TextResponseChoice{choice},
+	}
+	return &fullTextResponse
+}
+
+func StreamHandler(c *gin.Context, awsCli *bedrockruntime.Client) (*relaymodel.ErrorWithStatusCode, *relaymodel.Usage) {
+	createdTime := helper.GetTimestamp()
+	awsModelId, err := awsModelID(c.GetString(ctxkey.RequestModel))
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "awsModelID")), nil
+	}
+
+	awsReq := &bedrockruntime.InvokeModelWithResponseStreamInput{
+		ModelId:     aws.String(awsModelId),
+		Accept:      aws.String("application/json"),
+		ContentType: aws.String("application/json"),
+	}
+
+	llamaReq, ok := c.Get(ctxkey.ConvertedRequest)
+	if !ok {
+		return utils.WrapErr(errors.New("request not found")), nil
+	}
+
+	awsReq.Body, err = json.Marshal(llamaReq)
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "marshal request")), nil
+	}
+
+	awsResp, err := awsCli.InvokeModelWithResponseStream(c.Request.Context(), awsReq)
+	if err != nil {
+		return utils.WrapErr(errors.Wrap(err, "InvokeModelWithResponseStream")), nil
+	}
+	stream := awsResp.GetStream()
+	defer stream.Close()
+
+	c.Writer.Header().Set("Content-Type", "text/event-stream")
+	var usage relaymodel.Usage
+	c.Stream(func(w io.Writer) bool {
+		event, ok := <-stream.Events()
+		if !ok {
+			c.Render(-1, common.CustomEvent{Data: "data: [DONE]"})
+			return false
+		}
+
+		switch v := event.(type) {
+		case *types.ResponseStreamMemberChunk:
+			var llamaResp StreamResponse
+			err := json.NewDecoder(bytes.NewReader(v.Value.Bytes)).Decode(&llamaResp)
+			if err != nil {
+				logger.SysError("error unmarshalling stream response: " + err.Error())
+				return false
+			}
+
+			if llamaResp.PromptTokenCount > 0 {
+				usage.PromptTokens = llamaResp.PromptTokenCount
+			}
+			if llamaResp.StopReason == "stop" {
+				usage.CompletionTokens = llamaResp.GenerationTokenCount
+				usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+			}
+			response := StreamResponseLlama2OpenAI(&llamaResp)
+			response.Id = fmt.Sprintf("chatcmpl-%s", random.GetUUID())
+			response.Model = c.GetString(ctxkey.OriginalModel)
+			response.Created = createdTime
+			jsonStr, err := json.Marshal(response)
+			if err != nil {
+				logger.SysError("error marshalling stream response: " + err.Error())
+				return true
+			}
+			c.Render(-1, common.CustomEvent{Data: "data: " + string(jsonStr)})
+			return true
+		case *types.UnknownUnionMember:
+			fmt.Println("unknown tag:", v.Tag)
+			return false
+		default:
+			fmt.Println("union is nil or unknown type")
+			return false
+		}
+	})
+
+	return nil, &usage
+}
+
+func StreamResponseLlama2OpenAI(llamaResponse *StreamResponse) *openai.ChatCompletionsStreamResponse {
+	var choice openai.ChatCompletionsStreamResponseChoice
+	choice.Delta.Content = llamaResponse.Generation
+	choice.Delta.Role = "assistant"
+	finishReason := llamaResponse.StopReason
+	if finishReason != "null" {
+		choice.FinishReason = &finishReason
+	}
+	var openaiResponse openai.ChatCompletionsStreamResponse
+	openaiResponse.Object = "chat.completion.chunk"
+	openaiResponse.Choices = []openai.ChatCompletionsStreamResponseChoice{choice}
+	return &openaiResponse
+}

relay/adaptor/aws/llama3/main_test.go

@@ -0,0 +1,45 @@
+package aws_test
+
+import (
+	"testing"
+
+	aws "github.com/songquanpeng/one-api/relay/adaptor/aws/llama3"
+	relaymodel "github.com/songquanpeng/one-api/relay/model"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestRenderPrompt(t *testing.T) {
+	messages := []relaymodel.Message{
+		{
+			Role:    "user",
+			Content: "What's your name?",
+		},
+	}
+	prompt := aws.RenderPrompt(messages)
+	expected := `<|begin_of_text|><|start_header_id|>user<|end_header_id|>What's your name?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+`
+	assert.Equal(t, expected, prompt)
+
+	messages = []relaymodel.Message{
+		{
+			Role:    "system",
+			Content: "Your name is Kat. You are a detective.",
+		},
+		{
+			Role:    "user",
+			Content: "What's your name?",
+		},
+		{
+			Role:    "assistant",
+			Content: "Kat",
+		},
+		{
+			Role:    "user",
+			Content: "What's your job?",
+		},
+	}
+	prompt = aws.RenderPrompt(messages)
+	expected = `<|begin_of_text|><|start_header_id|>system<|end_header_id|>Your name is Kat. You are a detective.<|eot_id|><|start_header_id|>user<|end_header_id|>What's your name?<|eot_id|><|start_header_id|>assistant<|end_header_id|>Kat<|eot_id|><|start_header_id|>user<|end_header_id|>What's your job?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+`
+	assert.Equal(t, expected, prompt)
+}

relay/adaptor/aws/llama3/model.go

@@ -0,0 +1,29 @@
+package aws
+
+// Request is the request to AWS Llama3
+//
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-meta.html
+type Request struct {
+	Prompt      string  `json:"prompt"`
+	MaxGenLen   int     `json:"max_gen_len,omitempty"`
+	Temperature float64 `json:"temperature,omitempty"`
+	TopP        float64 `json:"top_p,omitempty"`
+}
+
+// Response is the response from AWS Llama3
+//
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-meta.html
+type Response struct {
+	Generation           string `json:"generation"`
+	PromptTokenCount     int    `json:"prompt_token_count"`
+	GenerationTokenCount int    `json:"generation_token_count"`
+	StopReason           string `json:"stop_reason"`
+}
+
+// {'generation': 'Hi', 'prompt_token_count': 15, 'generation_token_count': 1, 'stop_reason': None}
+type StreamResponse struct {
+	Generation           string `json:"generation"`
+	PromptTokenCount     int    `json:"prompt_token_count"`
+	GenerationTokenCount int    `json:"generation_token_count"`
+	StopReason           string `json:"stop_reason"`
+}

relay/adaptor/aws/registry.go

@@ -0,0 +1,39 @@
+package aws
+
+import (
+	claude "github.com/songquanpeng/one-api/relay/adaptor/aws/claude"
+	llama3 "github.com/songquanpeng/one-api/relay/adaptor/aws/llama3"
+	"github.com/songquanpeng/one-api/relay/adaptor/aws/utils"
+)
+
+type AwsModelType int
+
+const (
+	AwsClaude AwsModelType = iota + 1
+	AwsLlama3
+)
+
+var (
+	adaptors = map[string]AwsModelType{}
+)
+
+func init() {
+	for model := range claude.AwsModelIDMap {
+		adaptors[model] = AwsClaude
+	}
+	for model := range llama3.AwsModelIDMap {
+		adaptors[model] = AwsLlama3
+	}
+}
+
+func GetAdaptor(model string) utils.AwsAdapter {
+	adaptorType := adaptors[model]
+	switch adaptorType {
+	case AwsClaude:
+		return &claude.Adaptor{}
+	case AwsLlama3:
+		return &llama3.Adaptor{}
+	default:
+		return nil
+	}
+}

relay/adaptor/aws/utils/adaptor.go

@@ -0,0 +1,51 @@
+package utils
+
+import (
+	"errors"
+	"io"
+	"net/http"
+
+	"github.com/aws/aws-sdk-go-v2/aws"
+	"github.com/aws/aws-sdk-go-v2/credentials"
+	"github.com/aws/aws-sdk-go-v2/service/bedrockruntime"
+	"github.com/gin-gonic/gin"
+	"github.com/songquanpeng/one-api/relay/meta"
+	"github.com/songquanpeng/one-api/relay/model"
+)
+
+type AwsAdapter interface {
+	ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error)
+	DoResponse(c *gin.Context, awsCli *bedrockruntime.Client, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode)
+}
+
+type Adaptor struct {
+	Meta      *meta.Meta
+	AwsClient *bedrockruntime.Client
+}
+
+func (a *Adaptor) Init(meta *meta.Meta) {
+	a.Meta = meta
+	a.AwsClient = bedrockruntime.New(bedrockruntime.Options{
+		Region:      meta.Config.Region,
+		Credentials: aws.NewCredentialsCache(credentials.NewStaticCredentialsProvider(meta.Config.AK, meta.Config.SK, "")),
+	})
+}
+
+func (a *Adaptor) GetRequestURL(meta *meta.Meta) (string, error) {
+	return "", nil
+}
+
+func (a *Adaptor) SetupRequestHeader(c *gin.Context, req *http.Request, meta *meta.Meta) error {
+	return nil
+}
+
+func (a *Adaptor) ConvertImageRequest(request *model.ImageRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+	return request, nil
+}
+
+func (a *Adaptor) DoRequest(c *gin.Context, meta *meta.Meta, requestBody io.Reader) (*http.Response, error) {
+	return nil, nil
+}

relay/adaptor/aws/utils/utils.go

@@ -0,0 +1,16 @@
+package utils
+
+import (
+	"net/http"
+
+	relaymodel "github.com/songquanpeng/one-api/relay/model"
+)
+
+func WrapErr(err error) *relaymodel.ErrorWithStatusCode {
+	return &relaymodel.ErrorWithStatusCode{
+		StatusCode: http.StatusInternalServerError,
+		Error: relaymodel.Error{
+			Message: err.Error(),
+		},
+	}
+}

relay/adaptor/cloudflare/adaptor.go

@@ -10,6 +10,7 @@ import (
 	"github.com/songquanpeng/one-api/relay/adaptor"
 	"github.com/songquanpeng/one-api/relay/meta"
 	"github.com/songquanpeng/one-api/relay/model"
+	"github.com/songquanpeng/one-api/relay/relaymode"
 )
 
 type Adaptor struct {
@@ -28,7 +29,14 @@ func (a *Adaptor) Init(meta *meta.Meta) {
 }
 
 func (a *Adaptor) GetRequestURL(meta *meta.Meta) (string, error) {
-	return fmt.Sprintf("%s/client/v4/accounts/%s/ai/run/%s", meta.BaseURL, meta.Config.UserID, meta.ActualModelName), nil
+	switch meta.Mode {
+	case relaymode.ChatCompletions:
+		return fmt.Sprintf("%s/client/v4/accounts/%s/ai/v1/chat/completions", meta.BaseURL, meta.Config.UserID), nil
+	case relaymode.Embeddings:
+		return fmt.Sprintf("%s/client/v4/accounts/%s/ai/v1/embeddings", meta.BaseURL, meta.Config.UserID), nil
+	default:
+		return fmt.Sprintf("%s/client/v4/accounts/%s/ai/run/%s", meta.BaseURL, meta.Config.UserID, meta.ActualModelName), nil
+	}
 }
 
 func (a *Adaptor) SetupRequestHeader(c *gin.Context, req *http.Request, meta *meta.Meta) error {
@@ -41,7 +49,14 @@ func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.G
 	if request == nil {
 		return nil, errors.New("request is nil")
 	}
-	return ConvertRequest(*request), nil
+	switch relayMode {
+	case relaymode.Completions:
+		return ConvertCompletionsRequest(*request), nil
+	case relaymode.ChatCompletions, relaymode.Embeddings:
+		return request, nil
+	default:
+		return nil, errors.New("not implemented")
+	}
 }
 
 func (a *Adaptor) DoRequest(c *gin.Context, meta *meta.Meta, requestBody io.Reader) (*http.Response, error) {

relay/adaptor/cloudflare/main.go

@@ -3,11 +3,13 @@ package cloudflare
 import (
 	"bufio"
 	"encoding/json"
-	"github.com/songquanpeng/one-api/common/render"
 	"io"
 	"net/http"
 	"strings"
 
+	"github.com/songquanpeng/one-api/common/ctxkey"
+	"github.com/songquanpeng/one-api/common/render"
+
 	"github.com/gin-gonic/gin"
 	"github.com/songquanpeng/one-api/common"
 	"github.com/songquanpeng/one-api/common/helper"
@@ -16,57 +18,23 @@ import (
 	"github.com/songquanpeng/one-api/relay/model"
 )
 
-func ConvertRequest(textRequest model.GeneralOpenAIRequest) *Request {
-	var promptBuilder strings.Builder
-	for _, message := range textRequest.Messages {
-		promptBuilder.WriteString(message.StringContent())
-		promptBuilder.WriteString("\n") // 添加换行符来分隔每个消息
-	}
-
+func ConvertCompletionsRequest(textRequest model.GeneralOpenAIRequest) *Request {
+	p, _ := textRequest.Prompt.(string)
 	return &Request{
+		Prompt:      p,
 		MaxTokens:   textRequest.MaxTokens,
-		Prompt:      promptBuilder.String(),
 		Stream:      textRequest.Stream,
 		Temperature: textRequest.Temperature,
 	}
 }
 
-func ResponseCloudflare2OpenAI(cloudflareResponse *Response) *openai.TextResponse {
-	choice := openai.TextResponseChoice{
-		Index: 0,
-		Message: model.Message{
-			Role:    "assistant",
-			Content: cloudflareResponse.Result.Response,
-		},
-		FinishReason: "stop",
-	}
-	fullTextResponse := openai.TextResponse{
-		Object:  "chat.completion",
-		Created: helper.GetTimestamp(),
-		Choices: []openai.TextResponseChoice{choice},
-	}
-	return &fullTextResponse
-}
-
-func StreamResponseCloudflare2OpenAI(cloudflareResponse *StreamResponse) *openai.ChatCompletionsStreamResponse {
-	var choice openai.ChatCompletionsStreamResponseChoice
-	choice.Delta.Content = cloudflareResponse.Response
-	choice.Delta.Role = "assistant"
-	openaiResponse := openai.ChatCompletionsStreamResponse{
-		Object:  "chat.completion.chunk",
-		Choices: []openai.ChatCompletionsStreamResponseChoice{choice},
-		Created: helper.GetTimestamp(),
-	}
-	return &openaiResponse
-}
-
 func StreamHandler(c *gin.Context, resp *http.Response, promptTokens int, modelName string) (*model.ErrorWithStatusCode, *model.Usage) {
 	scanner := bufio.NewScanner(resp.Body)
 	scanner.Split(bufio.ScanLines)
 
 	common.SetEventStreamHeaders(c)
 	id := helper.GetResponseID(c)
-	responseModel := c.GetString("original_model")
+	responseModel := c.GetString(ctxkey.OriginalModel)
 	var responseText string
 
 	for scanner.Scan() {
@@ -77,22 +45,22 @@ func StreamHandler(c *gin.Context, resp *http.Response, promptTokens int, modelN
 		data = strings.TrimPrefix(data, "data: ")
 		data = strings.TrimSuffix(data, "\r")
 
-		var cloudflareResponse StreamResponse
-		err := json.Unmarshal([]byte(data), &cloudflareResponse)
+		if data == "[DONE]" {
+			break
+		}
+
+		var response openai.ChatCompletionsStreamResponse
+		err := json.Unmarshal([]byte(data), &response)
 		if err != nil {
 			logger.SysError("error unmarshalling stream response: " + err.Error())
 			continue
 		}
-
-		response := StreamResponseCloudflare2OpenAI(&cloudflareResponse)
-		if response == nil {
-			continue
+		for _, v := range response.Choices {
+			v.Delta.Role = "assistant"
+			responseText += v.Delta.StringContent()
 		}
-
-		responseText += cloudflareResponse.Response
 		response.Id = id
-		response.Model = responseModel
-
+		response.Model = modelName
 		err = render.ObjectData(c, response)
 		if err != nil {
 			logger.SysError(err.Error())
@@ -123,22 +91,25 @@ func Handler(c *gin.Context, resp *http.Response, promptTokens int, modelName st
 	if err != nil {
 		return openai.ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
 	}
-	var cloudflareResponse Response
-	err = json.Unmarshal(responseBody, &cloudflareResponse)
+	var response openai.TextResponse
+	err = json.Unmarshal(responseBody, &response)
 	if err != nil {
 		return openai.ErrorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError), nil
 	}
-	fullTextResponse := ResponseCloudflare2OpenAI(&cloudflareResponse)
-	fullTextResponse.Model = modelName
-	usage := openai.ResponseText2Usage(cloudflareResponse.Result.Response, modelName, promptTokens)
-	fullTextResponse.Usage = *usage
-	fullTextResponse.Id = helper.GetResponseID(c)
-	jsonResponse, err := json.Marshal(fullTextResponse)
+	response.Model = modelName
+	var responseText string
+	for _, v := range response.Choices {
+		responseText += v.Message.Content.(string)
+	}
+	usage := openai.ResponseText2Usage(responseText, modelName, promptTokens)
+	response.Usage = *usage
+	response.Id = helper.GetResponseID(c)
+	jsonResponse, err := json.Marshal(response)
 	if err != nil {
 		return openai.ErrorWrapper(err, "marshal_response_body_failed", http.StatusInternalServerError), nil
 	}
 	c.Writer.Header().Set("Content-Type", "application/json")
 	c.Writer.WriteHeader(resp.StatusCode)
-	_, err = c.Writer.Write(jsonResponse)
+	_, _ = c.Writer.Write(jsonResponse)
 	return nil, usage
 }

relay/adaptor/cloudflare/model.go

@@ -1,25 +1,13 @@
 package cloudflare
 
+import "github.com/songquanpeng/one-api/relay/model"
+
 type Request struct {
-	Lora        string  `json:"lora,omitempty"`
-	MaxTokens   int     `json:"max_tokens,omitempty"`
-	Prompt      string  `json:"prompt,omitempty"`
-	Raw         bool    `json:"raw,omitempty"`
-	Stream      bool    `json:"stream,omitempty"`
-	Temperature float64 `json:"temperature,omitempty"`
-}
-
-type Result struct {
-	Response string `json:"response"`
-}
-
-type Response struct {
-	Result   Result   `json:"result"`
-	Success  bool     `json:"success"`
-	Errors   []string `json:"errors"`
-	Messages []string `json:"messages"`
-}
-
-type StreamResponse struct {
-	Response string `json:"response"`
+	Messages    []model.Message `json:"messages,omitempty"`
+	Lora        string          `json:"lora,omitempty"`
+	MaxTokens   int             `json:"max_tokens,omitempty"`
+	Prompt      string          `json:"prompt,omitempty"`
+	Raw         bool            `json:"raw,omitempty"`
+	Stream      bool            `json:"stream,omitempty"`
+	Temperature float64         `json:"temperature,omitempty"`
 }

relay/adaptor/novita/constants.go

@@ -0,0 +1,19 @@
+package novita
+
+// https://novita.ai/llm-api
+
+var ModelList = []string{
+	"meta-llama/llama-3-8b-instruct",
+	"meta-llama/llama-3-70b-instruct",
+	"nousresearch/hermes-2-pro-llama-3-8b",
+	"nousresearch/nous-hermes-llama2-13b",
+	"mistralai/mistral-7b-instruct",
+	"cognitivecomputations/dolphin-mixtral-8x22b",
+	"sao10k/l3-70b-euryale-v2.1",
+	"sophosympatheia/midnight-rose-70b",
+	"gryphe/mythomax-l2-13b",
+	"Nous-Hermes-2-Mixtral-8x7B-DPO",
+	"lzlv_70b",
+	"teknium/openhermes-2.5-mistral-7b",
+	"microsoft/wizardlm-2-8x22b",
+}

relay/adaptor/novita/main.go

@@ -0,0 +1,15 @@
+package novita
+
+import (
+	"fmt"
+
+	"github.com/songquanpeng/one-api/relay/meta"
+	"github.com/songquanpeng/one-api/relay/relaymode"
+)
+
+func GetRequestURL(meta *meta.Meta) (string, error) {
+	if meta.Mode == relaymode.ChatCompletions {
+		return fmt.Sprintf("%s/chat/completions", meta.BaseURL), nil
+	}
+	return "", fmt.Errorf("unsupported relay mode %d for novita", meta.Mode)
+}

relay/adaptor/openai/adaptor.go

@@ -2,18 +2,20 @@ package openai
 
 import (
 	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
 	"github.com/Laisky/errors/v2"
 	"github.com/gin-gonic/gin"
 	"github.com/songquanpeng/one-api/relay/adaptor"
 	"github.com/songquanpeng/one-api/relay/adaptor/doubao"
 	"github.com/songquanpeng/one-api/relay/adaptor/minimax"
+	"github.com/songquanpeng/one-api/relay/adaptor/novita"
 	"github.com/songquanpeng/one-api/relay/channeltype"
 	"github.com/songquanpeng/one-api/relay/meta"
 	"github.com/songquanpeng/one-api/relay/model"
 	"github.com/songquanpeng/one-api/relay/relaymode"
-	"io"
-	"net/http"
-	"strings"
 )
 
 type Adaptor struct {
@@ -48,6 +50,8 @@ func (a *Adaptor) GetRequestURL(meta *meta.Meta) (string, error) {
 		return minimax.GetRequestURL(meta)
 	case channeltype.Doubao:
 		return doubao.GetRequestURL(meta)
+	case channeltype.Novita:
+		return novita.GetRequestURL(meta)
 	default:
 		return GetFullRequestURL(meta.BaseURL, meta.RequestURLPath, meta.ChannelType), nil
 	}

relay/adaptor/openai/compatible.go

@@ -10,6 +10,7 @@ import (
 	"github.com/songquanpeng/one-api/relay/adaptor/minimax"
 	"github.com/songquanpeng/one-api/relay/adaptor/mistral"
 	"github.com/songquanpeng/one-api/relay/adaptor/moonshot"
+	"github.com/songquanpeng/one-api/relay/adaptor/novita"
 	"github.com/songquanpeng/one-api/relay/adaptor/stepfun"
 	"github.com/songquanpeng/one-api/relay/adaptor/togetherai"
 	"github.com/songquanpeng/one-api/relay/channeltype"
@@ -28,6 +29,7 @@ var CompatibleChannels = []int{
 	channeltype.StepFun,
 	channeltype.DeepSeek,
 	channeltype.TogetherAI,
+	channeltype.Novita,
 }
 
 func GetCompatibleChannelMeta(channelType int) (string, []string) {
@@ -56,6 +58,8 @@ func GetCompatibleChannelMeta(channelType int) (string, []string) {
 		return "together.ai", togetherai.ModelList
 	case channeltype.Doubao:
 		return "doubao", doubao.ModelList
+	case channeltype.Novita:
+		return "novita", novita.ModelList
 	default:
 		return "openai", ModelList
 	}