Update api/ltx_server.py
Browse files- api/ltx_server.py +126 -263
api/ltx_server.py
CHANGED
|
@@ -1,27 +1,19 @@
|
|
| 1 |
# ltx_server.py — VideoService (beta 1.1)
|
| 2 |
# Sempre output_type="latent"; no final: VAE (bloco inteiro) → pixels → MP4.
|
| 3 |
# Ignora UserWarning/FutureWarning e injeta VAE no manager com dtype/device corretos.
|
| 4 |
-
|
| 5 |
# --- 0. WARNINGS E AMBIENTE ---
|
|
|
|
| 6 |
import warnings
|
| 7 |
warnings.filterwarnings("ignore", category=UserWarning)
|
| 8 |
warnings.filterwarnings("ignore", category=FutureWarning)
|
| 9 |
warnings.filterwarnings("ignore", message=".*")
|
| 10 |
-
|
| 11 |
from huggingface_hub import logging
|
| 12 |
-
|
| 13 |
logging.set_verbosity_error()
|
| 14 |
logging.set_verbosity_warning()
|
| 15 |
logging.set_verbosity_info()
|
| 16 |
logging.set_verbosity_debug()
|
| 17 |
-
|
| 18 |
-
|
| 19 |
LTXV_DEBUG=1
|
| 20 |
LTXV_FRAME_LOG_EVERY=8
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
# --- 1. IMPORTAÇÕES ---
|
| 25 |
import os, subprocess, shlex, tempfile
|
| 26 |
import torch
|
| 27 |
import json
|
|
@@ -44,12 +36,30 @@ import time
|
|
| 44 |
import traceback
|
| 45 |
from einops import rearrange
|
| 46 |
import torch.nn.functional as F
|
| 47 |
-
|
| 48 |
-
# Singletons (versões simples)
|
| 49 |
from managers.vae_manager import vae_manager_singleton
|
| 50 |
from tools.video_encode_tool import video_encode_tool_singleton
|
| 51 |
-
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
def _query_gpu_processes_via_nvml(device_index: int) -> List[Dict]:
|
| 54 |
try:
|
| 55 |
import psutil
|
|
@@ -83,7 +93,6 @@ def _query_gpu_processes_via_nvml(device_index: int) -> List[Dict]:
|
|
| 83 |
return results
|
| 84 |
except Exception:
|
| 85 |
return []
|
| 86 |
-
|
| 87 |
def _query_gpu_processes_via_nvidiasmi(device_index: int) -> List[Dict]:
|
| 88 |
cmd = f"nvidia-smi -i {device_index} --query-compute-apps=pid,process_name,used_memory --format=csv,noheader,nounits"
|
| 89 |
try:
|
|
@@ -107,9 +116,6 @@ def _query_gpu_processes_via_nvidiasmi(device_index: int) -> List[Dict]:
|
|
| 107 |
except Exception:
|
| 108 |
continue
|
| 109 |
return results
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
def calculate_new_dimensions(orig_w, orig_h, divisor=8):
|
| 114 |
"""
|
| 115 |
Calcula novas dimensões mantendo a proporção, garantindo que ambos os
|
|
@@ -143,8 +149,6 @@ def calculate_new_dimensions(orig_w, orig_h, divisor=8):
|
|
| 143 |
|
| 144 |
print(f"[Dimension Calc] Original: {orig_w}x{orig_h} -> Calculado: {new_w:.0f}x{new_h:.0f} -> Final (divisível por {divisor}): {final_w}x{final_h}")
|
| 145 |
return final_h, final_w # Retorna (altura, largura)
|
| 146 |
-
|
| 147 |
-
|
| 148 |
def handle_media_upload_for_dims(filepath, current_h, current_w):
|
| 149 |
"""
|
| 150 |
Esta função agora usará o novo cálculo robusto.
|
|
@@ -168,8 +172,6 @@ def handle_media_upload_for_dims(filepath, current_h, current_w):
|
|
| 168 |
except Exception as e:
|
| 169 |
print(f"Erro ao processar mídia para dimensões: {e}")
|
| 170 |
return gr.update(value=current_h), gr.update(value=current_w)
|
| 171 |
-
|
| 172 |
-
|
| 173 |
def _gpu_process_table(processes: List[Dict], current_pid: int) -> str:
|
| 174 |
if not processes:
|
| 175 |
return " - Processos ativos: (nenhum)\n"
|
|
@@ -180,52 +182,6 @@ def _gpu_process_table(processes: List[Dict], current_pid: int) -> str:
|
|
| 180 |
used_str = str(p["used_mb"]) if p.get("used_mb") is not None else "N/A"
|
| 181 |
lines.append(f" {star} {p['pid']} | {p['user']} | {p['name']} | {used_str}")
|
| 182 |
return "\n".join(lines) + "\n"
|
| 183 |
-
|
| 184 |
-
def run_setup():
|
| 185 |
-
setup_script_path = "setup.py"
|
| 186 |
-
if not os.path.exists(setup_script_path):
|
| 187 |
-
print("[DEBUG] 'setup.py' não encontrado. Pulando clonagem de dependências.")
|
| 188 |
-
return
|
| 189 |
-
try:
|
| 190 |
-
print("[DEBUG] Executando setup.py para dependências...")
|
| 191 |
-
subprocess.run([sys.executable, setup_script_path], check=True)
|
| 192 |
-
print("[DEBUG] Setup concluído com sucesso.")
|
| 193 |
-
except subprocess.CalledProcessError as e:
|
| 194 |
-
print(f"[DEBUG] ERRO no setup.py (code {e.returncode}). Abortando.")
|
| 195 |
-
sys.exit(1)
|
| 196 |
-
|
| 197 |
-
from api.ltx.inference import (
|
| 198 |
-
create_ltx_video_pipeline,
|
| 199 |
-
create_latent_upsampler,
|
| 200 |
-
load_image_to_tensor_with_resize_and_crop,
|
| 201 |
-
seed_everething,
|
| 202 |
-
calculate_padding,
|
| 203 |
-
load_media_file,
|
| 204 |
-
)
|
| 205 |
-
|
| 206 |
-
DEPS_DIR = Path("/data")
|
| 207 |
-
LTX_VIDEO_REPO_DIR = DEPS_DIR / "LTX-Video"
|
| 208 |
-
if not LTX_VIDEO_REPO_DIR.exists():
|
| 209 |
-
print(f"[DEBUG] Repositório não encontrado em {LTX_VIDEO_REPO_DIR}. Rodando setup...")
|
| 210 |
-
run_setup()
|
| 211 |
-
|
| 212 |
-
def add_deps_to_path():
|
| 213 |
-
repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
|
| 214 |
-
if str(LTX_VIDEO_REPO_DIR.resolve()) not in sys.path:
|
| 215 |
-
sys.path.insert(0, repo_path)
|
| 216 |
-
print(f"[DEBUG] Repo adicionado ao sys.path: {repo_path}")
|
| 217 |
-
|
| 218 |
-
add_deps_to_path()
|
| 219 |
-
|
| 220 |
-
# --- 3. IMPORTAÇÕES ESPECÍFICAS DO MODELO ---
|
| 221 |
-
|
| 222 |
-
from ltx_video.pipelines.pipeline_ltx_video import ConditioningItem, LTXMultiScalePipeline
|
| 223 |
-
from ltx_video.utils.skip_layer_strategy import SkipLayerStrategy
|
| 224 |
-
from ltx_video.models.autoencoders.vae_encode import un_normalize_latents, normalize_latents
|
| 225 |
-
from ltx_video.pipelines.pipeline_ltx_video import adain_filter_latent
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
# --- 4. FUNÇÕES HELPER DE LOG ---
|
| 229 |
def log_tensor_info(tensor, name="Tensor"):
|
| 230 |
if not isinstance(tensor, torch.Tensor):
|
| 231 |
print(f"\n[INFO] '{name}' não é tensor.")
|
|
@@ -240,12 +196,19 @@ def log_tensor_info(tensor, name="Tensor"):
|
|
| 240 |
except Exception:
|
| 241 |
pass
|
| 242 |
print("------------------------------------------\n")
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
class VideoService:
|
| 250 |
def __init__(self):
|
| 251 |
t0 = time.perf_counter()
|
|
@@ -366,57 +329,31 @@ class VideoService:
|
|
| 366 |
return yaml.safe_load(file)
|
| 367 |
|
| 368 |
def _load_models(self):
|
| 369 |
-
"""
|
| 370 |
-
Carrega os modelos de forma inteligente:
|
| 371 |
-
1. Tenta resolver o caminho do cache local (rápido, sem rede).
|
| 372 |
-
2. Se o arquivo não for encontrado localmente, baixa como fallback.
|
| 373 |
-
Garante que o serviço possa iniciar mesmo que o setup.py não tenha sido executado.
|
| 374 |
-
"""
|
| 375 |
t0 = time.perf_counter()
|
| 376 |
LTX_REPO = "Lightricks/LTX-Video"
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
# Se o arquivo estiver no cache, retorna o caminho instantaneamente (após uma verificação rápida de metadados).
|
| 385 |
-
# Se não estiver no cache, ela o baixa.
|
| 386 |
-
print(f"[DEBUG] Verificando {description}: {filename}...")
|
| 387 |
-
model_path = hf_hub_download(
|
| 388 |
-
repo_id=repo_id,
|
| 389 |
-
filename=filename,
|
| 390 |
-
# Forçar o uso de um cache específico se necessário
|
| 391 |
-
cache_dir=os.getenv("HF_HOME_CACHE"),
|
| 392 |
-
token=os.getenv("HF_TOKEN")
|
| 393 |
-
)
|
| 394 |
-
print(f"[DEBUG] Caminho do {description} resolvido com sucesso.")
|
| 395 |
-
return model_path
|
| 396 |
-
except Exception as e:
|
| 397 |
-
print("\n" + "="*80)
|
| 398 |
-
print(f"[ERRO CRÍTICO] Falha ao obter o modelo '{filename}'.")
|
| 399 |
-
print(f"Detalhe do erro: {e}")
|
| 400 |
-
print("Verifique sua conexão com a internet ou o estado do cache do Hugging Face.")
|
| 401 |
-
print("="*80 + "\n")
|
| 402 |
-
sys.exit(1)
|
| 403 |
-
|
| 404 |
-
# --- Checkpoint Principal ---
|
| 405 |
-
checkpoint_filename = self.config["checkpoint_path"]
|
| 406 |
-
distilled_model_path = get_or_download_model(
|
| 407 |
-
LTX_REPO, checkpoint_filename, "checkpoint principal"
|
| 408 |
)
|
| 409 |
self.config["checkpoint_path"] = distilled_model_path
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
spatial_upscaler_path =
|
| 414 |
-
LTX_REPO,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 415 |
)
|
| 416 |
self.config["spatial_upscaler_model_path"] = spatial_upscaler_path
|
|
|
|
| 417 |
|
| 418 |
-
|
| 419 |
-
print("\n[DEBUG] Construindo pipeline a partir dos caminhos resolvidos...")
|
| 420 |
pipeline = create_ltx_video_pipeline(
|
| 421 |
ckpt_path=self.config["checkpoint_path"],
|
| 422 |
precision=self.config["precision"],
|
|
@@ -434,7 +371,6 @@ class VideoService:
|
|
| 434 |
print("[DEBUG] Construindo latent_upsampler...")
|
| 435 |
latent_upsampler = create_latent_upsampler(self.config["spatial_upscaler_model_path"], device="cpu")
|
| 436 |
print("[DEBUG] Upsampler pronto.")
|
| 437 |
-
|
| 438 |
print(f"[DEBUG] _load_models() tempo total={time.perf_counter()-t0:.3f}s")
|
| 439 |
return pipeline, latent_upsampler
|
| 440 |
|
|
@@ -462,6 +398,8 @@ class VideoService:
|
|
| 462 |
pass
|
| 463 |
print(f"[DEBUG] FP8→BF16: params_promoted={p_cnt}, buffers_promoted={b_cnt}")
|
| 464 |
|
|
|
|
|
|
|
| 465 |
@torch.no_grad()
|
| 466 |
def _upsample_latents_internal(self, latents: torch.Tensor) -> torch.Tensor:
|
| 467 |
"""
|
|
@@ -478,8 +416,11 @@ class VideoService:
|
|
| 478 |
upsampled_latents = self.latent_upsampler(latents)
|
| 479 |
upsampled_latents = normalize_latents(upsampled_latents, self.pipeline.vae, vae_per_channel_normalize=True)
|
| 480 |
print(f"[DEBUG-UPSAMPLE] Shape de saída: {tuple(upsampled_latents.shape)}")
|
|
|
|
| 481 |
return upsampled_latents
|
| 482 |
|
|
|
|
|
|
|
| 483 |
def _apply_precision_policy(self):
|
| 484 |
prec = str(self.config.get("precision", "")).lower()
|
| 485 |
self.runtime_autocast_dtype = torch.float32
|
|
@@ -513,6 +454,7 @@ class VideoService:
|
|
| 513 |
print(f"[DEBUG] Cond shape={tuple(out.shape)} dtype={out.dtype} device={out.device}")
|
| 514 |
return out
|
| 515 |
|
|
|
|
| 516 |
def _dividir_latentes_por_tamanho(self, latents_brutos, num_latente_por_chunk: int, overlap: int = 1):
|
| 517 |
"""
|
| 518 |
Divide o tensor de latentes em chunks com tamanho definido em número de latentes.
|
|
@@ -543,10 +485,10 @@ class VideoService:
|
|
| 543 |
start = (num_latente_por_chunk*i)
|
| 544 |
end = (start+num_latente_por_chunk+overlap)
|
| 545 |
if i+1 < n_chunks:
|
| 546 |
-
chunk = latents_brutos[:, :, start:end, :, :].detach()
|
| 547 |
print(f"[DEBUG] chunk{i+1}[:, :, {start}:{end}, :, :] = {chunk.shape[2]}")
|
| 548 |
else:
|
| 549 |
-
chunk = latents_brutos[:, :, start:, :, :].detach()
|
| 550 |
print(f"[DEBUG] chunk{i+1}[:, :, {start}:, :, :] = {chunk.shape[2]}")
|
| 551 |
chunks.append(chunk)
|
| 552 |
i+=1
|
|
@@ -570,21 +512,6 @@ class VideoService:
|
|
| 570 |
result = subprocess.run(cmd, capture_output=True, text=True, check=True)
|
| 571 |
return int(result.stdout.strip())
|
| 572 |
|
| 573 |
-
|
| 574 |
-
|
| 575 |
-
def _dividir_latentes(self, latents_brutos):
|
| 576 |
-
total = latents_brutos.shape[2] # dimensão temporal (número de latentes)
|
| 577 |
-
|
| 578 |
-
#if total % 2 == 1: # ÍMPAR
|
| 579 |
-
# Ex: 11 → primeira 0..5, segunda 5..10
|
| 580 |
-
cut = total // 2
|
| 581 |
-
primeira = latents_brutos[:, :, :cut+1, :, :].detach()
|
| 582 |
-
segunda = latents_brutos[:, :, cut:, :, :].detach()
|
| 583 |
-
|
| 584 |
-
|
| 585 |
-
return primeira, segunda
|
| 586 |
-
|
| 587 |
-
|
| 588 |
def _gerar_lista_com_transicoes(self, pasta: str, video_paths: list[str], crossfade_frames: int = 8) -> list[str]:
|
| 589 |
"""
|
| 590 |
Gera uma nova lista de vídeos aplicando transições suaves (blend frame a frame)
|
|
@@ -662,25 +589,10 @@ class VideoService:
|
|
| 662 |
print(f"[DEBUG] Video podado {i+1} adicionado {self._get_total_frames(video_podado)} frames ✅")
|
| 663 |
|
| 664 |
|
|
|
|
| 665 |
print("===========CONCATECAO CAUSAL=============")
|
| 666 |
print(f"[DEBUG] {nova_lista}")
|
| 667 |
return nova_lista
|
| 668 |
-
|
| 669 |
-
def _concat_mp4s_no_reencode2(self, mp4_a: str, mp4_b: str, out_path: str):
|
| 670 |
-
# Concat demuxer do ffmpeg (sem reencode)
|
| 671 |
-
import tempfile, subprocess, shlex, os
|
| 672 |
-
with tempfile.NamedTemporaryFile("w", delete=False, suffix=".txt") as f:
|
| 673 |
-
f.write(f"file '{os.path.abspath(mp4_a)}'\n")
|
| 674 |
-
f.write(f"file '{os.path.abspath(mp4_b)}'\n")
|
| 675 |
-
list_path = f.name
|
| 676 |
-
cmd = f"ffmpeg -y -f concat -safe 0 -i {list_path} -c copy {out_path}"
|
| 677 |
-
print(f"[DEBUG] Concat: {cmd}")
|
| 678 |
-
try:
|
| 679 |
-
subprocess.check_call(shlex.split(cmd))
|
| 680 |
-
finally:
|
| 681 |
-
try: os.remove(list_path)
|
| 682 |
-
except Exception: pass
|
| 683 |
-
|
| 684 |
|
| 685 |
def _concat_mp4s_no_reencode(self, mp4_list: List[str], out_path: str):
|
| 686 |
"""
|
|
@@ -759,20 +671,14 @@ class VideoService:
|
|
| 759 |
if mode == "image-to-video":
|
| 760 |
start_tensor = self._prepare_conditioning_tensor(start_image_filepath, height, width, padding_values)
|
| 761 |
conditioning_items.append(ConditioningItem(start_tensor, 0, 1.0))
|
| 762 |
-
conditioning_items1.append(ConditioningItem(start_tensor, 0, 1.0))
|
| 763 |
-
|
| 764 |
if middle_image_filepath and middle_frame_number is not None:
|
| 765 |
middle_tensor = self._prepare_conditioning_tensor(middle_image_filepath, height, width, padding_values)
|
| 766 |
safe_middle_frame = max(0, min(int(middle_frame_number), actual_num_frames - 1))
|
| 767 |
conditioning_items.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight)))
|
| 768 |
-
conditioning_items1.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight)))
|
| 769 |
-
|
| 770 |
if end_image_filepath:
|
| 771 |
end_tensor = self._prepare_conditioning_tensor(end_image_filepath, height, width, padding_values)
|
| 772 |
last_frame_index = actual_num_frames - 1
|
| 773 |
conditioning_items.append(ConditioningItem(end_tensor, last_frame_index, float(end_image_weight)))
|
| 774 |
-
conditioning_items2.append(ConditioningItem(end_tensor, last_frame_index//2, 1.0))
|
| 775 |
-
|
| 776 |
print(f"[DEBUG] Conditioning items: {len(conditioning_items)}")
|
| 777 |
|
| 778 |
call_kwargs = {
|
|
@@ -789,7 +695,7 @@ class VideoService:
|
|
| 789 |
"decode_timestep": self.config["decode_timestep"],
|
| 790 |
"decode_noise_scale": self.config["decode_noise_scale"],
|
| 791 |
"stochastic_sampling": self.config["stochastic_sampling"],
|
| 792 |
-
"image_cond_noise_scale": 0.
|
| 793 |
"is_video": True,
|
| 794 |
"vae_per_channel_normalize": True,
|
| 795 |
"mixed_precision": (self.config["precision"] == "mixed_precision"),
|
|
@@ -797,12 +703,8 @@ class VideoService:
|
|
| 797 |
"enhance_prompt": False,
|
| 798 |
"skip_layer_strategy": SkipLayerStrategy.AttentionValues,
|
| 799 |
}
|
| 800 |
-
print(f"[DEBUG] output_type={call_kwargs['output_type']} skip_layer_strategy={call_kwargs['skip_layer_strategy']}")
|
| 801 |
-
|
| 802 |
latents = None
|
| 803 |
-
latents_list
|
| 804 |
-
results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
|
| 805 |
-
|
| 806 |
|
| 807 |
try:
|
| 808 |
ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
|
|
@@ -819,7 +721,6 @@ class VideoService:
|
|
| 819 |
downscale_factor = self.config.get("downscale_factor", 0.6666666)
|
| 820 |
vae_scale_factor = self.pipeline.vae_scale_factor # Geralmente 8
|
| 821 |
|
| 822 |
-
# --- <INÍCIO DA LÓGICA DE CÁLCULO EXATA> ---
|
| 823 |
# Replica a fórmula da LTXMultiScalePipeline
|
| 824 |
x_width = int(width_padded * downscale_factor)
|
| 825 |
downscaled_width = x_width - (x_width % vae_scale_factor)
|
|
@@ -852,8 +753,8 @@ class VideoService:
|
|
| 852 |
log_tensor_info(upsampled_latents, "Latentes Pós-Upscale")
|
| 853 |
print(f"[DEBUG] Upscale de Latentes concluído em {time.perf_counter() - t_upscale:.2f}s")
|
| 854 |
del base_latents; gc.collect(); torch.cuda.empty_cache()
|
| 855 |
-
|
| 856 |
|
|
|
|
| 857 |
latents_cpu_up = upsampled_latents.detach().to("cpu", non_blocking=True)
|
| 858 |
torch.cuda.empty_cache()
|
| 859 |
try:
|
|
@@ -861,18 +762,11 @@ class VideoService:
|
|
| 861 |
except Exception:
|
| 862 |
pass
|
| 863 |
|
| 864 |
-
|
| 865 |
-
|
| 866 |
-
|
| 867 |
-
lat_aup, lat_bup = self._dividir_latentes(latents_cpu_up)
|
| 868 |
-
print(f"[DEBUG] Partição Aup: {tuple(lat_aup.shape)}")
|
| 869 |
-
print(f"[DEBUG] Partição Bup: {tuple(lat_bup.shape)}")
|
| 870 |
|
| 871 |
-
|
| 872 |
-
|
| 873 |
-
#latents_parts_up = [latents_cpu_up]
|
| 874 |
-
|
| 875 |
-
par = 0
|
| 876 |
for latents in latents_parts_up:
|
| 877 |
|
| 878 |
# # --- ETAPA 3: REFINAMENTO DE TEXTURA (SECOND PASS) ---
|
|
@@ -886,39 +780,16 @@ class VideoService:
|
|
| 886 |
print(f"[DEBUG] Second Pass Dims: Target ({second_pass_width}x{second_pass_height})")
|
| 887 |
# --- <FIM DA LÓGICA DE CÁLCULO EXATA> ---
|
| 888 |
t_pass2 = time.perf_counter()
|
| 889 |
-
|
| 890 |
-
num_latent_frames_part = latents.shape[2]
|
| 891 |
-
|
| 892 |
-
|
| 893 |
-
vae_temporal_scale = self.pipeline.video_scale_factor # Geralmente 4 ou 8
|
| 894 |
-
num_pixel_frames_part = ((num_latent_frames_part - 1) * vae_temporal_scale) + 1
|
| 895 |
-
print(f"[DEBUG] Parte: {num_latent_frames_part - 1} latentes -> {num_pixel_frames_part} frames de pixel (alvo)")
|
| 896 |
-
|
| 897 |
second_pass_kwargs = call_kwargs.copy()
|
| 898 |
-
|
| 899 |
-
|
| 900 |
-
|
| 901 |
-
|
| 902 |
-
|
| 903 |
-
|
| 904 |
-
|
| 905 |
-
|
| 906 |
-
"latents": latents, # O tensor upscaled
|
| 907 |
-
"guidance_scale": float(guidance_scale),
|
| 908 |
-
**second_pass_config
|
| 909 |
-
})
|
| 910 |
-
else:
|
| 911 |
-
second_pass_kwargs.update({
|
| 912 |
-
"conditioning_items": conditioning_items2,
|
| 913 |
-
"output_type": "latent",
|
| 914 |
-
"width": second_pass_width,
|
| 915 |
-
"height": second_pass_height,
|
| 916 |
-
"num_frames": num_pixel_frames_part,
|
| 917 |
-
"latents": latents, # O tensor upscaled
|
| 918 |
-
"guidance_scale": float(guidance_scale),
|
| 919 |
-
**second_pass_config
|
| 920 |
-
})
|
| 921 |
-
par+=1
|
| 922 |
|
| 923 |
print(f"[DEBUG] Second Pass: Refinando em {width_padded}x{height_padded}...")
|
| 924 |
final_latents = self.pipeline(**second_pass_kwargs).images
|
|
@@ -943,70 +814,62 @@ class VideoService:
|
|
| 943 |
|
| 944 |
# --- ETAPA FINAL: DECODIFICAÇÃO E CODIFICAÇÃO MP4 ---
|
| 945 |
print("\n--- INICIANDO ETAPA FINAL: DECODIFICAÇÃO E MONTAGEM ---")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 946 |
|
| 947 |
-
|
| 948 |
-
results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
|
| 949 |
partes_mp4 = []
|
| 950 |
par = 0
|
| 951 |
-
|
| 952 |
-
|
| 953 |
-
|
| 954 |
-
|
| 955 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 956 |
try:
|
| 957 |
-
|
| 958 |
-
|
| 959 |
-
|
|
|
|
| 960 |
|
| 961 |
-
|
| 962 |
-
|
| 963 |
-
|
| 964 |
-
lat_a, lat_b = self._dividir_latentes(latents_cpu_vae)
|
| 965 |
-
print(f"[DEBUG] Partição A: {tuple(lat_a.shape)}")
|
| 966 |
-
print(f"[DEBUG] Partição B: {tuple(lat_b.shape)}")
|
| 967 |
-
|
| 968 |
-
latents_parts_vae = [lat_a, lat_b]
|
| 969 |
|
| 970 |
-
|
| 971 |
-
for latents in latents_parts_vae:
|
| 972 |
-
#print(f"[DEBUG] Partição {par}: {tuple(latents.shape)}")
|
| 973 |
-
|
| 974 |
-
par = par + 1
|
| 975 |
-
output_video_path = os.path.join(temp_dir, f"output_{used_seed}_{par}.mp4")
|
| 976 |
-
final_output_path = None
|
| 977 |
-
|
| 978 |
-
print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
|
| 979 |
-
# Usar manager com timestep por item; previne target_shape e rota NoneType.decode
|
| 980 |
-
pixel_tensor = vae_manager_singleton.decode(
|
| 981 |
-
latents.to(self.device, non_blocking=True),
|
| 982 |
-
decode_timestep=float(self.config.get("decode_timestep", 0.05))
|
| 983 |
-
)
|
| 984 |
-
log_tensor_info(pixel_tensor, "Pixel tensor (VAE saída)")
|
| 985 |
-
|
| 986 |
-
print("[DEBUG] Codificando MP4 a partir do tensor de pixels (bloco inteiro)...")
|
| 987 |
-
video_encode_tool_singleton.save_video_from_tensor(
|
| 988 |
-
pixel_tensor,
|
| 989 |
-
output_video_path,
|
| 990 |
-
fps=call_kwargs["frame_rate"],
|
| 991 |
-
progress_callback=progress_callback,
|
| 992 |
-
)
|
| 993 |
-
|
| 994 |
-
try:
|
| 995 |
-
candidate = os.path.join(results_dir, f"output_par_{par}.mp4")
|
| 996 |
-
shutil.move(output_video_path, candidate)
|
| 997 |
-
print(f"[DEBUG] MP4 parte {par} movido para {candidate}")
|
| 998 |
-
partes_mp4.append(candidate)
|
| 999 |
-
|
| 1000 |
-
except Exception as e:
|
| 1001 |
-
final_output_path = output_video_path
|
| 1002 |
-
print(f"[DEBUG] Falha no move; usando tmp como final: {e}")
|
| 1003 |
-
|
| 1004 |
total_partes = len(partes_mp4)
|
| 1005 |
if (total_partes>1):
|
| 1006 |
final_vid = os.path.join(results_dir, f"concat_fim_{used_seed}.mp4")
|
| 1007 |
-
|
| 1008 |
-
|
| 1009 |
-
self._concat_mp4s_no_reencode(partes_mp4, final_vid)
|
| 1010 |
else:
|
| 1011 |
final_vid = partes_mp4[0]
|
| 1012 |
|
|
|
|
| 1 |
# ltx_server.py — VideoService (beta 1.1)
|
| 2 |
# Sempre output_type="latent"; no final: VAE (bloco inteiro) → pixels → MP4.
|
| 3 |
# Ignora UserWarning/FutureWarning e injeta VAE no manager com dtype/device corretos.
|
|
|
|
| 4 |
# --- 0. WARNINGS E AMBIENTE ---
|
| 5 |
+
|
| 6 |
import warnings
|
| 7 |
warnings.filterwarnings("ignore", category=UserWarning)
|
| 8 |
warnings.filterwarnings("ignore", category=FutureWarning)
|
| 9 |
warnings.filterwarnings("ignore", message=".*")
|
|
|
|
| 10 |
from huggingface_hub import logging
|
|
|
|
| 11 |
logging.set_verbosity_error()
|
| 12 |
logging.set_verbosity_warning()
|
| 13 |
logging.set_verbosity_info()
|
| 14 |
logging.set_verbosity_debug()
|
|
|
|
|
|
|
| 15 |
LTXV_DEBUG=1
|
| 16 |
LTXV_FRAME_LOG_EVERY=8
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
import os, subprocess, shlex, tempfile
|
| 18 |
import torch
|
| 19 |
import json
|
|
|
|
| 36 |
import traceback
|
| 37 |
from einops import rearrange
|
| 38 |
import torch.nn.functional as F
|
|
|
|
|
|
|
| 39 |
from managers.vae_manager import vae_manager_singleton
|
| 40 |
from tools.video_encode_tool import video_encode_tool_singleton
|
| 41 |
+
DEPS_DIR = Path("/data")
|
| 42 |
+
LTX_VIDEO_REPO_DIR = DEPS_DIR / "LTX-Video"
|
| 43 |
+
if not LTX_VIDEO_REPO_DIR.exists():
|
| 44 |
+
print(f"[DEBUG] Repositório não encontrado em {LTX_VIDEO_REPO_DIR}. Rodando setup...")
|
| 45 |
+
run_setup()
|
| 46 |
+
def run_setup():
|
| 47 |
+
setup_script_path = "setup.py"
|
| 48 |
+
if not os.path.exists(setup_script_path):
|
| 49 |
+
print("[DEBUG] 'setup.py' não encontrado. Pulando clonagem de dependências.")
|
| 50 |
+
return
|
| 51 |
+
try:
|
| 52 |
+
print("[DEBUG] Executando setup.py para dependências...")
|
| 53 |
+
subprocess.run([sys.executable, setup_script_path], check=True)
|
| 54 |
+
print("[DEBUG] Setup concluído com sucesso.")
|
| 55 |
+
except subprocess.CalledProcessError as e:
|
| 56 |
+
print(f"[DEBUG] ERRO no setup.py (code {e.returncode}). Abortando.")
|
| 57 |
+
sys.exit(1)
|
| 58 |
+
def add_deps_to_path():
|
| 59 |
+
repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
|
| 60 |
+
if str(LTX_VIDEO_REPO_DIR.resolve()) not in sys.path:
|
| 61 |
+
sys.path.insert(0, repo_path)
|
| 62 |
+
print(f"[DEBUG] Repo adicionado ao sys.path: {repo_path}")
|
| 63 |
def _query_gpu_processes_via_nvml(device_index: int) -> List[Dict]:
|
| 64 |
try:
|
| 65 |
import psutil
|
|
|
|
| 93 |
return results
|
| 94 |
except Exception:
|
| 95 |
return []
|
|
|
|
| 96 |
def _query_gpu_processes_via_nvidiasmi(device_index: int) -> List[Dict]:
|
| 97 |
cmd = f"nvidia-smi -i {device_index} --query-compute-apps=pid,process_name,used_memory --format=csv,noheader,nounits"
|
| 98 |
try:
|
|
|
|
| 116 |
except Exception:
|
| 117 |
continue
|
| 118 |
return results
|
|
|
|
|
|
|
|
|
|
| 119 |
def calculate_new_dimensions(orig_w, orig_h, divisor=8):
|
| 120 |
"""
|
| 121 |
Calcula novas dimensões mantendo a proporção, garantindo que ambos os
|
|
|
|
| 149 |
|
| 150 |
print(f"[Dimension Calc] Original: {orig_w}x{orig_h} -> Calculado: {new_w:.0f}x{new_h:.0f} -> Final (divisível por {divisor}): {final_w}x{final_h}")
|
| 151 |
return final_h, final_w # Retorna (altura, largura)
|
|
|
|
|
|
|
| 152 |
def handle_media_upload_for_dims(filepath, current_h, current_w):
|
| 153 |
"""
|
| 154 |
Esta função agora usará o novo cálculo robusto.
|
|
|
|
| 172 |
except Exception as e:
|
| 173 |
print(f"Erro ao processar mídia para dimensões: {e}")
|
| 174 |
return gr.update(value=current_h), gr.update(value=current_w)
|
|
|
|
|
|
|
| 175 |
def _gpu_process_table(processes: List[Dict], current_pid: int) -> str:
|
| 176 |
if not processes:
|
| 177 |
return " - Processos ativos: (nenhum)\n"
|
|
|
|
| 182 |
used_str = str(p["used_mb"]) if p.get("used_mb") is not None else "N/A"
|
| 183 |
lines.append(f" {star} {p['pid']} | {p['user']} | {p['name']} | {used_str}")
|
| 184 |
return "\n".join(lines) + "\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
def log_tensor_info(tensor, name="Tensor"):
|
| 186 |
if not isinstance(tensor, torch.Tensor):
|
| 187 |
print(f"\n[INFO] '{name}' não é tensor.")
|
|
|
|
| 196 |
except Exception:
|
| 197 |
pass
|
| 198 |
print("------------------------------------------\n")
|
| 199 |
+
add_deps_to_path()
|
| 200 |
+
from ltx_video.pipelines.pipeline_ltx_video import ConditioningItem, LTXMultiScalePipeline
|
| 201 |
+
from ltx_video.utils.skip_layer_strategy import SkipLayerStrategy
|
| 202 |
+
from ltx_video.models.autoencoders.vae_encode import un_normalize_latents, normalize_latents
|
| 203 |
+
from ltx_video.pipelines.pipeline_ltx_video import adain_filter_latent
|
| 204 |
+
from api.ltx.inference import (
|
| 205 |
+
create_ltx_video_pipeline,
|
| 206 |
+
create_latent_upsampler,
|
| 207 |
+
load_image_to_tensor_with_resize_and_crop,
|
| 208 |
+
seed_everething,
|
| 209 |
+
calculate_padding,
|
| 210 |
+
load_media_file,
|
| 211 |
+
)
|
| 212 |
class VideoService:
|
| 213 |
def __init__(self):
|
| 214 |
t0 = time.perf_counter()
|
|
|
|
| 329 |
return yaml.safe_load(file)
|
| 330 |
|
| 331 |
def _load_models(self):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 332 |
t0 = time.perf_counter()
|
| 333 |
LTX_REPO = "Lightricks/LTX-Video"
|
| 334 |
+
print("[DEBUG] Baixando checkpoint principal...")
|
| 335 |
+
distilled_model_path = hf_hub_download(
|
| 336 |
+
repo_id=LTX_REPO,
|
| 337 |
+
filename=self.config["checkpoint_path"],
|
| 338 |
+
local_dir=os.getenv("HF_HOME"),
|
| 339 |
+
cache_dir=os.getenv("HF_HOME_CACHE"),
|
| 340 |
+
token=os.getenv("HF_TOKEN"),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 341 |
)
|
| 342 |
self.config["checkpoint_path"] = distilled_model_path
|
| 343 |
+
print(f"[DEBUG] Checkpoint em: {distilled_model_path}")
|
| 344 |
+
|
| 345 |
+
print("[DEBUG] Baixando upscaler espacial...")
|
| 346 |
+
spatial_upscaler_path = hf_hub_download(
|
| 347 |
+
repo_id=LTX_REPO,
|
| 348 |
+
filename=self.config["spatial_upscaler_model_path"],
|
| 349 |
+
local_dir=os.getenv("HF_HOME"),
|
| 350 |
+
cache_dir=os.getenv("HF_HOME_CACHE"),
|
| 351 |
+
token=os.getenv("HF_TOKEN")
|
| 352 |
)
|
| 353 |
self.config["spatial_upscaler_model_path"] = spatial_upscaler_path
|
| 354 |
+
print(f"[DEBUG] Upscaler em: {spatial_upscaler_path}")
|
| 355 |
|
| 356 |
+
print("[DEBUG] Construindo pipeline...")
|
|
|
|
| 357 |
pipeline = create_ltx_video_pipeline(
|
| 358 |
ckpt_path=self.config["checkpoint_path"],
|
| 359 |
precision=self.config["precision"],
|
|
|
|
| 371 |
print("[DEBUG] Construindo latent_upsampler...")
|
| 372 |
latent_upsampler = create_latent_upsampler(self.config["spatial_upscaler_model_path"], device="cpu")
|
| 373 |
print("[DEBUG] Upsampler pronto.")
|
|
|
|
| 374 |
print(f"[DEBUG] _load_models() tempo total={time.perf_counter()-t0:.3f}s")
|
| 375 |
return pipeline, latent_upsampler
|
| 376 |
|
|
|
|
| 398 |
pass
|
| 399 |
print(f"[DEBUG] FP8→BF16: params_promoted={p_cnt}, buffers_promoted={b_cnt}")
|
| 400 |
|
| 401 |
+
|
| 402 |
+
|
| 403 |
@torch.no_grad()
|
| 404 |
def _upsample_latents_internal(self, latents: torch.Tensor) -> torch.Tensor:
|
| 405 |
"""
|
|
|
|
| 416 |
upsampled_latents = self.latent_upsampler(latents)
|
| 417 |
upsampled_latents = normalize_latents(upsampled_latents, self.pipeline.vae, vae_per_channel_normalize=True)
|
| 418 |
print(f"[DEBUG-UPSAMPLE] Shape de saída: {tuple(upsampled_latents.shape)}")
|
| 419 |
+
|
| 420 |
return upsampled_latents
|
| 421 |
|
| 422 |
+
|
| 423 |
+
|
| 424 |
def _apply_precision_policy(self):
|
| 425 |
prec = str(self.config.get("precision", "")).lower()
|
| 426 |
self.runtime_autocast_dtype = torch.float32
|
|
|
|
| 454 |
print(f"[DEBUG] Cond shape={tuple(out.shape)} dtype={out.dtype} device={out.device}")
|
| 455 |
return out
|
| 456 |
|
| 457 |
+
|
| 458 |
def _dividir_latentes_por_tamanho(self, latents_brutos, num_latente_por_chunk: int, overlap: int = 1):
|
| 459 |
"""
|
| 460 |
Divide o tensor de latentes em chunks com tamanho definido em número de latentes.
|
|
|
|
| 485 |
start = (num_latente_por_chunk*i)
|
| 486 |
end = (start+num_latente_por_chunk+overlap)
|
| 487 |
if i+1 < n_chunks:
|
| 488 |
+
chunk = latents_brutos[:, :, start:end, :, :].clone().detach()
|
| 489 |
print(f"[DEBUG] chunk{i+1}[:, :, {start}:{end}, :, :] = {chunk.shape[2]}")
|
| 490 |
else:
|
| 491 |
+
chunk = latents_brutos[:, :, start:, :, :].clone().detach()
|
| 492 |
print(f"[DEBUG] chunk{i+1}[:, :, {start}:, :, :] = {chunk.shape[2]}")
|
| 493 |
chunks.append(chunk)
|
| 494 |
i+=1
|
|
|
|
| 512 |
result = subprocess.run(cmd, capture_output=True, text=True, check=True)
|
| 513 |
return int(result.stdout.strip())
|
| 514 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 515 |
def _gerar_lista_com_transicoes(self, pasta: str, video_paths: list[str], crossfade_frames: int = 8) -> list[str]:
|
| 516 |
"""
|
| 517 |
Gera uma nova lista de vídeos aplicando transições suaves (blend frame a frame)
|
|
|
|
| 589 |
print(f"[DEBUG] Video podado {i+1} adicionado {self._get_total_frames(video_podado)} frames ✅")
|
| 590 |
|
| 591 |
|
| 592 |
+
|
| 593 |
print("===========CONCATECAO CAUSAL=============")
|
| 594 |
print(f"[DEBUG] {nova_lista}")
|
| 595 |
return nova_lista
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 596 |
|
| 597 |
def _concat_mp4s_no_reencode(self, mp4_list: List[str], out_path: str):
|
| 598 |
"""
|
|
|
|
| 671 |
if mode == "image-to-video":
|
| 672 |
start_tensor = self._prepare_conditioning_tensor(start_image_filepath, height, width, padding_values)
|
| 673 |
conditioning_items.append(ConditioningItem(start_tensor, 0, 1.0))
|
|
|
|
|
|
|
| 674 |
if middle_image_filepath and middle_frame_number is not None:
|
| 675 |
middle_tensor = self._prepare_conditioning_tensor(middle_image_filepath, height, width, padding_values)
|
| 676 |
safe_middle_frame = max(0, min(int(middle_frame_number), actual_num_frames - 1))
|
| 677 |
conditioning_items.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight)))
|
|
|
|
|
|
|
| 678 |
if end_image_filepath:
|
| 679 |
end_tensor = self._prepare_conditioning_tensor(end_image_filepath, height, width, padding_values)
|
| 680 |
last_frame_index = actual_num_frames - 1
|
| 681 |
conditioning_items.append(ConditioningItem(end_tensor, last_frame_index, float(end_image_weight)))
|
|
|
|
|
|
|
| 682 |
print(f"[DEBUG] Conditioning items: {len(conditioning_items)}")
|
| 683 |
|
| 684 |
call_kwargs = {
|
|
|
|
| 695 |
"decode_timestep": self.config["decode_timestep"],
|
| 696 |
"decode_noise_scale": self.config["decode_noise_scale"],
|
| 697 |
"stochastic_sampling": self.config["stochastic_sampling"],
|
| 698 |
+
"image_cond_noise_scale": 0.01,
|
| 699 |
"is_video": True,
|
| 700 |
"vae_per_channel_normalize": True,
|
| 701 |
"mixed_precision": (self.config["precision"] == "mixed_precision"),
|
|
|
|
| 703 |
"enhance_prompt": False,
|
| 704 |
"skip_layer_strategy": SkipLayerStrategy.AttentionValues,
|
| 705 |
}
|
|
|
|
|
|
|
| 706 |
latents = None
|
| 707 |
+
latents_list[]
|
|
|
|
|
|
|
| 708 |
|
| 709 |
try:
|
| 710 |
ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
|
|
|
|
| 721 |
downscale_factor = self.config.get("downscale_factor", 0.6666666)
|
| 722 |
vae_scale_factor = self.pipeline.vae_scale_factor # Geralmente 8
|
| 723 |
|
|
|
|
| 724 |
# Replica a fórmula da LTXMultiScalePipeline
|
| 725 |
x_width = int(width_padded * downscale_factor)
|
| 726 |
downscaled_width = x_width - (x_width % vae_scale_factor)
|
|
|
|
| 753 |
log_tensor_info(upsampled_latents, "Latentes Pós-Upscale")
|
| 754 |
print(f"[DEBUG] Upscale de Latentes concluído em {time.perf_counter() - t_upscale:.2f}s")
|
| 755 |
del base_latents; gc.collect(); torch.cuda.empty_cache()
|
|
|
|
| 756 |
|
| 757 |
+
par = 0
|
| 758 |
latents_cpu_up = upsampled_latents.detach().to("cpu", non_blocking=True)
|
| 759 |
torch.cuda.empty_cache()
|
| 760 |
try:
|
|
|
|
| 762 |
except Exception:
|
| 763 |
pass
|
| 764 |
|
| 765 |
+
latents_parts_up = self._dividir_latentes_por_tamanho(latents_cpu_up,4,1)
|
| 766 |
+
temp_dir = tempfile.mkdtemp(prefix="ltxv_"); self._register_tmp_dir(temp_dir)
|
| 767 |
+
results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
|
|
|
|
|
|
|
|
|
|
| 768 |
|
| 769 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
| 770 |
for latents in latents_parts_up:
|
| 771 |
|
| 772 |
# # --- ETAPA 3: REFINAMENTO DE TEXTURA (SECOND PASS) ---
|
|
|
|
| 780 |
print(f"[DEBUG] Second Pass Dims: Target ({second_pass_width}x{second_pass_height})")
|
| 781 |
# --- <FIM DA LÓGICA DE CÁLCULO EXATA> ---
|
| 782 |
t_pass2 = time.perf_counter()
|
| 783 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 784 |
second_pass_kwargs = call_kwargs.copy()
|
| 785 |
+
second_pass_kwargs.update({
|
| 786 |
+
"output_type": "latent",
|
| 787 |
+
"width": second_pass_width,
|
| 788 |
+
"height": second_pass_height,
|
| 789 |
+
"latents": upsampled_latents, # O tensor upscaled
|
| 790 |
+
"guidance_scale": float(guidance_scale),
|
| 791 |
+
**second_pass_config
|
| 792 |
+
})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 793 |
|
| 794 |
print(f"[DEBUG] Second Pass: Refinando em {width_padded}x{height_padded}...")
|
| 795 |
final_latents = self.pipeline(**second_pass_kwargs).images
|
|
|
|
| 814 |
|
| 815 |
# --- ETAPA FINAL: DECODIFICAÇÃO E CODIFICAÇÃO MP4 ---
|
| 816 |
print("\n--- INICIANDO ETAPA FINAL: DECODIFICAÇÃO E MONTAGEM ---")
|
| 817 |
+
|
| 818 |
+
#latents_cpu = latents.detach().to("cpu", non_blocking=True)
|
| 819 |
+
#torch.cuda.empty_cache()
|
| 820 |
+
#try:
|
| 821 |
+
# torch.cuda.ipc_collect()
|
| 822 |
+
#except Exception:
|
| 823 |
+
# pass
|
| 824 |
+
|
| 825 |
+
latents_parts[]
|
| 826 |
+
for latents in latents_list:
|
| 827 |
+
latents_parts.append(self._dividir_latentes_por_tamanho(latents_cpu,4,1))
|
| 828 |
+
temp_dir = tempfile.mkdtemp(prefix="ltxv_"); self._register_tmp_dir(temp_dir)
|
| 829 |
+
results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
|
| 830 |
|
| 831 |
+
|
|
|
|
| 832 |
partes_mp4 = []
|
| 833 |
par = 0
|
| 834 |
+
for latents in latents_parts:
|
| 835 |
+
print(f"[DEBUG] Partição {par}: {tuple(latents.shape)}")
|
| 836 |
+
|
| 837 |
+
par = par + 1
|
| 838 |
+
output_video_path = os.path.join(temp_dir, f"output_{used_seed}_{par}.mp4")
|
| 839 |
+
final_output_path = None
|
| 840 |
+
|
| 841 |
+
print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
|
| 842 |
+
# Usar manager com timestep por item; previne target_shape e rota NoneType.decode
|
| 843 |
+
pixel_tensor = vae_manager_singleton.decode(
|
| 844 |
+
latents.to(self.device, non_blocking=True),
|
| 845 |
+
decode_timestep=float(self.config.get("decode_timestep", 0.05))
|
| 846 |
+
)
|
| 847 |
+
log_tensor_info(pixel_tensor, "Pixel tensor (VAE saída)")
|
| 848 |
+
|
| 849 |
+
print("[DEBUG] Codificando MP4 a partir do tensor de pixels (bloco inteiro)...")
|
| 850 |
+
video_encode_tool_singleton.save_video_from_tensor(
|
| 851 |
+
pixel_tensor,
|
| 852 |
+
output_video_path,
|
| 853 |
+
fps=call_kwargs["frame_rate"],
|
| 854 |
+
progress_callback=progress_callback
|
| 855 |
+
)
|
| 856 |
+
|
| 857 |
+
candidate = os.path.join(results_dir, f"output_par_{par}.mp4")
|
| 858 |
try:
|
| 859 |
+
shutil.move(output_video_path, candidate)
|
| 860 |
+
final_output_path = candidate
|
| 861 |
+
print(f"[DEBUG] MP4 parte {par} movido para {final_output_path}")
|
| 862 |
+
partes_mp4.append(final_output_path)
|
| 863 |
|
| 864 |
+
except Exception as e:
|
| 865 |
+
final_output_path = output_video_path
|
| 866 |
+
print(f"[DEBUG] Falha no move; usando tmp como final: {e}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 867 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 868 |
total_partes = len(partes_mp4)
|
| 869 |
if (total_partes>1):
|
| 870 |
final_vid = os.path.join(results_dir, f"concat_fim_{used_seed}.mp4")
|
| 871 |
+
partes_mp4_fade = self._gerar_lista_com_transicoes(pasta=results_dir, video_paths=partes_mp4, crossfade_frames=8)
|
| 872 |
+
self._concat_mp4s_no_reencode(partes_mp4_fade, final_vid)
|
|
|
|
| 873 |
else:
|
| 874 |
final_vid = partes_mp4[0]
|
| 875 |
|