Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
fix(openai): handle reasoning_effort (#11868)
Michael Yangd0cf6c82811c
diff --git a/api/types.go b/api/types.goindex 0f99de18c..0309ebbe3 100644--- a/api/types.go+++ b/api/types.go@@ -769,8 +769,8 @@ func (t *ThinkValue) IsString() bool {return ok}-// AsBool returns the value as a bool (true if enabled in any way)-func (t *ThinkValue) AsBool() bool {+// Bool returns the value as a bool (true if enabled in any way)+func (t *ThinkValue) Bool() bool {if t == nil || t.Value == nil {return false}@@ -786,8 +786,8 @@ func (t *ThinkValue) AsBool() bool {}}-// AsString returns the value as a string-func (t *ThinkValue) AsString() string {+// String returns the value as a string+func (t *ThinkValue) String() string {if t == nil || t.Value == nil {return ""}diff --git a/openai/openai.go b/openai/openai.goindex 50fdb81e9..13b9c425f 100644--- a/openai/openai.go+++ b/openai/openai.go@@ -103,6 +103,7 @@ type ChatCompletionRequest struct {ResponseFormat *ResponseFormat `json:"response_format"`Tools []api.Tool `json:"tools"`Reasoning *Reasoning `json:"reasoning,omitempty"`+ ReasoningEffort *string `json:"reasoning_effort,omitempty"`}type ChatCompletion struct {@@ -541,10 +542,6 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {options["top_p"] = 1.0}- if r.Reasoning != nil {- options["reasoning"] = *r.Reasoning.Effort- }-var format json.RawMessageif r.ResponseFormat != nil {switch strings.ToLower(strings.TrimSpace(r.ResponseFormat.Type)) {@@ -560,9 +557,15 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {var think *api.ThinkValueif r.Reasoning != nil {+ options["reasoning"] = *r.Reasoning.Effortthink = &api.ThinkValue{Value: *r.Reasoning.Effort,}+ } else if r.ReasoningEffort != nil {+ options["reasoning"] = *r.ReasoningEffort+ think = &api.ThinkValue{+ Value: *r.ReasoningEffort,+ }}return &api.ChatRequest{diff --git a/server/prompt.go b/server/prompt.goindex 5d6c3e27c..f1d8020ea 100644--- a/server/prompt.go+++ b/server/prompt.go@@ -44,8 +44,8 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.thinkVal := falsethinkLevel := ""if think != nil {- thinkVal = think.AsBool()- thinkLevel = think.AsString()+ thinkVal = think.Bool()+ thinkLevel = think.String()}var b bytes.Bufferif err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools, Think: thinkVal, ThinkLevel: thinkLevel, IsThinkSet: think != nil}); err != nil {@@ -105,8 +105,8 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.thinkVal := falsethinkLevel := ""if think != nil {- thinkVal = think.AsBool()- thinkLevel = think.AsString()+ thinkVal = think.Bool()+ thinkLevel = think.String()}if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[currMsgIdx:]...), Tools: tools, Think: thinkVal, ThinkLevel: thinkLevel, IsThinkSet: think != nil}); err != nil {return "", nil, errdiff --git a/server/routes.go b/server/routes.goindex d8d1e301c..3c044cd00 100644--- a/server/routes.go+++ b/server/routes.go@@ -205,7 +205,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {// Validate Think value: string values currently only allowed for gptoss modelsif req.Think != nil && req.Think.IsString() && !useHarmony {- c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("think value %q is not supported for this model", req.Think.AsString())})+ c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("think value %q is not supported for this model", req.Think.String())})return}@@ -213,7 +213,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {if req.Suffix != "" {caps = append(caps, model.CapabilityInsert)}- if req.Think != nil && req.Think.AsBool() {+ if req.Think != nil && req.Think.Bool() {caps = append(caps, model.CapabilityThinking)// TODO(drifkin): consider adding a warning if it's false and the model// doesn't support thinking. It's not strictly required, but it can be a@@ -288,10 +288,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {values.Messages = append(msgs, api.Message{Role: "user", Content: req.Prompt})}- values.Think = req.Think != nil && req.Think.AsBool()+ values.Think = req.Think != nil && req.Think.Bool()values.ThinkLevel = ""if req.Think != nil {- values.ThinkLevel = req.Think.AsString()+ values.ThinkLevel = req.Think.String()}values.IsThinkSet = req.Think != nil@@ -317,7 +317,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {var thinkingState *thinking.Parserif !useHarmony {openingTag, closingTag := thinking.InferTags(m.Template.Template)- if req.Think != nil && req.Think.AsBool() && openingTag != "" && closingTag != "" {+ if req.Think != nil && req.Think.Bool() && openingTag != "" && closingTag != "" {thinkingState = &thinking.Parser{OpeningTag: openingTag,ClosingTag: closingTag,@@ -1547,7 +1547,7 @@ func (s *Server) ChatHandler(c *gin.Context) {if len(req.Tools) > 0 {caps = append(caps, model.CapabilityTools)}- if req.Think != nil && req.Think.AsBool() {+ if req.Think != nil && req.Think.Bool() {caps = append(caps, model.CapabilityThinking)}@@ -1601,7 +1601,7 @@ func (s *Server) ChatHandler(c *gin.Context) {// Validate Think value: string values currently only allowed for gptoss modelsif req.Think != nil && req.Think.IsString() && !useHarmony {- c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("think value %q is not supported for this model", req.Think.AsString())})+ c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("think value %q is not supported for this model", req.Think.String())})return}@@ -1620,7 +1620,7 @@ func (s *Server) ChatHandler(c *gin.Context) {var thinkingState *thinking.ParseropeningTag, closingTag := thinking.InferTags(m.Template.Template)- if req.Think != nil && req.Think.AsBool() && openingTag != "" && closingTag != "" {+ if req.Think != nil && req.Think.Bool() && openingTag != "" && closingTag != "" {thinkingState = &thinking.Parser{OpeningTag: openingTag,ClosingTag: closingTag,