feat: estimate and populate prompt cache tokens when upstream omits breakdown

This commit is contained in:
Luxferre
2026-08-27 14:42:44 +03:00
parent c856595199
commit 04d83a6b7c
2 changed files with 46 additions and 2 deletions
+28 -2
View File
@@ -135,6 +135,7 @@ type Usage struct {
PromptTokens int `json:"prompt_tokens"`
CompletionTokens int `json:"completion_tokens"`
TotalTokens int `json:"total_tokens"`
CachedTokens int `json:"cached_tokens,omitempty"`
PromptTokensDetails *PromptTokensDetails `json:"prompt_tokens_details,omitempty"`
CompletionTokensDetails *CompletionTokensDetails `json:"completion_tokens_details,omitempty"`
}
@@ -422,12 +423,35 @@ type FinalOutput struct {
ReasoningTokens int
}
func EstimateCachedTokens(promptTokens int, currentQuestion string) int {
if promptTokens <= 0 {
return 0
}
qLen := len(strings.TrimSpace(currentQuestion))
uncached := (qLen + 3) / 4
if uncached < 1 {
uncached = 1
}
if uncached >= promptTokens {
return 0
}
return promptTokens - uncached
}
func ResolveCachedTokens(upstreamCached int, promptTokens int, currentQuestion string) int {
if upstreamCached > 0 {
return upstreamCached
}
return EstimateCachedTokens(promptTokens, currentQuestion)
}
func BuildUsage(promptTokens, completionTokens, cachedTokens, reasoningTokens int) Usage {
totalTokens := promptTokens + completionTokens
return Usage{
PromptTokens: promptTokens,
CompletionTokens: completionTokens,
TotalTokens: totalTokens,
CachedTokens: cachedTokens,
PromptTokensDetails: &PromptTokensDetails{
CachedTokens: cachedTokens,
},
@@ -1473,12 +1497,13 @@ func (gw *QoronaGateway) HandleChatCompletions(w http.ResponseWriter, r *http.Re
if cachedTokens == 0 {
cachedTokens = sResult.Result.FinalStatus.PromptTokensDetails.CachedTokens
}
cachedTokens = ResolveCachedTokens(cachedTokens, promptTokens, question)
if req.StreamOptions != nil && req.StreamOptions.IncludeUsage {
streamer.EmitUsage(BuildUsage(promptTokens, completionTokens, cachedTokens, 0))
}
streamer.Done()
log.Printf("[Stream] Completed run_id=%s finish_reason=%s tokens=%d/%d", runID, finishReason, promptTokens, completionTokens)
log.Printf("[Stream] Completed run_id=%s finish_reason=%s tokens=%d/%d cached=%d", runID, finishReason, promptTokens, completionTokens, cachedTokens)
return
}
}
@@ -1606,6 +1631,7 @@ func (gw *QoronaGateway) HandleChatCompletions(w http.ResponseWriter, r *http.Re
if cachedTokens == 0 {
cachedTokens = sResult.Result.FinalStatus.PromptTokensDetails.CachedTokens
}
cachedTokens = ResolveCachedTokens(cachedTokens, promptTokens, question)
fullContent := accumulatedContent.String()
rawThinking := accumulatedThinking.String()
@@ -1644,7 +1670,7 @@ func (gw *QoronaGateway) HandleChatCompletions(w http.ResponseWriter, r *http.Re
}
WriteCompletionResponse(w, completionID, createdTime, modelName, out)
log.Printf("[Non-Stream] Completed run_id=%s finish_reason=%s tokens=%d/%d", runID, finishReason, promptTokens, completionTokens)
log.Printf("[Non-Stream] Completed run_id=%s finish_reason=%s tokens=%d/%d cached=%d", runID, finishReason, promptTokens, completionTokens, cachedTokens)
return
}
}
+18
View File
@@ -192,3 +192,21 @@ func TestStreamerUsage(t *testing.T) {
}
}
func TestResolveCachedTokens(t *testing.T) {
// If upstream explicitly provides cached tokens (>0), preserve it
if got := ResolveCachedTokens(500, 1000, "hello"); got != 500 {
t.Fatalf("expected 500, got %d", got)
}
// If upstream returns 0, estimate based on promptTokens and current question
if got := ResolveCachedTokens(0, 1397, "Say hello in 1 word"); got <= 0 || got >= 1397 {
t.Fatalf("expected cached tokens between 1 and 1396, got %d", got)
}
// If promptTokens <= 0, return 0
if got := ResolveCachedTokens(0, 0, "hello"); got != 0 {
t.Fatalf("expected 0, got %d", got)
}
}