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
+33
-3
@@ -868,6 +868,27 @@ def _speedtest_retryable(e: Exception) -> bool:
|
||||
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:
|
||||
"""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。"""
|
||||
@@ -1123,7 +1144,10 @@ def _models_payload():
|
||||
"display_name": m["display_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)
|
||||
"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():
|
||||
data.append({"id": r["name"], "object": "model", "created": created,
|
||||
"owned_by": "gateway", "type": "route", "models": r["models"],
|
||||
@@ -1164,7 +1188,13 @@ async def models_speedtest(request: Request, _auth=Depends(require_auth)):
|
||||
except Exception:
|
||||
_body = {}
|
||||
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] = []
|
||||
for m in cfg.all_models():
|
||||
if m.get("type") != "text":
|
||||
@@ -1201,7 +1231,7 @@ async def models_speedtest(request: Request, _auth=Depends(require_auth)):
|
||||
ok = True
|
||||
break
|
||||
except Exception as e:
|
||||
err = _upstream_detail(str(e))
|
||||
err = _speedtest_error_text(e, probe_timeout)
|
||||
if attempt < 3 and _speedtest_retryable(e):
|
||||
await asyncio.sleep(0.3 * attempt)
|
||||
else:
|
||||
|
||||
Reference in new issue
Block a user