feat(relay):deepseek新增response端点
This commit is contained in:
qian committed
1 parent
a24f9d29db
commit
8a5ef91625
2 files changed
+111
-8
No files matched your search
@@ -415,16 +415,28 @@ func (a *Adaptor) handleClaudeNativeNonStream(ctx context.Context, resp *http.Re
|
||||
|
||||
// handleClaudeNativeStream 直通 Claude 流式响应
|
||||
func (a *Adaptor) handleClaudeNativeStream(ctx context.Context, resp *http.Response, info *common.RelayInfo, writer http.ResponseWriter) (*common.Usage, error) {
|
||||
defer resp.Body.Close()
|
||||
|
||||
if resp.StatusCode != http.StatusOK {
|
||||
body, _ := io.ReadAll(resp.Body)
|
||||
resp.Body.Close()
|
||||
return nil, constant.NewUpstreamError(resp.StatusCode, string(body), nil).WithRetryAfter(constant.RetryAfterFromHeader(resp.Header))
|
||||
}
|
||||
|
||||
// 创建可取消的上下文，用于在客户端断开时立即中止上游读取
|
||||
streamCtx, cancelStream := context.WithCancel(ctx)
|
||||
defer cancelStream()
|
||||
|
||||
// cleanup 函数：关闭上游连接，停止 token 生成
|
||||
cleanup := func() {
|
||||
cancelStream()
|
||||
if resp.Body != nil {
|
||||
resp.Body.Close()
|
||||
}
|
||||
}
|
||||
|
||||
helper.SetEventStreamHeaders(writer)
|
||||
writer = helper.NewSafeWriter(writer)
|
||||
defer helper.PingTicker(writer, 15*time.Second)()
|
||||
stopPing := helper.PingTicker(writer, 15*time.Second)
|
||||
defer stopPing()
|
||||
|
||||
reader := bufio.NewReaderSize(resp.Body, 64*1024)
|
||||
var usage dto.ClaudeUsage
|
||||
@@ -432,8 +444,10 @@ func (a *Adaptor) handleClaudeNativeStream(ctx context.Context, resp *http.Respo
|
||||
|
||||
for {
|
||||
select {
|
||||
case <-ctx.Done():
|
||||
info.StreamStatus.SetEndReason(common.StreamEndReasonClientGone, ctx.Err())
|
||||
case <-streamCtx.Done():
|
||||
// 客户端断开：立即关闭上游连接，停止生成
|
||||
cleanup()
|
||||
info.StreamStatus.SetEndReason(common.StreamEndReasonClientGone, streamCtx.Err())
|
||||
interruptedUsage := buildUsageFromClaude(&usage)
|
||||
helper.ApplyInterruptedUsageFallback(info, interruptedUsage, transferredTextLen)
|
||||
return interruptedUsage, common.ErrStreamInterrupted
|
||||
@@ -446,7 +460,12 @@ func (a *Adaptor) handleClaudeNativeStream(ctx context.Context, resp *http.Respo
|
||||
break
|
||||
}
|
||||
info.StreamStatus.SetEndReason(common.StreamEndReasonError, err)
|
||||
return buildUsageFromClaude(&usage), fmt.Errorf("stream read error: %w", err)
|
||||
// 已有部分输出时按部分成功处理（避免标记为完全失败）
|
||||
interruptedUsage := buildUsageFromClaude(&usage)
|
||||
if transferredTextLen > 0 {
|
||||
helper.ApplyInterruptedUsageFallback(info, interruptedUsage, transferredTextLen)
|
||||
}
|
||||
return interruptedUsage, fmt.Errorf("upstream stream interrupted: %w", err)
|
||||
}
|
||||
|
||||
if strings.HasPrefix(line, "data:") {
|
||||
@@ -503,6 +522,8 @@ func (a *Adaptor) handleClaudeNativeStream(ctx context.Context, resp *http.Respo
|
||||
}
|
||||
|
||||
if _, err := writer.Write([]byte(line)); err != nil {
|
||||
// 写入客户端失败：立即关闭上游连接，停止生成
|
||||
cleanup()
|
||||
info.StreamStatus.SetEndReason(common.StreamEndReasonClientGone, err)
|
||||
interruptedUsage := buildUsageFromClaude(&usage)
|
||||
helper.ApplyInterruptedUsageFallback(info, interruptedUsage, transferredTextLen)
|
||||
|
||||
@@ -26,7 +26,7 @@ func (a *Adaptor) Init(info *common.RelayInfo) {
|
||||
}
|
||||
|
||||
// GetRequestURL 构建上游请求 URL。
|
||||
// Completions(FIM) 模式使用 /beta/completions 端点，其余走标准 OpenAI 路径。
|
||||
// Completions(FIM) 模式使用 /beta/completions 端点，Responses 使用 /v1/responses，其余走标准 OpenAI 路径。
|
||||
func (a *Adaptor) GetRequestURL(info *common.RelayInfo) (string, error) {
|
||||
baseURL := strings.TrimSuffix(info.ChannelMeta.BaseURL, "/")
|
||||
|
||||
@@ -41,6 +41,8 @@ func (a *Adaptor) GetRequestURL(info *common.RelayInfo) (string, error) {
|
||||
return baseURL + "/v1/chat/completions", nil
|
||||
case constant.RelayModeClaudeMessages:
|
||||
return baseURL + "/anthropic/v1/messages", nil
|
||||
case constant.RelayModeResponses:
|
||||
return baseURL + "/v1/responses", nil
|
||||
case constant.RelayModeEmbeddings:
|
||||
return baseURL + "/v1/embeddings", nil
|
||||
default:
|
||||
@@ -56,13 +58,19 @@ func (a *Adaptor) SetupRequestHeader(header http.Header, info *common.RelayInfo)
|
||||
}
|
||||
|
||||
// ConvertRequest 转换请求体。
|
||||
// Claude 入站直接透传到 Anthropic 兼容端点，其他格式先转为 OpenAI 再做 DeepSeek 特有适配。
|
||||
// Claude 入站直接透传到 Anthropic 兼容端点，Responses 模式处理 reasoning_effort 映射，其他格式先转为 OpenAI 再做 DeepSeek 特有适配。
|
||||
func (a *Adaptor) ConvertRequest(ctx context.Context, info *common.RelayInfo, requestBody []byte) (io.Reader, error) {
|
||||
// Claude 入站：透传请求体，仅做模型映射和思考参数注入
|
||||
if info.InboundFormat == constant.RelayFormatClaude {
|
||||
return convertClaudeRequestForDeepSeek(requestBody, info)
|
||||
}
|
||||
|
||||
// Responses 模式：处理 DeepSeek V4 的 reasoning_effort 映射
|
||||
if constant.RelayMode(info.RelayMode) == constant.RelayModeResponses ||
|
||||
constant.RelayMode(info.RelayMode) == constant.RelayModeResponsesCompact {
|
||||
return convertResponsesRequestForDeepSeek(requestBody, info)
|
||||
}
|
||||
|
||||
// 非 OpenAI 格式先转换为 OpenAI
|
||||
if info.InboundFormat != "" && info.InboundFormat != constant.RelayFormatOpenAI {
|
||||
converted, err := openai.ConvertToOpenAI(requestBody, info)
|
||||
@@ -262,3 +270,77 @@ func convertClaudeRequestForDeepSeek(requestBody []byte, info *common.RelayInfo)
|
||||
}
|
||||
return bytes.NewReader(result), nil
|
||||
}
|
||||
|
||||
// convertResponsesRequestForDeepSeek 处理 Responses API 入站请求的 DeepSeek 兼容适配。
|
||||
// 功能：模型名映射 + 将 RelayInfo 中的思考后缀参数映射到 Responses API 的 reasoning.effort 字段。
|
||||
// 对齐 new-api commit 9724ef1b2 的实现逻辑。
|
||||
func convertResponsesRequestForDeepSeek(requestBody []byte, info *common.RelayInfo) (io.Reader, error) {
|
||||
var rawMap map[string]json.RawMessage
|
||||
if err := json.Unmarshal(requestBody, &rawMap); err != nil {
|
||||
return bytes.NewReader(requestBody), nil
|
||||
}
|
||||
|
||||
// 模型名映射
|
||||
if info.ChannelMeta.IsModelMapped {
|
||||
rawMap["model"] = json.RawMessage(`"` + info.ChannelMeta.UpstreamModelName + `"`)
|
||||
}
|
||||
|
||||
// 将思考后缀参数映射到 reasoning.effort（Responses API 格式）
|
||||
rawMap = injectResponsesReasoningEffort(rawMap, info)
|
||||
|
||||
result, err := json.Marshal(rawMap)
|
||||
if err != nil {
|
||||
return bytes.NewReader(requestBody), nil
|
||||
}
|
||||
return bytes.NewReader(result), nil
|
||||
}
|
||||
|
||||
// injectResponsesReasoningEffort 为 Responses API 请求注入 DeepSeek reasoning.effort 参数。
|
||||
//
|
||||
// Responses API 用 reasoning.effort 而非 Chat API 的 reasoning_effort 顶层字段：
|
||||
// - DeepSeek V4：仅支持 "max"，所有非空 effort 统一映射为 "max"
|
||||
// - DeepSeek V3/R1：支持 "none"/"high"/"max"，做兼容映射
|
||||
//
|
||||
// 注入优先级：客户端已显式设置 reasoning > 后缀路由注入 > 不干预（保留 DeepSeek 默认行为）
|
||||
func injectResponsesReasoningEffort(rawMap map[string]json.RawMessage, info *common.RelayInfo) map[string]json.RawMessage {
|
||||
// 客户端已显式设置 reasoning 字段 → 不干预
|
||||
if _, clientSet := rawMap["reasoning"]; clientSet {
|
||||
return rawMap
|
||||
}
|
||||
|
||||
// -nothinking 后缀：显式关闭思考
|
||||
if info.ThinkingDisabled {
|
||||
rawMap["reasoning"] = json.RawMessage(`{"effort":"none"}`)
|
||||
return rawMap
|
||||
}
|
||||
|
||||
// 确定目标 effort（来自 -thinking 或 -high/-low/-medium 后缀）
|
||||
effort := info.ReasoningEffort
|
||||
if info.ThinkingEnabled && effort == "" {
|
||||
effort = "max"
|
||||
}
|
||||
if effort == "" {
|
||||
return rawMap // 无后缀，不干预，保留 DeepSeek 默认行为
|
||||
}
|
||||
|
||||
modelName := info.ChannelMeta.UpstreamModelName
|
||||
|
||||
// V4 模型仅支持 max
|
||||
if isDeepSeekV4Model(modelName) {
|
||||
rawMap["reasoning"] = json.RawMessage(`{"effort":"max"}`)
|
||||
return rawMap
|
||||
}
|
||||
|
||||
// V3/R1：映射到支持的值（low/medium → high，xhigh/max → max）
|
||||
switch effort {
|
||||
case "low", "medium", "minimal":
|
||||
effort = "high"
|
||||
case "xhigh", "max":
|
||||
effort = "max"
|
||||
default:
|
||||
effort = "high"
|
||||
}
|
||||
effortJSON, _ := json.Marshal(effort)
|
||||
rawMap["reasoning"] = json.RawMessage(`{"effort":` + string(effortJSON) + `}`)
|
||||
return rawMap
|
||||
}
|
||||
Reference in new issue
Block a user