bizyengine 1.2.56__py3-none-any.whl → 1.2.57__py3-none-any.whl

bizyengine/misc/utils.py

@@ -1,14 +1,21 @@
+import asyncio
 import base64
+import concurrent.futures
 import json
+import logging
 import os
 import pickle
+import re
+import threading
+import time
 import urllib.parse
 import urllib.request
 import zlib
-from typing import List, Tuple, Union
+from typing import Callable, Dict, Generic, List, Optional, Tuple, TypeVar, Union
 
 import numpy as np
 
+from bizyengine.bizy_server.api_client import APIClient
 from bizyengine.core import pop_api_key_and_prompt_id
 from bizyengine.core.common import client
 from bizyengine.core.common.env_var import BIZYAIR_SERVER_ADDRESS
@@ -143,6 +150,18 @@ def get_llm_response(
     extra_data = pop_api_key_and_prompt_id(kwargs)
     headers = client.headers(api_key=extra_data["api_key"])
 
+    # 如果model已不可用，选择第一个可用model
+    if _MODELS_CACHE.get("llm_models") is None:
+        cache_models(extra_data["api_key"])
+    llm_models = _MODELS_CACHE.get("llm_models")
+    if llm_models is None:
+        logging.warning(f"No LLM models available, keeping the original model {model}")
+    elif model not in llm_models:
+        logging.warning(
+            f"Model {model} is not available, using the first available model {llm_models[0]}"
+        )
+        model = llm_models[0]
+
     payload = {
         "model": model,
         "messages": [
@@ -183,6 +202,18 @@ def get_vlm_response(
     extra_data = pop_api_key_and_prompt_id(kwargs)
     headers = client.headers(api_key=extra_data["api_key"])
 
+    # 如果model已不可用，选择第一个可用model
+    if _MODELS_CACHE.get("vlm_models") is None:
+        cache_models(extra_data["api_key"])
+    vlm_models = _MODELS_CACHE.get("vlm_models")
+    if vlm_models is None:
+        logging.warning(f"No VLM models available, keeping the original model {model}")
+    elif model not in vlm_models:
+        logging.warning(
+            f"Model {model} is not available, using the first available model {vlm_models[0]}"
+        )
+        model = vlm_models[0]
+
     messages = [
         {
             "role": "system",
@@ -230,3 +261,149 @@ def get_vlm_response(
         callback=None,
     )
     return response
+
+
+K = TypeVar("K")
+V = TypeVar("V")
+R = TypeVar("R")
+
+
+class TTLCache(Generic[K, V]):
+    """线程安全 TTL 内存缓存（仅依赖标准库）"""
+
+    def __init__(self, ttl_sec: float):
+        self.ttl = ttl_sec
+        self._data: Dict[K, tuple[V, float]] = {}
+        self._lock = threading.RLock()
+        self._stop_evt = threading.Event()
+        # 后台清扫线程
+        self._cleaner = threading.Thread(target=self._cleanup, daemon=True)
+        self._cleaner.start()
+
+    # ---------- 公共 API ----------
+    def set(self, key: K, value: V) -> None:
+        """写入/刷新键值"""
+        with self._lock:
+            self._data[key] = (value, time.time() + self.ttl)
+
+    def get(self, key: K) -> Optional[V]:
+        """读取键值；不存在或已过期返回 None"""
+        with self._lock:
+            val, expire = self._data.get(key, (None, 0))
+            if val is None or time.time() > expire:
+                self._data.pop(key, None)
+                return None
+            return val
+
+    def delete(self, key: K) -> None:
+        """手动删除"""
+        with self._lock:
+            self._data.pop(key, None)
+
+    def stop(self):
+        """停止后台线程（程序退出前调用）"""
+        self._stop_evt.set()
+        self._cleaner.join(timeout=self.ttl + 1)
+
+    # ---------- 内部 ----------
+    def _cleanup(self):
+        """周期清扫过期键"""
+        while not self._stop_evt.wait(self.ttl / 2):
+            with self._lock:
+                now = time.time()
+                for key, (_, expire) in list(self._data.items()):
+                    if now > expire:
+                        self._data.pop(key, None)
+
+
+class SingleFlight(Generic[R]):
+    """Python 版 singleflight.Group（线程安全）"""
+
+    def __init__(self):
+        self._lock = threading.Lock()
+        self._call_map: dict[str, SingleFlight._Call[R]] = {}
+
+    class _Call(Generic[R]):
+        __slots__ = ("mu", "done", "result", "err", "waiters")
+
+        def __init__(self):
+            self.mu = threading.Lock()
+            self.done = False
+            self.result: Optional[R] = None
+            self.err: Optional[BaseException] = None
+            self.waiters = 0
+
+    def do(
+        self, key: str, fn: Callable[[], R]
+    ) -> tuple[R, bool, Optional[BaseException]]:
+        """
+        返回值: (result, shared?, exception)
+        shared=True 表示本次未真正执行 fn，复用了别人结果
+        """
+        with self._lock:
+            call = self._call_map.get(key)
+            if call is None:  # 我是第一个
+                call = self._Call[R]()
+                call.waiters = 1
+                self._call_map[key] = call
+                first = True
+            else:  # 已有并发请求
+                call.waiters += 1
+                first = False
+
+        if first:  # 只有第一个真正执行
+            try:
+                result = fn()
+                with call.mu:
+                    call.result = result
+                    call.done = True
+            except BaseException as e:
+                with call.mu:
+                    call.err = e
+                    call.done = True
+                raise
+            finally:  # 把自己从 map 摘掉
+                with self._lock:
+                    if call.waiters == 0:
+                        self._call_map.pop(key, None)
+        else:  # 其它人阻塞等待
+            with call.mu:
+                while not call.done:
+                    call.mu.wait()
+        # 读取结果
+        with call.mu:
+            if call.err is not None:
+                return call.result, True, call.err
+            return call.result, True, None
+
+
+_MODELS_CACHE = TTLCache[str, list[str]](ttl_sec=600)
+_SF = SingleFlight[None]()
+
+
+def cache_models(request_api_key: str):
+    # TODO: 效果待验证，目前节点只会被ComfyUI串行执行，所以不会出现竞争
+    _SF.do("_cache_models", lambda: _cache_models(request_api_key))
+
+
+def _cache_models(request_api_key: str):
+    # ① 开一条新线程专门跑协程 - 应该不需要在prompt那层上锁，因为并发只有1
+    with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+        api_client = APIClient()
+        all_models = pool.submit(
+            asyncio.run, api_client.fetch_all_llm_models(request_api_key)
+        ).result()
+        if len(all_models) == 0:
+            return
+        llm_models = [
+            model
+            for model in all_models
+            if not (re.search(r"\d+(\.\d+)?v", model.lower()) or "vl" in model.lower())
+        ]
+        vlm_models = [
+            model
+            for model in all_models
+            if re.search(r"\d+(\.\d+)?v", model.lower()) or "vl" in model.lower()
+        ]
+        _MODELS_CACHE.set("llm_models", llm_models)
+        _MODELS_CACHE.set("vlm_models", vlm_models)

bizyengine/version.txt

@@ -1 +1 @@
-1.2.56
+1.2.57

{bizyengine-1.2.56.dist-info → bizyengine-1.2.57.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: bizyengine
-Version: 1.2.56
+Version: 1.2.57
 Summary: [a/BizyAir](https://github.com/siliconflow/BizyAir) Comfy Nodes that can run in any environment.
 Author-email: SiliconFlow <yaochi@siliconflow.cn>
 Project-URL: Repository, https://github.com/siliconflow/BizyAir

{bizyengine-1.2.56.dist-info → bizyengine-1.2.57.dist-info}/RECORD

@@ -1,5 +1,5 @@
 bizyengine/__init__.py,sha256=GP9V-JM07fz7uv_qTB43QEA2rKdrVJxi5I7LRnn_3ZQ,914
-bizyengine/version.txt,sha256=hVI0lnTLmTtdalpHou3oV96AsS0SmN76nA7QLjtnbz0,7
+bizyengine/version.txt,sha256=txvxJYLvzIKQ52dHqmS8wu-ETqexjhrTKOe0yuriLUE,7
 bizyengine/bizy_server/__init__.py,sha256=SP9oSblnPo4KQyh7yOGD26YCskFAcQHAZy04nQBNRIw,200
 bizyengine/bizy_server/api_client.py,sha256=Z7G5IjaEqSJkF6nLLw2R3bpgBAOi5ClQiUbel6NMXmE,43932
 bizyengine/bizy_server/errno.py,sha256=RIyvegX3lzpx_1L1q2XVvu3on0kvYgKiUQ8U3ZtyF68,16823
@@ -94,8 +94,8 @@ bizyengine/misc/nodes_controlnet_union_sdxl.py,sha256=fYyu_XMY7mcX1Ad9x30q1tYB8m
 bizyengine/misc/route_sam.py,sha256=-bMIR2QalfnszipGxSxvDAHGJa5gPSrjkYPb5baaRg4,1561
 bizyengine/misc/segment_anything.py,sha256=wNKYwlYPMszfwj23524geFZJjZaG4eye65SGaUnh77I,8941
 bizyengine/misc/supernode.py,sha256=STN9gaxfTSErH8OiHeZa47d8z-G9S0I7fXuJvHQOBFM,4532
-bizyengine/misc/utils.py,sha256=lYRtJxKvwCyAM8eQMvJvzxjLCwYnQGy0mjLsc5p0Dxo,6844
-bizyengine-1.2.56.dist-info/METADATA,sha256=5ra9vOgtx_1O1tprDSBt8sQD3uLPvCmVsMG95VpyX28,734
-bizyengine-1.2.56.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-bizyengine-1.2.56.dist-info/top_level.txt,sha256=2zapzqxX-we5cRyJkGf9bd5JinRtXp3-_uDI-xCAnc0,11
-bizyengine-1.2.56.dist-info/RECORD,,
+bizyengine/misc/utils.py,sha256=hLGARJ2fkckGBm6_noC38hGT4b6OXUA4dx68eCdTbzc,12913
+bizyengine-1.2.57.dist-info/METADATA,sha256=lgy5VTeQLkdhZzQc8xxX7-uJtTicw0XqR9gZQuYTvGM,734
+bizyengine-1.2.57.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+bizyengine-1.2.57.dist-info/top_level.txt,sha256=2zapzqxX-we5cRyJkGf9bd5JinRtXp3-_uDI-xCAnc0,11
+bizyengine-1.2.57.dist-info/RECORD,,