Upload ltx_server.py

api/ltx_server.py

@@ -1,5 +1,13 @@
-# ltx_server.py — VideoService (beta 1.0)
+# ltx_server.py — VideoService (beta 1.1)
 # Sempre output_type="latent"; no final: VAE (bloco inteiro) → pixels → MP4.
+# Ignora UserWarning/FutureWarning e injeta VAE no manager com dtype/device corretos.
+
+# --- 0. WARNINGS E AMBIENTE ---
+import warnings
+warnings.filterwarnings("ignore", category=UserWarning)
+warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore", message=".*torch.meshgrid.*indexing.*")
+warnings.filterwarnings("ignore", message=".*Accessing config attribute `in_channels`.*")
 
 # --- 1. IMPORTAÇÕES ---
 import torch
@@ -21,22 +29,9 @@ import contextlib
 import time
 import traceback
 
-
-from huggingface_hub import logging
-
-
-logging.set_verbosity_error()
-logging.set_verbosity_warning()
-logging.set_verbosity_info()
-logging.set_verbosity_debug()
-
-enable_progress_bars()
-LTXV_DEBUG=1
-LTXV_FRAME_LOG_EVERY=8
-
-# Singletons do projeto para VAE e Encoder
-from tools.video_encode_tool import video_encode_tool_singleton
-from managers.vae_manager import vae_manager_singleton
+# Singletons (versões simples)
+from vae_manager import vae_manager_singleton
+from video_encode_tool import video_encode_tool_singleton
 
 # --- 2. GERENCIAMENTO DE DEPENDÊNCIAS E SETUP ---
 def _query_gpu_processes_via_nvml(device_index: int) -> List[Dict]:
@@ -61,6 +56,7 @@ def _query_gpu_processes_via_nvml(device_index: int) -> List[Dict]:
             name = "unknown"
             user = "unknown"
             try:
+                import psutil
                 pr = psutil.Process(pid)
                 name = pr.name()
                 user = pr.username()
@@ -126,18 +122,6 @@ if not LTX_VIDEO_REPO_DIR.exists():
     print(f"[DEBUG] Repositório não encontrado em {LTX_VIDEO_REPO_DIR}. Rodando setup...")
     run_setup()
 
-
-from api.ltx.inference import (
-    create_ltx_video_pipeline,
-    create_latent_upsampler,
-    load_image_to_tensor_with_resize_and_crop,
-    seed_everething,
-    calculate_padding,
-    load_media_file,
-)
-
-    
-
 def add_deps_to_path():
     repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
     if str(LTX_VIDEO_REPO_DIR.resolve()) not in sys.path:
@@ -147,6 +131,14 @@ def add_deps_to_path():
 add_deps_to_path()
 
 # --- 3. 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, LTXMultiScalePipeline
 from ltx_video.utils.skip_layer_strategy import SkipLayerStrategy
 
@@ -191,6 +183,14 @@ class VideoService:
         self._apply_precision_policy()
         print(f"[DEBUG] runtime_autocast_dtype = {getattr(self, 'runtime_autocast_dtype', None)}")
 
+        # Injeta pipeline/vae no manager (impede vae=None)
+        vae_manager_singleton.attach_pipeline(
+            self.pipeline,
+            device=self.device,
+            autocast_dtype=self.runtime_autocast_dtype
+        )
+        print(f"[DEBUG] VAE manager conectado: has_vae={hasattr(self.pipeline, 'vae')} device={self.device}")
+
         if self.device == "cuda":
             torch.cuda.empty_cache()
             self._log_gpu_memory("Após carregar modelos")
@@ -513,7 +513,6 @@ class VideoService:
                     result = multi_scale_pipeline(**multi_scale_call_kwargs)
                 print(f"[DEBUG] multi_scale_pipeline tempo={time.perf_counter()-t_ms:.3f}s")
 
-                # Captura latentes
                 if hasattr(result, "latents"):
                     latents = result.latents
                 elif hasattr(result, "images") and isinstance(result.images, torch.Tensor):
@@ -561,24 +560,20 @@ class VideoService:
             output_video_path = os.path.join(temp_dir, f"output_{used_seed}.mp4")
             final_output_path = None
 
+            print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
+            # Usar manager com timestep por item; previne target_shape e rota NoneType.decode
             pixel_tensor = vae_manager_singleton.decode(
                 latents.to(self.device, non_blocking=True),
                 decode_timestep=float(self.config.get("decode_timestep", 0.05))
             )
-            
-
-
-            print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
-            # Se desejar “desocupar” a GPU antes do decode, pode-se mover p/ CPU e limpar:
-            # latents_cpu = latents.detach().to("cpu", non_blocking=True); torch.cuda.empty_cache(); torch.cuda.ipc_collect(); latents = latents_cpu.to(self.device)
-            pixel_tensor = vae_manager_singleton.decode(latents.to(self.device, non_blocking=True))
             log_tensor_info(pixel_tensor, "Pixel tensor (VAE saída)")
 
             print("[DEBUG] Codificando MP4 a partir do tensor de pixels (bloco inteiro)...")
             video_encode_tool_singleton.save_video_from_tensor(
                 pixel_tensor,
                 output_video_path,
-                fps=call_kwargs["frame_rate"]
+                fps=call_kwargs["frame_rate"],
+                progress_callback=progress_callback
             )
 
             candidate_final = os.path.join(results_dir, f"output_{used_seed}.mp4")
@@ -627,4 +622,3 @@ class VideoService:
 
 print("Criando instância do VideoService. O carregamento do modelo começará agora...")
 video_generation_service = VideoService()
-