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
e4065d31c3
commit
4507dc4b8e
14 files changed
+631
-13
No files matched your search
@@ -4,6 +4,54 @@
|
|||||||
|
|
||||||
---
|
---
|
||||||
|
|
||||||
|
## 2026-09-18 · 第二十五轮:模型参数硬约束 + 测速超时/错误文案(一轮全修)
|
||||||
|
|
||||||
|
**背景(两条线一起收)**:上一轮排查出两个"看起来都坏了"的模型,其实是两回事:
|
||||||
|
1. **`kimi-k2.7-code` 每次必挂** —— 上游只接受 `temperature=1`,而一键测速探针固定发 `temperature: 0`,于是每次都 400 `invalid temperature: only 1 is allowed for this model`。更严重的是:这不只影响测速 —— **前端对话页的温度滑块默认 0.7**,用户把这个模型拖到任何非 1 的值都会 400。
|
||||||
|
2. **`mimo-v2.5` 不是坏,是慢** —— 直连 2.2~3s、过网关 2.9~8.4s,偶发冲破探针 15s 超时。但旧文案只回 `str(e)`,两者在前端都显示成含糊的「调用失败」,无从区分。
|
||||||
|
|
||||||
|
### 改动
|
||||||
|
- **`app/config.py`**
|
||||||
|
- `ModelInfo` 新增 `params: dict`(模型级**强制参数**):发往上游前覆盖客户端传值。
|
||||||
|
- 新增 `MODEL_PARAM_KEYS` 白名单(`temperature/top_p/top_k/max_tokens/min_tokens/presence_penalty/frequency_penalty/repetition_penalty/seed/logit_bias/n`)。**结构性键(`model`/`messages`/`stream`/`tools`)刻意不放行** —— 放行等于把路由与协议语义交给配置文件,重开一条越权面。
|
||||||
|
- 新增 `_parse_model_params()`:非映射、空参数名、白名单外的键、非法值类型一律 `ConfigError`(**写错立刻报错,绝不静默失效**)。
|
||||||
|
- `all_models()` 暴露 `params`,供 `/v1/models` 与前端标注。
|
||||||
|
- **`app/adapters/openai_compat.py`**
|
||||||
|
- 新增 `model_params` 入参、`_forced_params(model)`(配置 > 运行期自学习)。
|
||||||
|
- `_build_payload()` **末尾**应用强制参数:在 `strip_protected(extra, CHAT_CORE_KEYS)` **之后**覆盖,客户端连同 `extra` 逃生舱都改不动。
|
||||||
|
- 新增 `_learn_from_upstream_error()`:上游 400 报文里出现 `invalid <参数>: only <X> is allowed` / `<参数> must be greater than <N>` 时,按上游要求修正并**重发一次**。这不是"重试"——报文照原样重发只会再吃一次 400。学到的值缓存进实例(后续请求直接按修正值组装,不再多一轮往返);学不到新东西就返回 `None`,绝不与同一约束来回打转;管理员已在配置里写死的参数**不被偷偷改**(值冲突时只打日志说明)。
|
||||||
|
- `chat` / `chat_stream` 拆出 `_post_chat` / `_stream_once`,各自在"开局就被 400 顶回"时走一次修正重发;**流式若已开始吐内容则不重发**(避免重复输出/串话)。
|
||||||
|
- **`app/adapters/__init__.py`**:`create_adapter` 把各模型 `params` 带进适配器 —— 对话 / 流式 / 批量 / 一键测速**所有入口自动一致**,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` 丢弃适配器缓存,约束不会滞留旧值。
|
||||||
|
- **`app/main.py`**
|
||||||
|
- `_models_payload()` 在 `/v1/models` 暴露 `params`(该端点自己构造 data 项,不是直接用 `all_models()` 的结果)。
|
||||||
|
- 测速探针超时 **15s → 30s**(≈实测最慢值的 3 倍)。该 timeout 只作用于探针自己创建的 adapter(适配器缓存 key 含 timeout),**不影响真实对话链路**的 `server.timeout`。
|
||||||
|
- 新增 `_speedtest_error_text()`:超时 → `上游超时(>30s 未响应)`;上游 4xx/5xx → 透出**上游原文**(如 `invalid temperature: only 1 is allowed for this model`)。
|
||||||
|
- **`app/settings.py`**
|
||||||
|
- `view_to_raw()` 的模型项改为「以文件原文为底」再覆盖面板管的 4 个键 —— 面板的模型编辑器不认 `params`,照它重建会**静默清掉**手写约束(管理员只是在页面上改了别的东西,上游硬约束就没了)。
|
||||||
|
- `config_to_view()` 回传 `params`,roundtrip 不丢。
|
||||||
|
- **`data/gateway.yaml`**:`kimi-k2.7-code` 写死 `params: {temperature: 1}`。
|
||||||
|
- **前端**
|
||||||
|
- `Models.vue`:模型行前缀挂锁图标(`title` 说明),模型清单「能力」列新增「参数固定」标签(列宽 120 → 190)。
|
||||||
|
- `Chat.vue`:温度 chip 加挂锁;参数抽屉顶部新增说明条「该模型参数由网关固定:… (你在这里调的值会被网关覆盖后再发送)」;模型下拉选项同步挂锁。
|
||||||
|
- `Icon.vue` 新增 `lock`;中英文案 `paramTag` / `paramLocked` 对称(726/726)。
|
||||||
|
|
||||||
|
### 验证
|
||||||
|
- **上游全量扫描**:20 个文本模型逐个以 `temperature=0` 探测,**只有 `kimi-k2.7-code` 被拒**(400 `invalid temperature: only 1 is allowed for this model`),改 `temperature=1` 复测 200。→ 约束只此一处,未凭猜测给别的模型加参数。
|
||||||
|
- **配置层**:非法 `params`(结构性键 / 非法值类型 / 非映射)全部被 `ConfigError` 拦下;真配置解析出 `opencode/kimi-k2.7-code → {'temperature': 1}`;`/v1/models` 暴露一致。
|
||||||
|
- **适配器层(MockTransport,7 个场景)**:① 配置强制覆盖客户端传值且只发 1 次;② 未配时 400 → 按上游要求修正重发(0 → 1);③ 学到后后续请求直接带修正值(1 次);④ 上游一直 400 → 只发 2 次就抛错;⑤ 配置值与上游要求冲突 → 尊重配置、不发第二枪;⑥ 非参数类 400(额度)→ 原样抛错,不误改参数;⑦ 流式开局 400 → 修正后重发成功。
|
||||||
|
- **隔离实例真机**(`/tmp/gw25-verify/project`,8899,生产 8808 全程未动):真 key 实测 `kimi-k2.7-code` 测速 **1668ms 成功**、`mimo-v2.5` **4691ms 成功**;主链路显式 `temperature=0` → **200 + 正常内容**(`finish_reason=stop`),流式同样 200。
|
||||||
|
- **真浏览器**(同一隔离实例):Models 页 `kimi-k2.7-code` 行出现橙色挂锁、「参数固定」标签、单行测速 **1566 ms**;Chat 页选中该模型后温度 chip 带挂锁、抽屉顶部出现固定说明条。
|
||||||
|
- **单测**:**168 passed / 0 failed**(+10:参数解析/白名单拦截/真配置锁定/强制覆盖/自动修正重发/不无限重试/配置冲突/流式修正/测速文案/面板保留 params);前端 `vue-tsc --noEmit` 通过、`check:i18n` 726/726 对称、`check:svg` 通过、`npm run build` 成功(在副本内构建,未覆盖生产 dist)。
|
||||||
|
|
||||||
|
### 上线(待老板批准)
|
||||||
|
- 含 Python 代码改动(`config` / `adapters` / `main` / `settings`)→ 需 `supervisorctl restart api-gateway:api-gateway_00` 才生效(会踢网页会话)。
|
||||||
|
- 前端需 `npm run build` 覆盖 `frontend/dist`(挂锁/提示条才会出现在线上)。
|
||||||
|
- CNB 推送新增提交。
|
||||||
|
|
||||||
|
**一句话**:给模型加「强制参数」机制并给上游硬约束的 `kimi-k2.7-code` 写死 `temperature: 1`(同时兜一道"上游报什么就改成什么"的自学习重发),把探针超时从 15s 放宽到 30s 并让失败文案说清是超时还是上游拒绝 —— 于是"必挂的模型"变 1.5s 成功、"慢的模型"不再被误判成坏。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
## 2026-09-18 · OpenCode Go 会话 id 升级为「对话级稳定」(方案 A)
|
## 2026-09-18 · OpenCode Go 会话 id 升级为「对话级稳定」(方案 A)
|
||||||
|
|
||||||
**背景**:上一轮 `x-opencode-session` 是每请求随机 UUID —— 能过上游 400 强校验,但每次 id 都变,吃不到文档所述的提示词缓存。本轮按老板要求实现**稳定会话**:同一段对话共用同一个 session id。
|
**背景**:上一轮 `x-opencode-session` 是每请求随机 UUID —— 能过上游 400 强校验,但每次 id 都变,吃不到文档所述的提示词缓存。本轮按老板要求实现**稳定会话**:同一段对话共用同一个 session id。
|
||||||
|
|||||||
@@ -87,10 +87,20 @@ providers:
|
|||||||
video_model_name: "agnes-video-2.5-flash"
|
video_model_name: "agnes-video-2.5-flash"
|
||||||
models:
|
models:
|
||||||
- { id: "agnes-3.0-flash", display_name: "Agnes 3.0 Flash", type: "text" }
|
- { id: "agnes-3.0-flash", display_name: "Agnes 3.0 Flash", type: "text" }
|
||||||
|
# params(可选,第二十五轮):该模型发往上游前**强制覆盖**的参数,用于应对上游硬约束。
|
||||||
|
# kimi-k2.7-code 只接受 temperature=1,写死后客户端传什么值都按 1 发(含一键测速探针)。
|
||||||
|
- id: "kimi-k2.7-code"
|
||||||
|
display_name: "Kimi K2.7 Code"
|
||||||
|
type: "text"
|
||||||
|
params: { temperature: 1 }
|
||||||
```
|
```
|
||||||
|
|
||||||
认证优先级:**环境变量 > yaml 里的 `token` 字段**。
|
认证优先级:**环境变量 > yaml 里的 `token` 字段**。
|
||||||
|
|
||||||
|
**模型级强制参数 `models[].params`(第二十五轮)**:网关在组装完请求(**在 `extra` 逃生舱之后**)才应用它,所以客户端改不动 —— 目的就是"上游只接受某个值,别让用户踩雷"。允许的键:`temperature`、`top_p`、`top_k`、`max_tokens`、`min_tokens`、`presence_penalty`、`frequency_penalty`、`repetition_penalty`、`seed`、`logit_bias`、`n`;`model` / `messages` / `stream` / `tools` 这类**结构性键被刻意排除**(放行等于让配置文件改写路由与协议语义)。键名写错或值类型不对 → 启动/保存配置时直接 `ConfigError`,不会静默失效。`GET /v1/models` 会回 `params`,前端据此标挂锁。
|
||||||
|
|
||||||
|
此外还有一层运行时兜底:上游若回 400 说某参数"只允许 X"(`invalid temperature: only 1 is allowed for this model`)或"必须大于 N",网关会自动按上游要求修正并**重发一次**,并把这个值记在适配器实例上(后续请求不再多跑一轮);日志会提示建议把它写进 `params`。管理员已写死的值不会被自动改掉(冲突时只记日志)。
|
||||||
|
|
||||||
### 路由别名 + 按「可用 + 速度」动态排序 + 熔断
|
### 路由别名 + 按「可用 + 速度」动态排序 + 熔断
|
||||||
(第十一轮 #4:`wjd` 不再按配置里的数组优先级,改为运行时挑可用且最快的文本模型先调用;详见下方路由说明)
|
(第十一轮 #4:`wjd` 不再按配置里的数组优先级,改为运行时挑可用且最快的文本模型先调用;详见下方路由说明)
|
||||||
|
|
||||||
@@ -1183,7 +1193,7 @@ python smoke_test.py --url http://127.0.0.1:8808
|
|||||||
python live_check.py --url http://127.0.0.1:8808
|
python live_check.py --url http://127.0.0.1:8808
|
||||||
```
|
```
|
||||||
|
|
||||||
> 当前状态:单元测试 **137 passed / 0 failed**(`python tests/test_gateway.py`,不需要 pytest);冒烟 5/5 PASS;活体检查 14 PASS / 0 FAIL / 0 WARN。
|
> 当前状态:单元测试 **168 passed / 0 failed**(`python tests/test_gateway.py`,不需要 pytest);冒烟 5/5 PASS;活体检查 14 PASS / 0 FAIL / 0 WARN。
|
||||||
|
|
||||||
## 备注
|
## 备注
|
||||||
|
|
||||||
|
|||||||
@@ -30,4 +30,10 @@ def create_adapter(
|
|||||||
timeout=timeout,
|
timeout=timeout,
|
||||||
media=pconfig.media,
|
media=pconfig.media,
|
||||||
requires_session_header=pconfig.requires_session_header,
|
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
|
from __future__ import annotations
|
||||||
|
|
||||||
import json
|
import json
|
||||||
|
import logging
|
||||||
|
import re
|
||||||
import uuid
|
import uuid
|
||||||
from typing import Any, AsyncIterator
|
from typing import Any, AsyncIterator
|
||||||
|
|
||||||
|
import httpx
|
||||||
|
|
||||||
from app.adapters.base import BaseAdapter, json_or_raise
|
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
|
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):
|
class OpenAICompatAdapter(BaseAdapter):
|
||||||
"""适用于所有 OpenAI 兼容 /chat/completions 端点的适配器。"""
|
"""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,
|
timeout: int = 120,
|
||||||
media=None,
|
media=None,
|
||||||
requires_session_header: bool = False,
|
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)
|
super().__init__(name=name, base_url=base_url, timeout=timeout, media=media)
|
||||||
self._headers: dict[str, str] = headers or {}
|
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),
|
# 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。
|
# 所以在每次请求时动态生成一个 UUID 会话头。
|
||||||
self._requires_session = requires_session_header
|
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]:
|
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。
|
"""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。
|
# 以及供应商特有参数)必须放行 —— 见 models.CHAT_CORE_KEYS 的说明。
|
||||||
if extra:
|
if extra:
|
||||||
payload.update(strip_protected(extra, CHAT_CORE_KEYS))
|
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
|
return payload
|
||||||
|
|
||||||
async def chat(
|
async def chat(
|
||||||
@@ -105,6 +200,18 @@ class OpenAICompatAdapter(BaseAdapter):
|
|||||||
model, messages, temperature, top_p, max_tokens,
|
model, messages, temperature, top_p, max_tokens,
|
||||||
stream, stop, tools, tool_choice, extra,
|
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
|
# 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:
|
async with self._lease() as client:
|
||||||
resp = await client.post("/chat/completions", json=payload, headers=self._req_headers(session_id))
|
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
|
# 统一返回格式
|
||||||
return {
|
return {
|
||||||
"provider": self.name,
|
"provider": self.name,
|
||||||
"model": model,
|
"model": payload.get("model", ""),
|
||||||
"choices": data.get("choices", []),
|
"choices": data.get("choices", []),
|
||||||
"usage": data.get("usage"),
|
"usage": data.get("usage"),
|
||||||
"raw": data,
|
"raw": data,
|
||||||
@@ -138,6 +245,27 @@ class OpenAICompatAdapter(BaseAdapter):
|
|||||||
model, messages, temperature, top_p, max_tokens,
|
model, messages, temperature, top_p, max_tokens,
|
||||||
True, stop, tools, tool_choice, extra,
|
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
|
# 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 self._lease() as client:
|
||||||
async with client.stream(
|
async with client.stream(
|
||||||
|
|||||||
+63
-2
@@ -15,6 +15,19 @@ class ConfigError(ValueError):
|
|||||||
|
|
||||||
|
|
||||||
# ── 模型 / 供应商 ──
|
# ── 模型 / 供应商 ──
|
||||||
|
# 模型级「强制参数」(providers.<key>.models[].params)允许出现的键。
|
||||||
|
#
|
||||||
|
# 只放采样 / 长度类参数:它们不影响"调哪个模型、走什么协议",被覆盖最坏也只是改变生成风格。
|
||||||
|
# 而 model / messages / stream / tools / stop 这类**结构性**键一旦允许出现在这里,
|
||||||
|
# 等于把路由与协议语义交给配置文件,等于重开一条越权面(见 app/models.py 的
|
||||||
|
# PROTECTED_PAYLOAD_KEYS 说明),故一律不放行。
|
||||||
|
MODEL_PARAM_KEYS: frozenset[str] = frozenset({
|
||||||
|
"temperature", "top_p", "top_k", "min_tokens", "max_tokens",
|
||||||
|
"presence_penalty", "frequency_penalty", "repetition_penalty",
|
||||||
|
"seed", "logit_bias", "n",
|
||||||
|
})
|
||||||
|
|
||||||
|
|
||||||
@dataclass
|
@dataclass
|
||||||
class ModelInfo:
|
class ModelInfo:
|
||||||
id: str
|
id: str
|
||||||
@@ -25,6 +38,14 @@ class ModelInfo:
|
|||||||
# 其余文本模型(ark-code 等)文档未声明图像输入 → 默认 False。
|
# 其余文本模型(ark-code 等)文档未声明图像输入 → 默认 False。
|
||||||
# 用途:① /v1/models 暴露给前端标注/告警;② 带图请求走路由别名时把有视觉能力的模型排到前面。
|
# 用途:① /v1/models 暴露给前端标注/告警;② 带图请求走路由别名时把有视觉能力的模型排到前面。
|
||||||
vision: bool = False
|
vision: bool = False
|
||||||
|
# 第二十五轮(模型参数硬约束): 该模型**必须用**的上游参数,发送前强制覆盖客户端传值。
|
||||||
|
# 背景:OpenCode Go 的 kimi-k2.7-code 只接受 temperature=1,客户端传 0(含一键测速的
|
||||||
|
# 探针,它固定发 temperature=0)会被上游 400:
|
||||||
|
# "invalid temperature: only 1 is allowed for this model"
|
||||||
|
# 配置成 {"temperature": 1} 后,网关无论收到什么值都按 1 发 —— 用户在前端把温度滑块
|
||||||
|
# 拉到 0 也不会再 400(值被网关改成上游唯一允许的那个)。
|
||||||
|
# 键必须在 MODEL_PARAM_KEYS 白名单内,否则 validate_config 直接报错(不静默失效)。
|
||||||
|
params: dict[str, Any] = field(default_factory=dict)
|
||||||
|
|
||||||
|
|
||||||
@dataclass
|
@dataclass
|
||||||
@@ -225,6 +246,10 @@ class GatewayConfig:
|
|||||||
"vision": m.vision,
|
"vision": m.vision,
|
||||||
"provider": pkey,
|
"provider": pkey,
|
||||||
"provider_name": pconfig.name,
|
"provider_name": pconfig.name,
|
||||||
|
# 第二十五轮: 暴露模型级强制参数(如 {"temperature": 1})。
|
||||||
|
# 前端据此标注「该模型温度已由网关固定」,API 调用方也能提前知道
|
||||||
|
# 自己传的采样参数会被上游要求覆盖,省掉一轮「传了没生效」的困惑。
|
||||||
|
"params": dict(m.params or {}),
|
||||||
})
|
})
|
||||||
return result
|
return result
|
||||||
|
|
||||||
@@ -296,6 +321,40 @@ def _opt_int(v: Any, field_name: str, default: int, minimum: int | None = None)
|
|||||||
return val
|
return val
|
||||||
|
|
||||||
|
|
||||||
|
def _parse_model_params(v: Any, where: str) -> dict[str, Any]:
|
||||||
|
"""解析模型级「强制参数」(第二十五轮)。
|
||||||
|
|
||||||
|
形如 `params: {temperature: 1}` —— 网关发往上游前用这些值**覆盖**客户端传的同名参数。
|
||||||
|
用来应对上游的硬约束(某模型只接受 temperature=1 之类):否则客户端传别的值就 400,
|
||||||
|
而且用户在前端拉个滑块就踩雷,根本不可能自己猜出上游要什么。
|
||||||
|
|
||||||
|
容错口径(宁可报错,不可静默失效):
|
||||||
|
- 缺失 / None → {}(不是错误,只是这个模型没有约束);
|
||||||
|
- 不是映射 → ConfigError;
|
||||||
|
- 参数名不在 MODEL_PARAM_KEYS 白名单内 → ConfigError(防止把 model/messages/stream
|
||||||
|
这类结构性键配进来,那等于让配置文件改写路由语义);
|
||||||
|
- 值类型必须是标量 / list / dict。
|
||||||
|
"""
|
||||||
|
if v is None:
|
||||||
|
return {}
|
||||||
|
if not isinstance(v, dict):
|
||||||
|
raise ConfigError(
|
||||||
|
f"{where} 必须是映射(如 {{temperature: 1}}),得到 {type(v).__name__}: {v!r}")
|
||||||
|
out: dict[str, Any] = {}
|
||||||
|
for k, val in v.items():
|
||||||
|
key = str(k).strip()
|
||||||
|
if not key:
|
||||||
|
raise ConfigError(f"{where} 出现空参数名")
|
||||||
|
if key not in MODEL_PARAM_KEYS:
|
||||||
|
raise ConfigError(
|
||||||
|
f"{where}.{key} 不是允许的模型参数(允许:{', '.join(sorted(MODEL_PARAM_KEYS))})")
|
||||||
|
if val is None or isinstance(val, (bool, int, float, str, list, dict)):
|
||||||
|
out[key] = val
|
||||||
|
else:
|
||||||
|
raise ConfigError(f"{where}.{key} 的值类型不支持: {type(val).__name__}")
|
||||||
|
return out
|
||||||
|
|
||||||
|
|
||||||
def _parse_media(pval: dict[str, Any]) -> MediaConfig:
|
def _parse_media(pval: dict[str, Any]) -> MediaConfig:
|
||||||
m = pval.get("media", {}) or {}
|
m = pval.get("media", {}) or {}
|
||||||
return MediaConfig(
|
return MediaConfig(
|
||||||
@@ -423,8 +482,10 @@ def parse_config(raw: dict[str, Any]) -> GatewayConfig:
|
|||||||
)
|
)
|
||||||
models = [
|
models = [
|
||||||
ModelInfo(id=m.get("id", ""), display_name=m.get("display_name", ""),
|
ModelInfo(id=m.get("id", ""), display_name=m.get("display_name", ""),
|
||||||
type=m.get("type", "text"), vision=_opt_bool(m.get("vision"), "providers.*.models.*.vision", False))
|
type=m.get("type", "text"), vision=_opt_bool(m.get("vision"), "providers.*.models.*.vision", False),
|
||||||
for m in pval.get("models", [])
|
params=_parse_model_params(
|
||||||
|
m.get("params"), f"providers.{pkey}.models[{idx}].params"))
|
||||||
|
for idx, m in enumerate(pval.get("models", []))
|
||||||
]
|
]
|
||||||
providers[pkey] = ProviderConfig(
|
providers[pkey] = ProviderConfig(
|
||||||
key=pkey,
|
key=pkey,
|
||||||
|
|||||||
+33
-3
@@ -868,6 +868,27 @@ def _speedtest_retryable(e: Exception) -> bool:
|
|||||||
return status == 429 or (status is not None and status >= 500)
|
return status == 429 or (status is not None and status >= 500)
|
||||||
|
|
||||||
|
|
||||||
|
def _speedtest_error_text(e: Exception, timeout: float) -> str:
|
||||||
|
"""测速探针的失败文案(第二十五轮):把「超时」和「上游原文」这两个关键信息说清楚。
|
||||||
|
|
||||||
|
修复前只回 `str(e)`,形如 `Client error '400 Bad Request' for url ...` ——
|
||||||
|
上游真正的原因(如 `invalid temperature: only 1 is allowed for this model`)
|
||||||
|
根本看不到;超时更是只显示成含糊的「调用失败」。结果是:
|
||||||
|
「参数被上游拒绝」(每次都失败)与「模型响应慢」(偶发失败)
|
||||||
|
在前端长得一模一样,只能靠猜 —— 排查这两个模型时就是这么被误导的。
|
||||||
|
"""
|
||||||
|
if isinstance(e, (httpx.TimeoutException, TimeoutError)):
|
||||||
|
return f"上游超时(>{int(timeout)}s 未响应)"
|
||||||
|
resp = getattr(e, "response", None)
|
||||||
|
if resp is not None:
|
||||||
|
detail = _upstream_detail(getattr(resp, "text", "") or "")
|
||||||
|
if detail:
|
||||||
|
return detail
|
||||||
|
return f"上游返回 HTTP {getattr(resp, 'status_code', '?')}(无响应正文)"
|
||||||
|
text = str(e)
|
||||||
|
return _upstream_detail(text) or text[:200] or "调用失败"
|
||||||
|
|
||||||
|
|
||||||
def _clear_auth_fails(request: Request) -> None:
|
def _clear_auth_fails(request: Request) -> None:
|
||||||
"""登录成功后清掉该来源的失败计数:否则同一出口 IP(办公室 NAT)被前人失败连带,
|
"""登录成功后清掉该来源的失败计数:否则同一出口 IP(办公室 NAT)被前人失败连带,
|
||||||
真人输对密码后紧接一次手误就吃 429。"""
|
真人输对密码后紧接一次手误就吃 429。"""
|
||||||
@@ -1123,7 +1144,10 @@ def _models_payload():
|
|||||||
"display_name": m["display_name"],
|
"display_name": m["display_name"],
|
||||||
"type": m["type"], "provider": m["provider"], "provider_name": m["provider_name"],
|
"type": m["type"], "provider": m["provider"], "provider_name": m["provider_name"],
|
||||||
# 第十三轮: 暴露「是否支持图像输入」,前端据此标注/告警(对话页带图时提示换模型)
|
# 第十三轮: 暴露「是否支持图像输入」,前端据此标注/告警(对话页带图时提示换模型)
|
||||||
"vision": bool(m.get("vision"))})
|
"vision": bool(m.get("vision")),
|
||||||
|
# 第二十五轮: 暴露模型级强制参数(如 {"temperature": 1})。前端据此标挂锁,
|
||||||
|
# API 调用方也能提前知道自己传的采样参数会被上游要求覆盖。
|
||||||
|
"params": dict(m.get("params") or {})})
|
||||||
for r in cfg.all_routes():
|
for r in cfg.all_routes():
|
||||||
data.append({"id": r["name"], "object": "model", "created": created,
|
data.append({"id": r["name"], "object": "model", "created": created,
|
||||||
"owned_by": "gateway", "type": "route", "models": r["models"],
|
"owned_by": "gateway", "type": "route", "models": r["models"],
|
||||||
@@ -1164,7 +1188,13 @@ async def models_speedtest(request: Request, _auth=Depends(require_auth)):
|
|||||||
except Exception:
|
except Exception:
|
||||||
_body = {}
|
_body = {}
|
||||||
only = _body.get("model") if isinstance(_body, dict) else None
|
only = _body.get("model") if isinstance(_body, dict) else None
|
||||||
probe_timeout = 15.0
|
# 探针超时(第二十五轮: 15s → 30s):
|
||||||
|
# 15s 对"慢的推理模型"太紧 —— 实测 mimo-v2.5 过网关 2.9s~8.4s、agnes-2.5-flash 11s,
|
||||||
|
# 上游一排队就冲破 15s,被记成「调用失败」,和「模型真的坏了」完全分不清。
|
||||||
|
# 30s ≈ 实测最慢值的 3 倍,仍远小于真实对话链路用的 server.timeout。
|
||||||
|
# 注意:该 timeout 只作用于**本探针自己创建的 adapter**(适配器缓存 key 含 timeout),
|
||||||
|
# 改它不影响用户对话的超时行为。
|
||||||
|
probe_timeout = 30.0
|
||||||
results: list[dict] = []
|
results: list[dict] = []
|
||||||
for m in cfg.all_models():
|
for m in cfg.all_models():
|
||||||
if m.get("type") != "text":
|
if m.get("type") != "text":
|
||||||
@@ -1201,7 +1231,7 @@ async def models_speedtest(request: Request, _auth=Depends(require_auth)):
|
|||||||
ok = True
|
ok = True
|
||||||
break
|
break
|
||||||
except Exception as e:
|
except Exception as e:
|
||||||
err = _upstream_detail(str(e))
|
err = _speedtest_error_text(e, probe_timeout)
|
||||||
if attempt < 3 and _speedtest_retryable(e):
|
if attempt < 3 and _speedtest_retryable(e):
|
||||||
await asyncio.sleep(0.3 * attempt)
|
await asyncio.sleep(0.3 * attempt)
|
||||||
else:
|
else:
|
||||||
|
|||||||
+18
-3
@@ -89,7 +89,10 @@ def config_to_view(cfg: GatewayConfig) -> dict[str, Any]:
|
|||||||
"requires_session_header": bool(p.requires_session_header),
|
"requires_session_header": bool(p.requires_session_header),
|
||||||
"media": {k: getattr(p.media, k) for k in MEDIA_KEYS},
|
"media": {k: getattr(p.media, k) for k in MEDIA_KEYS},
|
||||||
"models": [
|
"models": [
|
||||||
{"id": m.id, "display_name": m.display_name, "type": m.type, "vision": bool(m.vision)}
|
{"id": m.id, "display_name": m.display_name, "type": m.type, "vision": bool(m.vision),
|
||||||
|
# 第二十五轮: 回传模型级强制参数。面板不提供编辑入口(在 YAML 全文编辑器里改),
|
||||||
|
# 但必须一路带着走 —— 否则读回来再提交一次就会把它弄丢。
|
||||||
|
"params": dict(m.params or {})}
|
||||||
for m in p.models
|
for m in p.models
|
||||||
],
|
],
|
||||||
}
|
}
|
||||||
@@ -220,14 +223,26 @@ def view_to_raw(cfg: GatewayConfig, raw: dict[str, Any], view: dict[str, Any]) -
|
|||||||
media_in = _as_dict(pv.get("media"), f"providers.{key}.media")
|
media_in = _as_dict(pv.get("media"), f"providers.{key}.media")
|
||||||
base_media = _as_dict(base_p.get("media"), f"providers.{key}.media")
|
base_media = _as_dict(base_p.get("media"), f"providers.{key}.media")
|
||||||
models = []
|
models = []
|
||||||
|
# 第二十五轮: 模型项要「以文件原文为底」再覆盖面板管得着的键。
|
||||||
|
# 面板的模型编辑器只认 id/display_name/type/vision 四个字段,若照着它重建,
|
||||||
|
# 手写在 YAML 里的 models[].params(模型级强制参数,如 temperature: 1)
|
||||||
|
# 会被**静默清掉** —— 管理员只是在页面上改了别的东西,上游硬约束就没了,
|
||||||
|
# 之后又变成"这个模型怎么老是 400"。按 id 对齐原文、保留其余键即可。
|
||||||
|
base_models: dict[str, dict[str, Any]] = {}
|
||||||
|
for bm in (base_p.get("models") or []):
|
||||||
|
if isinstance(bm, dict) and str(bm.get("id") or "").strip():
|
||||||
|
base_models[str(bm["id"]).strip()] = bm
|
||||||
for i, m in enumerate(pv.get("models") or []):
|
for i, m in enumerate(pv.get("models") or []):
|
||||||
m = _as_dict(m, f"providers.{key}.models[{i}]")
|
m = _as_dict(m, f"providers.{key}.models[{i}]")
|
||||||
models.append({
|
mid = str(m.get("id") or "").strip()
|
||||||
"id": str(m.get("id") or "").strip(),
|
item = dict(base_models.get(mid) or {})
|
||||||
|
item.update({
|
||||||
|
"id": mid,
|
||||||
"display_name": str(m.get("display_name") or "").strip(),
|
"display_name": str(m.get("display_name") or "").strip(),
|
||||||
"type": str(m.get("type") or "text").strip(),
|
"type": str(m.get("type") or "text").strip(),
|
||||||
"vision": bool(m.get("vision", False)),
|
"vision": bool(m.get("vision", False)),
|
||||||
})
|
})
|
||||||
|
models.append(item)
|
||||||
providers[key] = {
|
providers[key] = {
|
||||||
"name": str(pv.get("name") or key).strip(),
|
"name": str(pv.get("name") or key).strip(),
|
||||||
"base_url": str(pv.get("base_url") or "").strip(),
|
"base_url": str(pv.get("base_url") or "").strip(),
|
||||||
|
|||||||
@@ -86,6 +86,12 @@ providers:
|
|||||||
- id: "kimi-k2.7-code"
|
- id: "kimi-k2.7-code"
|
||||||
display_name: "Kimi K2.7 Code"
|
display_name: "Kimi K2.7 Code"
|
||||||
type: "text"
|
type: "text"
|
||||||
|
# 上游硬约束:该模型**只接受 temperature=1**,传 0 / 0.7 等一律 400
|
||||||
|
# "invalid temperature: only 1 is allowed for this model"。
|
||||||
|
# 写死后网关无论收到什么温度都按 1 发(前端温度滑块、API 传参、一键测速探针
|
||||||
|
# 统一生效);客户端不传 temperature 时也会显式按 1 发送,与上游默认值一致。
|
||||||
|
params:
|
||||||
|
temperature: 1
|
||||||
- id: "kimi-k2.6"
|
- id: "kimi-k2.6"
|
||||||
display_name: "Kimi K2.6"
|
display_name: "Kimi K2.6"
|
||||||
type: "text"
|
type: "text"
|
||||||
|
|||||||
@@ -66,6 +66,8 @@ const ICONS: Record<string, string> = {
|
|||||||
inbox: '<path d="M3.4 11.6 6 4.6h8l2.6 7v4.8H3.4z"/><path d="M3.4 11.6h3.2l.9 1.8h5l.9-1.8h3.2"/>',
|
inbox: '<path d="M3.4 11.6 6 4.6h8l2.6 7v4.8H3.4z"/><path d="M3.4 11.6h3.2l.9 1.8h5l.9-1.8h3.2"/>',
|
||||||
// 第二十四轮:账号停用(禁用符号「圆 + 斜杠」),启用配已有的 play(圆 + 三角)
|
// 第二十四轮:账号停用(禁用符号「圆 + 斜杠」),启用配已有的 play(圆 + 三角)
|
||||||
ban: '<circle cx="10" cy="10" r="7"/><path d="M5.1 5.1l9.8 9.8"/>',
|
ban: '<circle cx="10" cy="10" r="7"/><path d="M5.1 5.1l9.8 9.8"/>',
|
||||||
|
// 第二十五轮:模型参数被上游固定(挂锁 = 这参数你说了不算,由网关按上游要求锁定)
|
||||||
|
lock: '<rect x="4.6" y="9" width="10.8" height="7.2" rx="2"/><path d="M7 9V6.8a3 3 0 0 1 6 0V9"/>',
|
||||||
}
|
}
|
||||||
|
|
||||||
const body = computed(() => ICONS[props.name] || '')
|
const body = computed(() => ICONS[props.name] || '')
|
||||||
|
|||||||
@@ -254,6 +254,7 @@ export default {
|
|||||||
imgMax: 'At most {n} images per message',
|
imgMax: 'At most {n} images per message',
|
||||||
imgPickErr: 'Image processing failed: {msg}',
|
imgPickErr: 'Image processing failed: {msg}',
|
||||||
visionWarn: 'Model {model} does not accept image input — images will be ignored or rejected. Switch to a vision-capable model.',
|
visionWarn: 'Model {model} does not accept image input — images will be ignored or rejected. Switch to a vision-capable model.',
|
||||||
|
paramLocked: 'Parameters pinned by the gateway: {p} (upstream hard limit; what you set here is overridden before sending)',
|
||||||
useVision: 'Switch to {model}',
|
useVision: 'Switch to {model}',
|
||||||
waiting: 'Awaiting input…',
|
waiting: 'Awaiting input…',
|
||||||
send: 'Send',
|
send: 'Send',
|
||||||
@@ -384,6 +385,8 @@ export default {
|
|||||||
loadErr: 'Failed to load models',
|
loadErr: 'Failed to load models',
|
||||||
capability: 'Capability',
|
capability: 'Capability',
|
||||||
visionTag: 'Vision',
|
visionTag: 'Vision',
|
||||||
|
paramTag: 'Params locked',
|
||||||
|
paramLocked: 'Parameters pinned by the gateway: {p} (upstream hard limit; values from clients and the speed test are overridden)',
|
||||||
allTextModels: 'All text models',
|
allTextModels: 'All text models',
|
||||||
providerCard: 'Providers',
|
providerCard: 'Providers',
|
||||||
keyOk: 'Key configured',
|
keyOk: 'Key configured',
|
||||||
|
|||||||
@@ -254,6 +254,7 @@ export default {
|
|||||||
imgMax: '一条消息最多 {n} 张图片',
|
imgMax: '一条消息最多 {n} 张图片',
|
||||||
imgPickErr: '图片处理失败:{msg}',
|
imgPickErr: '图片处理失败:{msg}',
|
||||||
visionWarn: '当前模型 {model} 不支持图像输入,图片会被忽略或直接报错 —— 请换用带视觉能力的模型。',
|
visionWarn: '当前模型 {model} 不支持图像输入,图片会被忽略或直接报错 —— 请换用带视觉能力的模型。',
|
||||||
|
paramLocked: '该模型参数由网关固定:{p}(上游硬约束,你在这里调的值会被网关覆盖后再发送)',
|
||||||
useVision: '改用 {model}',
|
useVision: '改用 {model}',
|
||||||
waiting: '等待输入…',
|
waiting: '等待输入…',
|
||||||
send: '发送',
|
send: '发送',
|
||||||
@@ -384,6 +385,8 @@ export default {
|
|||||||
loadErr: '加载模型失败',
|
loadErr: '加载模型失败',
|
||||||
capability: '能力',
|
capability: '能力',
|
||||||
visionTag: '视觉',
|
visionTag: '视觉',
|
||||||
|
paramTag: '参数固定',
|
||||||
|
paramLocked: '该模型参数由网关固定:{p}(上游硬约束,客户端与测速探针传的值都会被覆盖)',
|
||||||
allTextModels: '全部文本模型',
|
allTextModels: '全部文本模型',
|
||||||
providerCard: '供应商',
|
providerCard: '供应商',
|
||||||
keyOk: 'Key 已配置',
|
keyOk: 'Key 已配置',
|
||||||
|
|||||||
@@ -67,7 +67,7 @@
|
|||||||
<span class="chip-sel" :class="{ on: params.stream }" @click="paramsOpen = true">
|
<span class="chip-sel" :class="{ on: params.stream }" @click="paramsOpen = true">
|
||||||
<Icon name="bolt" :size="13" />{{ t('chat.streaming') }}
|
<Icon name="bolt" :size="13" />{{ t('chat.streaming') }}
|
||||||
</span>
|
</span>
|
||||||
<span class="chip-sel" @click="paramsOpen = true">{{ t('chat.tempShort') }} {{ params.temperature }}</span>
|
<span class="chip-sel" @click="paramsOpen = true" :title="currentPinNote || undefined">{{ t('chat.tempShort') }} {{ params.temperature }}<Icon v-if="currentPinNote" name="lock" :size="11" class="plock" /></span>
|
||||||
<span class="chip-sel" @click="paramsOpen = true">
|
<span class="chip-sel" @click="paramsOpen = true">
|
||||||
{{ t('chat.maxShort') }} {{ params.max_tokens ?? t('common.none') }}
|
{{ t('chat.maxShort') }} {{ params.max_tokens ?? t('common.none') }}
|
||||||
</span>
|
</span>
|
||||||
@@ -213,6 +213,8 @@
|
|||||||
<span class="opt">
|
<span class="opt">
|
||||||
<span>{{ modelLabel(m) }}</span>
|
<span>{{ modelLabel(m) }}</span>
|
||||||
<Icon v-if="m.vision" name="eye" :size="13" />
|
<Icon v-if="m.vision" name="eye" :size="13" />
|
||||||
|
<!-- 第二十五轮:参数被上游锁定的模型(网关会覆盖客户端传值) -->
|
||||||
|
<Icon v-if="pinned(m)" name="lock" :size="13" class="plock" :title="pinNote(m)" />
|
||||||
</span>
|
</span>
|
||||||
</el-option>
|
</el-option>
|
||||||
</el-option-group>
|
</el-option-group>
|
||||||
@@ -236,6 +238,12 @@
|
|||||||
|
|
||||||
<!-- 参数抽屉:窄屏按屏宽百分比,避免 330px 在小屏上顶到边 -->
|
<!-- 参数抽屉:窄屏按屏宽百分比,避免 330px 在小屏上顶到边 -->
|
||||||
<el-drawer v-model="paramsOpen" :title="t('chat.params')" :size="isMobile ? '86%' : '330px'" :close-on-click-modal="true">
|
<el-drawer v-model="paramsOpen" :title="t('chat.params')" :size="isMobile ? '86%' : '330px'" :close-on-click-modal="true">
|
||||||
|
<!-- 第二十五轮:当前模型有上游硬约束的参数时,先说清楚"这里的值会被覆盖后再发",
|
||||||
|
否则用户会以为调了温度没生效(kimi-k2.7-code 只接受 temperature=1 就是这种情况)。 -->
|
||||||
|
<div v-if="currentPinNote" class="pin-note">
|
||||||
|
<Icon name="lock" :size="13" class="plock" />
|
||||||
|
<span>{{ currentPinNote }}</span>
|
||||||
|
</div>
|
||||||
<div class="param">
|
<div class="param">
|
||||||
<div class="plabel">{{ t('chat.temp') }}</div>
|
<div class="plabel">{{ t('chat.temp') }}</div>
|
||||||
<el-slider v-model="params.temperature" :min="0" :max="2" :step="0.05" show-input />
|
<el-slider v-model="params.temperature" :min="0" :max="2" :step="0.05" show-input />
|
||||||
@@ -330,6 +338,22 @@ const params = reactive({
|
|||||||
system: '',
|
system: '',
|
||||||
})
|
})
|
||||||
const paramsOpen = ref(false)
|
const paramsOpen = ref(false)
|
||||||
|
|
||||||
|
// ── 模型级强制参数(第二十五轮)──
|
||||||
|
// 网关 /v1/models 会回 params(如 kimi-k2.7-code → {temperature: 1}):这些参数由网关
|
||||||
|
// 按上游硬约束覆盖后才发出去。不提示的话,用户把温度从 0.7 调到 0 也看不出区别,
|
||||||
|
// 只会以为"网关把我参数吞了"。这里只做标注,参数本身在 gateway.yaml 里改。
|
||||||
|
function pinned(m: any): boolean {
|
||||||
|
return Object.keys((m && m.params) || {}).length > 0
|
||||||
|
}
|
||||||
|
function pinNote(m: any): string {
|
||||||
|
const ps = (m && m.params) || {}
|
||||||
|
return t('chat.paramLocked', { p: Object.keys(ps).map((k) => `${k}=${ps[k]}`).join(', ') })
|
||||||
|
}
|
||||||
|
const currentPinNote = computed(() => {
|
||||||
|
const m = textModels.value.find((x) => x.id === params.model)
|
||||||
|
return m ? pinNote(m) : ''
|
||||||
|
})
|
||||||
const streaming = ref(false)
|
const streaming = ref(false)
|
||||||
const messages = ref<ChatMessage[]>([])
|
const messages = ref<ChatMessage[]>([])
|
||||||
|
|
||||||
@@ -825,6 +849,14 @@ watch(messages, scrollToBottom, { deep: true })
|
|||||||
/* ── 参数抽屉 ── */
|
/* ── 参数抽屉 ── */
|
||||||
.param { margin-bottom: 18px; }
|
.param { margin-bottom: 18px; }
|
||||||
.plabel { font-size: 13px; color: var(--muted); margin-bottom: 6px; }
|
.plabel { font-size: 13px; color: var(--muted); margin-bottom: 6px; }
|
||||||
|
/* 第二十五轮:模型参数被上游锁定时的标注(挂锁 + 抽屉顶部说明条) */
|
||||||
|
.plock { color: var(--warn); margin-left: 4px; }
|
||||||
|
.pin-note {
|
||||||
|
display: flex; align-items: flex-start; gap: 6px;
|
||||||
|
margin: 0 0 14px; padding: 8px 10px; border-radius: 8px;
|
||||||
|
font-size: 12px; line-height: 1.5; color: var(--warn); background: var(--warn-soft);
|
||||||
|
}
|
||||||
|
.pin-note .plock { margin-left: 0; margin-top: 2px; }
|
||||||
|
|
||||||
/* ── 窄屏 ── */
|
/* ── 窄屏 ── */
|
||||||
@media (max-width: 768px) {
|
@media (max-width: 768px) {
|
||||||
|
|||||||
@@ -37,6 +37,9 @@
|
|||||||
<div v-for="m in textModelRows" :key="m.id" class="st-item">
|
<div v-for="m in textModelRows" :key="m.id" class="st-item">
|
||||||
<div class="st-row">
|
<div class="st-row">
|
||||||
<Icon v-if="m.vision" name="eye" :size="12" class="vision" />
|
<Icon v-if="m.vision" name="eye" :size="12" class="vision" />
|
||||||
|
<!-- 第二十五轮:该模型的参数被上游固定(如 temperature 只能是 1),
|
||||||
|
挂锁标注 + 悬浮说明,省掉「为什么我传的参数不生效」这类困惑。 -->
|
||||||
|
<Icon v-if="m.paramNote" name="lock" :size="12" class="plock" :title="m.paramNote" />
|
||||||
<span class="code">{{ m.id }}</span>
|
<span class="code">{{ m.id }}</span>
|
||||||
<span class="muted-2 st-prov">{{ m.provider }}</span>
|
<span class="muted-2 st-prov">{{ m.provider }}</span>
|
||||||
<span style="flex:1"></span>
|
<span style="flex:1"></span>
|
||||||
@@ -97,12 +100,15 @@
|
|||||||
<el-table-column :label="t('models.provider')" min-width="150">
|
<el-table-column :label="t('models.provider')" min-width="150">
|
||||||
<template #default="{ row }">{{ row.provider_name || '—' }}</template>
|
<template #default="{ row }">{{ row.provider_name || '—' }}</template>
|
||||||
</el-table-column>
|
</el-table-column>
|
||||||
<el-table-column :label="t('models.capability')" width="120">
|
<el-table-column :label="t('models.capability')" width="190">
|
||||||
<template #default="{ row }">
|
<template #default="{ row }">
|
||||||
<span v-if="row.vision" class="pill ok cap">
|
<span v-if="row.vision" class="pill ok cap">
|
||||||
<Icon name="eye" :size="12" />{{ t('models.visionTag') }}
|
<Icon name="eye" :size="12" />{{ t('models.visionTag') }}
|
||||||
</span>
|
</span>
|
||||||
<span v-else class="muted-2">—</span>
|
<span v-if="paramKeys(row).length" class="pill warn cap" :title="paramNote(row)">
|
||||||
|
<Icon name="lock" :size="12" />{{ t('models.paramTag') }}
|
||||||
|
</span>
|
||||||
|
<span v-if="!row.vision && !paramKeys(row).length" class="muted-2">—</span>
|
||||||
</template>
|
</template>
|
||||||
</el-table-column>
|
</el-table-column>
|
||||||
<el-table-column :label="t('models.keySet')" width="120">
|
<el-table-column :label="t('models.keySet')" width="120">
|
||||||
@@ -231,9 +237,23 @@ const textModelRows = computed(() =>
|
|||||||
id: m.id,
|
id: m.id,
|
||||||
provider: m.provider_name || '—',
|
provider: m.provider_name || '—',
|
||||||
vision: !!m.vision,
|
vision: !!m.vision,
|
||||||
|
// 模型级强制参数(第二十五轮):被上游固定住的采样参数,用挂锁标注出来。
|
||||||
|
paramNote: paramKeys(m).length ? paramNote(m) : '',
|
||||||
})),
|
})),
|
||||||
)
|
)
|
||||||
|
|
||||||
|
// ── 模型级强制参数(第二十五轮)──
|
||||||
|
// 后端 /v1/models 在 params 里给出「该模型发往上游前会被强制覆盖的参数」
|
||||||
|
// (例:kimi-k2.7-code 只接受 temperature=1)。这里只做展示,不提供编辑入口。
|
||||||
|
function paramKeys(m: any): string[] {
|
||||||
|
return Object.keys((m && m.params) || {})
|
||||||
|
}
|
||||||
|
function paramNote(m: any): string {
|
||||||
|
const ps = (m && m.params) || {}
|
||||||
|
const txt = paramKeys(m).map((k) => `${k}=${ps[k]}`).join(', ')
|
||||||
|
return t('models.paramLocked', { p: txt })
|
||||||
|
}
|
||||||
|
|
||||||
const keySet = computed(() => {
|
const keySet = computed(() => {
|
||||||
const ks: Record<string, boolean> = {}
|
const ks: Record<string, boolean> = {}
|
||||||
for (const p of Object.values(health.value.providers || {}) as any) ks[p.name] = !!p.api_key_set
|
for (const p of Object.values(health.value.providers || {}) as any) ks[p.name] = !!p.api_key_set
|
||||||
@@ -286,6 +306,7 @@ onMounted(async () => {
|
|||||||
.chain-item { display: inline-flex; align-items: center; gap: 6px; }
|
.chain-item { display: inline-flex; align-items: center; gap: 6px; }
|
||||||
.arrow { color: var(--muted-2); }
|
.arrow { color: var(--muted-2); }
|
||||||
.vision { color: var(--primary); }
|
.vision { color: var(--primary); }
|
||||||
|
.plock { color: var(--warn); }
|
||||||
.prov { padding: 10px 0; border-bottom: 1px solid var(--line-2); }
|
.prov { padding: 10px 0; border-bottom: 1px solid var(--line-2); }
|
||||||
.prov:last-child { border-bottom: 0; }
|
.prov:last-child { border-bottom: 0; }
|
||||||
.prov-top { display: flex; align-items: center; gap: 8px; }
|
.prov-top { display: flex; align-items: center; gap: 8px; }
|
||||||
|
|||||||
@@ -3779,6 +3779,259 @@ def test_opencode_session_header_from_conversation_id():
|
|||||||
assert "x-opencode-session" not in ag._req_headers("conv-abc")
|
assert "x-opencode-session" not in ag._req_headers("conv-abc")
|
||||||
|
|
||||||
|
|
||||||
|
def test_api_conversation_id_flows_to_session_header():
|
||||||
|
"""API 调用(HTTP body)传 conversation_id → 适配器 session_id;且不透传上游 payload。
|
||||||
|
|
||||||
|
覆盖老板问的「非前端、纯 API 调用是否支持稳定会话」:只要请求体带
|
||||||
|
conversation_id,后端就把它作为 OpenCode Go 的 x-opencode-session 透传。
|
||||||
|
"""
|
||||||
|
from fastapi.testclient import TestClient
|
||||||
|
import app.main as m
|
||||||
|
|
||||||
|
captured = {}
|
||||||
|
|
||||||
|
class Capture:
|
||||||
|
async def chat(self, model=None, messages=None, session_id=None, extra=None, **kw):
|
||||||
|
captured["session_id"] = session_id
|
||||||
|
captured["extra"] = extra
|
||||||
|
return {"id": "c1", "object": "chat.completion", "model": model,
|
||||||
|
"choices": [{"message": {"role": "assistant", "content": "ok"},
|
||||||
|
"finish_reason": "stop"}],
|
||||||
|
"usage": {"prompt_tokens": 2, "completion_tokens": 1}}
|
||||||
|
|
||||||
|
orig = m._get_adapter
|
||||||
|
m._get_adapter = lambda model_id, timeout=None, cfg=None: (
|
||||||
|
Capture(), "agnes", m._cfg().providers["agnes"])
|
||||||
|
try:
|
||||||
|
client = TestClient(m.app)
|
||||||
|
# ① 带 conversation_id → 原样落为 session_id;且不作为上游参数透传
|
||||||
|
r = client.post("/v1/chat/completions", headers={"Authorization": _AUTOBEARER},
|
||||||
|
json={"model": "agnes-3.0-flash",
|
||||||
|
"messages": [{"role": "user", "content": "hi"}],
|
||||||
|
"conversation_id": "conv-api-42"})
|
||||||
|
assert r.status_code == 200, r.text
|
||||||
|
assert captured["session_id"] == "conv-api-42"
|
||||||
|
assert not captured["extra"] or "conversation_id" not in captured["extra"]
|
||||||
|
# ② 不传 → session_id 为 None(由适配器回退随机 UUID)
|
||||||
|
r2 = client.post("/v1/chat/completions", headers={"Authorization": _AUTOBEARER},
|
||||||
|
json={"model": "agnes-3.0-flash",
|
||||||
|
"messages": [{"role": "user", "content": "hi"}]})
|
||||||
|
assert r2.status_code == 200, r2.text
|
||||||
|
assert captured["session_id"] is None
|
||||||
|
finally:
|
||||||
|
m._get_adapter = orig
|
||||||
|
|
||||||
|
|
||||||
|
# ── 第二十五轮:模型级强制参数(上游硬约束)+ 测速超时/错误文案 ──
|
||||||
|
#
|
||||||
|
# 起因:kimi-k2.7-code 上游只接受 temperature=1,测速探针固定发 0 → 每次必 400;
|
||||||
|
# 而 mimo-v2.5 是慢(偶发冲破 15s 探针超时),却被显示成同样的「调用失败」。
|
||||||
|
# 两条修法各自的回归如下。
|
||||||
|
|
||||||
|
def test_model_params_parsed_from_config():
|
||||||
|
"""gateway.yaml 的 models[].params 要被解析出来,并随 all_models 暴露。"""
|
||||||
|
from app.config import parse_config
|
||||||
|
cfg = parse_config({"providers": {"p": {"base_url": "http://x/v1", "models": [
|
||||||
|
{"id": "m1", "params": {"temperature": 1, "top_p": 0.9}},
|
||||||
|
{"id": "m2"}]}}})
|
||||||
|
assert cfg.providers["p"].models[0].params == {"temperature": 1, "top_p": 0.9}
|
||||||
|
assert cfg.providers["p"].models[1].params == {}
|
||||||
|
am = {m["id"]: m for m in cfg.all_models()}
|
||||||
|
assert am["m1"]["params"] == {"temperature": 1, "top_p": 0.9}
|
||||||
|
|
||||||
|
|
||||||
|
def test_model_params_rejects_bad_keys_and_types():
|
||||||
|
"""params 的键必须在白名单内:结构性键(model/messages/stream)绝不允许 ——
|
||||||
|
否则等于让配置文件改写路由与协议语义(回到越权面)。写错必须报错,不能静默失效。"""
|
||||||
|
from app.config import parse_config, ConfigError
|
||||||
|
bads = [{"model": "evil"}, {"messages": []}, {"stream": True}, {"temperature": object()}, [1, 2]]
|
||||||
|
for bad in bads:
|
||||||
|
try:
|
||||||
|
parse_config({"providers": {"p": {"base_url": "http://x/v1",
|
||||||
|
"models": [{"id": "m", "params": bad}]}}})
|
||||||
|
except ConfigError:
|
||||||
|
continue
|
||||||
|
raise AssertionError(f"非法 params 未被拦截: {bad!r}")
|
||||||
|
|
||||||
|
|
||||||
|
def test_real_config_locks_kimi_temperature():
|
||||||
|
"""真配置口径:kimi-k2.7-code 只接受 temperature=1,必须在配置里写死。"""
|
||||||
|
cfg = load_config(CONFIG)
|
||||||
|
m = [x for x in cfg.providers["opencode"].models if x.id == "kimi-k2.7-code"][0]
|
||||||
|
assert m.params.get("temperature") == 1, m.params
|
||||||
|
|
||||||
|
|
||||||
|
def test_model_params_override_client_value():
|
||||||
|
"""配了 params 的模型:客户端传什么值都被覆盖,且只发一次请求(不做无谓重发)。"""
|
||||||
|
from app.adapters.openai_compat import OpenAICompatAdapter
|
||||||
|
seen = []
|
||||||
|
|
||||||
|
def handler(req):
|
||||||
|
seen.append(json.loads(req.content))
|
||||||
|
return httpx.Response(200, json={"id": "1", "choices": [], "usage": {}})
|
||||||
|
|
||||||
|
ad = OpenAICompatAdapter("t", "http://mock.test/v1",
|
||||||
|
model_params={"locked": {"temperature": 1}})
|
||||||
|
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
|
||||||
|
base_url="http://mock.test/v1")
|
||||||
|
|
||||||
|
async def run():
|
||||||
|
await ad.chat("locked", [{"role": "user", "content": "hi"}], temperature=0)
|
||||||
|
await ad.close()
|
||||||
|
|
||||||
|
asyncio.run(run())
|
||||||
|
assert seen[0]["temperature"] == 1, seen[0]
|
||||||
|
assert len(seen) == 1
|
||||||
|
|
||||||
|
|
||||||
|
def test_model_param_constraint_auto_corrected_and_retried():
|
||||||
|
"""上游 400 说「只允许 temperature=1」→ 网关按 1 重发一次(而非原样重试)。"""
|
||||||
|
from app.adapters.openai_compat import OpenAICompatAdapter
|
||||||
|
seen = []
|
||||||
|
|
||||||
|
def handler(req):
|
||||||
|
seen.append(json.loads(req.content))
|
||||||
|
if len(seen) == 1:
|
||||||
|
return httpx.Response(400, json={"error": {"message":
|
||||||
|
"Error from provider (Console Go): [invalid_request_error] "
|
||||||
|
"invalid temperature: only 1 is allowed for this model"}})
|
||||||
|
return httpx.Response(200, json={"id": "1", "choices": [
|
||||||
|
{"message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], "usage": {}})
|
||||||
|
|
||||||
|
ad = OpenAICompatAdapter("t", "http://mock.test/v1")
|
||||||
|
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
|
||||||
|
base_url="http://mock.test/v1")
|
||||||
|
|
||||||
|
async def run():
|
||||||
|
r = await ad.chat("kimi-k2.7-code", [{"role": "user", "content": "hi"}], temperature=0)
|
||||||
|
await ad.close()
|
||||||
|
return r
|
||||||
|
|
||||||
|
r = asyncio.run(run())
|
||||||
|
assert len(seen) == 2, seen
|
||||||
|
assert seen[0]["temperature"] == 0 and seen[1]["temperature"] == 1, seen
|
||||||
|
assert r["choices"][0]["message"]["content"] == "ok"
|
||||||
|
|
||||||
|
|
||||||
|
def test_model_param_constraint_no_infinite_retry():
|
||||||
|
"""上游始终 400 → 只修正重发一次就抛错(绝不能与同一约束来回打转)。"""
|
||||||
|
from app.adapters.openai_compat import OpenAICompatAdapter
|
||||||
|
seen = []
|
||||||
|
|
||||||
|
def handler(req):
|
||||||
|
seen.append(json.loads(req.content))
|
||||||
|
return httpx.Response(400, json={"error": {"message":
|
||||||
|
"invalid temperature: only 1 is allowed for this model"}})
|
||||||
|
|
||||||
|
ad = OpenAICompatAdapter("t", "http://mock.test/v1")
|
||||||
|
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
|
||||||
|
base_url="http://mock.test/v1")
|
||||||
|
|
||||||
|
async def run():
|
||||||
|
try:
|
||||||
|
await ad.chat("m", [{"role": "user", "content": "hi"}], temperature=0)
|
||||||
|
finally:
|
||||||
|
await ad.close()
|
||||||
|
|
||||||
|
try:
|
||||||
|
asyncio.run(run())
|
||||||
|
raise AssertionError("应抛出 httpx.HTTPStatusError")
|
||||||
|
except httpx.HTTPStatusError:
|
||||||
|
pass
|
||||||
|
assert len(seen) == 2, f"应只发 2 次(原始 + 修正),实际 {len(seen)}"
|
||||||
|
|
||||||
|
|
||||||
|
def test_model_param_conflict_respects_configured_value():
|
||||||
|
"""管理员写死的值与上游要求冲突 → 尊重配置,不偷偷改、也不发无谓的第二枪。"""
|
||||||
|
from app.adapters.openai_compat import OpenAICompatAdapter
|
||||||
|
seen = []
|
||||||
|
|
||||||
|
def handler(req):
|
||||||
|
seen.append(json.loads(req.content))
|
||||||
|
return httpx.Response(400, json={"error": {"message":
|
||||||
|
"invalid temperature: only 1 is allowed for this model"}})
|
||||||
|
|
||||||
|
ad = OpenAICompatAdapter("t", "http://mock.test/v1", model_params={"m": {"temperature": 0.5}})
|
||||||
|
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
|
||||||
|
base_url="http://mock.test/v1")
|
||||||
|
|
||||||
|
async def run():
|
||||||
|
try:
|
||||||
|
await ad.chat("m", [{"role": "user", "content": "hi"}], temperature=0.5)
|
||||||
|
finally:
|
||||||
|
await ad.close()
|
||||||
|
|
||||||
|
try:
|
||||||
|
asyncio.run(run())
|
||||||
|
raise AssertionError("应抛出 httpx.HTTPStatusError")
|
||||||
|
except httpx.HTTPStatusError:
|
||||||
|
pass
|
||||||
|
assert len(seen) == 1, seen
|
||||||
|
assert seen[0]["temperature"] == 0.5, seen[0]
|
||||||
|
|
||||||
|
|
||||||
|
def test_stream_param_constraint_auto_corrected():
|
||||||
|
"""流式路径同样自修复:开局 400 → 按上游要求修正后重发。"""
|
||||||
|
from app.adapters.openai_compat import OpenAICompatAdapter
|
||||||
|
seen = []
|
||||||
|
|
||||||
|
def handler(req):
|
||||||
|
seen.append(json.loads(req.content))
|
||||||
|
if len(seen) == 1:
|
||||||
|
return httpx.Response(400, json={"error": {"message":
|
||||||
|
"invalid temperature: only 1 is allowed for this model"}})
|
||||||
|
return httpx.Response(200, text='data: {"choices":[{"delta":{"content":"hi"}}]}\n\ndata: [DONE]\n\n',
|
||||||
|
headers={"content-type": "text/event-stream"})
|
||||||
|
|
||||||
|
ad = OpenAICompatAdapter("t", "http://mock.test/v1")
|
||||||
|
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
|
||||||
|
base_url="http://mock.test/v1")
|
||||||
|
|
||||||
|
async def run():
|
||||||
|
out = [c async for c in ad.chat_stream("kimi-k2.7-code",
|
||||||
|
[{"role": "user", "content": "hi"}], temperature=0)]
|
||||||
|
await ad.close()
|
||||||
|
return out
|
||||||
|
|
||||||
|
out = asyncio.run(run())
|
||||||
|
assert len(seen) == 2, seen
|
||||||
|
assert seen[0]["temperature"] == 0 and seen[1]["temperature"] == 1, seen
|
||||||
|
assert out, "流式应至少收到一个数据块"
|
||||||
|
|
||||||
|
|
||||||
|
def test_speedtest_error_text_surfaces_timeout_and_upstream_detail():
|
||||||
|
"""测速失败文案:超时要说清是超时,400 要透出上游原文。
|
||||||
|
(本轮排查时,"参数被上游拒绝" 与 "模型太慢" 在旧文案里长得一模一样。)"""
|
||||||
|
import app.main as m
|
||||||
|
t = m._speedtest_error_text(httpx.ReadTimeout("timed out"), 30.0)
|
||||||
|
assert "超时" in t and "30" in t, t
|
||||||
|
|
||||||
|
req = httpx.Request("POST", "http://x/v1/chat/completions")
|
||||||
|
resp = httpx.Response(400, request=req, json={"error": {
|
||||||
|
"message": "invalid temperature: only 1 is allowed for this model"}})
|
||||||
|
e = httpx.HTTPStatusError("Client error '400 Bad Request'", request=req, response=resp)
|
||||||
|
t2 = m._speedtest_error_text(e, 30.0)
|
||||||
|
assert "only 1 is allowed" in t2, t2
|
||||||
|
|
||||||
|
|
||||||
|
def test_settings_panel_preserves_model_params():
|
||||||
|
"""管理页面保存配置时,手写在 YAML 里的 models[].params 不能被面板冲掉。
|
||||||
|
(面板的模型编辑器只认 id/display_name/type/vision,照它重建就会静默丢掉 params。)"""
|
||||||
|
from app import settings as gw_settings
|
||||||
|
from app.config import parse_config
|
||||||
|
raw = {"providers": {"p": {"name": "P", "base_url": "http://x/v1", "models": [
|
||||||
|
{"id": "m", "display_name": "M", "type": "text", "vision": False,
|
||||||
|
"params": {"temperature": 1}}]}}}
|
||||||
|
cfg = parse_config(raw)
|
||||||
|
view = gw_settings.config_to_view(cfg)
|
||||||
|
# 面板读出来就带着 params
|
||||||
|
assert view["providers"]["p"]["models"][0]["params"] == {"temperature": 1}
|
||||||
|
# 提交回去(模拟面板原样回传)→ params 必须还在
|
||||||
|
out = gw_settings.view_to_raw(cfg, raw, {"providers": view["providers"]})
|
||||||
|
assert out["providers"]["p"]["models"][0]["params"] == {"temperature": 1}, \
|
||||||
|
out["providers"]["p"]["models"][0]
|
||||||
|
|
||||||
|
|
||||||
# 直接运行(非 pytest)
|
# 直接运行(非 pytest)
|
||||||
if __name__ == "__main__":
|
if __name__ == "__main__":
|
||||||
fns = [v for k, v in list(globals().items()) if k.startswith("test_") and callable(v)]
|
fns = [v for k, v in list(globals().items()) if k.startswith("test_") and callable(v)]
|
||||||
|
|||||||
Reference in new issue
Block a user