oadmin/new-api
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #1231 from RedwindA/feat/gemini-budget-in-name

Browse Source 
feat(Gemini): implement thinking budget control in model name
This commit is contained in:
Calcium-Ion
2025-06-17 00:03:53 +08:00
committed by GitHub
parent 6b7295bbdf 1294d286ee
commit 246ca40aac
5 changed files with 74 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
relay/channel/gemini/adaptor.go
View File

@@ -72,8 +72,11 @@ func (a *Adaptor) Init(info *relaycommon.RelayInfo) {
func (a *Adaptor) GetRequestURL(info *relaycommon.RelayInfo) (string, error) {
if model_setting.GetGeminiSettings().ThinkingAdapterEnabled {
// suffix -thinking and -nothinking
if strings.HasSuffix(info.OriginModelName, "-thinking") {
// 新增逻辑：处理 -thinking-<budget> 格式
if strings.Contains(info.OriginModelName, "-thinking-") {
parts := strings.Split(info.UpstreamModelName, "-thinking-")
info.UpstreamModelName = parts[0]
} else if strings.HasSuffix(info.OriginModelName, "-thinking") { // 旧的适配
info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-thinking")
} else if strings.HasSuffix(info.OriginModelName, "-nothinking") {
info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-nothinking")

43
relay/channel/gemini/relay-gemini.go
View File

@@ -12,6 +12,7 @@ import (
"one-api/relay/helper"
"one-api/service"
"one-api/setting/model_setting"
"strconv"
"strings"
"unicode/utf8"
@@ -36,6 +37,13 @@ var geminiSupportedMimeTypes = map[string]bool{
"video/flv": true,
}
// Gemini 允许的思考预算范围
const (
pro25MinBudget = 128
pro25MaxBudget = 32768
flash25MaxBudget = 24576
)
// Setting safety to the lowest possible values since Gemini is already powerless enough
func CovertGemini2OpenAI(textRequest dto.GeneralOpenAIRequest, info *relaycommon.RelayInfo) (*GeminiChatRequest, error) {
@@ -57,7 +65,40 @@ func CovertGemini2OpenAI(textRequest dto.GeneralOpenAIRequest, info *relaycommon
}
if model_setting.GetGeminiSettings().ThinkingAdapterEnabled {
if strings.HasSuffix(info.OriginModelName, "-thinking") {
// 新增逻辑：处理 -thinking-<budget> 格式
if strings.Contains(info.OriginModelName, "-thinking-") {
parts := strings.SplitN(info.OriginModelName, "-thinking-", 2)
if len(parts) == 2 && parts[1] != "" {
if budgetTokens, err := strconv.Atoi(parts[1]); err == nil {
// 从模型名称成功解析预算
isNew25Pro := strings.HasPrefix(info.OriginModelName, "gemini-2.5-pro") &&
!strings.HasPrefix(info.OriginModelName, "gemini-2.5-pro-preview-05-06") &&
!strings.HasPrefix(info.OriginModelName, "gemini-2.5-pro-preview-03-25")
if isNew25Pro {
// 新的2.5pro模型：ThinkingBudget范围为128-32768
if budgetTokens < pro25MinBudget {
budgetTokens = pro25MinBudget
} else if budgetTokens > pro25MaxBudget {
budgetTokens = pro25MaxBudget
}
} else {
// 其他模型：ThinkingBudget范围为0-24576
if budgetTokens < 0 {
budgetTokens = 0
} else if budgetTokens > flash25MaxBudget {
budgetTokens = flash25MaxBudget
}
}
geminiRequest.GenerationConfig.ThinkingConfig = &GeminiThinkingConfig{
ThinkingBudget: common.GetPointer(budgetTokens),
IncludeThoughts: true,
}
}
// 如果解析失败，则不设置ThinkingConfig，静默处理
}
} else if strings.HasSuffix(info.OriginModelName, "-thinking") { // 保留旧逻辑以兼容
// 硬编码不支持 ThinkingBudget 的旧模型
unsupportedModels := []string{
"gemini-2.5-pro-preview-05-06",

19
setting/operation_setting/model-ratio.go
View File

@@ -142,6 +142,11 @@ var defaultModelRatio = map[string]float64{
"gemini-2.5-flash-preview-04-17": 0.075,
"gemini-2.5-flash-preview-04-17-thinking": 0.075,
"gemini-2.5-flash-preview-04-17-nothinking": 0.075,
"gemini-2.5-flash-preview-05-20": 0.075,
"gemini-2.5-flash-preview-05-20-thinking": 0.075,
"gemini-2.5-flash-preview-05-20-nothinking": 0.075,
"gemini-2.5-flash-thinking-*": 0.075, // 用于为后续所有2.5 flash thinking budget 模型设置默认倍率
"gemini-2.5-pro-thinking-*": 0.625, // 用于为后续所有2.5 pro thinking budget 模型设置默认倍率
"text-embedding-004": 0.001,
"chatglm_turbo": 0.3572, // ￥0.005 / 1k tokens
"chatglm_pro": 0.7143, // ￥0.01 / 1k tokens
@@ -342,10 +347,20 @@ func UpdateModelRatioByJSONString(jsonStr string) error {
return json.Unmarshal([]byte(jsonStr), &modelRatioMap)
}
// 处理带有思考预算的模型名称，方便统一定价
func handleThinkingBudgetModel(name, prefix, wildcard string) string {
if strings.HasPrefix(name, prefix) && strings.Contains(name, "-thinking-") {
return wildcard
}
return name
}
func GetModelRatio(name string) (float64, bool) {
modelRatioMapMutex.RLock()
defer modelRatioMapMutex.RUnlock()
name = handleThinkingBudgetModel(name, "gemini-2.5-flash", "gemini-2.5-flash-thinking-*")
name = handleThinkingBudgetModel(name, "gemini-2.5-pro", "gemini-2.5-pro-thinking-*")
if strings.HasPrefix(name, "gpt-4-gizmo") {
name = "gpt-4-gizmo-*"
}
@@ -470,9 +485,9 @@ func getHardcodedCompletionModelRatio(name string) (float64, bool) {
return 4, true
} else if strings.HasPrefix(name, "gemini-2.0") {
return 4, true
} else if strings.HasPrefix(name, "gemini-2.5-pro-preview") {
} else if strings.HasPrefix(name, "gemini-2.5-pro") { // 移除preview来增加兼容性，这里假设正式版的倍率和preview一致
return 8, true
} else if strings.HasPrefix(name, "gemini-2.5-flash-preview") {
} else if strings.HasPrefix(name, "gemini-2.5-flash") { // 同上
if strings.HasSuffix(name, "-nothinking") {
return 4, false
} else {

6
web/src/i18n/locales/en.json
View File

@@ -1373,6 +1373,12 @@
"示例": "Example",
"缺省 MaxTokens": "Default MaxTokens",
"启用Claude思考适配（-thinking后缀）": "Enable Claude thinking adaptation (-thinking suffix)",
"和Claude不同，默认情况下Gemini的思考模型会自动决定要不要思考，就算不开启适配模型也可以正常使用，": "Unlike Claude, Gemini's thinking model automatically decides whether to think by default, and can be used normally even without enabling the adaptation model.",
"如果您需要计费，推荐设置无后缀模型价格按思考价格设置。": "If you need billing, it is recommended to set the no-suffix model price according to the thinking price.",
"支持使用 gemini-2.5-pro-preview-06-05-thinking-128 格式来精确传递思考预算。": "Supports using gemini-2.5-pro-preview-06-05-thinking-128 format to precisely pass thinking budget.",
"启用Gemini思考后缀适配": "Enable Gemini thinking suffix adaptation",
"适配-thinking、-thinking-预算数字和-nothinking后缀": "Adapt -thinking, -thinking-budgetNumber, and -nothinking suffixes",
"思考预算占比": "Thinking budget ratio",
"Claude思考适配 BudgetTokens = MaxTokens * BudgetTokens 百分比": "Claude thinking adaptation BudgetTokens = MaxTokens * BudgetTokens percentage",
"思考适配 BudgetTokens 百分比": "Thinking adaptation BudgetTokens percentage",
"0.1-1之间的小数": "Decimal between 0.1 and 1",

7
web/src/pages/Setting/Model/SettingGeminiModel.js
View File

@@ -173,7 +173,8 @@ export default function SettingGeminiModel(props) {
<Text>
{t(
"和Claude不同，默认情况下Gemini的思考模型会自动决定要不要思考，就算不开启适配模型也可以正常使用，" +
"如果您需要计费，推荐设置无后缀模型价格按思考价格设置"
"如果您需要计费，推荐设置无后缀模型价格按思考价格设置。" +
"支持使用 gemini-2.5-pro-preview-06-05-thinking-128 格式来精确传递思考预算。"
)}
</Text>
</Col>
@@ -183,7 +184,7 @@ export default function SettingGeminiModel(props) {
<Form.Switch
label={t('启用Gemini思考后缀适配')}
field={'gemini.thinking_adapter_enabled'}
extraText={"适配-thinking和-nothinking后缀"}
extraText={t('适配 -thinking、-thinking-预算数字 和 -nothinking 后缀')}
onChange={(value) =>
setInputs({
...inputs,
@@ -205,7 +206,7 @@ export default function SettingGeminiModel(props) {
<Row>
<Col xs={24} sm={12} md={8} lg={8} xl={8}>
<Form.InputNumber
label={t('请求模型带-thinking后缀的BudgetTokens数（超出24576的部分将被忽略）')}
label={t('思考预算占比')}
field={'gemini.thinking_adapter_budget_tokens_percentage'}
initValue={''}
extraText={t('0.1-1之间的小数')}