mirror of
https://github.com/open-webui/open-webui.git
synced 2026-09-02 12:14:48 +02:00
b8f279b8fb
The Redis-backed model registry skips its write when the content signature matches what is already stored. That skip has never worked across processes. Two of the values it hashes come out of Python sets, and set iteration order varies with each process's hash seed, so every worker computed a different signature for identical content and every worker rewrote the whole registry on every refresh. Sorting both makes the signature depend on content alone. Measured on a 120 model registry, 522 KiB serialized: a refresh whose content already matches drops from GET, HKEYS, HSET and SET at 5.1 ms to a single GET at 2.2 ms per worker, and the 522 KiB write leaves the wire entirely. Verified across 12 child processes with 12 distinct hash seeds: 12 different signatures before, 1 after. Filter execution order is unaffected, because the filter pipeline re-sorts by priority and id before running.
536 lines
21 KiB
Python
536 lines
21 KiB
Python
import asyncio
|
||
import copy
|
||
import logging
|
||
import sys
|
||
|
||
from aiocache import cached
|
||
from fastapi import Request
|
||
from open_webui.config import (
|
||
BYPASS_ADMIN_ACCESS_CONTROL,
|
||
DEFAULT_ARENA_MODEL,
|
||
)
|
||
from open_webui.env import BYPASS_MODEL_ACCESS_CONTROL, ENABLE_PLUGINS, GLOBAL_LOG_LEVEL
|
||
from open_webui.functions import get_function_models
|
||
from open_webui.models.access_grants import AccessGrants
|
||
from open_webui.models.config import Config
|
||
from open_webui.models.functions import Functions
|
||
from open_webui.models.groups import Groups
|
||
from open_webui.models.models import Models
|
||
from open_webui.utils.chat_variables import get_chat_variables_schema
|
||
from open_webui.models.users import UserModel
|
||
from open_webui.routers import ollama, openai
|
||
from open_webui.socket.utils import RedisDict
|
||
from open_webui.utils.access_control import has_access, has_base_model_access
|
||
from open_webui.utils.plugin import (
|
||
get_functions_cache,
|
||
get_function_module_from_cache,
|
||
)
|
||
|
||
logging.basicConfig(stream=sys.stdout, level=GLOBAL_LOG_LEVEL)
|
||
log = logging.getLogger(__name__)
|
||
|
||
|
||
async def fetch_ollama_models(request: Request, user: UserModel = None):
|
||
raw_ollama_models = await ollama.get_all_models(request, user=user)
|
||
return [
|
||
{
|
||
'id': model['model'],
|
||
'name': model['name'],
|
||
'object': 'model',
|
||
'created': 0,
|
||
'owned_by': 'ollama',
|
||
'ollama': model,
|
||
'loaded': 'expires_at' in model,
|
||
'connection_type': model.get('connection_type', 'local'),
|
||
'tags': model.get('tags', []),
|
||
}
|
||
for model in raw_ollama_models['models']
|
||
]
|
||
|
||
|
||
async def fetch_openai_models(request: Request, user: UserModel = None):
|
||
openai_response = await openai.get_all_models(request, user=user)
|
||
return openai_response['data']
|
||
|
||
|
||
async def get_all_base_models(request: Request, user: UserModel = None):
|
||
config = await Config.get_many('openai.enable', 'ollama.enable')
|
||
openai_task = fetch_openai_models(request, user) if config.get('openai.enable') else asyncio.sleep(0, result=[])
|
||
ollama_task = fetch_ollama_models(request, user) if config.get('ollama.enable') else asyncio.sleep(0, result=[])
|
||
function_task = get_function_models(request)
|
||
|
||
openai_models, ollama_models, function_models = await asyncio.gather(openai_task, ollama_task, function_task)
|
||
|
||
return function_models + openai_models + ollama_models
|
||
|
||
|
||
async def get_all_models(request, refresh: bool = False, user: UserModel = None):
|
||
config = await Config.get_many(
|
||
'models.base_models_cache',
|
||
'evaluation.arena.enable',
|
||
'evaluation.arena.models',
|
||
'models.default_metadata',
|
||
)
|
||
if refresh:
|
||
await openai.get_all_models.cache.clear()
|
||
await ollama.get_all_models.cache.clear()
|
||
|
||
if (
|
||
request.app.state.MODELS
|
||
and request.app.state.BASE_MODELS
|
||
and (config.get('models.base_models_cache') and not refresh)
|
||
):
|
||
base_models = request.app.state.BASE_MODELS
|
||
else:
|
||
base_models = await get_all_base_models(request, user=user)
|
||
if base_models:
|
||
request.app.state.BASE_MODELS = base_models
|
||
else:
|
||
base_models = request.app.state.BASE_MODELS
|
||
|
||
# deep copy the base models to avoid modifying the original list
|
||
models = [model.copy() for model in base_models]
|
||
|
||
# If there are no models, return an empty list
|
||
if len(models) == 0:
|
||
return []
|
||
|
||
# Add arena models
|
||
if config.get('evaluation.arena.enable'):
|
||
arena_models = []
|
||
arena_config = config.get('evaluation.arena.models') or []
|
||
if len(arena_config) > 0:
|
||
arena_models = [
|
||
{
|
||
'id': model['id'],
|
||
'name': model['name'],
|
||
'info': {
|
||
'meta': model['meta'],
|
||
},
|
||
'object': 'model',
|
||
'created': 0,
|
||
'owned_by': 'arena',
|
||
'arena': True,
|
||
}
|
||
for model in arena_config
|
||
]
|
||
else:
|
||
# Add default arena model
|
||
arena_models = [
|
||
{
|
||
'id': DEFAULT_ARENA_MODEL['id'],
|
||
'name': DEFAULT_ARENA_MODEL['name'],
|
||
'info': {
|
||
'meta': DEFAULT_ARENA_MODEL['meta'],
|
||
},
|
||
'object': 'model',
|
||
'created': 0,
|
||
'owned_by': 'arena',
|
||
'arena': True,
|
||
}
|
||
]
|
||
models = models + arena_models
|
||
|
||
# One query per type: the global sets are subsets of the active sets, so
|
||
# deriving them from the same rows halves the function-table queries.
|
||
if ENABLE_PLUGINS:
|
||
active_actions = await Functions.get_active_function_ids_by_type('action')
|
||
global_action_ids = {function_id for function_id, is_global in active_actions if is_global}
|
||
enabled_action_ids = {function_id for function_id, _ in active_actions}
|
||
|
||
active_filters = await Functions.get_active_function_ids_by_type('filter')
|
||
global_filter_ids = {function_id for function_id, is_global in active_filters if is_global}
|
||
enabled_filter_ids = {function_id for function_id, _ in active_filters}
|
||
else:
|
||
global_action_ids = set()
|
||
enabled_action_ids = set()
|
||
global_filter_ids = set()
|
||
enabled_filter_ids = set()
|
||
|
||
custom_models = await Models.get_all_models()
|
||
|
||
# Single O(1) lookup: Ollama base names first, then exact IDs (exact wins).
|
||
base_model_lookup = {}
|
||
for model in models:
|
||
if model.get('owned_by') == 'ollama':
|
||
base_model_lookup.setdefault(model['id'].split(':')[0], model)
|
||
base_model_lookup[model['id']] = model
|
||
|
||
existing_ids = {m['id'] for m in models}
|
||
|
||
for custom_model in custom_models:
|
||
if custom_model.base_model_id is None:
|
||
# Override applied directly to a base model (shares the same ID)
|
||
model = base_model_lookup.get(custom_model.id)
|
||
|
||
if model:
|
||
if custom_model.is_active:
|
||
model['name'] = custom_model.name
|
||
model['info'] = custom_model.model_dump()
|
||
schema = get_chat_variables_schema(custom_model.params.model_dump().get('system'))
|
||
if schema:
|
||
model['info'].setdefault('meta', {})['chat_variables_schema'] = schema
|
||
|
||
action_ids = []
|
||
filter_ids = []
|
||
|
||
if 'info' in model:
|
||
if 'meta' in model['info']:
|
||
if ENABLE_PLUGINS:
|
||
action_ids.extend(model['info']['meta'].get('actionIds', []))
|
||
filter_ids.extend(model['info']['meta'].get('filterIds', []))
|
||
|
||
if 'params' in model['info']:
|
||
del model['info']['params']
|
||
|
||
model['action_ids'] = action_ids
|
||
model['filter_ids'] = filter_ids
|
||
else:
|
||
models = [m for m in models if m is not model]
|
||
|
||
elif custom_model.is_active:
|
||
if custom_model.id in existing_ids:
|
||
continue
|
||
|
||
owned_by = 'openai'
|
||
connection_type = None
|
||
pipe = None
|
||
|
||
base_model = base_model_lookup.get(custom_model.base_model_id)
|
||
if base_model is None:
|
||
base_model = base_model_lookup.get(custom_model.base_model_id.split(':')[0])
|
||
if base_model:
|
||
owned_by = base_model.get('owned_by', 'unknown')
|
||
if 'pipe' in base_model:
|
||
pipe = base_model['pipe']
|
||
connection_type = base_model.get('connection_type', None)
|
||
|
||
model = {
|
||
'id': f'{custom_model.id}',
|
||
'name': custom_model.name,
|
||
'object': 'model',
|
||
'created': custom_model.created_at,
|
||
'owned_by': owned_by,
|
||
'connection_type': connection_type,
|
||
'preset': True,
|
||
**({'pipe': pipe} if pipe is not None else {}),
|
||
**({'provider': base_model.get('provider')} if base_model and base_model.get('provider') else {}),
|
||
**({'loaded': base_model.get('loaded')} if base_model and base_model.get('loaded') is not None else {}),
|
||
}
|
||
|
||
info = custom_model.model_dump()
|
||
schema = get_chat_variables_schema(custom_model.params.model_dump().get('system'))
|
||
if schema:
|
||
info.setdefault('meta', {})['chat_variables_schema'] = schema
|
||
if 'params' in info:
|
||
# Remove params to avoid exposing sensitive info
|
||
del info['params']
|
||
|
||
model['info'] = info
|
||
|
||
action_ids = []
|
||
filter_ids = []
|
||
|
||
if custom_model.meta:
|
||
meta = custom_model.meta.model_dump()
|
||
|
||
if ENABLE_PLUGINS and 'actionIds' in meta:
|
||
action_ids.extend(meta['actionIds'])
|
||
|
||
if ENABLE_PLUGINS and 'filterIds' in meta:
|
||
filter_ids.extend(meta['filterIds'])
|
||
|
||
model['action_ids'] = action_ids
|
||
model['filter_ids'] = filter_ids
|
||
|
||
models.append(model)
|
||
|
||
# Process action_ids to get the actions
|
||
def get_action_items_from_module(function, module):
|
||
actions = []
|
||
if hasattr(module, 'actions'):
|
||
actions = module.actions
|
||
return [
|
||
{
|
||
'id': f'{function.id}.{action["id"]}',
|
||
'name': action.get('name', f'{function.name} ({action["id"]})'),
|
||
'description': function.meta.description,
|
||
'icon': action.get(
|
||
'icon_url',
|
||
function.meta.manifest.get('icon_url', None)
|
||
or getattr(module, 'icon_url', None)
|
||
or getattr(module, 'icon', None),
|
||
),
|
||
}
|
||
for action in actions
|
||
]
|
||
else:
|
||
return [
|
||
{
|
||
'id': function.id,
|
||
'name': function.name,
|
||
'description': function.meta.description,
|
||
'icon': function.meta.manifest.get('icon_url', None)
|
||
or getattr(module, 'icon_url', None)
|
||
or getattr(module, 'icon', None),
|
||
}
|
||
]
|
||
|
||
# Process filter_ids to get the filters
|
||
def get_filter_items_from_module(function, module):
|
||
return [
|
||
{
|
||
'id': function.id,
|
||
'name': function.name,
|
||
'description': function.meta.description,
|
||
'icon': function.meta.manifest.get('icon_url', None)
|
||
or getattr(module, 'icon_url', None)
|
||
or getattr(module, 'icon', None),
|
||
'has_user_valves': hasattr(module, 'UserValves'),
|
||
}
|
||
]
|
||
|
||
# Batch-prefetch all needed function records to avoid N+1 queries
|
||
all_function_ids = set()
|
||
for model in models:
|
||
all_function_ids.update(model.get('action_ids', []))
|
||
all_function_ids.update(model.get('filter_ids', []))
|
||
all_function_ids.update(global_action_ids)
|
||
all_function_ids.update(global_filter_ids)
|
||
|
||
functions_by_id = {f.id: f for f in await Functions.get_functions_by_ids(list(all_function_ids))}
|
||
|
||
# Pre-warm the function module cache once per unique function ID.
|
||
# This ensures each function's DB freshness check runs exactly once,
|
||
# not once per (model × function) pair.
|
||
# Only attempt to load functions that actually exist in the local DB;
|
||
# imported/custom model configs may reference tools or filters the user
|
||
# hasn't installed, and trying to load those would cause persistent
|
||
# "Failed to load function module" log spam on every model refresh.
|
||
for function_id, function in functions_by_id.items():
|
||
try:
|
||
await get_function_module_from_cache(request, function_id, function=function)
|
||
except Exception as e:
|
||
log.debug('Failed to load function module for %s: %s', function_id, e)
|
||
|
||
# Apply global model defaults to all models
|
||
# Per-model overrides take precedence over global defaults
|
||
default_metadata = config.get('models.default_metadata') or {}
|
||
|
||
if default_metadata:
|
||
for model in models:
|
||
info = model.get('info')
|
||
|
||
if info is None:
|
||
model['info'] = {'meta': copy.deepcopy(default_metadata)}
|
||
continue
|
||
|
||
meta = info.setdefault('meta', {})
|
||
for key, value in default_metadata.items():
|
||
if key == 'capabilities':
|
||
# Merge capabilities: defaults as base, per-model overrides win
|
||
existing = meta.get('capabilities') or {}
|
||
meta['capabilities'] = {**value, **existing}
|
||
elif meta.get(key) is None:
|
||
meta[key] = copy.deepcopy(value)
|
||
|
||
# Batch-fetch all function valves in one query to avoid N+1 DB hits
|
||
# inside get_action_priority (previously called per action × per model).
|
||
all_function_valves = await Functions.get_function_valves_by_ids(list(all_function_ids))
|
||
functions_cache = get_functions_cache(request)
|
||
|
||
# Global actions and filters appear in every model, so priorities and item
|
||
# lists are memoized across the loop instead of rebuilt per model.
|
||
action_priorities = {}
|
||
|
||
def get_action_priority(action_id):
|
||
if action_id in action_priorities:
|
||
return action_priorities[action_id]
|
||
priority = 0
|
||
try:
|
||
function_module = functions_cache.get(action_id)
|
||
if function_module and hasattr(function_module, 'Valves'):
|
||
valves_db = all_function_valves.get(action_id)
|
||
valves = function_module.Valves(**(valves_db if valves_db else {}))
|
||
priority = getattr(valves, 'priority', 0)
|
||
except Exception:
|
||
priority = 0
|
||
action_priorities[action_id] = priority
|
||
return priority
|
||
|
||
action_items_by_id = {}
|
||
filter_items_by_id = {}
|
||
|
||
for model in models:
|
||
action_ids = [
|
||
action_id
|
||
for action_id in set(model.pop('action_ids', [])) | global_action_ids
|
||
if action_id in enabled_action_ids
|
||
]
|
||
action_ids.sort(key=lambda aid: (get_action_priority(aid), aid))
|
||
|
||
filter_ids = [
|
||
filter_id
|
||
for filter_id in set(model.pop('filter_ids', [])) | global_filter_ids
|
||
if filter_id in enabled_filter_ids
|
||
]
|
||
# Set order varies per process, and an unstable order defeats the RedisDict content signature.
|
||
filter_ids.sort()
|
||
|
||
model['actions'] = []
|
||
for action_id in action_ids:
|
||
items = action_items_by_id.get(action_id)
|
||
if items is None:
|
||
action_function = functions_by_id.get(action_id)
|
||
if action_function is None:
|
||
log.info('Action not found: %s', action_id)
|
||
action_items_by_id[action_id] = []
|
||
continue
|
||
|
||
function_module = functions_cache.get(action_id)
|
||
if function_module is None:
|
||
log.info('Failed to load action module: %s', action_id)
|
||
action_items_by_id[action_id] = []
|
||
continue
|
||
items = get_action_items_from_module(action_function, function_module)
|
||
action_items_by_id[action_id] = items
|
||
# Shallow copies keep per-model item dicts independent, as before
|
||
model['actions'].extend({**item} for item in items)
|
||
|
||
model['filters'] = []
|
||
for filter_id in filter_ids:
|
||
items = filter_items_by_id.get(filter_id)
|
||
if items is None:
|
||
filter_function = functions_by_id.get(filter_id)
|
||
if filter_function is None:
|
||
log.info('Filter not found: %s', filter_id)
|
||
filter_items_by_id[filter_id] = []
|
||
continue
|
||
|
||
function_module = functions_cache.get(filter_id)
|
||
if function_module is None:
|
||
log.info('Failed to load filter module: %s', filter_id)
|
||
filter_items_by_id[filter_id] = []
|
||
continue
|
||
if getattr(function_module, 'toggle', None):
|
||
items = get_filter_items_from_module(filter_function, function_module)
|
||
else:
|
||
items = []
|
||
filter_items_by_id[filter_id] = items
|
||
model['filters'].extend({**item} for item in items)
|
||
|
||
log.debug('get_all_models() returned %s models', len(models))
|
||
|
||
models_dict = {model['id']: model for model in models}
|
||
if isinstance(request.app.state.MODELS, RedisDict):
|
||
try:
|
||
request.app.state.MODELS.set(models_dict)
|
||
except Exception as e:
|
||
log.warning(f'Failed to update Redis model cache, using in-process cache: {e}')
|
||
request.app.state.MODELS = models_dict
|
||
else:
|
||
request.app.state.MODELS = models_dict
|
||
|
||
return models
|
||
|
||
|
||
async def check_model_access(user, model, model_info=None, db=None):
|
||
if model.get('arena'):
|
||
meta = model.get('info', {}).get('meta', {})
|
||
access_grants = meta.get('access_grants', [])
|
||
if not await has_access(
|
||
user.id,
|
||
permission='read',
|
||
access_grants=access_grants,
|
||
db=db,
|
||
):
|
||
raise Exception('Model not found')
|
||
else:
|
||
# Callers that already fetched the row (chat completion entry) pass it in
|
||
if model_info is None or model_info.id != model.get('id'):
|
||
model_info = await Models.get_model_by_id(model.get('id'), db=db)
|
||
if not model_info:
|
||
raise Exception('Model not found')
|
||
|
||
# One group-membership fetch shared by the direct check and every
|
||
# base-model hop; skipped when no check below needs it.
|
||
user_group_ids = None
|
||
if user.id != model_info.user_id or model_info.base_model_id:
|
||
user_group_ids = {group.id for group in await Groups.get_groups_by_member_id(user.id, db=db)}
|
||
|
||
if not (
|
||
user.id == model_info.user_id
|
||
or await AccessGrants.has_access(
|
||
user_id=user.id,
|
||
resource_type='model',
|
||
resource_id=model_info.id,
|
||
permission='read',
|
||
user_group_ids=user_group_ids,
|
||
db=db,
|
||
)
|
||
):
|
||
raise Exception('Model not found')
|
||
|
||
# Enforce access on chained base models
|
||
if not await has_base_model_access(
|
||
user.id, model_info, user_role=user.role, user_group_ids=user_group_ids, db=db
|
||
):
|
||
raise Exception('Model not found')
|
||
|
||
|
||
async def get_filtered_models(models, user, db=None):
|
||
# Filter out models that the user does not have access to
|
||
if (
|
||
user.role == 'user' or (user.role == 'admin' and not BYPASS_ADMIN_ACCESS_CONTROL)
|
||
) and not BYPASS_MODEL_ACCESS_CONTROL:
|
||
model_infos = {}
|
||
for model in models:
|
||
if model.get('arena'):
|
||
continue
|
||
info = model.get('info')
|
||
if info:
|
||
model_infos[model['id']] = info
|
||
|
||
user_group_ids = {group.id for group in await Groups.get_groups_by_member_id(user.id, db=db)}
|
||
|
||
# Batch-fetch accessible resource IDs in a single query instead of N has_access calls
|
||
accessible_model_ids = await AccessGrants.get_accessible_resource_ids(
|
||
user_id=user.id,
|
||
resource_type='model',
|
||
resource_ids=list(model_infos.keys()),
|
||
permission='read',
|
||
user_group_ids=user_group_ids,
|
||
db=db,
|
||
)
|
||
|
||
filtered_models = []
|
||
for model in models:
|
||
if model.get('arena'):
|
||
meta = model.get('info', {}).get('meta', {})
|
||
access_grants = meta.get('access_grants', [])
|
||
if await has_access(
|
||
user.id,
|
||
permission='read',
|
||
access_grants=access_grants,
|
||
user_group_ids=user_group_ids,
|
||
):
|
||
filtered_models.append(model)
|
||
continue
|
||
|
||
model_info = model_infos.get(model['id'])
|
||
if model_info:
|
||
if (
|
||
(user.role == 'admin' and BYPASS_ADMIN_ACCESS_CONTROL)
|
||
or user.id == model_info.get('user_id')
|
||
or model['id'] in accessible_model_ids
|
||
):
|
||
filtered_models.append(model)
|
||
elif user.role == 'admin':
|
||
# No DB entry means no access control configured yet;
|
||
# only admins can see unconfigured models.
|
||
filtered_models.append(model)
|
||
|
||
return filtered_models
|
||
else:
|
||
return models
|