import gc import os import time from typing import Dict, List, Callable, Union from copy import deepcopy from collections import OrderedDict import re import importlib from utils.logger import logger as LOGGER from utils import shared GPUINTENSIVE_SET = {'cuda', 'mps', 'xpu', 'privateuseone'} def register_hooks(hooks_registered: OrderedDict, callbacks: Union[List, Callable, Dict]): if callbacks is None: return if isinstance(callbacks, (Dict, OrderedDict)): for k, v in callbacks.items(): hooks_registered[k] = v else: nhooks = len(hooks_registered) if isinstance(callbacks, Callable): callbacks = [callbacks] for callback in callbacks: hk = 'hook_' + str(nhooks).zfill(2) while True: if hk not in hooks_registered: break hk = hk + '_' + str(time.time_ns()) hooks_registered[hk] = callback nhooks += 1 def patch_module_params(cfg_param, module_params, module_name: str = ''): # cfg_param = config_params[module_key] cfg_key_set = set(cfg_param.keys()) module_key_set = set(module_params.keys()) for ck in cfg_key_set: if ck not in module_key_set: LOGGER.warning(f'Found invalid {module_name} config: {ck}') cfg_param.pop(ck) for mk in module_key_set: if mk not in cfg_key_set: if not mk.startswith('__') and mk != 'description': LOGGER.info(f'Found new {module_name} config: {mk}') cfg_param[mk] = module_params[mk] else: mparam = module_params[mk] cparam = cfg_param[mk] if isinstance(mparam, dict): tgt_type = mparam.get('data_type', type(mparam['value'])) if isinstance(cparam, dict): if 'value' in cparam: v = cparam['value'] elif isinstance(mparam['value'], dict): for k in mparam['value']: if k in cparam: mparam['value'][k] = cparam[k] v = mparam['value'] else: v = mparam['value'] else: v = cparam valid = True if tgt_type != type(v): try: v = tgt_type(v) except: valid = False LOGGER.warning(f'Invalid param value {v} for defined dtype: {tgt_type}, it will be set to default value: {mparam}') if valid: mparam['value'] = v cfg_param[mk] = mparam else: if type(cparam) != type(mparam): if not isinstance(mparam, dict) and isinstance(cparam, dict): cparam = cparam['value'] try: cfg_param[mk] = type(mparam)(cparam) except ValueError: LOGGER.warning(f'Invalid param value {cparam} for defined dtype: {type(mparam)}, it will be set to default value: {mparam}') cfg_param[mk] = mparam cfg_key_list = list(cfg_param.keys()) module_key_list = list(module_params.keys()) if cfg_key_list != module_key_list: new_params = {key: cfg_param[key] for key in module_key_list} cfg_param.clear() cfg_param.update(new_params) module_key_set = set(module_params.keys()) cfg_param['__param_patched'] = True return cfg_param def merge_config_module_params(config_params: Dict, module_keys: List, get_module: Callable) -> Dict: for module_key in module_keys: module_params = get_module(module_key).params if module_key not in config_params or config_params[module_key] is None: config_params[module_key] = module_params else: patch_module_params(config_params[module_key], module_params, module_key) return config_params def standardize_module_params(params): if params is None: return for k, v in params.items(): if not isinstance(v, dict): v = {'value': v} if 'data_type' not in v: v['data_type'] = type(v['value']) params[k] = v class BaseModule: params: Dict = None logger = LOGGER _preprocess_hooks: OrderedDict = None _postprocess_hooks: OrderedDict = None download_file_list: List = None download_file_on_load = False _load_model_keys: set = None def __init__(self, **params) -> None: standardize_module_params(self.params) if self.params is not None and '__param_patched' not in params: params = patch_module_params(params, self.params, self) if params: if self.params is None: self.params = params else: self.params.update(params) @classmethod def register_postprocess_hooks(cls, callbacks: Union[List, Callable]): """ these hooks would be shared among all objects inherited from the same super class """ assert cls._postprocess_hooks is not None register_hooks(cls._postprocess_hooks, callbacks) @classmethod def register_preprocess_hooks(cls, callbacks: Union[List, Callable, Dict]): """ these hooks would be shared among all objects inherited from the same super class """ assert cls._preprocess_hooks is not None register_hooks(cls._preprocess_hooks, callbacks) def get_param_value(self, param_key: str): assert self.params is not None and param_key in self.params p = self.params[param_key] if isinstance(p, dict): return p['value'] return p def set_param_value(self, param_key: str, param_value, convert_dtype=True): assert self.params is not None and param_key in self.params p = self.params[param_key] if isinstance(p, dict): if convert_dtype: try: val_type = p.get('data_type', type(p['value'])) param_value = val_type(param_value) except ValueError: dtype = type(p['value']) self.logger.warning(f'Invalid param value {param_value} for defined dtype: {dtype}') p['value'] = param_value else: if convert_dtype: try: param_value = type(p)(param_value) except ValueError: self.logger.warning(f'Invalid param value {param_value} for defined dtype: {type(p)}, revert to original value {p}') param_value = p self.params[param_key] = param_value def updateParam(self, param_key: str, param_content): self.set_param_value(param_key, param_content) @property def low_vram_mode(self): if 'low vram mode' in self.params: return self.get_param_value('low vram mode') return False def is_cpu_intensive(self)->bool: if self.params is not None and 'device' in self.params: return self.params['device']['value'] == 'cpu' return False def is_gpu_intensive(self) -> bool: if self.params is not None and 'device' in self.params: return self.params['device']['value'] in GPUINTENSIVE_SET return False def is_computational_intensive(self) -> bool: if self.params is not None and 'device' in self.params: return True return False def unload_model(self, empty_cache=False): model_deleted = False if self._load_model_keys is not None: for k in self._load_model_keys: if hasattr(self, k): model = getattr(self, k) if model is not None: if hasattr(model, 'unload_model'): model.unload_model(empty_cache=False) del model setattr(self, k, None) model_deleted = True if empty_cache and model_deleted: soft_empty_cache() return model_deleted def load_model(self): # TODO: check and download files self._load_model() return def _load_model(self): return def all_model_loaded(self): if self._load_model_keys is None: return True for k in self._load_model_keys: if not hasattr(self, k) or getattr(self, k) is None: return False return True def __del__(self): self.unload_model() @property def debug_mode(self): return shared.DEBUG def flush(self, param_key: str): return None os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' import torch DEFAULT_DEVICE = 'cpu' AVAILABLE_DEVICES = ['cpu'] if hasattr(torch, 'cuda') and torch.cuda.is_available(): DEFAULT_DEVICE = 'cuda' AVAILABLE_DEVICES.append(DEFAULT_DEVICE) if hasattr(torch, 'xpu') and torch.xpu.is_available(): DEFAULT_DEVICE = 'xpu' if torch.xpu.is_available() else 'cpu' AVAILABLE_DEVICES.append(DEFAULT_DEVICE) if hasattr(torch, 'backends') and hasattr(torch.backends, 'mps') and torch.backends.mps.is_available(): DEFAULT_DEVICE = 'mps' AVAILABLE_DEVICES.append(DEFAULT_DEVICE) try: import torch_directml if hasattr(torch, 'privateuseone') and torch_directml.device_count() > 0: torch.dml = torch_directml DEFAULT_DEVICE = f'privateuseone:{torch.dml.default_device()}' AVAILABLE_DEVICES += [f"privateuseone:{d}" for d in range(torch.dml.device_count())] except: # directml is not supported pass BF16_SUPPORTED = DEFAULT_DEVICE == 'cuda' and torch.cuda.is_bf16_supported() or DEFAULT_DEVICE == 'xpu' and torch.xpu.is_bf16_supported() def is_nvidia(): if DEFAULT_DEVICE == 'cuda': if torch.version.cuda: return True return False def is_intel(): if DEFAULT_DEVICE == 'xpu': if torch.version.xpu: return True return False def soft_empty_cache(): gc.collect() if DEFAULT_DEVICE == 'cuda': torch.cuda.empty_cache() torch.cuda.ipc_collect() elif DEFAULT_DEVICE == 'xpu': torch.xpu.empty_cache() # torch.xpu.ipc_collect() elif DEFAULT_DEVICE == 'mps': torch.mps.empty_cache() def DEVICE_SELECTOR(not_supported:list[str]=[]): return deepcopy( { 'type': 'selector', 'options': [opt for opt in AVAILABLE_DEVICES if all(device not in opt for device in not_supported)], 'value': DEFAULT_DEVICE if not any(DEFAULT_DEVICE in device for device in not_supported) else 'cpu' } ) TORCH_DTYPE_MAP = { 'fp32': torch.float32, 'fp16': torch.float16, 'bf16': torch.bfloat16, } MODULE_SCRIPTS = { 'translator': {'module_dir': 'modules/translators', 'module_pattern': r'trans_(.*?).py'}, 'textdetector': {'module_dir': 'modules/textdetector', 'module_pattern': r'detector_(.*?).py'}, 'inpainter': {'module_dir': 'modules/inpaint', 'module_pattern': r'inpaint_(.*?).py'}, 'ocr': {'module_dir': 'modules/ocr', 'module_pattern': r'ocr_(.*?).py'}, } def init_module_registries(target_modules=None): def _load_module(module_dir: str, module_pattern: str): modules = os.listdir(module_dir) pattern = re.compile(module_pattern) module_path = module_dir.replace('/', '.') if not module_path.endswith('.'): module_path += '.' for module_name in modules: if pattern.match(module_name) is not None: try: module = module_path + module_name.replace('.py', '') importlib.import_module(module) except Exception as e: LOGGER.warning(f'Failed to import {module}: {e}') if target_modules is None: target_modules = MODULE_SCRIPTS if isinstance(target_modules, str): target_modules = [target_modules] for k in target_modules: _load_module(**MODULE_SCRIPTS[k]) def init_textdetector_registries(): init_module_registries('textdetector') def init_inpainter_registries(): init_module_registries('inpainter') def init_ocr_registries(): init_module_registries('ocr') def init_translator_registries(): init_module_registries('translator')