hz4th_coder/llm-proxy
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

v2.1.1: Auto配置改为模型粒度拖动排序 + 历史缓存复用 + 对话生图

Browse Source 
- Auto配置编辑: 粒度从提供商改为具体模型, 原生HTML5拖动排序, 修复已保存模型不显示勾选(字符串/对象混拼bug)
- 历史缓存: 带历史上下文的auto请求按首条消息识别会话, 优先复用上次模型命中上游前缀缓存
  (可配置: prefer_cache_model + cache_ttl_seconds, 系统配置页开关 + /api/admin/routing)
- 对话生图: chat页选 auto-image/Qwen-Image 直接调生图接口并在对话内展示图片
- 修复: Flask模板缓存导致改模板不生效 -> TEMPLATES_AUTO_RELOAD
- 修复: 自动回退/显式有序模型列表路由, 空列表回退提供商能力选择
This commit is contained in:
hz4th_coder
2026-08-28 12:16:25 +08:00
parent 178968ba1a
commit d387d5ba09
7 changed files with 538 additions and 143 deletions
Download Patch File Download Diff File Expand all files Collapse all files
app.py
+218 -15
View File
@@ -14,6 +14,7 @@ import requests
import json
import time
import random
import hashlib
import logging
from datetime import datetime, date
from pathlib import Path
@@ -25,17 +26,18 @@ import uuid
sys.path.insert(0, str(Path(__file__).parent))
from config.settings import (
get_providers, get_model_aliases, get_auto_profiles, get_auto_profile,
SERVER_CONFIG, LOG_CONFIG, RETRY_CONFIG, CAPABILITY_DEFS,
SERVER_CONFIG, LOG_CONFIG, RETRY_CONFIG, CAPABILITY_DEFS, ROUTING_CONFIG,
load_config, save_config, get_provider, add_provider, update_provider,
delete_provider, update_priority, update_model_alias, delete_model_alias,
add_auto_profile, update_auto_profile, delete_auto_profile,
add_auto_profile, update_auto_profile, delete_auto_profile, load_routing_config, save_routing_config,
DEFAULT_PROVIDERS, DEFAULT_MODEL_ALIASES, DEFAULT_AUTO_PROFILES
)
app = Flask(__name__, template_folder='templates')
app.config['TEMPLATES_AUTO_RELOAD'] = True # 模板修改即时生效（无需重启）
CORS(app)
VERSION = "2.1.0"
VERSION = "2.1.1"
# 数据目录和统计文件
DATA_DIR = Path(__file__).parent / 'data'
@@ -49,6 +51,56 @@ LOGS_DIR.mkdir(exist_ok=True)
stats_lock = threading.Lock()
chats_lock = threading.Lock()
# 历史上下文→模型缓存（用于 auto 请求优先复用上次模型，命中前缀缓存）
history_model_cache = {}
history_cache_lock = threading.Lock()
def history_fingerprint(data):
"""计算会话指纹：用首条消息作为会话锚点（同一对话各轮次首条消息不变，可跨轮次复用模型）"""
messages = data.get('messages', []) if isinstance(data, dict) else []
if not messages:
return None
try:
s = json.dumps(messages[0], ensure_ascii=False, sort_keys=True)
except Exception:
return None
return hashlib.md5(s.encode('utf-8')).hexdigest()
def has_history(data):
"""请求是否带历史上下文"""
messages = data.get('messages', []) if isinstance(data, dict) else []
return len(messages) >= 2
def get_cached_history_model(fp, capability=None):
"""获取历史前缀对应的上次模型（未过期且可用才返回）"""
if not fp:
return None
with history_cache_lock:
entry = history_model_cache.get(fp)
if not entry:
return None
routing = load_routing_config()
ttl = routing.get('cache_ttl_seconds', 3600)
if time.time() - entry.get('ts', 0) > ttl:
with history_cache_lock:
history_model_cache.pop(fp, None)
return None
provider = find_provider_for_model(entry.get('model', ''), capability)
if not provider:
return None
return {'provider': provider, 'model': entry['model']}
def remember_history_model(fp, provider_name, model):
"""记录某历史前缀使用的模型"""
if not fp:
return
with history_cache_lock:
history_model_cache[fp] = {'provider': provider_name, 'model': model, 'ts': time.time()}
# 提供商状态缓存
provider_status = {}
@@ -307,14 +359,34 @@ def find_provider_for_model(model_name, capability=None, exclude=None):
def get_auto_provider(profile_name='auto', capability=None, exclude=None):
"""获取auto模式下可用的提供商与模型（按能力过滤）"""
"""获取auto模式下可用的提供商与模型
优先级：
1. 配置了显式有序 models 列表 → 按列表顺序逐个找能托管该模型的提供商
2. 否则按提供商能力 + 优先级选择（旧逻辑回退）
"""
refresh_config()
exclude = exclude or set()
profile = _cached_auto_profiles.get(profile_name, _cached_auto_profiles.get('auto', {}))
req_cap = profile.get('capability') or capability or 'text'
allowed_providers = profile.get('providers', ['*'])
strategy = profile.get('strategy', 'priority')
ordered_models = profile.get('models') or []
# 方式1：显式有序模型列表（粒度=具体模型，用户可拖动排序）
if ordered_models:
candidates = []
for mname in ordered_models:
provider = find_provider_for_model(mname, req_cap, exclude=exclude)
if provider:
candidates.append((provider, mname))
if candidates:
if strategy == 'random':
return random.choice(candidates)
return candidates[0]
return None, None # 显式列表全不可用时，不悄悄换成别的模型
# 方式2：按提供商能力 + 优先级
candidates = []
for provider in sorted_providers():
if not provider['enabled']:
@@ -383,6 +455,14 @@ def detect_capability(data):
return 'text'
def is_image_gen_request(model):
"""判断请求是否为图片生成（auto-image 配置或具备 image_gen 能力的模型）"""
resolved = resolve_model_name(model)
if is_auto_model(resolved):
return _cached_auto_profiles.get(resolved, {}).get('capability') == 'image_gen'
return find_provider_for_model(resolved, 'image_gen') is not None
# ============ 上游转发 ============
def build_headers(provider, content_type='application/json', extra=None):
@@ -563,14 +643,32 @@ def chat_completions():
request_model = model
capability = detect_capability(data)
provider, resolved_model = get_provider_for_model(model, capability)
# 历史上下文缓存：auto 请求始终记录会话锚点（首条消息），带历史时优先复用上次模型
history_fp = None
routing_cfg = load_routing_config()
if routing_cfg.get('prefer_cache_model', True) and is_auto_model(model):
history_fp = history_fingerprint(data)
if has_history(data):
cached = get_cached_history_model(history_fp, capability)
if cached:
provider = cached['provider']
resolved_model = cached['model']
request_provider = provider['name']
logger.info(f"History cache hit: model={model} -> provider={provider['name']}, resolved_model={resolved_model}, stream={stream}, capability={capability}")
else:
provider, resolved_model = get_provider_for_model(model, capability)
else:
provider, resolved_model = get_provider_for_model(model, capability)
else:
provider, resolved_model = get_provider_for_model(model, capability)
if not provider:
increment_stats(model, 'unknown', success=False, error=f'No provider for model: {model}')
return jsonify({"error": {"message": f"No available provider for model: {model} (capability: {capability})", "type": "invalid_request_error"}}), 400
request_provider = provider['name']
logger.info(f"Request: model={model} -> provider={provider['name']}, resolved_model={resolved_model}, stream={stream}, capability={capability}")
if not history_fp or not routing_cfg.get('prefer_cache_model', True):
logger.info(f"Request: model={model} -> provider={provider['name']}, resolved_model={resolved_model}, stream={stream}, capability={capability}")
last_error = None
tried_providers = set()
@@ -585,6 +683,7 @@ def chat_completions():
if stream:
increment_stats(model, provider['name'], success=True, tokens=0)
remember_history_model(history_fp, provider['name'], resolved_model)
return Response(
stream_with_context(stream_response(response)),
content_type='text/event-stream',
@@ -595,6 +694,7 @@ def chat_completions():
usage = result.get('usage', {})
request_tokens = usage.get('total_tokens', 0)
increment_stats(model, provider['name'], success=True, tokens=request_tokens)
remember_history_model(history_fp, provider['name'], resolved_model)
return jsonify(result)
else:
@@ -1384,20 +1484,38 @@ def api_admin_delete_alias(alias):
# ============ 后台管理 API：Auto配置 ============
def _auto_profile_public(name, profile):
"""Auto配置对外结构（含能力与匹配模型预览）"""
"""Auto配置对外结构（含能力、有序模型列表与可挑选模型详情）"""
providers = get_providers()
allowed_providers = profile.get('providers', ['*'])
capability = profile.get('capability', 'text')
ordered_models = profile.get('models') or []
provider_details = []
matched_models = []
# 收集所有具备该能力、且在候选提供商范围内的模型（供编辑器挑选）
matched = []
seen = set()
for p in sorted(providers, key=lambda x: x['priority']):
if not ('*' in allowed_providers or p.get('id') in allowed_providers or p['name'] in allowed_providers):
continue
models = provider_capability_models(p, capability)
if models:
provider_details.append({'id': p.get('id'), 'name': p['name'], 'priority': p['priority'], 'selected': True, 'matched_models': models})
matched_models.extend(models)
for m in provider_models(p):
if model_has_capability(p, m['name'], capability) and m['name'] not in seen:
seen.add(m['name'])
matched.append({
'name': m['name'],
'provider_id': p.get('id'),
'provider_name': p['name'],
'provider_priority': p['priority'],
'capabilities': m['capabilities'],
})
# 已保存的有序模型排前面，其余可挑选模型按提供商优先级追加
by_name = {x['name']: x for x in matched}
sorted_matched = []
for mname in ordered_models:
if mname in by_name:
sorted_matched.append(by_name[mname])
for x in matched:
if x['name'] not in ordered_models:
sorted_matched.append(x)
return {
'name': name,
@@ -1407,8 +1525,9 @@ def _auto_profile_public(name, profile):
'capability_label': CAPABILITY_DEFS.get(capability, capability),
'strategy': profile.get('strategy', 'priority'),
'providers': allowed_providers,
'provider_details': provider_details,
'matched_models': matched_models,
'models': ordered_models, # 有序的具体模型列表（优先级顺序）
'matched_models': [x['name'] for x in matched],
'matched_models_detail': sorted_matched, # 每个模型的提供商详情（编辑器用）
}
@@ -1452,6 +1571,7 @@ def api_admin_add_auto_profile():
'name': data.get('display_name', data['name']),
'description': data.get('description', ''),
'capability': capability,
'models': data.get('models', []), # 有序的具体模型列表（优先级顺序）
'providers': data.get('providers', ['*']),
'strategy': data.get('strategy', 'priority'),
}
@@ -1478,6 +1598,8 @@ def api_admin_update_auto_profile(profile_name):
if data['capability'] not in CAPABILITY_DEFS:
return jsonify({'error': f'Invalid capability: {data["capability"]}'}), 400
profile_data['capability'] = data['capability']
if 'models' in data:
profile_data['models'] = data['models']
if 'providers' in data:
profile_data['providers'] = data['providers']
if 'strategy' in data:
@@ -1538,11 +1660,41 @@ def api_admin_config():
'model_aliases': aliases,
'auto_profiles': get_auto_profiles(),
'retry_config': RETRY_CONFIG,
'routing_config': load_routing_config(),
'capabilities': CAPABILITY_DEFS,
'server_config': {'port': SERVER_CONFIG['port']}
})
@app.route('/api/admin/routing', methods=['GET'])
def api_admin_routing_get():
"""获取路由缓存配置"""
return jsonify(load_routing_config())
@app.route('/api/admin/routing', methods=['PUT'])
def api_admin_routing_put():
"""更新路由缓存配置"""
data = request.get_json()
if not data:
return jsonify({'error': 'Invalid request body'}), 400
update = {}
if 'prefer_cache_model' in data:
update['prefer_cache_model'] = bool(data['prefer_cache_model'])
if 'cache_ttl_seconds' in data:
try:
update['cache_ttl_seconds'] = max(0, int(data['cache_ttl_seconds']))
except:
pass
if not update:
return jsonify({'error': 'No valid fields'}), 400
result = save_routing_config(update)
return jsonify({'success': True, 'routing_config': result})
# ============ 后台管理 API：对话 ============
def load_chats():
@@ -1665,6 +1817,57 @@ def api_admin_chat_send():
try:
proxy_url = f"http://localhost:{SERVER_CONFIG['port']}/v1/chat/completions"
# 图片生成模型（auto-image / Qwen-Image）→ 走生图接口
if is_image_gen_request(model):
img_resp = requests.post(f"http://localhost:{SERVER_CONFIG['port']}/v1/images/generations", json={
'model': model,
'prompt': user_message,
'n': 1,
}, timeout=180)
if img_resp.status_code == 200:
img_result = img_resp.json()
img_url = None
img_b64 = None
if img_result.get('data') and len(img_result['data']) > 0:
img_url = img_result['data'][0].get('url')
img_b64 = img_result['data'][0].get('b64_json')
used_model = img_result.get('model', model)
with chats_lock:
data = load_chats()
for c in data['chats']:
if c['id'] == chat_id:
c['messages'].append({
'role': 'assistant',
'content': '🖼️ 图片生成成功',
'image_url': img_url,
'image_b64': img_b64,
'model': used_model,
'time': datetime.now().isoformat()
})
if len(c['messages']) == 2:
c['title'] = user_message[:30] + ('...' if len(user_message) > 30 else '')
c['updated_at'] = datetime.now().isoformat()
break
save_chats(data)
return jsonify({
'success': True,
'chat_id': chat_id,
'response': '🖼️ 图片生成成功',
'model': used_model,
'image_url': img_url,
'image_b64': img_b64,
'title': chat.get('title', '新对话')
})
else:
try:
error_msg = img_resp.json().get('error', {}).get('message', 'Unknown error')
except:
error_msg = f'HTTP {img_resp.status_code}'
return jsonify({'error': error_msg}), img_resp.status_code
messages = []
for msg in chat['messages'][-20:]:
messages.append({'role': msg['role'], 'content': msg['content']})