ai-gateway/relay/channel/openai/token.go

package openai

import (
	"errors"
	"fmt"
	"github.com/pkoukk/tiktoken-go"
	"github.com/songquanpeng/one-api/common"
	"github.com/songquanpeng/one-api/common/config"
	"github.com/songquanpeng/one-api/common/image"
	"github.com/songquanpeng/one-api/common/logger"
	"github.com/songquanpeng/one-api/relay/model"
	"math"
	"strings"
)

// tokenEncoderMap won't grow after initialization
var tokenEncoderMap = map[string]*tiktoken.Tiktoken{}
var defaultTokenEncoder *tiktoken.Tiktoken

func InitTokenEncoders() {
	logger.SysLog("initializing token encoders")
	gpt35TokenEncoder, err := tiktoken.EncodingForModel("gpt-3.5-turbo")
	if err != nil {
		logger.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 {
		logger.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
		}
	}
	logger.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 {
			logger.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 config.ApproximateTokenEnabled {
		return int(float64(len(text)) * 0.38)
	}
	return len(tokenEncoder.Encode(text, nil, nil))
}

func CountTokenMessages(messages []model.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 {
							logger.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)
}