fix: 模型级强制参数(kimi-k2.7-code 锁 temperature=1)+ 测速超时/错误文案
- config: ModelInfo.params + MODEL_PARAM_KEYS 白名单 + _parse_model_params(写错即 ConfigError) - adapters: _forced_params 在 payload 组装后强制覆盖;_learn_from_upstream_error 按上游 400 要求修正并重发一次(非原样重试);流式已吐内容则不重发 - main: /v1/models 暴露 params;探针超时 15s→30s;_speedtest_error_text 说清超时/上游原文 - settings: 面板保存保留手写 models[].params(原先会被静默清掉) - gateway.yaml: kimi-k2.7-code params.temperature=1 - 前端: Models 挂锁+参数固定标签;Chat 温度 chip 挂锁+抽屉提示;新增 lock 图标;中英对称 - tests: +10 → 168 passed / 0 failed
This commit is contained in:
1 parent
43a4f8613b
commit
840c5c83aa
14 files changed
+631
-13
No files matched your search
@@ -30,4 +30,10 @@ def create_adapter(
|
||||
timeout=timeout,
|
||||
media=pconfig.media,
|
||||
requires_session_header=pconfig.requires_session_header,
|
||||
# 第二十五轮: 把模型级强制参数(providers.*.models[].params)随适配器带下去。
|
||||
# 适配器按 model id 查表,在**组装完 payload 之后**覆盖,客户端无法绕过。
|
||||
# 放在适配器而非每个调用点:所有入口(对话 / 流式 / 批量 / 一键测速)自动一致,
|
||||
# 不会出现"主链路修了、测速还漏着"的偏差。
|
||||
# 配置变更时 main._apply_config 会丢弃整个适配器缓存,约束不会滞留旧值。
|
||||
model_params={m.id: dict(m.params) for m in pconfig.models if m.params},
|
||||
)
|
||||
@@ -11,13 +11,31 @@ Agnes / 火山方舟 两家均使用 OpenAI 兼容 API
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import logging
|
||||
import re
|
||||
import uuid
|
||||
from typing import Any, AsyncIterator
|
||||
|
||||
import httpx
|
||||
|
||||
from app.adapters.base import BaseAdapter, json_or_raise
|
||||
from app.config import MODEL_PARAM_KEYS
|
||||
from app.models import CHAT_CORE_KEYS, PROTECTED_VIDEO_KEYS, strip_protected
|
||||
|
||||
log = logging.getLogger("api-gateway")
|
||||
|
||||
# ── 上游"某参数只允许取某个值"的报错形态(第二十五轮)──
|
||||
# 各家措辞不同,这里收录已实测到的两类:
|
||||
# OpenCode Go: "Error from provider (Console Go): [invalid_request_error]
|
||||
# invalid temperature: only 1 is allowed for this model"
|
||||
# "max_tokens must be greater than 2"
|
||||
# 参数名不写死 —— 同类约束也可能出现在 top_p / max_tokens 上。
|
||||
_PARAM_ONLY_RE = re.compile(
|
||||
r"invalid\s+([a-z_][a-z0-9_]*)\s*:\s*only\s+([-+]?\d+(?:\.\d+)?)\s+is allowed", re.I)
|
||||
_PARAM_MIN_RE = re.compile(
|
||||
r"([a-z_][a-z0-9_]*)\s+must be greater than\s+(\d+)", re.I)
|
||||
|
||||
|
||||
|
||||
class OpenAICompatAdapter(BaseAdapter):
|
||||
"""适用于所有 OpenAI 兼容 /chat/completions 端点的适配器。"""
|
||||
@@ -30,6 +48,7 @@ class OpenAICompatAdapter(BaseAdapter):
|
||||
timeout: int = 120,
|
||||
media=None,
|
||||
requires_session_header: bool = False,
|
||||
model_params: dict[str, dict[str, Any]] | None = None,
|
||||
):
|
||||
super().__init__(name=name, base_url=base_url, timeout=timeout, media=media)
|
||||
self._headers: dict[str, str] = headers or {}
|
||||
@@ -37,6 +56,77 @@ class OpenAICompatAdapter(BaseAdapter):
|
||||
# 不能放静态 headers(_get_adapter 会缓存适配器、冻结 header),
|
||||
# 所以在每次请求时动态生成一个 UUID 会话头。
|
||||
self._requires_session = requires_session_header
|
||||
# 模型级强制参数:model id → {参数名: 强制值}(来自 gateway.yaml 的 models[].params)
|
||||
self._model_params: dict[str, dict[str, Any]] = {
|
||||
k: dict(v) for k, v in (model_params or {}).items() if v
|
||||
}
|
||||
# 运行期自学习到的参数约束(进程内,不落盘):见 _learn_from_upstream_error。
|
||||
# 仅作兜底 —— 管理员在 gateway.yaml 写死 params 后就不再需要它。
|
||||
self._learned_params: dict[str, dict[str, Any]] = {}
|
||||
|
||||
# ── 模型级强制参数 / 上游参数约束自学习(第二十五轮) ──
|
||||
|
||||
def _forced_params(self, model: str) -> dict[str, Any]:
|
||||
"""该模型发往上游前**必须使用**的参数值。
|
||||
|
||||
两层,配置优先:`gateway.yaml` 的 `models[].params` > 运行期自学习值。
|
||||
配置优先是刻意的 —— 管理员写死的口径是最终口径,自学习绝不能把它改掉。
|
||||
"""
|
||||
out = dict(self._learned_params.get(model) or {})
|
||||
out.update(self._model_params.get(model) or {})
|
||||
return out
|
||||
|
||||
def _learn_from_upstream_error(self, model: str, exc: Exception) -> dict[str, Any] | None:
|
||||
"""上游 400 说"这个参数只允许 X"→ 记下 X,并返回修正后的参数集供重发。
|
||||
|
||||
返回 None = 报文里没识别出可修正的参数(调用方按原样抛错)。
|
||||
|
||||
这不是「重试」:报文一模一样地重发只会再吃一次同样的 400,必须先把参数改成
|
||||
上游允许的值。学到之后会缓存到本适配器实例,后续请求直接按修正值组装 payload,
|
||||
不再多跑这一轮往返;配置变更时 main 会丢弃适配器缓存,学习值随之失效。
|
||||
"""
|
||||
resp = getattr(exc, "response", None)
|
||||
if resp is None or getattr(resp, "status_code", None) != 400:
|
||||
return None
|
||||
try:
|
||||
text = resp.text or ""
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
fixes: dict[str, Any] = {}
|
||||
for m in _PARAM_ONLY_RE.finditer(text):
|
||||
name = m.group(1).lower()
|
||||
if name in MODEL_PARAM_KEYS:
|
||||
raw = m.group(2)
|
||||
fixes[name] = float(raw) if "." in raw else int(raw)
|
||||
for m in _PARAM_MIN_RE.finditer(text):
|
||||
name = m.group(1).lower()
|
||||
if name in MODEL_PARAM_KEYS and name not in fixes:
|
||||
fixes[name] = int(m.group(2)) + 1 # “must be greater than N” → 发 N+1
|
||||
if not fixes:
|
||||
return None
|
||||
|
||||
cfg_params = self._model_params.get(model) or {}
|
||||
learned = dict(self._learned_params.get(model) or {})
|
||||
changed: dict[str, Any] = {}
|
||||
for k, v in fixes.items():
|
||||
if k in cfg_params:
|
||||
# 管理员已在 params 里写死该参数 → 不偷偷覆盖配置;但冲突必须说出来,
|
||||
# 否则用户只会看到「明明配了却还是 400」,无从下手。
|
||||
if cfg_params[k] != v:
|
||||
log.warning("模型 %s 的 params.%s=%r 与上游要求(%r)冲突,仍按配置发送",
|
||||
model, k, cfg_params[k], v)
|
||||
continue
|
||||
if learned.get(k) != v:
|
||||
learned[k] = v
|
||||
changed[k] = v
|
||||
if not changed:
|
||||
return None # 没学到新东西 → 不重发,避免在同一约束上来回打转
|
||||
self._learned_params[model] = learned
|
||||
log.warning("上游限制了模型 %s 的参数:%s —— 已自动修正并按修正值重发一次;"
|
||||
"建议在 gateway.yaml 该模型下写死 params,省掉以后每次这一轮额外往返",
|
||||
model, changed)
|
||||
return self._forced_params(model)
|
||||
|
||||
def _req_headers(self, session_id: str | None = None) -> dict[str, str]:
|
||||
"""每次请求的 headers。需要会话头的供应商动态注入 x-opencode-session。
|
||||
@@ -85,6 +175,11 @@ class OpenAICompatAdapter(BaseAdapter):
|
||||
# 以及供应商特有参数)必须放行 —— 见 models.CHAT_CORE_KEYS 的说明。
|
||||
if extra:
|
||||
payload.update(strip_protected(extra, CHAT_CORE_KEYS))
|
||||
# 第二十五轮: 模型级强制参数放在**最后**覆盖 —— 客户端(含 extra 逃生舱)改不动它。
|
||||
# 这是应对"上游只接受某个值"的最终口径,优先级高于客户端的一切输入。
|
||||
forced = self._forced_params(model)
|
||||
if forced:
|
||||
payload.update(forced)
|
||||
return payload
|
||||
|
||||
async def chat(
|
||||
@@ -105,6 +200,18 @@ class OpenAICompatAdapter(BaseAdapter):
|
||||
model, messages, temperature, top_p, max_tokens,
|
||||
stream, stop, tools, tool_choice, extra,
|
||||
)
|
||||
try:
|
||||
return await self._post_chat(payload, session_id)
|
||||
except httpx.HTTPStatusError as e:
|
||||
# 上游嫌参数(如 temperature)不合法时,按它要求的值修正后**重发一次**。
|
||||
# fixed 为 None(不是参数约束 / 已学过仍失败)则原样抛错,不做无谓重发。
|
||||
fixed = self._learn_from_upstream_error(model, e)
|
||||
if fixed is None:
|
||||
raise
|
||||
return await self._post_chat({**payload, **fixed}, session_id)
|
||||
|
||||
async def _post_chat(self, payload: dict[str, Any], session_id: str | None) -> dict[str, Any]:
|
||||
"""执行一次非流式 /chat/completions 请求(不含参数修正逻辑)。"""
|
||||
# Bug 4 修复: 用 lease 占用 client,热重载 close 会等本次请求结束才真正关连接
|
||||
async with self._lease() as client:
|
||||
resp = await client.post("/chat/completions", json=payload, headers=self._req_headers(session_id))
|
||||
@@ -114,7 +221,7 @@ class OpenAICompatAdapter(BaseAdapter):
|
||||
# 统一返回格式
|
||||
return {
|
||||
"provider": self.name,
|
||||
"model": model,
|
||||
"model": payload.get("model", ""),
|
||||
"choices": data.get("choices", []),
|
||||
"usage": data.get("usage"),
|
||||
"raw": data,
|
||||
@@ -138,6 +245,27 @@ class OpenAICompatAdapter(BaseAdapter):
|
||||
model, messages, temperature, top_p, max_tokens,
|
||||
True, stop, tools, tool_choice, extra,
|
||||
)
|
||||
started = False
|
||||
try:
|
||||
async for chunk in self._stream_once(payload, session_id):
|
||||
started = True
|
||||
yield chunk
|
||||
return
|
||||
except httpx.HTTPStatusError as e:
|
||||
# 已经开始吐内容就不能重发(会重复输出 / 串话);只有"开局就被上游 400 顶回来"
|
||||
# 才允许按上游要求修正参数重来一次。
|
||||
if started:
|
||||
raise
|
||||
fixed = self._learn_from_upstream_error(model, e)
|
||||
if fixed is None:
|
||||
raise
|
||||
payload = {**payload, **fixed}
|
||||
async for chunk in self._stream_once(payload, session_id):
|
||||
yield chunk
|
||||
|
||||
async def _stream_once(self, payload: dict[str, Any],
|
||||
session_id: str | None) -> AsyncIterator[dict]:
|
||||
"""执行一次流式 /chat/completions(不含参数修正逻辑)。"""
|
||||
# Bug 4 修复: lease 在整个流式期间持有占用,热重载/关闭不会掐断在途流
|
||||
async with self._lease() as client:
|
||||
async with client.stream(
|
||||
|
||||
Reference in new issue
Block a user