feat: add runner failover rules and cache affinity
This commit is contained in:
@@ -259,6 +259,9 @@ func NormalizeChatCompletionResult(result map[string]any) map[string]any {
|
||||
if result == nil {
|
||||
return nil
|
||||
}
|
||||
if usage, ok := result["usage"].(map[string]any); ok && len(usage) > 0 {
|
||||
result["usage"] = NormalizeChatCompletionUsage(usage)
|
||||
}
|
||||
choices, _ := result["choices"].([]any)
|
||||
for _, rawChoice := range choices {
|
||||
choice, _ := rawChoice.(map[string]any)
|
||||
@@ -280,6 +283,9 @@ func NormalizeChatCompletionStreamEvent(event map[string]any) map[string]any {
|
||||
if event == nil {
|
||||
return nil
|
||||
}
|
||||
if usage, ok := event["usage"].(map[string]any); ok && len(usage) > 0 {
|
||||
event["usage"] = NormalizeChatCompletionUsage(usage)
|
||||
}
|
||||
choices, _ := event["choices"].([]any)
|
||||
for _, rawChoice := range choices {
|
||||
choice, _ := rawChoice.(map[string]any)
|
||||
@@ -295,6 +301,33 @@ func NormalizeChatCompletionStreamEvent(event map[string]any) map[string]any {
|
||||
return event
|
||||
}
|
||||
|
||||
func NormalizeChatCompletionUsage(usage map[string]any) map[string]any {
|
||||
if len(usage) == 0 {
|
||||
return usage
|
||||
}
|
||||
normalized := usageFromOpenAIUsage(usage)
|
||||
out := cloneMapAny(usage)
|
||||
if normalized.InputTokens > 0 && intFromAny(out["prompt_tokens"]) == 0 {
|
||||
out["prompt_tokens"] = normalized.InputTokens
|
||||
}
|
||||
if normalized.OutputTokens > 0 && intFromAny(out["completion_tokens"]) == 0 {
|
||||
out["completion_tokens"] = normalized.OutputTokens
|
||||
}
|
||||
if normalized.TotalTokens > 0 && intFromAny(out["total_tokens"]) == 0 {
|
||||
out["total_tokens"] = normalized.TotalTokens
|
||||
}
|
||||
if normalized.CachedInputTokens > 0 {
|
||||
promptDetails, _ := firstPresent(out["prompt_tokens_details"], out["promptTokensDetails"]).(map[string]any)
|
||||
normalizedPromptDetails := cloneMapAny(promptDetails)
|
||||
if normalizedPromptDetails == nil {
|
||||
normalizedPromptDetails = map[string]any{}
|
||||
}
|
||||
normalizedPromptDetails["cached_tokens"] = normalized.CachedInputTokens
|
||||
out["prompt_tokens_details"] = normalizedPromptDetails
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func normalizeToolCallsContainer(container map[string]any, stream bool) {
|
||||
if container == nil {
|
||||
return
|
||||
@@ -877,27 +910,36 @@ func sortedStreamToolCalls(toolCalls map[int]map[string]any) []any {
|
||||
|
||||
func usageFromOpenAI(result map[string]any) Usage {
|
||||
usage, _ := result["usage"].(map[string]any)
|
||||
return usageFromOpenAIUsage(usage)
|
||||
}
|
||||
|
||||
func usageFromOpenAIUsage(usage map[string]any) Usage {
|
||||
input := intFromAny(firstPresent(usage["prompt_tokens"], usage["input_tokens"]))
|
||||
output := intFromAny(firstPresent(usage["completion_tokens"], usage["output_tokens"]))
|
||||
total := intFromAny(usage["total_tokens"])
|
||||
if total == 0 {
|
||||
total = input + output
|
||||
}
|
||||
cachedInput := cachedInputTokensFromOpenAIUsage(usage)
|
||||
cachedInput, cachedInputKnown := cachedInputTokensValueFromOpenAIUsage(usage)
|
||||
if cachedInput > input && input > 0 {
|
||||
cachedInput = input
|
||||
}
|
||||
return Usage{InputTokens: input, OutputTokens: output, CachedInputTokens: cachedInput, TotalTokens: total}
|
||||
return Usage{InputTokens: input, OutputTokens: output, CachedInputTokens: cachedInput, CachedInputTokensKnown: cachedInputKnown, TotalTokens: total}
|
||||
}
|
||||
|
||||
func cachedInputTokensFromOpenAIUsage(usage map[string]any) int {
|
||||
cachedInput, _ := cachedInputTokensValueFromOpenAIUsage(usage)
|
||||
return cachedInput
|
||||
}
|
||||
|
||||
func cachedInputTokensValueFromOpenAIUsage(usage map[string]any) (int, bool) {
|
||||
if len(usage) == 0 {
|
||||
return 0
|
||||
return 0, false
|
||||
}
|
||||
promptDetails, _ := firstPresent(usage["prompt_tokens_details"], usage["promptTokensDetails"]).(map[string]any)
|
||||
inputDetails, _ := firstPresent(usage["input_tokens_details"], usage["inputTokensDetails"]).(map[string]any)
|
||||
usageMetadata, _ := usage["usageMetadata"].(map[string]any)
|
||||
return intFromAny(firstPresent(
|
||||
value, ok := firstPresentValue(
|
||||
promptDetails["cached_tokens"],
|
||||
promptDetails["cachedTokens"],
|
||||
promptDetails["cache_read_input_tokens"],
|
||||
@@ -918,7 +960,11 @@ func cachedInputTokensFromOpenAIUsage(usage map[string]any) int {
|
||||
usage["cachedContentTokenCount"],
|
||||
usageMetadata["cached_content_token_count"],
|
||||
usageMetadata["cachedContentTokenCount"],
|
||||
))
|
||||
)
|
||||
if !ok {
|
||||
return 0, false
|
||||
}
|
||||
return intFromAny(value), true
|
||||
}
|
||||
|
||||
func requestIDFromHTTPResponse(resp *http.Response) string {
|
||||
@@ -979,12 +1025,17 @@ func firstNonEmptyString(values ...any) string {
|
||||
}
|
||||
|
||||
func firstPresent(values ...any) any {
|
||||
value, _ := firstPresentValue(values...)
|
||||
return value
|
||||
}
|
||||
|
||||
func firstPresentValue(values ...any) (any, bool) {
|
||||
for _, value := range values {
|
||||
if value != nil {
|
||||
return value
|
||||
return value, true
|
||||
}
|
||||
}
|
||||
return nil
|
||||
return nil, false
|
||||
}
|
||||
|
||||
func errorMessage(raw []byte, fallback string) string {
|
||||
|
||||
Reference in New Issue
Block a user