229 lines
10 KiB
Python
229 lines
10 KiB
Python
"""
|
||
LLM bridge for pipeline handlers — 统一收敛到模型治理模块（pipeline-llm）推理 API。
|
||
|
||
2026-09-04 改造：产线平台所有模型调用切换到 /pipeline-llm/api/v1（OpenAI 兼容，
|
||
分类照 llmage）。本模块不再直查旧 `llm` 表——签名保持不变，内部改为：
|
||
签发内部短期 token（机构隔离，真 key 不出进程）
|
||
→ HTTP 自调用本进程 /pipeline-llm/api/v1/chat/completions
|
||
→ 门禁链（限流/限额/主备容错/端点轮转/预授权）+ 双维度记账由端点侧执行
|
||
|
||
自调用走 127.0.0.1:<本进程端口>（web/worker 进程都起 HTTP，worker 端口 9090+N）。
|
||
每次自调用带机构/用户上下文，用量流水可归属到真实用户。
|
||
"""
|
||
|
||
import json
|
||
import logging
|
||
import time
|
||
|
||
logger = logging.getLogger("pipeline.llm_bridge")
|
||
|
||
# 内部自调用 token 缓存：(org_id, user_id, model_name) -> {token, calls, expires_at}
|
||
# 每次 LLM 调用都签发新 token 会让 tokens 表膨胀，故按上下文缓存复用；
|
||
# 接近调用上限或临近过期时换新（上限/过期由签发侧强制，本地计数只是提前量）。
|
||
_token_cache: dict = {}
|
||
_TOKEN_MAX_LOCAL_CALLS = 400 # token max_calls=500，留 100 余量防竞态
|
||
_TOKEN_REFRESH_BEFORE_EXPIRY = 600 # 距过期不足 10 分钟即换新
|
||
_TOKEN_TTL_HOURS = 8
|
||
|
||
|
||
async def _self_base_url():
|
||
"""本进程推理 API 基址（自调用，不出本机）。"""
|
||
port = 9090
|
||
try:
|
||
from ahserver.serverenv import ServerEnv
|
||
p = getattr(ServerEnv(), 'port', None)
|
||
if p:
|
||
port = int(p)
|
||
except (TypeError, ValueError):
|
||
port = 9090
|
||
except Exception:
|
||
port = 9090
|
||
return "http://127.0.0.1:%d/pipeline-llm/api/v1" % port
|
||
|
||
|
||
async def _get_internal_token(org_id, user_id, model_name):
|
||
"""取/发内部短期 token。失败抛 ValueError（消息真实可行动）。"""
|
||
key = (org_id or '0', user_id or '', model_name or '')
|
||
now = time.time()
|
||
ent = _token_cache.get(key)
|
||
if ent and ent['calls'] < _TOKEN_MAX_LOCAL_CALLS \
|
||
and ent['expires_at'] - now > _TOKEN_REFRESH_BEFORE_EXPIRY:
|
||
ent['calls'] += 1
|
||
return ent['token']
|
||
from .llm_proxy import create_llm_token
|
||
ok, token = await create_llm_token(
|
||
org_id or '0', project_id='', task_id='', model_name=model_name or '',
|
||
purpose='internal_bridge', ttl_hours=_TOKEN_TTL_HOURS,
|
||
max_calls=500, created_by=user_id or '')
|
||
if not ok:
|
||
raise ValueError('内部 LLM token 签发失败：%s（模型治理模块未就绪或机构标识缺失）' % token)
|
||
_token_cache[key] = {
|
||
'token': token, 'calls': 1,
|
||
'expires_at': now + _TOKEN_TTL_HOURS * 3600,
|
||
}
|
||
return token
|
||
|
||
|
||
async def _http_chat(payload, org_id, user_id, model_name, timeout: int = 0):
|
||
"""POST 本进程推理端点。返回上游响应 dict；失败抛 ValueError（消息真实可行动）。
|
||
|
||
timeout：客户端等待秒数（0=缺省 330）。异步生成模型（视频等）端点侧最长
|
||
等 900 秒，调用方须同步放大客户端超时，否则客户端先断连。
|
||
"""
|
||
import aiohttp
|
||
|
||
token = await _get_internal_token(org_id, user_id, model_name)
|
||
base = await _self_base_url()
|
||
# ⚠️ "Bearer " 前缀用拼接构造——字面量写在源码里会被脱敏工具替换成 ***
|
||
# （2026-09-04 实测：Authorization 头变成 "***plk-..." 致端点校验失败）
|
||
_BEARER = 'Bea' + 'rer '
|
||
headers = {"Authorization": _BEARER + token, "Content-Type": "application/json"}
|
||
_total = int(timeout) if timeout and int(timeout) > 0 else 330
|
||
try:
|
||
async with aiohttp.ClientSession() as session:
|
||
async with session.post(
|
||
base + "/chat/completions", headers=headers, json=payload,
|
||
timeout=aiohttp.ClientTimeout(total=_total, connect=30),
|
||
) as resp:
|
||
text = await resp.text()
|
||
status = resp.status
|
||
except Exception as e:
|
||
raise ValueError(
|
||
'LLM 推理端点不可达（%s/chat/completions）：%s。'
|
||
'请检查本进程服务是否正常' % (base, e))
|
||
try:
|
||
data = json.loads(text)
|
||
except Exception:
|
||
raise ValueError('LLM 推理端点返回非 JSON（HTTP %s）：%s' % (status, text[:200]))
|
||
if isinstance(data, dict) and data.get('error'):
|
||
err = data['error']
|
||
msg = err.get('message', '') if isinstance(err, dict) else str(err)
|
||
raise ValueError(msg or 'LLM 推理失败（无详情）')
|
||
return data
|
||
|
||
|
||
def _no_llm_error(model_name=None, org_id=None) -> ValueError:
|
||
"""兜底错误（正常路径错误消息来自端点侧，这里只防解析异常）。"""
|
||
return ValueError(
|
||
"模型「%s」调用失败：机构 %s 未完成模型治理接入（无容错策略或模型未注册）。"
|
||
"请在模型治理→组织容错策略配置主/备模型。" % (model_name or '(缺省)', org_id or '(未指定)'))
|
||
|
||
|
||
async def llm_call(prompt: str, model: str = None, temperature: float = 0.7,
|
||
org_id: str = None, user_id: str = None, purpose: str = '',
|
||
timeout: int = 0) -> str:
|
||
"""Call LLM and return text response.
|
||
|
||
统一走模型治理推理 API（门禁链 + 双维度记账）。
|
||
org_id 为空 = 系统级（'0'），与旧语义（不过滤机构）等价。
|
||
purpose='utility'：辅助任务（分类/选择/摘要），治理层按用途选模型链
|
||
（机构策略配置辅助模型优先），经 payload 的 _purpose 键透传到端点。
|
||
"""
|
||
# 兼容旧优先级：harnessed_agent（若宿主加载了独立推理后端）
|
||
try:
|
||
from ahserver.serverenv import ServerEnv
|
||
env = ServerEnv()
|
||
fn = getattr(env, 'llm_chat', None)
|
||
if callable(fn):
|
||
result = await fn(prompt, model=model, temperature=temperature)
|
||
if isinstance(result, dict):
|
||
return result.get("content", result.get("text", str(result)))
|
||
return str(result)
|
||
except Exception:
|
||
pass
|
||
|
||
payload = {
|
||
"model": model or '',
|
||
"messages": [{"role": "user", "content": prompt}],
|
||
"temperature": temperature,
|
||
}
|
||
if purpose:
|
||
payload["_purpose"] = purpose
|
||
if timeout:
|
||
payload["_timeout"] = int(timeout)
|
||
data = await _http_chat(payload, org_id or '0', user_id or '', model or '')
|
||
try:
|
||
return data["choices"][0]["message"]["content"]
|
||
except (KeyError, IndexError, TypeError) as e:
|
||
raise ValueError('LLM 响应缺 choices: %s' % (
|
||
json.dumps(data, ensure_ascii=False, default=str)[:300])) from e
|
||
|
||
|
||
async def call_llm(tenant_id: str, prompt: str, model: str = None, temperature: float = 0.7) -> str:
|
||
"""SDLC handler interface — delegates to llm_call."""
|
||
return await llm_call(prompt, model=model, temperature=temperature)
|
||
|
||
|
||
async def llm_call_msgs(messages: list, model: str = None, temperature: float = 0.7,
|
||
org_id: str = None, user_id: str = None, purpose: str = '',
|
||
timeout: int = 0) -> str:
|
||
"""Call LLM with full message array (system/user/assistant).
|
||
|
||
purpose='utility'：辅助任务（分类/选择/摘要），治理层按用途选模型链。
|
||
timeout：单次上游调用超时秒数（0=用端点默认；上限 900，超长文本提取用）。
|
||
"""
|
||
payload = {"model": model or '', "messages": messages, "temperature": temperature}
|
||
if purpose:
|
||
payload["_purpose"] = purpose
|
||
if timeout:
|
||
payload["_timeout"] = int(timeout)
|
||
data = await _http_chat(payload, org_id or '0', user_id or '', model or '')
|
||
try:
|
||
return data["choices"][0]["message"]["content"]
|
||
except (KeyError, IndexError, TypeError) as e:
|
||
raise ValueError('LLM 响应缺 choices: %s' % (
|
||
json.dumps(data, ensure_ascii=False, default=str)[:300])) from e
|
||
|
||
|
||
async def llm_call_msgs_native(messages: list, tools: list = None, model: str = None,
|
||
temperature: float = 0.7, org_id: str = None,
|
||
user_id: str = None, purpose: str = '') -> dict:
|
||
"""Native function calling. 传入 tools JSON schema，返回 message dict。
|
||
|
||
Returns:
|
||
{"content": str, "tool_calls": [{"id","type","function":{"name","arguments"}}]}
|
||
当模型返回 tool_calls 时，content 通常为空字符串。
|
||
purpose='utility'：辅助任务（分类/选择/摘要），治理层按用途选模型链。
|
||
"""
|
||
payload = {"model": model or '', "messages": messages, "temperature": temperature}
|
||
if tools:
|
||
payload["tools"] = tools
|
||
payload["tool_choice"] = "auto"
|
||
if purpose:
|
||
payload["_purpose"] = purpose
|
||
data = await _http_chat(payload, org_id or '0', user_id or '', model or '')
|
||
try:
|
||
msg = data["choices"][0]["message"]
|
||
except (KeyError, IndexError, TypeError) as e:
|
||
raise ValueError('LLM 响应缺 choices: %s' % (
|
||
json.dumps(data, ensure_ascii=False, default=str)[:300])) from e
|
||
return {
|
||
"content": msg.get("content") or "",
|
||
"tool_calls": msg.get("tool_calls") or [],
|
||
}
|
||
|
||
|
||
async def llm_infer(payload: dict, model: str = None, org_id: str = None,
|
||
user_id: str = None, timeout: int = 0) -> dict:
|
||
"""通用推理（全能力，2026-09-07）：透传任意 payload 到统一推理端点，
|
||
返回上游响应 dict（OpenAI 兼容 choices；生成类另带 media/output/task_id）。
|
||
|
||
与 llm_call* 的区别：不假设 messages 结构，调用方自己组包——供 agent 的
|
||
invoke_model 工具调用非对话能力（t2i/t2v/i2v/tts/asr 等）。payload 里的
|
||
messages 由调用方按能力构造；媒体参数按三数组契约（image_files/audio_files/
|
||
video_files），inference 层的 _normalize_media_aliases 会归一旧别名。
|
||
|
||
model 空 = 端点按机构策略缺省模型；owner/机构归属硬校验在治理层执行。
|
||
timeout：单次调用超时秒数（生成类慢，可传大值；上限 900，端点侧封顶）。
|
||
失败抛 ValueError（消息真实可行动，直接展示给 agent/用户）。
|
||
"""
|
||
body = dict(payload or {})
|
||
if model:
|
||
body["model"] = model
|
||
if timeout:
|
||
body["_timeout"] = int(timeout)
|
||
# 客户端等待须覆盖端点侧预算（_timeout 端点封顶 900）+ 余量，否则客户端先断连
|
||
client_timeout = min(int(timeout or 0), 900) + 60 if timeout else 0
|
||
return await _http_chat(body, org_id or '0', user_id or '', model or '',
|
||
timeout=client_timeout)
|