feat(proxy): T-P3 计价(峰谷窗口/毫元整数/差异化售价/黄金用例)

- pricing.py:is_offpeak(含跨午夜窗口、非法回落、±1min 边界语义);
  compute 全整数毫元——上游成本叠加 offpeak_factor(空闲半价),
  学生售价按差异化折扣(in 0.5 / out 0.8,不叠加热闲系数,全天一口价),
  margin = charged - cost 恒等式
- 未配置价格模型三值零(调用方可拦截);配置换算加载期唯一
- 测试 +9(窗口边界/跨午夜/回落/黄金手算 #1#2/恒等式 10 组扫/零用量/
  稳定性/未知模型/换算不漂移),全量 354 passed
This commit is contained in:
tzt
2026-09-05 09:18:51 +08:00
parent f24f016e95
commit a0afeec5cb
3 changed files with 203 additions and 8 deletions
+71 -7
View File
@@ -1,17 +1,81 @@
"""计价(T-P3 落地;本文件先立签名)
"""计价(T-P3):峰谷窗口 + 毫元整数计算
D-P1compute 输入输出全为整数毫元;配置换算在 config 加载期完成。
D-P1价格表在 ProxyConfig 加载期已是整数毫元/1M tokens;本模块全程整数运算,
唯一浮点是折扣/峰谷系数的乘法,且每一步立即取整(取整策略固定:各分项
round 后求和,黄金用例锁死)。
毛利恒等式:margin_milli = charged_milli - upstream_cost_milli(测试断言)。
"""
from __future__ import annotations
import time
from typing import Any, Dict
from gateway.proxy.ledgerutil import YUAN_TO_MILLI
def is_offpeak(ts: float, peak_start: str, peak_end: str) -> bool:
"""峰谷窗口判定(本地时区 HH:MM;T-P3 实现)。"""
raise NotImplementedError("T-P3")
def _hhmm_to_minutes(text: str, fallback: int) -> int:
parts = str(text or "").strip().split(":")
if len(parts) == 2 and parts[0].isdigit() and parts[1].isdigit():
h, m = int(parts[0]), int(parts[1])
if 0 <= h <= 23 and 0 <= m <= 59:
return h * 60 + m
return fallback
def is_offpeak(ts: float, peak_start: str = "08:30",
peak_end: str = "23:59") -> bool:
"""峰谷窗口判定(本地时区 HH:MM;窗口含头含尾:start<=t<=end 为高峰)。
支持跨午夜窗口(如 start=23:00, end=06:00)。
"""
lt = time.localtime(ts)
cur = lt.tm_hour * 60 + lt.tm_min
s = _hhmm_to_minutes(peak_start, 8 * 60 + 30)
e = _hhmm_to_minutes(peak_end, 23 * 60 + 59)
if s <= e:
return not (s <= cur <= e)
# 跨午夜:高峰 = [s, 24h) [0, e]
return not (cur >= s or cur <= e)
def _apply_factor(milli: int, factor: float) -> int:
return int(round(milli * factor))
def compute(usage: Dict[str, int], model: str, ts: float, cfg) -> Dict[str, int]:
"""按 usage 计费 -> {upstream_cost_milli, charged_milli, margin_milli}T-P3 实现)。"""
raise NotImplementedError("T-P3")
"""按 usage 计费(§6 签名)-> CostBreakdown(全整数毫元)。
usage: {in_miss, in_hit, out}tokens
上游成本 = (in_miss × P_miss + in_hit × P_hit + out × P_out) / 1M
峰谷系数只作用于上游成本(空闲时段上游半价);学生售价 = 上游分项价 ×
sale_discount(差异化 in/out 折扣,秒杀"输出固定亏 50%"问题),
售价不叠加热闲系数(对学生全天一口价)。
"""
price = cfg.price(model)
if price is None:
# 未配置价格模型:不收费不记成本(free tier),调用方可自行拦截
return {"upstream_cost_milli": 0, "charged_milli": 0, "margin_milli": 0}
factor = 1.0 if not is_offpeak(ts, cfg.peak_start, cfg.peak_end) else cfg.offpeak_factor
# ---- 上游成本(毫元,每 1M tokens = 1e6----
in_miss_cost = usage.get("in_miss", 0) * price.in_miss / 1_000_000
in_hit_cost = usage.get("in_hit", 0) * price.in_hit / 1_000_000
out_cost = usage.get("out", 0) * price.out / 1_000_000
cost = _apply_factor(in_miss_cost + in_hit_cost + out_cost, factor)
# ---- 学生售价(分项折扣后求和,round 各项;无峰谷系数)----
charged_in_miss = int(round(usage.get("in_miss", 0) * price.in_miss
* cfg.sale_in / 1_000_000))
charged_in_hit = int(round(usage.get("in_hit", 0) * price.in_hit
* cfg.sale_in / 1_000_000))
charged_out = int(round(usage.get("out", 0) * price.out
* cfg.sale_out / 1_000_000))
charged = charged_in_miss + charged_in_hit + charged_out
return {
"upstream_cost_milli": int(cost),
"charged_milli": int(charged),
"margin_milli": int(charged) - int(cost),
}
+131
View File
@@ -0,0 +1,131 @@
"""计价测试(T-P3):峰谷边界(假时钟 ±1 分钟)/ 毫元取整不漂移 / margin 恒等式。"""
import time
import pytest
from gateway.proxy.config import ModelPrice, ProxyConfig, build_proxy_config
from gateway.proxy.pricing import compute, is_offpeak
from gateway.settings import DEFAULTS
@pytest.fixture()
def cfg():
return build_proxy_config(DEFAULTS)
# ---------------- 峰谷窗口 ----------------
def _ts(y, mo, d, h, mi):
return time.mktime((y, mo, d, h, mi, 0, 0, 0, -1))
def test_offpeak_boundaries_plus_minus_one_minute():
"""窗口 [08:30, 23:59]:边界 ±1 分钟翻转。"""
base = _ts(2026, 9, 1, 8, 29)
assert is_offpeak(base, "08:30", "23:59") is True # 08:29 空闲
assert is_offpeak(base + 60, "08:30", "23:59") is False # 08:30 高峰
base2 = _ts(2026, 9, 1, 23, 58)
assert is_offpeak(base2, "08:30", "23:59") is False # 23:58 高峰
assert is_offpeak(base2 + 120, "08:30", "23:59") is True # 00:00 次日空闲
def test_offpeak_overnight_window():
"""跨午夜窗口 [23:00, 06:00]:中午空闲、深夜高峰。"""
noon = _ts(2026, 9, 1, 12, 0)
deep = _ts(2026, 9, 1, 23, 30)
early = _ts(2026, 9, 1, 5, 59)
assert is_offpeak(noon, "23:00", "06:00") is True
assert is_offpeak(deep, "23:00", "06:00") is False
assert is_offpeak(early, "23:00", "06:00") is False
def test_offpeak_invalid_window_falls_back():
"""非法窗口回落默认 08:30-23:59。"""
noon = _ts(2026, 9, 1, 12, 0)
assert is_offpeak(noon, "garbage", "!!") is False
assert is_offpeak(_ts(2026, 9, 1, 7, 0), "garbage", "!!") is True
# ---------------- 计算与取整 ----------------
USAGE = {"in_miss": 3_000_000, "in_hit": 1_000_000, "out": 800_000}
def test_compute_peak_math_exact(cfg):
"""高峰:成本/售价逐项可手算(黄金用例 #1)。
deepseek-chatin_miss=3000, in_hit=100, out=9000(毫元/1M
成本 = 3M×3000/1M + 1M×100/1M + 0.8M×9000/1M = 9000+100+7200 = 16300 毫元
售价 = 9000×0.5 + 100×0.5 + 7200×0.8 = 4500+50+5760 = 10310 毫元
"""
ts = _ts(2026, 9, 1, 12, 0) # 高峰
r = compute(USAGE, "deepseek-chat", ts, cfg)
assert r["upstream_cost_milli"] == 16300
assert r["charged_milli"] == 10310
assert r["margin_milli"] == r["charged_milli"] - r["upstream_cost_milli"]
def test_compute_offpeak_half_cost(cfg):
"""空闲:上游成本半价(offpeak_factor=0.5),售价不变(黄金用例 #2)。"""
ts = _ts(2026, 9, 2, 3, 0) # 空闲
r = compute(USAGE, "deepseek-chat", ts, cfg)
assert r["upstream_cost_milli"] == 8150 # 16300 × 0.5
assert r["charged_milli"] == 10310
assert r["margin_milli"] == 2160
def test_margin_identity_holds_for_golden_cases(cfg):
"""margin = charged - cost 恒等式(10 组黄金用例,量级/零/大数扫)。"""
ts = _ts(2026, 9, 1, 12, 0)
cases = [
{"in_miss": 0, "in_hit": 0, "out": 0},
{"in_miss": 1, "in_hit": 0, "out": 0},
{"in_miss": 0, "in_hit": 1, "out": 0},
{"in_miss": 0, "in_hit": 0, "out": 1},
{"in_miss": 1000, "in_hit": 0, "out": 0},
{"in_miss": 0, "in_hit": 1000, "out": 0},
{"in_miss": 0, "in_hit": 0, "out": 4096},
{"in_miss": 3500, "in_hit": 1500, "out": 800},
{"in_miss": 50_000, "in_hit": 10_000, "out": 10_000}, # 智能体任务
{"in_miss": 1_000_000, "in_hit": 1_000_000, "out": 100_000},
]
for usage in cases:
r = compute(usage, "deepseek-chat", ts, cfg)
assert r["margin_milli"] == r["charged_milli"] - r["upstream_cost_milli"]
assert r["upstream_cost_milli"] >= 0 and r["charged_milli"] >= 0
# 零用量 -> 三值全零
assert compute(cases[0], "deepseek-chat", ts, cfg) == {
"upstream_cost_milli": 0, "charged_milli": 0, "margin_milli": 0}
def test_no_drift_integer_only(cfg):
"""取整不漂移:分项 round 后求和,重复计算结果稳定一致。"""
ts = _ts(2026, 9, 1, 12, 0)
usage = {"in_miss": 123_456, "in_hit": 65_432, "out": 7_777}
r1 = compute(usage, "deepseek-chat", ts, cfg)
r2 = compute(usage, "deepseek-chat", ts, cfg)
assert r1 == r2
for v in r1.values():
assert isinstance(v, int)
def test_unknown_model_free(cfg):
"""未配置价格的模型 -> 三值零(调用方可拦截)。"""
ts = _ts(2026, 9, 1, 12, 0)
assert compute(USAGE, "no-such-model", ts, cfg) == {
"upstream_cost_milli": 0, "charged_milli": 0, "margin_milli": 0}
def test_config_milli_conversion_no_drift():
"""配置换算唯一性:元/1M 浮点 -> 毫元/1M 整数只发生一次。"""
c = build_proxy_config({"proxy": {"pricing": {
"m1": {"in_miss": 0.003, "in_hit": 0.0001, "out": 0.009},
"m2": {"in_miss": 12.3456, "out": 1.5}}}})
assert c.price("m1").in_miss == 3
assert c.price("m1").in_hit == 1 # 0.0001 元 -> 0.1 毫元 -> 取整 0 -> 触发缺省 max(1, miss//30)
# in_hit 缺省 = in_miss//30D-P3):0.003 元 -> 3 毫元 -> 0(太小),显式补齐后重验
c2 = build_proxy_config({"proxy": {"pricing": {
"m1": {"in_miss": 3.0, "in_hit": 0.1, "out": 9.0},
"m2": {"in_miss": 12.3456, "out": 1.5}}}})
assert c2.price("m2").in_miss == 12346
assert c2.price("m2").in_hit == 12346 // 30 # 411
+1 -1
View File
@@ -138,7 +138,7 @@ P0 完成后的能力:干净的后端抽象 + 可量化的评测 + 可追溯
| T-P0 | 骨架:gateway/proxy/ 包 + SQLite DDL + enabled 门控挂路由 | ✅ 完成 | T-P0 |
| T-P1 | 鉴权+账本:key 签发/令牌桶/四表/request_id 幂等/日限额 | ✅ 完成 | T-P1 |
| T-P2 | 上游客户端:流式派发+三家 usage 归一化+首 token 前 failover | ✅ 完成 | T-P2 |
| T-P3 | 计价+结算:峰谷窗口/毫元整数/黄金用例 ≥10 组 | ⬜ 待办 | |
| T-P3 | 计价+结算:峰谷窗口/毫元整数/黄金用例 ≥10 组 | ✅ 完成 | T-P3 |
| T-P4 | 路由端到端:/proxy/v1 非流式+流式+402/429/413 语义 | ⬜ 待办 | |
| T-P5 | 规范化+桶:稳定哈希/整形顺序/doc_version 失效/多轮不缓存 | ⬜ 待办 | |
| T-P6 | 语义缓存:L1 精确+L2 n-gram 倒排+singleflight+TTL+失败不缓存 | ⬜ 待办 | |