# -*- coding: utf-8 -*- # Copyright (c) Alibaba, Inc. and its affiliates. import copy import os.path import random from collections import OrderedDict import torch import torch.nn.functional as F from PIL.Image import Image from scepter.modules.model.network.diffusion.diffusion import GaussianDiffusion from scepter.modules.model.network.diffusion.schedules import noise_schedule from scepter.modules.model.registry import (BACKBONES, EMBEDDERS, MODELS, TOKENIZERS, DIFFUSIONS) from scepter.modules.utils.distribute import we from scepter.modules.utils.file_system import FS from scepter.modules.utils.config import Config from scepter.studio.utils.env import get_available_memory from .control_inference import ControlInference from .tuner_inference import TunerInference def get_model(model_tuple): assert 'model' in model_tuple return model_tuple['model'] class DiffusionInference(): ''' define vae, unet, text-encoder, tuner, refiner components support to load the components dynamicly. create and load model when run this model at the first time. ''' def __init__(self, logger=None): self.logger = logger self.is_redefine_paras = True self.loaded_model = {} self.loaded_model_name = [ 'diffusion_model', 'first_stage_model', 'cond_stage_model' ] self.diffusion_insclass = GaussianDiffusion self.tuner_infer = TunerInference(self.logger) self.control_infer = ControlInference(self.logger) def init_from_cfg(self, cfg): self.name = cfg.NAME self.is_default = cfg.get('IS_DEFAULT', False) module_paras = self.load_default(cfg.get('DEFAULT_PARAS', None)) assert cfg.have('MODEL') if self.is_redefine_paras: cfg.MODEL = self.redefine_paras(cfg.MODEL) if 'DIFFUSION' in cfg.MODEL: self.diffusion = DIFFUSIONS.build(cfg.MODEL.DIFFUSION, logger=self.logger) else: self.diffusion = self.load_schedule(cfg.MODEL.SCHEDULE) self.diffusion_model = self.infer_model( cfg.MODEL.DIFFUSION_MODEL, module_paras.get( 'DIFFUSION_MODEL', None)) if cfg.MODEL.have('DIFFUSION_MODEL') else None self.first_stage_model = self.infer_model( cfg.MODEL.FIRST_STAGE_MODEL, module_paras.get( 'FIRST_STAGE_MODEL', None)) if cfg.MODEL.have('FIRST_STAGE_MODEL') else None self.cond_stage_model = self.infer_model( cfg.MODEL.COND_STAGE_MODEL, module_paras.get( 'COND_STAGE_MODEL', None)) if cfg.MODEL.have('COND_STAGE_MODEL') else None self.refiner_cond_model = self.infer_model( cfg.MODEL.REFINER_COND_MODEL, module_paras.get( 'REFINER_COND_MODEL', None)) if cfg.MODEL.have('REFINER_COND_MODEL') else None self.refiner_diffusion_model = self.infer_model( cfg.MODEL.REFINER_MODEL, module_paras.get( 'REFINER_MODEL', None)) if cfg.MODEL.have('REFINER_MODEL') else None self.tokenizer = TOKENIZERS.build( cfg.MODEL.TOKENIZER, logger=self.logger) if cfg.MODEL.have('TOKENIZER') else None if self.tokenizer is not None: self.cond_stage_model['cfg'].KWARGS = { 'vocab_size': self.tokenizer.vocab_size } def redefine_paras(self, cfg): if cfg.get('PRETRAINED_MODEL', None): with FS.get_from(cfg.PRETRAINED_MODEL, wait_finish=True) as local_path: if local_path.endswith('safetensors'): from safetensors.torch import load_file as load_safetensors sd = load_safetensors(local_path) else: if 'weights_only' in torch.load.__code__.co_varnames: sd = torch.load(local_path, map_location='cpu', weights_only=True) else: sd = torch.load(local_path, map_location='cpu', weights_only=True) first_stage_model_path = os.path.join( os.path.dirname(local_path), 'first_stage_model.pth') cond_stage_model_path = os.path.join( os.path.dirname(local_path), 'cond_stage_model.pth') diffusion_model_path = os.path.join( os.path.dirname(local_path), 'diffusion_model.pth') if (not os.path.exists(first_stage_model_path) or not os.path.exists(cond_stage_model_path) or not os.path.exists(diffusion_model_path)): self.logger.info( 'Now read the whole model and rearrange the modules, it may take several mins.' ) first_stage_model = OrderedDict() cond_stage_model = OrderedDict() diffusion_model = OrderedDict() for k, v in sd.items(): if k.startswith('first_stage_model.'): first_stage_model[k.replace( 'first_stage_model.', '')] = v elif k.startswith('conditioner.'): cond_stage_model[k.replace('conditioner.', '')] = v elif k.startswith('cond_stage_model.'): if k.startswith('cond_stage_model.model.'): cond_stage_model[k.replace( 'cond_stage_model.model.', '')] = v else: cond_stage_model[k.replace( 'cond_stage_model.', '')] = v elif k.startswith('model.diffusion_model.'): diffusion_model[k.replace('model.diffusion_model.', '')] = v else: continue if cfg.have('FIRST_STAGE_MODEL'): with open(first_stage_model_path + 'cache', 'wb') as f: torch.save(first_stage_model, f) os.rename(first_stage_model_path + 'cache', first_stage_model_path) self.logger.info( 'First stage model has been processed.') if cfg.have('COND_STAGE_MODEL'): with open(cond_stage_model_path + 'cache', 'wb') as f: torch.save(cond_stage_model, f) os.rename(cond_stage_model_path + 'cache', cond_stage_model_path) self.logger.info( 'Cond stage model has been processed.') if cfg.have('DIFFUSION_MODEL'): with open(diffusion_model_path + 'cache', 'wb') as f: torch.save(diffusion_model, f) os.rename(diffusion_model_path + 'cache', diffusion_model_path) self.logger.info('Diffusion model has been processed.') if not cfg.FIRST_STAGE_MODEL.get('PRETRAINED_MODEL', None): cfg.FIRST_STAGE_MODEL.PRETRAINED_MODEL = first_stage_model_path else: cfg.FIRST_STAGE_MODEL.RELOAD_MODEL = first_stage_model_path if not cfg.COND_STAGE_MODEL.get('PRETRAINED_MODEL', None): cfg.COND_STAGE_MODEL.PRETRAINED_MODEL = cond_stage_model_path else: cfg.COND_STAGE_MODEL.RELOAD_MODEL = cond_stage_model_path if not cfg.DIFFUSION_MODEL.get('PRETRAINED_MODEL', None): cfg.DIFFUSION_MODEL.PRETRAINED_MODEL = diffusion_model_path else: cfg.DIFFUSION_MODEL.RELOAD_MODEL = diffusion_model_path return cfg def init_from_modules(self, modules): for k, v in modules.items(): self.__setattr__(k, v) def infer_model(self, cfg, module_paras=None): module = { 'model': None, 'cfg': cfg, 'device': 'offline', 'name': cfg.NAME, 'function_info': {}, 'paras': {} } if module_paras is None: return module function_info = {} paras = { k.lower(): v for k, v in module_paras.get('PARAS', {}).items() } for function in module_paras.get('FUNCTION', []): input_dict = {} for inp in function.get('INPUT', []): if inp.lower() in self.input: input_dict[inp.lower()] = self.input[inp.lower()] function_info[function.NAME] = { 'dtype': function.get('DTYPE', 'float32'), 'input': input_dict } module['paras'] = paras module['function_info'] = function_info return module def init_from_ckpt(self, path, model, ignore_keys=list()): if path.endswith('safetensors'): from safetensors.torch import load_file as load_safetensors sd = load_safetensors(path) else: sd = torch.load(path, map_location='cpu', weights_only=True) new_sd = OrderedDict() for k, v in sd.items(): ignored = False for ik in ignore_keys: if ik in k: if we.rank == 0: self.logger.info( 'Ignore key {} from state_dict.'.format(k)) ignored = True break if not ignored: new_sd[k] = v missing, unexpected = model.load_state_dict(new_sd, strict=False) if we.rank == 0: self.logger.info( f'Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys' ) if len(missing) > 0: self.logger.info(f'Missing Keys:\n {missing}') if len(unexpected) > 0: self.logger.info(f'\nUnexpected Keys:\n {unexpected}') def load(self, module): if module['device'] == 'offline': from scepter.modules.utils.import_utils import LazyImportModule if (LazyImportModule.get_module_type(('MODELS', module['cfg'].NAME)) or module['cfg'].NAME in MODELS.class_map): model = MODELS.build(module['cfg'], logger=self.logger).eval() elif (LazyImportModule.get_module_type(('BACKBONES', module['cfg'].NAME)) or module['cfg'].NAME in BACKBONES.class_map): model = BACKBONES.build(module['cfg'], logger=self.logger).eval() elif (LazyImportModule.get_module_type(('EMBEDDERS', module['cfg'].NAME)) or module['cfg'].NAME in EMBEDDERS.class_map): model = EMBEDDERS.build(module['cfg'], logger=self.logger).eval() else: raise NotImplementedError if 'DTYPE' in module['cfg'] and module['cfg']['DTYPE'] is not None: model = model.to(getattr(torch, module['cfg'].DTYPE)) if module['cfg'].get('RELOAD_MODEL', None): self.init_from_ckpt(module['cfg'].RELOAD_MODEL, model) module['model'] = model module['device'] = 'cpu' if module['device'] == 'cpu': module['device'] = we.device_id module['model'] = module['model'].to(we.device_id) return module def unload(self, module): if module is None: return module mem = get_available_memory() free_mem = int(mem['available'] / (1024**2)) total_mem = int(mem['total'] / (1024**2)) if free_mem < 0.5 * total_mem: if module['model'] is not None: module['model'] = module['model'].to('cpu') del module['model'] module['model'] = None module['device'] = 'offline' print('delete module') else: if module['model'] is not None: module['model'] = module['model'].to('cpu') module['device'] = 'cpu' else: module['device'] = 'offline' if torch.cuda.is_available(): torch.cuda.empty_cache() torch.cuda.ipc_collect() return module def dynamic_load(self, module=None, name=''): self.logger.info('Loading {} model'.format(name)) if name == 'all': for subname in self.loaded_model_name: self.loaded_model[subname] = self.dynamic_load( getattr(self, subname), subname) elif name in self.loaded_model_name: if name in self.loaded_model: if module['cfg'] != self.loaded_model[name]['cfg']: self.unload(self.loaded_model[name]) module = self.load(module) self.loaded_model[name] = module return module elif module['device'] == 'cpu' or module['device'] == 'offline': module = self.load(module) return module else: return module else: module = self.load(module) self.loaded_model[name] = module return module else: return self.load(module) def dynamic_unload(self, module=None, name='', skip_loaded=False): self.logger.info('Unloading {} model'.format(name)) if name == 'all': for name, module in self.loaded_model.items(): module = self.unload(self.loaded_model[name]) self.loaded_model[name] = module elif name in self.loaded_model_name: if name in self.loaded_model: if not skip_loaded: module = self.unload(self.loaded_model[name]) self.loaded_model[name] = module else: self.unload(module) else: self.unload(module) def load_default(self, cfg): module_paras = {} if cfg is not None: self.paras = cfg.PARAS self.input_cfg = {k.lower(): v for k, v in cfg.INPUT.items()} self.input = {k.lower(): dict(v).get('DEFAULT', None) if isinstance(v, (dict, OrderedDict, Config)) else v for k, v in cfg.INPUT.items()} self.output = {k.lower(): v for k, v in cfg.OUTPUT.items()} module_paras = cfg.MODULES_PARAS return module_paras def load_schedule(self, cfg): parameterization = cfg.get('PARAMETERIZATION', 'eps') assert parameterization in [ 'eps', 'x0', 'v', 'rf' ], 'currently only supporting "eps" and "x0" and "v" and "rf"' num_timesteps = cfg.get('TIMESTEPS', 1000) schedule_args = { k.lower(): v for k, v in cfg.get('SCHEDULE_ARGS', { 'NAME': 'logsnr_cosine_interp', 'SCALE_MIN': 2.0, 'SCALE_MAX': 4.0 }).items() } zero_terminal_snr = cfg.get('ZERO_TERMINAL_SNR', False) if zero_terminal_snr: assert parameterization == 'v', 'Now zero_terminal_snr only support v-prediction mode.' sigmas = noise_schedule(schedule=schedule_args.pop('name'), n=num_timesteps, zero_terminal_snr=zero_terminal_snr, **schedule_args) diffusion = self.diffusion_insclass(sigmas=sigmas, prediction_type=parameterization) return diffusion def get_batch(self, value_dict, num_samples=1): batch = {} batch_uc = {} N = num_samples device = we.device_id for key in value_dict: if key == 'prompt': if not self.tokenizer: batch['prompt'] = value_dict['prompt'] batch_uc['prompt'] = value_dict['negative_prompt'] else: batch['tokens'] = self.tokenizer(value_dict['prompt']).to( we.device_id) batch_uc['tokens'] = self.tokenizer( value_dict['negative_prompt']).to(we.device_id) elif key == 'original_size_as_tuple': batch['original_size_as_tuple'] = (torch.tensor( value_dict['original_size_as_tuple']).to(device).repeat( N, 1)) elif key == 'crop_coords_top_left': batch['crop_coords_top_left'] = (torch.tensor( value_dict['crop_coords_top_left']).to(device).repeat( N, 1)) elif key == 'aesthetic_score': batch['aesthetic_score'] = (torch.tensor( [value_dict['aesthetic_score']]).to(device).repeat(N, 1)) batch_uc['aesthetic_score'] = (torch.tensor([ value_dict['negative_aesthetic_score'] ]).to(device).repeat(N, 1)) elif key == 'target_size_as_tuple': batch['target_size_as_tuple'] = (torch.tensor( value_dict['target_size_as_tuple']).to(device).repeat( N, 1)) elif key == 'image': batch[key] = self.load_image(value_dict[key], num_samples=N) else: batch[key] = value_dict[key] for key in batch.keys(): if key not in batch_uc and isinstance(batch[key], torch.Tensor): batch_uc[key] = torch.clone(batch[key]) return batch, batch_uc def load_image(self, image, num_samples=1): if isinstance(image, torch.Tensor): pass elif isinstance(image, Image): pass elif isinstance(image, Image): pass def get_function_info(self, module, function_name=None): all_function = module['function_info'] if function_name in all_function: return function_name, all_function[function_name]['dtype'] if function_name is None and len(all_function) == 1: for k, v in all_function.items(): return k, v['dtype'] def encode_first_stage(self, x, **kwargs): _, dtype = self.get_function_info(self.first_stage_model, 'encode') with torch.autocast('cuda', enabled=dtype == 'float16', dtype=getattr(torch, dtype)): z = get_model(self.first_stage_model).encode(x) return self.first_stage_model['paras']['scale_factor'] * z def decode_first_stage(self, z): _, dtype = self.get_function_info(self.first_stage_model, 'decode') with torch.autocast('cuda', enabled=dtype == 'float16', dtype=getattr(torch, dtype)): z = 1. / self.first_stage_model['paras']['scale_factor'] * z return get_model(self.first_stage_model).decode(z) @torch.no_grad() def __call__(self, input, num_samples=1, intermediate_callback=None, refine_strength=0, img_to_img_strength=0, cat_uc=True, tuner_model=None, control_model=None, **kwargs): value_input = copy.deepcopy(self.input) value_input.update(input) print(value_input) height, width = value_input['target_size_as_tuple'] value_output = copy.deepcopy(self.output) batch, batch_uc = self.get_batch(value_input, num_samples=1) # register tuner if tuner_model is not None and tuner_model != '' and len( tuner_model) > 0: if not isinstance(tuner_model, list): tuner_model = [tuner_model] self.dynamic_load(self.diffusion_model, 'diffusion_model') self.dynamic_load(self.cond_stage_model, 'cond_stage_model') self.tuner_infer.register_tuner(tuner_model, self.diffusion_model, self.cond_stage_model) self.dynamic_unload(self.diffusion_model, 'diffusion_model', skip_loaded=True) self.dynamic_unload(self.cond_stage_model, 'cond_stage_model', skip_loaded=True) # register control if control_model is not None and control_model != '': self.dynamic_load(self.diffusion_model, 'diffusion_model') hints = ControlInference.get_control_input( control_model, kwargs.pop('control_cond_image', None), height, width) self.control_infer.register_controllers(control_model, self.diffusion_model) self.dynamic_unload(self.diffusion_model, 'diffusion_model', skip_loaded=True) else: hints = None # first stage encode image = input.pop('image', None) if image is not None and img_to_img_strength > 0: # run image2image b, c, ori_width, ori_height = image.shape if not (ori_width == width and ori_height == height): image = F.interpolate(image, (width, height), mode='bicubic') self.dynamic_load(self.first_stage_model, 'first_stage_model') input_latent = self.encode_first_stage(image) self.dynamic_unload(self.first_stage_model, 'first_stage_model', skip_loaded=True) else: input_latent = None if 'input_latent' in value_output and input_latent is not None: value_output['input_latent'] = input_latent # cond stage self.dynamic_load(self.cond_stage_model, 'cond_stage_model') function_name, dtype = self.get_function_info(self.cond_stage_model) with torch.autocast('cuda', enabled=dtype == 'float16', dtype=getattr(torch, dtype)): if self.tokenizer: if not hasattr(get_model(self.cond_stage_model), 'tokenizer'): setattr(get_model(self.cond_stage_model), 'tokenizer', self.tokenizer) context = getattr(get_model(self.cond_stage_model), function_name)(batch['tokens']) null_context = getattr(get_model(self.cond_stage_model), function_name)(batch_uc['tokens']) else: context = getattr(get_model(self.cond_stage_model), function_name)(batch) null_context = getattr(get_model(self.cond_stage_model), function_name)(batch_uc) self.dynamic_unload(self.cond_stage_model, 'cond_stage_model', skip_loaded=True) if refine_strength > 0 and self.refiner_diffusion_model is not None: assert self.refiner_cond_model is not None self.refiner_cond_model = self.load(self.refiner_cond_model) function_name, dtype = self.get_function_info( self.refiner_cond_model) with torch.autocast('cuda', enabled=dtype == 'float16', dtype=getattr(torch, dtype)): if self.tokenizer: refine_context = getattr( get_model(self.refiner_cond_model), function_name)(batch['tokens']) refine_null_context = getattr( get_model(self.refiner_cond_model), function_name)(batch_uc['tokens']) else: refine_context = getattr( get_model(self.refiner_cond_model), function_name)(batch) refine_null_context = getattr( get_model(self.refiner_cond_model), function_name)(batch_uc) self.refiner_cond_model = self.unload(self.refiner_cond_model) # get noise seed = kwargs.pop('seed', -1) g = torch.Generator(device=we.device_id) seed = seed if seed >= 0 else random.randint(0, 2**32 - 1) g.manual_seed(seed) if 'seed' in value_output: value_output['seed'] = seed for sample_id in range(num_samples): if self.diffusion_model is not None: noise = torch.empty( 1, 4, height // self.first_stage_model['paras']['size_factor'], width // self.first_stage_model['paras']['size_factor'], device=we.device_id).normal_(generator=g) self.dynamic_load(self.diffusion_model, 'diffusion_model') # UNet use input n_prompt function_name, dtype = self.get_function_info( self.diffusion_model) with torch.autocast('cuda', enabled=dtype == 'float16', dtype=getattr(torch, dtype)): latent = self.diffusion.sample( noise=noise, x=input_latent, denoising_strength=img_to_img_strength if input_latent is not None else 1.0, refine_strength=refine_strength, solver=value_input.get('sample', 'ddim'), model=get_model(self.diffusion_model), model_kwargs=[{ 'cond': context, 'hint': hints }, { 'cond': null_context, 'hint': hints }], steps=value_input.get('sample_steps', 50), guide_scale=value_input.get('guide_scale', 7.5), guide_rescale=value_input.get('guide_rescale', 0.5), discretization=value_input.get('discretization', 'trailing'), show_progress=True, seed=seed, condition_fn=None, clamp=None, sharpness=value_input.get('sharpness', 0.0), percentile=None, t_max=None, t_min=None, discard_penultimate_step=None, intermediate_callback=intermediate_callback, cat_uc=value_input.get('cat_uc', cat_uc), **kwargs) self.dynamic_unload(self.diffusion_model, 'diffusion_model', skip_loaded=True) # apply refiner if refine_strength > 0 and self.refiner_diffusion_model is not None: assert self.refiner_diffusion_model is not None # decode intermidiet latent before refine self.first_stage_model = self.load(self.first_stage_model) before_refiner_samples = self.decode_first_stage( latent).float() self.first_stage_model = self.unload(self.first_stage_model) before_refiner_samples = torch.clamp( (before_refiner_samples + 1.0) / 2.0, min=0.0, max=1.0) if 'before_refine_images' in value_output: if value_output['before_refine_images'] is None or ( isinstance(value_output['before_refine_images'], list) and len(value_output['before_refine_images']) < 1): value_output['before_refine_images'] = [] value_output['before_refine_images'].append( before_refiner_samples) self.refiner_model = self.load(self.refiner_diffusion_model) function_name, dtype = self.get_function_info( self.refiner_model) with torch.autocast('cuda', enabled=dtype == 'float16', dtype=getattr(torch, dtype)): latent = self.diffusion.sample( noise=noise, x=latent, denoising_strength=img_to_img_strength if input_latent is not None else 1.0, refine_strength=refine_strength, refine_stage=True, solver=value_input.get('refine_sample', 'ddim'), model=get_model(self.refiner_model), model_kwargs=[{ 'cond': refine_context }, { 'cond': refine_null_context }], steps=value_input.get('refine_sample_steps', 50), guide_scale=value_input.get('refine_guide_scale', 7.5), guide_rescale=value_input.get('refine_guide_rescale', 0.5), discretization=value_input.get('refine_discretization', 'trailing'), show_progress=True, seed=seed, condition_fn=None, clamp=None, percentile=None, t_max=None, t_min=None, discard_penultimate_step=None, return_intermediate=None, intermediate_callback=intermediate_callback, cat_uc=cat_uc, **kwargs) self.refiner_model = self.unload(self.refiner_model) if 'latent' in value_output: if value_output['latent'] is None or ( isinstance(value_output['latent'], list) and len(value_output['latent']) < 1): value_output['latent'] = [] value_output['latent'].append(latent) self.dynamic_load(self.first_stage_model, 'first_stage_model') x_samples = self.decode_first_stage(latent).float() self.dynamic_unload(self.first_stage_model, 'first_stage_model', skip_loaded=True) images = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) if 'images' in value_output: if value_output['images'] is None or ( isinstance(value_output['images'], list) and len(value_output['images']) < 1): value_output['images'] = [] value_output['images'].append(images) for k, v in value_output.items(): if isinstance(v, list): value_output[k] = torch.cat(v, dim=0) if isinstance(v, torch.Tensor): value_output[k] = v.cpu() # unregister tuner if tuner_model is not None and tuner_model != '' and len( tuner_model) > 0: self.tuner_infer.unregister_tuner(tuner_model, self.diffusion_model, self.cond_stage_model) # unregister control if control_model is not None and control_model != '': self.control_infer.unregister_controllers(control_model, self.diffusion_model) return value_output