fix: 缓存 token 计费按上游协议区分口径
- 新增 ComputeCost:OpenAI 系 prompt 含缓存读需扣减(异常数据钳制为 0); Anthropic input_tokens 不含缓存,按原值计费,缓存写按输入价 ×1.25(原实现误用输出价) - recordUsage 传入上游协议(用量语义跟随解析它的上游响应,而非客户端协议) - 补充单元测试覆盖两种协议口径与边界情况
This commit is contained in:
@@ -204,7 +204,7 @@ func (g *Gateway) Dispatch(c *gin.Context, req *Request) {
|
||||
g.writeError(c, http.StatusServiceUnavailable, "no enabled channels for model: "+req.Model)
|
||||
g.recordUsage(req, nil, nil, usage.Event{
|
||||
IsError: true, ErrorCode: "no_channel",
|
||||
}, convert.TokenUsage{})
|
||||
}, convert.TokenUsage{}, "")
|
||||
return
|
||||
}
|
||||
|
||||
@@ -273,7 +273,7 @@ func (g *Gateway) Dispatch(c *gin.Context, req *Request) {
|
||||
IsError: true,
|
||||
ErrorCode: "upstream_error",
|
||||
LatencyMS: int(time.Since(start).Milliseconds()),
|
||||
}, convert.TokenUsage{})
|
||||
}, convert.TokenUsage{}, targetFormat)
|
||||
continue // 可重试:换下一个渠道
|
||||
}
|
||||
|
||||
@@ -289,7 +289,7 @@ func (g *Gateway) Dispatch(c *gin.Context, req *Request) {
|
||||
IsError: true,
|
||||
ErrorCode: fmt.Sprintf("upstream_%d", resp.StatusCode),
|
||||
LatencyMS: int(time.Since(start).Milliseconds()),
|
||||
}, convert.TokenUsage{})
|
||||
}, convert.TokenUsage{}, targetFormat)
|
||||
// 429/5xx 可换渠道重试;4xx 直接透传
|
||||
if resp.StatusCode == http.StatusTooManyRequests || resp.StatusCode >= 500 {
|
||||
lastErrStatus, lastErrBody = resp.StatusCode, string(body)
|
||||
@@ -318,13 +318,13 @@ func (g *Gateway) Dispatch(c *gin.Context, req *Request) {
|
||||
IsError: true,
|
||||
ErrorCode: errCode,
|
||||
LatencyMS: int(time.Since(start).Milliseconds()),
|
||||
}, tok)
|
||||
}, tok, targetFormat)
|
||||
return
|
||||
}
|
||||
// 成功记录:用量 + 定价计费。
|
||||
g.recordUsage(req, cand, ch, usage.Event{
|
||||
LatencyMS: int(time.Since(start).Milliseconds()),
|
||||
}, tok)
|
||||
}, tok, targetFormat)
|
||||
return
|
||||
}
|
||||
|
||||
@@ -349,9 +349,10 @@ func rewriteModel(body []byte, upstreamModel string) []byte {
|
||||
return out
|
||||
}
|
||||
|
||||
// recordUsage 汇总一次请求的用量事件并异步落库。tok 为从上游响应提取的用量。
|
||||
// recordUsage 汇总一次请求的用量事件并异步落库。tok 为从上游响应提取的用量,
|
||||
// 其 token 语义由 upstreamProto(渠道实际使用的上游协议)决定。
|
||||
// cand/ch 可为 nil(无可用渠道的失败场景)。
|
||||
func (g *Gateway) recordUsage(req *Request, cand *channel.Candidate, ch *store.Channel, ev usage.Event, tok convert.TokenUsage) {
|
||||
func (g *Gateway) recordUsage(req *Request, cand *channel.Candidate, ch *store.Channel, ev usage.Event, tok convert.TokenUsage, upstreamProto string) {
|
||||
if g.usageRec == nil {
|
||||
return
|
||||
}
|
||||
@@ -380,7 +381,8 @@ func (g *Gateway) recordUsage(req *Request, cand *channel.Candidate, ch *store.C
|
||||
}
|
||||
}
|
||||
// 定价与成本(价格按每百万 token 的 USD 单价)。
|
||||
// 成本口径:非缓存输入 × 输入价 + 缓存读 × 缓存价 + 缓存写与输出 × 输出价。
|
||||
// 成本口径按上游协议区分(详见 ComputeCost):OpenAI 系 prompt 含缓存读需扣减;
|
||||
// Anthropic 的 input_tokens 不含缓存,缓存写按输入价 ×1.25。
|
||||
if ev.ModelID != 0 {
|
||||
if m, err := g.modelDAO.GetByID(ev.ModelID); err == nil {
|
||||
ev.InputPrice = m.InputPrice
|
||||
@@ -389,10 +391,8 @@ func (g *Gateway) recordUsage(req *Request, cand *channel.Candidate, ch *store.C
|
||||
}
|
||||
}
|
||||
if !ev.IsError {
|
||||
ev.Cost = (float64(ev.PromptTokens-ev.CacheReadTokens)*ev.InputPrice +
|
||||
float64(ev.CacheReadTokens)*ev.CacheReadPrice +
|
||||
float64(ev.CacheCreationTokens)*ev.OutputPrice +
|
||||
float64(ev.CompletionTokens)*ev.OutputPrice) / 1e6
|
||||
ev.Cost = ComputeCost(upstreamProto, tok.InputTokens, tok.OutputTokens, tok.CacheReadTokens, tok.CacheCreationTokens,
|
||||
ev.InputPrice, ev.OutputPrice, ev.CacheReadPrice)
|
||||
}
|
||||
g.usageRec.Record(ev)
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user