Update managers/seedvr_manager.py

managers/seedvr_manager.py

@@ -2,13 +2,14 @@
 #
 # Copyright (C) 2025 Carlos Rodrigues dos Santos
 #
-# Version: 2.3.2
+# Version: 2.3.3
 #
-# Esta versão implementa uma correção robusta para o FileNotFoundError da configuração do VAE,
-# antecipando a falha, carregando as configurações manualmente e fundindo-as para
-# contornar o caminho fixo problemático na biblioteca externa.
+# This version adds a monkey patch to disable torch.distributed.barrier calls
+# within the SeedVR library, allowing it to run in a single-GPU inference mode
+# without raising a "process group not initialized" error.
 
 import torch
+import torch.distributed as dist
 import os
 import gc
 import logging
@@ -25,29 +26,29 @@ from tools.tensor_utils import wavelet_reconstruction
 
 logger = logging.getLogger(__name__)
 
-# --- Gerenciamento de Dependências ---
+# --- Dependency Management ---
 DEPS_DIR = Path("./deps")
 SEEDVR_REPO_DIR = DEPS_DIR / "SeedVR"
 SEEDVR_REPO_URL = "https://github.com/ByteDance-Seed/SeedVR.git"
 VAE_CONFIG_URL = "https://raw.githubusercontent.com/ByteDance-Seed/SeedVR/main/models/video_vae_v3/s8_c16_t4_inflation_sd3.yaml"
 
 def setup_seedvr_dependencies():
-    """Garante que o repositório do SeedVR seja clonado e esteja disponível no sys.path."""
+    """Ensures the SeedVR repository is cloned and available in the sys.path."""
     if not SEEDVR_REPO_DIR.exists():
-        logger.info(f"Repositório SeedVR não encontrado em '{SEEDVR_REPO_DIR}'. Clonando do GitHub...")
+        logger.info(f"SeedVR repository not found at '{SEEDVR_REPO_DIR}'. Cloning from GitHub...")
         try:
             DEPS_DIR.mkdir(exist_ok=True)
             subprocess.run(["git", "clone", "--depth", "1", SEEDVR_REPO_URL, str(SEEDVR_REPO_DIR)], check=True, capture_output=True, text=True)
-            logger.info("Repositório SeedVR clonado com sucesso.")
+            logger.info("SeedVR repository cloned successfully.")
         except subprocess.CalledProcessError as e:
-            logger.error(f"Falha ao clonar o repositório SeedVR. Git stderr: {e.stderr}")
-            raise RuntimeError("Não foi possível clonar a dependência necessária do SeedVR do GitHub.")
+            logger.error(f"Failed to clone SeedVR repository. Git stderr: {e.stderr}")
+            raise RuntimeError("Could not clone the required SeedVR dependency from GitHub.")
     else:
-        logger.info("Repositório SeedVR local encontrado.")
+        logger.info("Found local SeedVR repository.")
     
     if str(SEEDVR_REPO_DIR.resolve()) not in sys.path:
         sys.path.insert(0, str(SEEDVR_REPO_DIR.resolve()))
-        logger.info(f"Adicionado '{SEEDVR_REPO_DIR.resolve()}' ao sys.path.")
+        logger.info(f"Added '{SEEDVR_REPO_DIR.resolve()}' to sys.path.")
 
 setup_seedvr_dependencies()
 
@@ -61,27 +62,29 @@ from torchvision.transforms import Compose, Lambda, Normalize
 from torchvision.io.video import read_video
 from omegaconf import OmegaConf
 
+
 def _load_file_from_url(url, model_dir='./', file_name=None):
     os.makedirs(model_dir, exist_ok=True)
     filename = file_name or os.path.basename(urlparse(url).path)
     cached_file = os.path.abspath(os.path.join(model_dir, filename))
     if not os.path.exists(cached_file):
-        logger.info(f'Baixando: "{url}" para {cached_file}')
+        logger.info(f'Downloading: "{url}" to {cached_file}')
         download_url_to_file(url, cached_file, hash_prefix=None, progress=True)
     return cached_file
 
 class SeedVrManager:
-    """Gerencia o modelo SeedVR para tarefas de Masterização HD."""
+    """Manages the SeedVR model for HD Mastering tasks."""
     def __init__(self, workspace_dir="deformes_workspace"):
         self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
         self.runner = None
         self.workspace_dir = workspace_dir
         self.is_initialized = False
-        logger.info("SeedVrManager inicializado. O modelo será carregado sob demanda.")
+        self._original_barrier = None # To store the original distributed barrier function
+        logger.info("SeedVrManager initialized. Model will be loaded on demand.")
 
     def _download_models_and_configs(self):
-        """Baixa os checkpoints necessários E o arquivo de configuração do VAE que pode estar faltando."""
-        logger.info("Verificando e baixando modelos e configurações do SeedVR2...")
+        """Downloads the necessary checkpoints AND the missing VAE config file."""
+        logger.info("Verifying and downloading SeedVR2 models and configs...")
         ckpt_dir = SEEDVR_REPO_DIR / 'ckpts'
         config_dir = SEEDVR_REPO_DIR / 'configs' / 'vae'
         ckpt_dir.mkdir(exist_ok=True)
@@ -96,13 +99,19 @@ class SeedVrManager:
         }
         for key, url in pretrain_model_urls.items():
             _load_file_from_url(url=url, model_dir=str(ckpt_dir))
-        logger.info("Modelos e configurações do SeedVR2 baixados com sucesso.")
+        logger.info("SeedVR2 models and configs downloaded successfully.")
 
     def _initialize_runner(self, model_version: str):
-        """Carrega e configura o modelo SeedVR, com uma correção robusta para o caminho da config do VAE."""
+        """Loads and configures the SeedVR model, with patches for single-GPU inference."""
         if self.runner is not None: return
         self._download_models_and_configs()
-        logger.info(f"Inicializando o executor do SeedVR2 {model_version}...")
+
+        if dist.is_available() and not dist.is_initialized():
+            logger.info("Applying patch to disable torch.distributed.barrier for single-GPU inference.")
+            self._original_barrier = dist.barrier
+            dist.barrier = lambda *args, **kwargs: None
+        
+        logger.info(f"Initializing SeedVR2 {model_version} runner...")
         if model_version == '3B':
             config_path = SEEDVR_REPO_DIR / 'configs_3b' / 'main.yaml'
             checkpoint_path = SEEDVR_REPO_DIR / 'ckpts' / 'seedvr2_ema_3b.pth'
@@ -110,17 +119,17 @@ class SeedVrManager:
             config_path = SEEDVR_REPO_DIR / 'configs_7b' / 'main.yaml'
             checkpoint_path = SEEDVR_REPO_DIR / 'ckpts' / 'seedvr2_ema_7b.pth'
         else:
-            raise ValueError(f"Versão do modelo SeedVR não suportada: {model_version}")
+            raise ValueError(f"Unsupported SeedVR model version: {model_version}")
 
         try:
             config = load_config(str(config_path))
         except FileNotFoundError:
-            logger.warning("FileNotFoundError esperado capturado. Carregando config manualmente.")
+            logger.warning("Caught expected FileNotFoundError. Loading config manually.")
             config = OmegaConf.load(str(config_path))
             correct_vae_config_path = SEEDVR_REPO_DIR / 'configs' / 'vae' / 's8_c16_t4_inflation_sd3.yaml'
             vae_config = OmegaConf.load(str(correct_vae_config_path))
             config.vae = vae_config
-            logger.info("Configuração carregada e corrigida manualmente com sucesso.")
+            logger.info("Configuration loaded and patched manually.")
 
         self.runner = VideoDiffusionInfer(config)
         OmegaConf.set_readonly(self.runner.config, False)
@@ -129,20 +138,25 @@ class SeedVrManager:
         if hasattr(self.runner.vae, "set_memory_limit"):
             self.runner.vae.set_memory_limit(**self.runner.config.vae.memory_limit)
         self.is_initialized = True
-        logger.info(f"Executor para SeedVR2 {model_version} inicializado e pronto.")
+        logger.info(f"Runner for SeedVR2 {model_version} initialized and ready.")
     
     def _unload_runner(self):
-        """Remove o executor da VRAM para liberar recursos."""
+        """Unloads the runner from VRAM and restores any applied patches."""
         if self.runner is not None:
             del self.runner; self.runner = None
             gc.collect(); torch.cuda.empty_cache()
             self.is_initialized = False
-            logger.info("Executor do SeedVR2 descarregado da VRAM.")
+            logger.info("SeedVR runner unloaded from VRAM.")
+        
+        if self._original_barrier is not None:
+            logger.info("Restoring original torch.distributed.barrier function.")
+            dist.barrier = self._original_barrier
+            self._original_barrier = None
 
     def process_video(self, input_video_path: str, output_video_path: str, prompt: str,
                       model_version: str = '3B', steps: int = 50, seed: int = 666, 
                       progress: gr.Progress = None) -> str:
-        """Aplica o aprimoramento HD a um vídeo usando a lógica do SeedVR."""
+        """Applies HD enhancement to a video using the SeedVR logic."""
         try:
             self._initialize_runner(model_version)
             set_seed(seed, same_across_ranks=True)
@@ -185,10 +199,10 @@ class SeedVrManager:
             final_sample = final_sample.clip(-1, 1).mul_(0.5).add_(0.5).mul_(255).round()
             final_sample_np = final_sample.to(torch.uint8).cpu().numpy()
             mediapy.write_video(output_video_path, final_sample_np, fps=24)
-            logger.info(f"Vídeo Masterizado em HD salvo em: {output_video_path}")
+            logger.info(f"HD Mastered video saved to: {output_video_path}")
             return output_video_path
         finally:
             self._unload_runner()
 
-# --- Instância Singleton ---
+# --- Singleton Instance ---
 seedvr_manager_singleton = SeedVrManager()