# ltx_manager_helpers.py # Copyright (C) 4 de Agosto de 2025 Carlos Rodrigues dos Santos # # Este programa é software livre: você pode redistribuí-lo e/ou modificá-lo # sob os termos da Licença Pública Geral Affero GNU... # AVISO DE PATENTE PENDENTE: Consulte NOTICE.md. import torch import gc import os import yaml import logging import huggingface_hub import time import threading import json from typing import Optional, List from optimization import optimize_ltx_worker, can_optimize_fp8 from hardware_manager import hardware_manager from inference import create_ltx_video_pipeline, calculate_padding from ltx_video.pipelines.pipeline_ltx_video import LatentConditioningItem, LTXMultiScalePipeline logger = logging.getLogger(__name__) class LtxWorker: """ Representa uma única instância da pipeline LTX-Video em um dispositivo específico. Gerencia o carregamento do modelo para a CPU e a movimentação de/para a GPU. """ def __init__(self, device_id, ltx_config_file): self.cpu_device = torch.device('cpu') self.device = torch.device(device_id if torch.cuda.is_available() else 'cpu') logger.info(f"LTX Worker ({self.device}): Inicializando com config '{ltx_config_file}'...") with open(ltx_config_file, "r") as file: self.config = yaml.safe_load(file) self.is_distilled = "distilled" in self.config.get("checkpoint_path", "") models_dir = "downloaded_models_gradio" logger.info(f"LTX Worker ({self.device}): Carregando modelo para a CPU...") model_path = os.path.join(models_dir, self.config["checkpoint_path"]) if not os.path.exists(model_path): model_path = huggingface_hub.hf_hub_download( repo_id="Lightricks/LTX-Video", filename=self.config["checkpoint_path"], local_dir=models_dir, local_dir_use_symlinks=False ) self.pipeline = create_ltx_video_pipeline( ckpt_path=model_path, precision=self.config["precision"], text_encoder_model_name_or_path=self.config["text_encoder_model_name_or_path"], sampler=self.config["sampler"], device='cpu' ) logger.info(f"LTX Worker ({self.device}): Modelo pronto na CPU. É um modelo destilado? {self.is_distilled}") def to_gpu(self): """Move o pipeline para a GPU designada E OTIMIZA SE POSSÍVEL.""" if self.device.type == 'cpu': return logger.info(f"LTX Worker: Movendo pipeline para a GPU {self.device}...") self.pipeline.to(self.device) if self.device.type == 'cuda' and can_optimize_fp8(): logger.info(f"LTX Worker ({self.device}): GPU com suporte a FP8 detectada. Iniciando otimização...") optimize_ltx_worker(self) logger.info(f"LTX Worker ({self.device}): Otimização concluída.") elif self.device.type == 'cuda': logger.info(f"LTX Worker ({self.device}): Otimização FP8 não suportada ou desativada.") def to_cpu(self): """Move o pipeline de volta para a CPU e libera a memória da GPU.""" if self.device.type == 'cpu': return logger.info(f"LTX Worker: Descarregando pipeline da GPU {self.device}...") self.pipeline.to('cpu') gc.collect() if torch.cuda.is_available(): torch.cuda.empty_cache() def generate_video_fragment_internal(self, **kwargs): """Invoca a pipeline de geração.""" return self.pipeline(**kwargs).images class LtxPoolManager: """ Gerencia um pool de LtxWorkers para otimizar o uso de múltiplas GPUs. MODO "HOT START": Mantém todos os modelos carregados na VRAM para latência mínima. """ def __init__(self, device_ids, ltx_config_file): logger.info(f"LTX POOL MANAGER: Criando workers para os dispositivos: {device_ids}") self.workers = [LtxWorker(dev_id, ltx_config_file) for dev_id in device_ids] self.current_worker_index = 0 self.lock = threading.Lock() if all(w.device.type == 'cuda' for w in self.workers): logger.info("LTX POOL MANAGER: MODO HOT START ATIVADO. Pré-aquecendo todas as GPUs...") for worker in self.workers: worker.to_gpu() logger.info("LTX POOL MANAGER: Todas as GPUs estão quentes e prontas.") else: logger.info("LTX POOL MANAGER: Operando em modo CPU ou misto. O pré-aquecimento de GPU foi ignorado.") def _get_next_worker(self): with self.lock: worker = self.workers[self.current_worker_index] self.current_worker_index = (self.current_worker_index + 1) % len(self.workers) return worker def _prepare_pipeline_params(self, worker: LtxWorker, **kwargs) -> dict: """Prepara o dicionário de parâmetros para a pipeline, tratando casos especiais como modelos destilados.""" pipeline_params = { "height": kwargs['height'], "width": kwargs['width'], "num_frames": kwargs['video_total_frames'], "frame_rate": kwargs.get('video_fps', 24), "generator": torch.Generator(device=worker.device).manual_seed(int(time.time()) + kwargs.get('current_fragment_index', 0)), "is_video": True, "vae_per_channel_normalize": True, "prompt": kwargs.get('motion_prompt', ""), "negative_prompt": kwargs.get('negative_prompt', "blurry, distorted, static, bad quality"), "guidance_scale": kwargs.get('guidance_scale', 1.0), "stg_scale": kwargs.get('stg_scale', 0.0), "rescaling_scale": kwargs.get('rescaling_scale', 0.15), "num_inference_steps": kwargs.get('num_inference_steps', 20), "output_type": "latent" } if 'latents' in kwargs: pipeline_params["latents"] = kwargs['latents'].to(worker.device, dtype=worker.pipeline.transformer.dtype) if 'strength' in kwargs: pipeline_params["strength"] = kwargs['strength'] if 'conditioning_items_data' in kwargs: final_conditioning_items = [] for item in kwargs['conditioning_items_data']: item.latent_tensor = item.latent_tensor.to(worker.device) final_conditioning_items.append(item) pipeline_params["conditioning_items"] = final_conditioning_items if worker.is_distilled: logger.info(f"Worker {worker.device} está usando um modelo destilado. Usando timesteps fixos.") fixed_timesteps = worker.config.get("first_pass", {}).get("timesteps") pipeline_params["timesteps"] = fixed_timesteps if fixed_timesteps: pipeline_params["num_inference_steps"] = len(fixed_timesteps) return pipeline_params def generate_latent_fragment(self, **kwargs) -> (torch.Tensor, tuple): worker_to_use = self._get_next_worker() try: # [CORREÇÃO] A lógica de padding é específica para a geração do zero. height, width = kwargs['height'], kwargs['width'] padded_h, padded_w = ((height - 1) // 32 + 1) * 32, ((width - 1) // 32 + 1) * 32 padding_vals = calculate_padding(height, width, padded_h, padded_w) kwargs['height'], kwargs['width'] = padded_h, padded_w pipeline_params = self._prepare_pipeline_params(worker_to_use, **kwargs) logger.info(f"Iniciando GERAÇÃO em {worker_to_use.device} com shape {padded_w}x{padded_h}") if isinstance(worker_to_use.pipeline, LTXMultiScalePipeline): result = worker_to_use.pipeline.video_pipeline(**pipeline_params).images else: result = worker_to_use.generate_video_fragment_internal(**pipeline_params) return result, padding_vals except Exception as e: logger.error(f"LTX POOL MANAGER: Erro durante a geração em {worker_to_use.device}: {e}", exc_info=True) raise e finally: if worker_to_use and worker_to_use.device.type == 'cuda': with torch.cuda.device(worker_to_use.device): gc.collect(); torch.cuda.empty_cache() def refine_latents(self, latents_to_refine: torch.Tensor, **kwargs) -> (torch.Tensor, tuple): worker_to_use = self._get_next_worker() try: # [CORREÇÃO] A lógica de dimensionamento para refinamento deriva da forma do latente. _b, _c, _f, latent_h, latent_w = latents_to_refine.shape vae_scale_factor = worker_to_use.pipeline.vae_scale_factor # Garante que as dimensões correspondam EXATAMENTE ao latente fornecido. kwargs['height'] = latent_h * vae_scale_factor kwargs['width'] = latent_w * vae_scale_factor kwargs['video_total_frames'] = kwargs.get('video_total_frames', _f * worker_to_use.pipeline.video_scale_factor) kwargs['latents'] = latents_to_refine kwargs['strength'] = kwargs.get('denoise_strength', 0.4) kwargs['num_inference_steps'] = int(kwargs.get('refine_steps', 10)) pipeline_params = self._prepare_pipeline_params(worker_to_use, **kwargs) logger.info(f"Iniciando REFINAMENTO em {worker_to_use.device} com shape {kwargs['width']}x{kwargs['height']}") pipeline_to_call = worker_to_use.pipeline.video_pipeline if isinstance(worker_to_use.pipeline, LTXMultiScalePipeline) else worker_to_use.pipeline result = pipeline_to_call(**pipeline_params).images return result, None except torch.cuda.OutOfMemoryError as e: logger.error(f"FALHA DE MEMÓRIA DURANTE O REFINAMENTO em {worker_to_use.device}: {e}") logger.warning("Limpando VRAM e retornando None para sinalizar a falha.") gc.collect(); torch.cuda.empty_cache() return None, None except Exception as e: logger.error(f"LTX POOL MANAGER: Erro inesperado durante o refinamento em {worker_to_use.device}: {e}", exc_info=True) raise e finally: if worker_to_use and worker_to_use.device.type == 'cuda': with torch.cuda.device(worker_to_use.device): gc.collect(); torch.cuda.empty_cache() # --- Instanciação Singleton --- logger.info("Lendo config.yaml para inicializar o LTX Pool Manager...") with open("config.yaml", 'r') as f: config = yaml.safe_load(f) ltx_gpus_required = config['specialists']['ltx']['gpus_required'] ltx_device_ids = hardware_manager.allocate_gpus('LTX', ltx_gpus_required) ltx_config_path = config['specialists']['ltx']['config_file'] ltx_manager_singleton = LtxPoolManager(device_ids=ltx_device_ids, ltx_config_file=ltx_config_path) logger.info("Especialista de Vídeo (LTX) pronto.")