ext/new-api
2
0
Fork 0
mirror of https://github.com/QuantumNous/new-api.git synced 2026-09-13 07:40:56 +00:00
Code Activity

fix: preserve Qwen thinking_budget passthrough (#5836)

Browse Source 
* fix: preserve qwen thinking budget

* test: address qwen thinking budget review comments

* chore: remove unreachable adaptor code

* test: cover zero Qwen thinking budgets
This commit is contained in:
Scott
2026-07-29 17:45:14 +08:00
committed by GitHub
parent cb4c8c02f8
commit 66ee6b8f98
20 changed files with 345 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
relaykit/dto/openai_request.go
+26
View File
@@ -89,6 +89,7 @@ type GeneralOpenAIRequest struct {
// Ali Qwen Params
VlHighResolutionImages json.RawMessage `json:"vl_high_resolution_images,omitempty"`
EnableThinking json.RawMessage `json:"enable_thinking,omitempty"`
ThinkingBudget json.RawMessage `json:"thinking_budget,omitempty"`
ChatTemplateKwargs json.RawMessage `json:"chat_template_kwargs,omitempty"`
EnableSearch json.RawMessage `json:"enable_search,omitempty"`
// ollama Params
@@ -107,6 +108,14 @@ type GeneralOpenAIRequest struct {
ReasoningSplit json.RawMessage `json:"reasoning_split,omitempty"`
}
func (r GeneralOpenAIRequest) MarshalJSON() ([]byte, error) {
type Alias GeneralOpenAIRequest
if !IsQwenThinkingBudgetModel(r.Model) {
r.ThinkingBudget = nil
}
return kitutil.Marshal((*Alias)(&r))
}
func (r *GeneralOpenAIRequest) GetTokenCountMeta() *types.TokenCountMeta {
var tokenCountMeta types.TokenCountMeta
var texts = make([]string, 0)
@@ -222,6 +231,14 @@ func IsOpenAIGPT5Model(modelName string) bool {
return strings.HasPrefix(modelName, "gpt-5")
}
func IsQwenThinkingBudgetModel(modelName string) bool {
normalized := strings.ToLower(strings.TrimSpace(modelName))
return strings.HasPrefix(normalized, "qwen") ||
strings.Contains(normalized, "/qwen") ||
strings.HasPrefix(normalized, "qwq") ||
strings.Contains(normalized, "/qwq")
}
func (r *GeneralOpenAIRequest) GetSystemRoleName() string {
if IsOpenAIReasoningOModel(r.Model) {
if !strings.HasPrefix(r.Model, "o1-mini") && !strings.HasPrefix(r.Model, "o1-preview") {
@@ -880,10 +897,19 @@ type OpenAIResponsesRequest struct {
ClientMetadata json.RawMessage `json:"client_metadata,omitempty"`
// qwen
EnableThinking json.RawMessage `json:"enable_thinking,omitempty"`
ThinkingBudget json.RawMessage `json:"thinking_budget,omitempty"`
// perplexity
Preset json.RawMessage `json:"preset,omitempty"`
}
func (r OpenAIResponsesRequest) MarshalJSON() ([]byte, error) {
type Alias OpenAIResponsesRequest
if !IsQwenThinkingBudgetModel(r.Model) {
r.ThinkingBudget = nil
}
return kitutil.Marshal((*Alias)(&r))
}
func (r *OpenAIResponsesRequest) GetTokenCountMeta() *types.TokenCountMeta {
var fileMeta = make([]*types.FileMeta, 0)
var texts = make([]string, 0)