feat(sense): T-G1 Embedder（int8 量化/降级 fail-closed + /v1/embeddings 透传）
- embedder.py：embed() 调 llama-server /v1/embeddings -> 对称 per-vector int8 量化（scale=127/max|v|，1024 维余弦扰动 ~1e-4 << 1e-2 验收）； 超时/连接/非200/空形状 -> EmbedderDown（D-G4 fail-closed）； quantize_int8/cosine_int8 辅助（int8 直接点积，scale 正标量不改方向） - routes.py 拆双路由组：/sense 前缀组 + /v1 无前缀组（/v1/embeddings OpenAI 兼容透传，embedder 不可用 503）；build_sense_routers 返回列表， api.py 逐个 include；build_sense_router 保留向后兼容 - 测试 +6：量化余弦误差 200 组 <1e-2/范围/零向量/int8 余弦/ 正常量化/三种 fail-closed/透传端点（503+形状+400），全量 381 passed - 待真机项：llama-server embedder 端点冒烟（需用户配置 embedder.base_url）
This commit is contained in:
tzt
@@ -1,17 +1,23 @@
|
||||
"""语义分析器与三级任务分级（T-G0…T-G8）：/v1/route + /sense/admin/*。
|
||||
"""语义分析器与三级任务分级（T-G0…T-G8）：/v1/route + /v1/embeddings + /sense/admin/*。
|
||||
|
||||
组装唯一入口：build_sense_router(cfg) -> APIRouter；
|
||||
由 gateway.api 在 sense.enabled 时 include（D-G7 关闭时不注册任何路由）。
|
||||
灰度：mode=collect（只记不决策）-> shadow（决策只比对）-> live（真分流）；
|
||||
collect 攒满 min_labels 前不得开 live（晋升门 §6）。
|
||||
组装唯一入口：build_sense_routers(cfg) -> [APIRouter, APIRouter]
|
||||
（/sense 前缀组 + /v1 无前缀组）；由 gateway.api 在 sense.enabled 时逐个 include
|
||||
（D-G7 关闭时不注册任何路由）。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import List
|
||||
|
||||
from fastapi import APIRouter
|
||||
|
||||
from gateway.sense.config import SenseConfig, build_sense_config # noqa: F401
|
||||
|
||||
|
||||
def build_sense_router(cfg: SenseConfig) -> APIRouter:
|
||||
from gateway.sense.routes import build_sense_router as _build
|
||||
def build_sense_routers(cfg: SenseConfig) -> List[APIRouter]:
|
||||
from gateway.sense.routes import build_sense_routers as _build
|
||||
return _build(cfg)
|
||||
|
||||
|
||||
def build_sense_router(cfg: SenseConfig) -> APIRouter:
|
||||
"""向后兼容：仅 /sense 前缀组（/v1 组见 build_sense_routers）。"""
|
||||
return build_sense_routers(cfg)[0]
|
||||
|
||||
@@ -0,0 +1,84 @@
|
||||
"""Embedder 客户端（T-G1）：llama-server /v1/embeddings + int8 量化 + 降级（D-G4）。
|
||||
|
||||
- int8 量化：对称 per-vector scale（scale = max|v| / 127），q = round(v × scale)。
|
||||
1024 维余弦扰动 ~1e-4（验收 <1e-2）；存储 1024 维 fp32 4KB -> int8 1KB（D-G6）。
|
||||
- 降级（D-G4）：超时/连接失败/非 200/形状不符 -> raise EmbedderDown，
|
||||
调用方（grader）退化为规则门 + 默认 T2。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import httpx
|
||||
|
||||
from gateway.sense.config import EmbedderCfg
|
||||
from gateway.sense.errors import EmbedderDown
|
||||
|
||||
_client: Optional[httpx.AsyncClient] = None
|
||||
|
||||
|
||||
def get_client() -> httpx.AsyncClient:
|
||||
global _client
|
||||
if _client is None:
|
||||
_client = httpx.AsyncClient(timeout=5.0)
|
||||
return _client
|
||||
|
||||
|
||||
async def close_client() -> None:
|
||||
global _client
|
||||
if _client is not None:
|
||||
await _client.aclose()
|
||||
_client = None
|
||||
|
||||
|
||||
def quantize_int8(vec: list[float]) -> list[int]:
|
||||
"""对称 per-vector int8 量化：q = round(v × scale)，scale = 127 / max|v|。"""
|
||||
m = max((abs(v) for v in vec), default=0.0)
|
||||
if m == 0:
|
||||
return [0] * len(vec)
|
||||
scale = 127.0 / m
|
||||
return [max(-127, min(127, int(round(v * scale)))) for v in vec]
|
||||
|
||||
|
||||
def dequantize_int8(q: list[int], scale: float) -> list[float]:
|
||||
"""反量化（需要保留 scale 时使用；检索场景直接用 int8 余弦即可）。"""
|
||||
if scale == 0:
|
||||
return [0.0] * len(q)
|
||||
return [x / scale for x in q]
|
||||
|
||||
|
||||
def cosine_int8(a: list[int], b: list[int]) -> float:
|
||||
"""int8 向量余弦相似度（无 scale 版本：scale 为正标量不改变方向）。"""
|
||||
n = min(len(a), len(b))
|
||||
if n == 0:
|
||||
return 0.0
|
||||
dot = na = nb = 0
|
||||
for i in range(n):
|
||||
x, y = a[i], b[i]
|
||||
dot += x * y
|
||||
na += x * x
|
||||
nb += y * y
|
||||
if na == 0 or nb == 0:
|
||||
return 0.0
|
||||
return max(-1.0, min(1.0, dot / ((na ** 0.5) * (nb ** 0.5))))
|
||||
|
||||
|
||||
async def embed(text: str, cfg: EmbedderCfg) -> list[int]:
|
||||
"""调用 llama-server /v1/embeddings，返回 int8 量化向量（§7 签名）。
|
||||
|
||||
失败（超时/连接/非 200/形状不符）-> raise EmbedderDown（D-G4 降级信号）。
|
||||
"""
|
||||
try:
|
||||
client = get_client()
|
||||
resp = await client.post(
|
||||
cfg.base_url.rstrip("/") + "/embeddings",
|
||||
json={"model": cfg.model, "input": text},
|
||||
timeout=cfg.timeout_s)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
vec = data["data"][0]["embedding"]
|
||||
if not isinstance(vec, list) or not vec:
|
||||
raise ValueError("空向量")
|
||||
return quantize_int8([float(v) for v in vec])
|
||||
except EmbedderDown:
|
||||
raise
|
||||
except Exception as e: # noqa: BLE001
|
||||
raise EmbedderDown(f"embedder 不可用: {type(e).__name__}: {e}") from e
|
||||
+40
-9
@@ -1,21 +1,52 @@
|
||||
"""Sense 路由（T-G0 骨架：健康检查；/v1/route 与 /v1/embeddings 在 T-G1/G5 落地）。"""
|
||||
"""Sense 路由（T-G1）：/sense 前缀组 + /v1 无前缀组（embeddings 透传）。
|
||||
|
||||
/v1/route（T-G5 grader）后续加入无前缀组；/sense/admin/*（T-G3+）加入前缀组。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from fastapi import APIRouter
|
||||
from typing import List
|
||||
|
||||
from fastapi import APIRouter, Request
|
||||
from fastapi.responses import JSONResponse
|
||||
|
||||
from gateway.sense.config import SenseConfig
|
||||
from gateway.sense.embedder import embed
|
||||
from gateway.sense.errors import EmbedderDown
|
||||
|
||||
|
||||
def build_sense_router(cfg: SenseConfig) -> APIRouter:
|
||||
"""组装 sense 面路由（唯一组装点）。"""
|
||||
router = APIRouter(prefix="/sense")
|
||||
def build_sense_routers(cfg: SenseConfig) -> List[APIRouter]:
|
||||
"""组装 sense 路由组：[/sense 前缀组, /v1 无前缀组]。"""
|
||||
main = APIRouter(prefix="/sense")
|
||||
v1 = APIRouter()
|
||||
|
||||
@router.get("/health", tags=["sense"])
|
||||
@main.get("/health", tags=["sense"])
|
||||
async def health():
|
||||
"""sense 面健康检查（含灰度状态，供看板/运维）。"""
|
||||
return {"enabled": cfg.enabled, "mode": cfg.mode,
|
||||
"embedder": cfg.embedder.base_url}
|
||||
|
||||
# /v1/route（T-G5 grader）、/v1/embeddings（T-G1）、/sense/admin/*（T-G3+）后续追加。
|
||||
# 注意 /v1/route 不带 /sense 前缀——由 api.py 单独 include 一个无前缀子路由（T-G5）。
|
||||
return router
|
||||
@v1.post("/v1/embeddings", tags=["sense"])
|
||||
async def embeddings(request: Request):
|
||||
"""OpenAI 兼容透传 embedder（客户端/代理共用）。
|
||||
|
||||
请求：{"model"?, "input": str}；响应：{"object":"list","data":[{"object":
|
||||
"embedding","index":0,"embedding":[int8...]}],"model":...}。
|
||||
embedder 不可用 -> 503（D-G4 fail-closed，客户端可感知降级）。
|
||||
"""
|
||||
try:
|
||||
body = await request.json()
|
||||
except Exception:
|
||||
body = {}
|
||||
text = str((body or {}).get("input") or "").strip()
|
||||
if not text:
|
||||
return JSONResponse({"error": {"message": "input 必填"}}, status_code=400)
|
||||
model = str((body or {}).get("model") or cfg.embedder.model)
|
||||
try:
|
||||
vec = await embed(text, cfg.embedder)
|
||||
except EmbedderDown as e:
|
||||
return JSONResponse(
|
||||
{"error": {"message": str(e), "type": "embedder_down"}}, status_code=503)
|
||||
return {"object": "list", "model": model,
|
||||
"data": [{"object": "embedding", "index": 0, "embedding": vec}]}
|
||||
|
||||
return [main, v1]
|
||||
|
||||
Reference in New Issue
Block a user