# video_service.py import torch import numpy as np import random import os import yaml from pathlib import Path import imageio import tempfile import sys import subprocess import threading import time from huggingface_hub import hf_hub_download # --- LÓGICA DE SETUP E DEPENDÊNCIAS --- def run_setup(): setup_script_path = "setup.py" if not os.path.exists(setup_script_path): print("AVISO: script 'setup.py' não encontrado. Pulando a clonagem de dependências.") return try: print("--- Executando setup.py para garantir que as dependências estão presentes ---") subprocess.run([sys.executable, setup_script_path], check=True) print("--- Setup concluído com sucesso ---") except subprocess.CalledProcessError as e: print(f"ERRO CRÍTICO DURANTE O SETUP: 'setup.py' falhou com código {e.returncode}.") sys.exit(1) DEPS_DIR = Path("./deps") LTX_VIDEO_REPO_DIR = DEPS_DIR / "LTX-Video" if not LTX_VIDEO_REPO_DIR.exists(): run_setup() def add_deps_to_path(): if not LTX_VIDEO_REPO_DIR.exists(): raise FileNotFoundError(f"Repositório LTX-Video não encontrado em '{LTX_VIDEO_REPO_DIR}'. Execute o setup.") if str(LTX_VIDEO_REPO_DIR.resolve()) not in sys.path: sys.path.insert(0, str(LTX_VIDEO_REPO_DIR.resolve())) add_deps_to_path() # Importações específicas do modelo from inference import ( create_ltx_video_pipeline, create_latent_upsampler, load_image_to_tensor_with_resize_and_crop, seed_everething, calculate_padding, load_media_file ) from ltx_video.pipelines.pipeline_ltx_video import ConditioningItem from ltx_video.utils.skip_layer_strategy import SkipLayerStrategy # --- CONFIGURAÇÃO DA DISTRIBUIÇÃO DE GPUS --- GPU_MAPPING = [ {'base': 'cuda:0', 'upscaler': 'cuda:2'}, {'base': 'cuda:1', 'upscaler': 'cuda:3'} ] class VideoService: def __init__(self): print("Inicializando VideoService (modo Lazy Loading)...") self.models_loaded = False self.workers = None self.config = self._load_config() self.models_dir = "downloaded_models" self.loading_lock = threading.Lock() # Para evitar que múltiplos usuários iniciem o carregamento ao mesmo tempo def _ensure_models_are_loaded(self): """Verifica se os modelos estão carregados e os carrega se não estiverem.""" with self.loading_lock: if not self.models_loaded: print("Primeira requisição recebida. Iniciando carregamento dos modelos...") if torch.cuda.is_available() and torch.cuda.device_count() < 4: raise RuntimeError(f"Este serviço está configurado para 4 GPUs, mas apenas {torch.cuda.device_count()} foram encontradas.") self._download_model_files() self.workers = self._initialize_workers() self.models_loaded = True print(f"Modelos carregados com sucesso. {len(self.workers)} workers prontos.") def _load_config(self): config_file_path = LTX_VIDEO_REPO_DIR / "configs" / "ltxv-13b-0.9.8-distilled.yaml" with open(config_file_path, "r") as file: return yaml.safe_load(file) def _download_model_files(self): Path(self.models_dir).mkdir(parents=True, exist_ok=True) LTX_REPO = "Lightricks/LTX-Video" print("Baixando arquivos de modelo (se necessário)...") self.distilled_model_path = hf_hub_download(repo_id=LTX_REPO, filename=self.config["checkpoint_path"], local_dir=self.models_dir) self.spatial_upscaler_path = hf_hub_download(repo_id=LTX_REPO, filename=self.config["spatial_upscaler_model_path"], local_dir=self.models_dir) print("Download de modelos concluído.") def _load_models_for_worker(self, base_device, upscaler_device): print(f"Carregando modelo base para {base_device} e upscaler para {upscaler_device}") pipeline = create_ltx_video_pipeline( ckpt_path=self.distilled_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", enhance_prompt=False, prompt_enhancer_image_caption_model_name_or_path=self.config["prompt_enhancer_image_caption_model_name_or_path"], prompt_enhancer_llm_model_name_or_path=self.config["prompt_enhancer_llm_model_name_or_path"], ) latent_upsampler = create_latent_upsampler(self.spatial_upscaler_path, device="cpu") pipeline.to(base_device) latent_upsampler.to(upscaler_device) return pipeline, latent_upsampler def _initialize_workers(self): workers = [] for i, mapping in enumerate(GPU_MAPPING): print(f"--- Inicializando Worker {i} ---") pipeline, latent_upsampler = self._load_models_for_worker(mapping['base'], mapping['upscaler']) workers.append({"id": i, "base_pipeline": pipeline, "latent_upsampler": latent_upsampler, "devices": mapping, "lock": threading.Lock()}) return workers def _acquire_worker(self): while True: for worker in self.workers: if worker["lock"].acquire(blocking=False): print(f"Worker {worker['id']} adquirido para uma nova tarefa.") return worker time.sleep(0.1) def generate(self, prompt, negative_prompt, input_image_filepath=None, input_video_filepath=None, height=512, width=704, mode="text-to-video", duration=2.0, frames_to_use=9, seed=42, randomize_seed=True, guidance_scale=1.0, # Ignorado, mas mantido por compatibilidade improve_texture=True, progress_callback=None): # A MÁGICA DO LAZY LOADING ACONTECE AQUI self._ensure_models_are_loaded() worker = self._acquire_worker() base_device = worker['devices']['base'] upscaler_device = worker['devices']['upscaler'] try: # ... (todo o resto do código da função generate permanece exatamente o mesmo) ... if mode == "image-to-video" and not input_image_filepath: raise ValueError("Caminho da imagem é obrigatório para o modo image-to-video") if mode == "video-to-video" and not input_video_filepath: raise ValueError("Caminho do vídeo é obrigatório para o modo video-to-video") used_seed = random.randint(0, 2**32 - 1) if randomize_seed else int(seed) seed_everething(used_seed) FPS = 24.0; MAX_NUM_FRAMES = 257 target_frames_rounded = round(duration * FPS) n_val = round((float(target_frames_rounded) - 1.0) / 8.0) actual_num_frames = max(9, min(MAX_NUM_FRAMES, int(n_val * 8 + 1))) height_padded = ((height - 1) // 32 + 1) * 32 width_padded = ((width - 1) // 32 + 1) * 32 padding_values = calculate_padding(height, width, height_padded, width_padded) pad_left, pad_right, pad_top, pad_bottom = padding_values call_kwargs_base = { "prompt": prompt, "negative_prompt": negative_prompt, "num_frames": actual_num_frames, "frame_rate": int(FPS), "decode_timestep": 0.05, "decode_noise_scale": self.config["decode_noise_scale"], "stochastic_sampling": self.config["stochastic_sampling"], "image_cond_noise_scale": 0.025, "is_video": True, "vae_per_channel_normalize": True, "mixed_precision": (self.config["precision"] == "mixed_precision"), "offload_to_cpu": False, "enhance_prompt": False, "skip_layer_strategy": SkipLayerStrategy.AttentionValues } result_tensor = None if improve_texture: downscale_factor = self.config.get("downscale_factor", 0.5) downscaled_height_ideal = int(height_padded * downscale_factor); downscaled_width_ideal = int(width_padded * downscale_factor) downscaled_height = ((downscaled_height_ideal - 1) // 32 + 1) * 32; downscaled_width = ((downscaled_width_ideal - 1) // 32 + 1) * 32 # --- PASSE 1 --- first_pass_kwargs = call_kwargs_base.copy() first_pass_kwargs.update({ "height": downscaled_height, "width": downscaled_width, "generator": torch.Generator(device=base_device).manual_seed(used_seed), "output_type": "latent", "guidance_scale": 1.0, "timesteps": self.config["first_pass"]["timesteps"], "stg_scale": self.config["first_pass"]["stg_scale"], "rescaling_scale": self.config["first_pass"]["rescaling_scale"], "skip_block_list": self.config["first_pass"]["skip_block_list"] }) if mode == "image-to-video": padding_low_res = calculate_padding(downscaled_height, downscaled_width, downscaled_height, downscaled_width) media_tensor_low_res = load_image_to_tensor_with_resize_and_crop(input_image_filepath, downscaled_height, downscaled_width) media_tensor_low_res = torch.nn.functional.pad(media_tensor_low_res, padding_low_res) first_pass_kwargs["conditioning_items"] = [ConditioningItem(media_tensor_low_res.to(base_device), 0, 1.0)] print(f"Worker {worker['id']}: Iniciando passe 1 em {base_device}") with torch.no_grad(): low_res_latents = worker['base_pipeline'](**first_pass_kwargs).images low_res_latents = low_res_latents.to(upscaler_device) with torch.no_grad(): high_res_latents = worker['latent_upsampler'](low_res_latents) high_res_latents = high_res_latents.to(base_device) # --- PASSE 2 --- second_pass_kwargs = call_kwargs_base.copy() high_res_h, high_res_w = downscaled_height * 2, downscaled_width * 2 second_pass_kwargs.update({ "height": high_res_h, "width": high_res_w, "latents": high_res_latents, "generator": torch.Generator(device=base_device).manual_seed(used_seed), "output_type": "pt", "image_cond_noise_scale": 0.0, "guidance_scale": 1.0, "timesteps": self.config["second_pass"]["timesteps"], "stg_scale": self.config["second_pass"]["stg_scale"], "rescaling_scale": self.config["second_pass"]["rescaling_scale"], "skip_block_list": self.config["second_pass"]["skip_block_list"], "tone_map_compression_ratio": self.config["second_pass"].get("tone_map_compression_ratio", 0.0) }) if mode == "image-to-video": padding_high_res = calculate_padding(high_res_h, high_res_w, high_res_h, high_res_w) media_tensor_high_res = load_image_to_tensor_with_resize_and_crop(input_image_filepath, high_res_h, high_res_w) media_tensor_high_res = torch.nn.functional.pad(media_tensor_high_res, padding_high_res) second_pass_kwargs["conditioning_items"] = [ConditioningItem(media_tensor_high_res.to(base_device), 0, 1.0)] print(f"Worker {worker['id']}: Iniciando passe 2 em {base_device}") with torch.no_grad(): result_tensor = worker['base_pipeline'](**second_pass_kwargs).images else: # Passe Único single_pass_kwargs = call_kwargs_base.copy() first_pass_config = self.config["first_pass"] single_pass_kwargs.update({ "height": height_padded, "width": width_padded, "output_type": "pt", "generator": torch.Generator(device=base_device).manual_seed(used_seed), "guidance_scale": 1.0, **first_pass_config }) if mode == "image-to-video": media_tensor_final = load_image_to_tensor_with_resize_and_crop(input_image_filepath, height_padded, width_padded) media_tensor_final = torch.nn.functional.pad(media_tensor_final, padding_values) single_pass_kwargs["conditioning_items"] = [ConditioningItem(media_tensor_final.to(base_device), 0, 1.0)] elif mode == "video-to-video": single_pass_kwargs["media_items"] = load_media_file(media_path=input_video_filepath, height=height_padded, width=width_padded, max_frames=int(frames_to_use), padding=padding_values).to(base_device) print(f"Worker {worker['id']}: Iniciando passe único em {base_device}") with torch.no_grad(): result_tensor = worker['base_pipeline'](**single_pass_kwargs).images if result_tensor.shape[-2:] != (height, width): num_frames_final = result_tensor.shape[2] videos_tensor = result_tensor.permute(0, 2, 1, 3, 4).reshape(-1, result_tensor.shape[1], result_tensor.shape[3], result_tensor.shape[4]) videos_resized = torch.nn.functional.interpolate(videos_tensor, size=(height, width), mode='bilinear', align_corners=False) result_tensor = videos_resized.reshape(result_tensor.shape[0], num_frames_final, result_tensor.shape[1], height, width).permute(0, 2, 1, 3, 4) result_tensor = result_tensor[:, :, :actual_num_frames, (pad_top if pad_top > 0 else None):(-pad_bottom if pad_bottom > 0 else None), (pad_left if pad_left > 0 else None):(-pad_right if pad_right > 0 else None)] video_np = (result_tensor[0].permute(1, 2, 3, 0).cpu().float().numpy() * 255).astype(np.uint8) temp_dir = tempfile.mkdtemp() output_video_path = os.path.join(temp_dir, f"output_{used_seed}.mp4") with imageio.get_writer(output_video_path, fps=call_kwargs_base["frame_rate"], codec='libx264', quality=8) as writer: for i, frame in enumerate(video_np): writer.append_data(frame) if progress_callback: progress_callback(i + 1, len(video_np)) return output_video_path, used_seed except Exception as e: print(f"!!!!!!!! ERRO no Worker {worker['id']} !!!!!!!!\n{e}") raise e finally: print(f"Worker {worker['id']}: Tarefa finalizada. Limpando cache e liberando worker...") with torch.cuda.device(base_device): torch.cuda.empty_cache() with torch.cuda.device(upscaler_device): torch.cuda.empty_cache() worker["lock"].release() # A instância do serviço é criada aqui, mas os modelos só serão carregados no primeiro clique. video_generation_service = VideoService()