feat: support replicate chat models

.github/workflows/ci.yml

@@ -1,13 +1,13 @@
 name: CI
 
 # This setup assumes that you run the unit tests with code coverage in the same
-# workflow that will also print the coverage report as comment to the pull request. 
+# workflow that will also print the coverage report as comment to the pull request.
 # Therefore, you need to trigger this workflow when a pull request is (re)opened or
 # when new code is pushed to the branch of the pull request. In addition, you also
-# need to trigger this workflow when new code is pushed to the main branch because 
+# need to trigger this workflow when new code is pushed to the main branch because
 # we need to upload the code coverage results as artifact for the main branch as
 # well since it will be the baseline code coverage.
-# 
+#
 # We do not want to trigger the workflow for pushes to *any* branch because this
 # would trigger our jobs twice on pull requests (once from "push" event and once
 # from "pull_request->synchronize")
@@ -31,7 +31,7 @@ jobs:
         with:
           go-version: ^1.22
 
-      # When you execute your unit tests, make sure to use the "-coverprofile" flag to write a 
+      # When you execute your unit tests, make sure to use the "-coverprofile" flag to write a
       # coverage profile to a file. You will need the name of the file (e.g. "coverage.txt")
       # in the next step as well as the next job.
       - name: Test

.gitignore

@@ -9,4 +9,5 @@ logs
 data
 /web/node_modules
 cmd.md
-.env
+.env
+/one-api

common/render/render.go

@@ -3,9 +3,10 @@ package render
 import (
 	"encoding/json"
 	"fmt"
+	"strings"
+
 	"github.com/gin-gonic/gin"
 	"github.com/songquanpeng/one-api/common"
-	"strings"
 )
 
 func StringData(c *gin.Context, str string) {

monitor/manage.go

@@ -34,7 +34,7 @@
 		strings.Contains(lowerMessage, "credit") ||
 		strings.Contains(lowerMessage, "balance") ||
 		strings.Contains(lowerMessage, "permission denied") ||
-  	strings.Contains(lowerMessage, "organization has been restricted") || // groq
+		strings.Contains(lowerMessage, "organization has been restricted") || // groq
 		strings.Contains(lowerMessage, "已欠费") {
 		return true
 	}

relay/adaptor.go

@@ -16,6 +16,7 @@ import (
 	"github.com/songquanpeng/one-api/relay/adaptor/openai"
 	"github.com/songquanpeng/one-api/relay/adaptor/palm"
 	"github.com/songquanpeng/one-api/relay/adaptor/proxy"
+	"github.com/songquanpeng/one-api/relay/adaptor/replicate"
 	"github.com/songquanpeng/one-api/relay/adaptor/tencent"
 	"github.com/songquanpeng/one-api/relay/adaptor/vertexai"
 	"github.com/songquanpeng/one-api/relay/adaptor/xunfei"
@@ -61,6 +62,8 @@ func GetAdaptor(apiType int) adaptor.Adaptor {
 		return &vertexai.Adaptor{}
 	case apitype.Proxy:
 		return &proxy.Adaptor{}
+	case apitype.Replicate:
+		return &replicate.Adaptor{}
 	}
 	return nil
 }

relay/adaptor/ollama/main.go

@@ -31,8 +31,8 @@
 			TopP:             request.TopP,
 			FrequencyPenalty: request.FrequencyPenalty,
 			PresencePenalty:  request.PresencePenalty,
-			NumPredict:  	  request.MaxTokens,
-			NumCtx:  	  request.NumCtx,
+			NumPredict:       request.MaxTokens,
+			NumCtx:           request.NumCtx,
 		},
 		Stream: request.Stream,
 	}
@@ -122,7 +122,7 @@
 	for scanner.Scan() {
 		data := scanner.Text()
 		if strings.HasPrefix(data, "}") {
-		    data = strings.TrimPrefix(data, "}") + "}"
+			data = strings.TrimPrefix(data, "}") + "}"
 		}
 
 		var ollamaResponse ChatResponse

relay/adaptor/openai/helper.go

@@ -2,15 +2,16 @@
 
 import (
 	"fmt"
+	"strings"
+
 	"github.com/songquanpeng/one-api/relay/channeltype"
 	"github.com/songquanpeng/one-api/relay/model"
-	"strings"
 )
 
-func ResponseText2Usage(responseText string, modeName string, promptTokens int) *model.Usage {
+func ResponseText2Usage(responseText string, modelName string, promptTokens int) *model.Usage {
 	usage := &model.Usage{}
 	usage.PromptTokens = promptTokens
-	usage.CompletionTokens = CountTokenText(responseText, modeName)
+	usage.CompletionTokens = CountTokenText(responseText, modelName)
 	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
 	return usage
 }

relay/adaptor/openai/util.go

@@ -1,8 +1,16 @@
 package openai
 
-import "github.com/songquanpeng/one-api/relay/model"
+import (
+	"context"
+	"fmt"
+
+	"github.com/songquanpeng/one-api/common/logger"
+	"github.com/songquanpeng/one-api/relay/model"
+)
 
 func ErrorWrapper(err error, code string, statusCode int) *model.ErrorWithStatusCode {
+	logger.Error(context.TODO(), fmt.Sprintf("[%s]%+v", code, err))
+
 	Error := model.Error{
 		Message: err.Error(),
 		Type:    "one_api_error",

relay/adaptor/replicate/adaptor.go

@@ -0,0 +1,136 @@
+package replicate
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"slices"
+	"strings"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/pkg/errors"
+	"github.com/songquanpeng/one-api/common/logger"
+	"github.com/songquanpeng/one-api/relay/adaptor"
+	"github.com/songquanpeng/one-api/relay/adaptor/openai"
+	"github.com/songquanpeng/one-api/relay/meta"
+	"github.com/songquanpeng/one-api/relay/model"
+	"github.com/songquanpeng/one-api/relay/relaymode"
+)
+
+type Adaptor struct {
+	meta *meta.Meta
+}
+
+// ConvertImageRequest implements adaptor.Adaptor.
+func (*Adaptor) ConvertImageRequest(request *model.ImageRequest) (any, error) {
+	return DrawImageRequest{
+		Input: ImageInput{
+			Steps:           25,
+			Prompt:          request.Prompt,
+			Guidance:        3,
+			Seed:            int(time.Now().UnixNano()),
+			SafetyTolerance: 5,
+			NImages:         1, // replicate will always return 1 image
+			Width:           1440,
+			Height:          1440,
+			AspectRatio:     "1:1",
+		},
+	}, nil
+}
+
+func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error) {
+	if !request.Stream {
+		// TODO: support non-stream mode
+		return nil, errors.Errorf("replicate models only support stream mode now, please set stream=true")
+	}
+
+	// Build the prompt from OpenAI messages
+	var promptBuilder strings.Builder
+	for _, message := range request.Messages {
+		switch msgCnt := message.Content.(type) {
+		case string:
+			promptBuilder.WriteString(message.Role)
+			promptBuilder.WriteString(": ")
+			promptBuilder.WriteString(msgCnt)
+			promptBuilder.WriteString("\n")
+		default:
+		}
+	}
+
+	replicateRequest := ReplicateChatRequest{
+		Input: ChatInput{
+			Prompt:           promptBuilder.String(),
+			MaxTokens:        request.MaxTokens,
+			Temperature:      1.0,
+			TopP:             1.0,
+			PresencePenalty:  0.0,
+			FrequencyPenalty: 0.0,
+		},
+	}
+
+	// Map optional fields
+	if request.Temperature != nil {
+		replicateRequest.Input.Temperature = *request.Temperature
+	}
+	if request.TopP != nil {
+		replicateRequest.Input.TopP = *request.TopP
+	}
+	if request.PresencePenalty != nil {
+		replicateRequest.Input.PresencePenalty = *request.PresencePenalty
+	}
+	if request.FrequencyPenalty != nil {
+		replicateRequest.Input.FrequencyPenalty = *request.FrequencyPenalty
+	}
+	if request.MaxTokens > 0 {
+		replicateRequest.Input.MaxTokens = request.MaxTokens
+	} else if request.MaxTokens == 0 {
+		replicateRequest.Input.MaxTokens = 500
+	}
+
+	return replicateRequest, nil
+}
+
+func (a *Adaptor) Init(meta *meta.Meta) {
+	a.meta = meta
+}
+
+func (a *Adaptor) GetRequestURL(meta *meta.Meta) (string, error) {
+	if !slices.Contains(ModelList, meta.OriginModelName) {
+		return "", errors.Errorf("model %s not supported", meta.OriginModelName)
+	}
+
+	return fmt.Sprintf("https://api.replicate.com/v1/models/%s/predictions", meta.OriginModelName), nil
+}
+
+func (a *Adaptor) SetupRequestHeader(c *gin.Context, req *http.Request, meta *meta.Meta) error {
+	adaptor.SetupCommonRequestHeader(c, req, meta)
+	req.Header.Set("Authorization", "Bearer "+meta.APIKey)
+	return nil
+}
+
+func (a *Adaptor) DoRequest(c *gin.Context, meta *meta.Meta, requestBody io.Reader) (*http.Response, error) {
+	logger.Info(c, "send request to replicate")
+	return adaptor.DoRequestHelper(a, c, meta, requestBody)
+}
+
+func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode) {
+	switch meta.Mode {
+	case relaymode.ImagesGenerations:
+		err, usage = ImageHandler(c, resp)
+	case relaymode.ChatCompletions:
+		err, usage = ChatHandler(c, resp)
+	default:
+		err = openai.ErrorWrapper(errors.New("not implemented"), "not_implemented", http.StatusInternalServerError)
+	}
+
+	return
+}
+
+func (a *Adaptor) GetModelList() []string {
+	return ModelList
+}
+
+func (a *Adaptor) GetChannelName() string {
+	return "replicate"
+}