mirror of
https://github.com/songquanpeng/one-api.git
synced 2025-09-19 10:06:37 +08:00
feat: enhance reasoning token handling in OpenAI adaptor
This commit is contained in:
parent
abe6afacca
commit
a45a3d6241
@ -66,6 +66,9 @@ func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*model.E
|
||||
if choice.Delta.Reasoning != nil {
|
||||
reasoningText += *choice.Delta.Reasoning
|
||||
}
|
||||
if choice.Delta.ReasoningContent != nil {
|
||||
reasoningText += *choice.Delta.ReasoningContent
|
||||
}
|
||||
|
||||
responseText += conv.AsString(choice.Delta.Content)
|
||||
}
|
||||
@ -99,6 +102,12 @@ func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*model.E
|
||||
return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), "", nil
|
||||
}
|
||||
|
||||
// If there is no reasoning tokens in the completion, we should count the reasoning tokens in the response.
|
||||
if len(reasoningText) > 0 &&
|
||||
(usage.CompletionTokensDetails == nil || usage.CompletionTokensDetails.ReasoningTokens == 0) {
|
||||
usage.CompletionTokens += CountToken(reasoningText)
|
||||
}
|
||||
|
||||
return nil, reasoningText + responseText, usage
|
||||
}
|
||||
|
||||
|
@ -104,6 +104,9 @@ func postConsumeQuota(ctx context.Context, usage *relaymodel.Usage, meta *meta.M
|
||||
|
||||
completionRatio := billingratio.GetCompletionRatio(textRequest.Model, meta.ChannelType)
|
||||
promptTokens := usage.PromptTokens
|
||||
// It appears that DeepSeek's official service automatically merges ReasoningTokens into CompletionTokens,
|
||||
// but the behavior of third-party providers may differ, so for now we do not add them manually.
|
||||
// completionTokens := usage.CompletionTokens + usage.CompletionTokensDetails.ReasoningTokens
|
||||
completionTokens := usage.CompletionTokens
|
||||
quota = int64(math.Ceil((float64(promptTokens) + float64(completionTokens)*completionRatio) * ratio))
|
||||
if ratio != 0 && quota <= 0 {
|
||||
|
Loading…
Reference in New Issue
Block a user