File size: 49,362 Bytes
2891de7 4ac877f 2891de7 8362ace 99beccb 8362ace 5e7dd18 ac23084 a6e974e ac23084 a6e974e ac23084 1797675 ac23084 bd507dd 33de423 5e7dd18 ac23084 2891de7 4eef400 4ac877f 1797675 cb3f487 2891de7 cb3f487 9b10e93 cb3f487 9b10e93 cb3f487 4ac877f cb3f487 ac23084 5e7dd18 ac23084 5e7dd18 ac23084 5e7dd18 ac23084 5e7dd18 ac23084 8362ace eb62b92 ac23084 5e7dd18 ac23084 5e7dd18 ac23084 5e7dd18 ac23084 1797675 8362ace 1797675 ac23084 1797675 5e7dd18 1797675 5e7dd18 1797675 5e7dd18 1797675 ac23084 5e7dd18 ac23084 5e7dd18 1797675 5e7dd18 bd507dd 9b10e93 33de423 1797675 5e7dd18 1797675 33de423 5e7dd18 33de423 2891de7 1797675 5e7dd18 bd507dd 9b10e93 5e7dd18 bd507dd 5e7dd18 bd507dd 31d7902 5e7dd18 9b10e93 31d7902 5e7dd18 9b10e93 31d7902 5e7dd18 9b10e93 5e7dd18 31d7902 9b10e93 5e7dd18 31d7902 5e7dd18 31d7902 9b10e93 5e7dd18 31d7902 5e7dd18 31d7902 5e7dd18 31d7902 5e7dd18 bd507dd ac23084 33de423 c9e392e 33de423 bda9780 33de423 5e7dd18 33de423 0f11d75 5e7dd18 ac23084 1797675 5e7dd18 ac23084 9b10e93 bd507dd 1797675 5e7dd18 bd507dd 9b10e93 bd507dd 1797675 5e7dd18 bd507dd 5e7dd18 bd507dd 5e7dd18 bd507dd 1797675 5e7dd18 1797675 5e7dd18 8ce4529 bd507dd 33de423 c825b23 5e7dd18 c825b23 33de423 5e7dd18 33de423 5e7dd18 33de423 9b10e93 33de423 9b10e93 33de423 5e7dd18 bda9780 33de423 5e7dd18 33de423 c825b23 5e7dd18 c825b23 5e7dd18 c825b23 5e7dd18 33de423 bda9780 1797675 5e7dd18 1797675 5e7dd18 1797675 cf601a2 2d8b03b cf601a2 e46b077 cf601a2 e46b077 cf601a2 79febce c8c0470 79febce c8c0470 79febce c8c0470 79febce c8c0470 cccc7f5 35d232d cf11167 35d232d c2ee5af 35d232d cf11167 35d232d cf11167 c2ee5af 35d232d c2ee5af 35d232d 85f06ad bd507dd 4ac877f 5e7dd18 bd507dd 5e7dd18 1797675 9b10e93 1797675 9b10e93 1797675 c9e392e 1797675 5e7dd18 bd507dd 1797675 5e7dd18 bd507dd 1797675 bd507dd 1797675 5e7dd18 1797675 9b10e93 1797675 bd507dd 9b10e93 bd507dd 1797675 bd507dd 0f11d75 bd507dd 1797675 bd507dd c7230e1 1797675 9b10e93 1797675 5e7dd18 bd507dd 5e7dd18 1797675 2e1e83b 85f06ad 5e7dd18 85f06ad 5e7dd18 6077505 29c78f7 6077505 3a06c45 6077505 2acde31 6077505 29c78f7 bb90730 29c78f7 6077505 29c78f7 6077505 29c78f7 6077505 85f06ad 29c78f7 85f06ad 29c78f7 6077505 29c78f7 6077505 85f06ad 2e1e83b 85f06ad 5e7dd18 9b10e93 5e7dd18 9b10e93 5e7dd18 85f06ad 9b10e93 1797675 85f06ad 1259f38 85f06ad 1259f38 8b8ae24 a600797 8b8ae24 1259f38 9b10e93 bd507dd 1259f38 85f06ad 1259f38 85f06ad 1259f38 85f06ad 1259f38 cf48f76 1259f38 33ef6a8 1259f38 415e712 d178c3a bd507dd cf601a2 1259f38 f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 0d469b1 f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 53be96b f70421a 5e7dd18 bd507dd 53be96b c344381 bd507dd 85f06ad bd507dd 85f06ad 31d7902 bd507dd 5e7dd18 ac23084 53be96b 1797675 85f06ad |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 |
# 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=".*")
from huggingface_hub import logging
logging.set_verbosity_error()
logging.set_verbosity_warning()
logging.set_verbosity_info()
logging.set_verbosity_debug()
LTXV_DEBUG=1
LTXV_FRAME_LOG_EVERY=8
# --- 1. IMPORTAÇÕES ---
import torch
import numpy as np
import random
import os
import shlex
import yaml
from typing import List, Dict
from pathlib import Path
import imageio
import tempfile
from huggingface_hub import hf_hub_download
import sys
import subprocess
import gc
import shutil
import contextlib
import time
import traceback
# Singletons (versões simples)
from managers.vae_manager import vae_manager_singleton
from tools.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]:
try:
import psutil
import pynvml as nvml
nvml.nvmlInit()
handle = nvml.nvmlDeviceGetHandleByIndex(device_index)
try:
procs = nvml.nvmlDeviceGetComputeRunningProcesses_v3(handle)
except Exception:
procs = nvml.nvmlDeviceGetComputeRunningProcesses(handle)
results = []
for p in procs:
pid = int(p.pid)
used_mb = None
try:
if getattr(p, "usedGpuMemory", None) is not None and p.usedGpuMemory not in (0,):
used_mb = max(0, int(p.usedGpuMemory) // (1024 * 1024))
except Exception:
used_mb = None
name = "unknown"
user = "unknown"
try:
import psutil
pr = psutil.Process(pid)
name = pr.name()
user = pr.username()
except Exception:
pass
results.append({"pid": pid, "name": name, "user": user, "used_mb": used_mb})
nvml.nvmlShutdown()
return results
except Exception:
return []
def _query_gpu_processes_via_nvidiasmi(device_index: int) -> List[Dict]:
cmd = f"nvidia-smi -i {device_index} --query-compute-apps=pid,process_name,used_memory --format=csv,noheader,nounits"
try:
out = subprocess.check_output(shlex.split(cmd), stderr=subprocess.STDOUT, text=True, timeout=2.0)
except Exception:
return []
results = []
for line in out.strip().splitlines():
parts = [p.strip() for p in line.split(",")]
if len(parts) >= 3:
try:
pid = int(parts[0]); name = parts[1]; used_mb = int(parts[2])
user = "unknown"
try:
import psutil
pr = psutil.Process(pid)
user = pr.username()
except Exception:
pass
results.append({"pid": pid, "name": name, "user": user, "used_mb": used_mb})
except Exception:
continue
return results
def _gpu_process_table(processes: List[Dict], current_pid: int) -> str:
if not processes:
return " - Processos ativos: (nenhum)\n"
processes = sorted(processes, key=lambda x: (x.get("used_mb") or 0), reverse=True)
lines = [" - Processos ativos (PID | USER | NAME | VRAM MB):"]
for p in processes:
star = "*" if p["pid"] == current_pid else " "
used_str = str(p["used_mb"]) if p.get("used_mb") is not None else "N/A"
lines.append(f" {star} {p['pid']} | {p['user']} | {p['name']} | {used_str}")
return "\n".join(lines) + "\n"
def run_setup():
setup_script_path = "setup.py"
if not os.path.exists(setup_script_path):
print("[DEBUG] 'setup.py' não encontrado. Pulando clonagem de dependências.")
return
try:
print("[DEBUG] Executando setup.py para dependências...")
subprocess.run([sys.executable, setup_script_path], check=True)
print("[DEBUG] Setup concluído com sucesso.")
except subprocess.CalledProcessError as e:
print(f"[DEBUG] ERRO no setup.py (code {e.returncode}). Abortando.")
sys.exit(1)
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,
)
DEPS_DIR = Path("/data")
LTX_VIDEO_REPO_DIR = DEPS_DIR / "LTX-Video"
if not LTX_VIDEO_REPO_DIR.exists():
print(f"[DEBUG] Repositório não encontrado em {LTX_VIDEO_REPO_DIR}. Rodando setup...")
run_setup()
def add_deps_to_path():
repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
if str(LTX_VIDEO_REPO_DIR.resolve()) not in sys.path:
sys.path.insert(0, repo_path)
print(f"[DEBUG] Repo adicionado ao sys.path: {repo_path}")
add_deps_to_path()
# --- 3. IMPORTAÇÕES ESPECÍFICAS DO MODELO ---
from ltx_video.pipelines.pipeline_ltx_video import ConditioningItem, LTXMultiScalePipeline
from ltx_video.utils.skip_layer_strategy import SkipLayerStrategy
# --- 4. FUNÇÕES HELPER DE LOG ---
def log_tensor_info(tensor, name="Tensor"):
if not isinstance(tensor, torch.Tensor):
print(f"\n[INFO] '{name}' não é tensor.")
return
print(f"\n--- Tensor: {name} ---")
print(f" - Shape: {tuple(tensor.shape)}")
print(f" - Dtype: {tensor.dtype}")
print(f" - Device: {tensor.device}")
if tensor.numel() > 0:
try:
print(f" - Min: {tensor.min().item():.4f} Max: {tensor.max().item():.4f} Mean: {tensor.mean().item():.4f}")
except Exception:
pass
print("------------------------------------------\n")
# --- 5. CLASSE PRINCIPAL DO SERVIÇO ---
class VideoService:
def __init__(self):
t0 = time.perf_counter()
print("[DEBUG] Inicializando VideoService...")
self.debug = os.getenv("LTXV_DEBUG", "1") == "1"
self.frame_log_every = int(os.getenv("LTXV_FRAME_LOG_EVERY", "8"))
self.config = self._load_config()
print(f"[DEBUG] Config carregada (precision={self.config.get('precision')}, sampler={self.config.get('sampler')})")
self.device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[DEBUG] Device selecionado: {self.device}")
self.last_memory_reserved_mb = 0.0
self._tmp_dirs = set(); self._tmp_files = set(); self._last_outputs = []
self.pipeline, self.latent_upsampler = self._load_models()
print(f"[DEBUG] Pipeline e Upsampler carregados. Upsampler ativo? {bool(self.latent_upsampler)}")
print(f"[DEBUG] Movendo modelos para {self.device}...")
self.pipeline.to(self.device)
if self.latent_upsampler:
self.latent_upsampler.to(self.device)
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")
print(f"[DEBUG] VideoService pronto. boot_time={time.perf_counter()-t0:.3f}s")
def _log_gpu_memory(self, stage_name: str):
if self.device != "cuda":
return
device_index = torch.cuda.current_device() if torch.cuda.is_available() else 0
current_reserved_b = torch.cuda.memory_reserved(device_index)
current_reserved_mb = current_reserved_b / (1024 ** 2)
total_memory_b = torch.cuda.get_device_properties(device_index).total_memory
total_memory_mb = total_memory_b / (1024 ** 2)
peak_reserved_mb = torch.cuda.max_memory_reserved(device_index) / (1024 ** 2)
delta_mb = current_reserved_mb - getattr(self, "last_memory_reserved_mb", 0.0)
processes = _query_gpu_processes_via_nvml(device_index) or _query_gpu_processes_via_nvidiasmi(device_index)
print(f"\n--- [LOG GPU] {stage_name} (cuda:{device_index}) ---")
print(f" - Reservado: {current_reserved_mb:.2f} MB / {total_memory_mb:.2f} MB (Δ={delta_mb:+.2f} MB)")
if peak_reserved_mb > getattr(self, "last_memory_reserved_mb", 0.0):
print(f" - Pico reservado (nesta fase): {peak_reserved_mb:.2f} MB")
print(_gpu_process_table(processes, os.getpid()), end="")
print("--------------------------------------------------\n")
self.last_memory_reserved_mb = current_reserved_mb
def _register_tmp_dir(self, d: str):
if d and os.path.isdir(d):
self._tmp_dirs.add(d); print(f"[DEBUG] Registrado tmp dir: {d}")
def _register_tmp_file(self, f: str):
if f and os.path.exists(f):
self._tmp_files.add(f); print(f"[DEBUG] Registrado tmp file: {f}")
def finalize(self, keep_paths=None, extra_paths=None, clear_gpu=True):
print("[DEBUG] Finalize: iniciando limpeza...")
keep = set(keep_paths or []); extras = set(extra_paths or [])
removed_files = 0
for f in list(self._tmp_files | extras):
try:
if f not in keep and os.path.isfile(f):
os.remove(f); removed_files += 1; print(f"[DEBUG] Removido arquivo tmp: {f}")
except Exception as e:
print(f"[DEBUG] Falha removendo arquivo {f}: {e}")
finally:
self._tmp_files.discard(f)
removed_dirs = 0
for d in list(self._tmp_dirs):
try:
if d not in keep and os.path.isdir(d):
shutil.rmtree(d, ignore_errors=True); removed_dirs += 1; print(f"[DEBUG] Removido diretório tmp: {d}")
except Exception as e:
print(f"[DEBUG] Falha removendo diretório {d}: {e}")
finally:
self._tmp_dirs.discard(d)
print(f"[DEBUG] Finalize: arquivos removidos={removed_files}, dirs removidos={removed_dirs}")
gc.collect()
try:
if clear_gpu and torch.cuda.is_available():
torch.cuda.empty_cache()
try:
torch.cuda.ipc_collect()
except Exception:
pass
except Exception as e:
print(f"[DEBUG] Finalize: limpeza GPU falhou: {e}")
try:
self._log_gpu_memory("Após finalize")
except Exception as e:
print(f"[DEBUG] Log GPU pós-finalize falhou: {e}")
def _load_config(self):
base = LTX_VIDEO_REPO_DIR / "configs"
candidates = [
base / "ltxv-13b-0.9.8-dev-fp8.yaml",
base / "ltxv-13b-0.9.8-distilled-fp8.yaml",
base / "ltxv-13b-0.9.8-distilled.yaml",
]
for cfg in candidates:
if cfg.exists():
print(f"[DEBUG] Config selecionada: {cfg}")
with open(cfg, "r") as file:
return yaml.safe_load(file)
cfg = base / "ltxv-13b-0.9.8-distilled-fp8.yaml"
print(f"[DEBUG] Config fallback: {cfg}")
with open(cfg, "r") as file:
return yaml.safe_load(file)
def _load_models(self):
t0 = time.perf_counter()
LTX_REPO = "Lightricks/LTX-Video"
print("[DEBUG] Baixando checkpoint principal...")
distilled_model_path = hf_hub_download(
repo_id=LTX_REPO,
filename=self.config["checkpoint_path"],
local_dir=os.getenv("HF_HOME"),
cache_dir=os.getenv("HF_HOME_CACHE"),
token=os.getenv("HF_TOKEN"),
)
self.config["checkpoint_path"] = distilled_model_path
print(f"[DEBUG] Checkpoint em: {distilled_model_path}")
print("[DEBUG] Baixando upscaler espacial...")
spatial_upscaler_path = hf_hub_download(
repo_id=LTX_REPO,
filename=self.config["spatial_upscaler_model_path"],
local_dir=os.getenv("HF_HOME"),
cache_dir=os.getenv("HF_HOME_CACHE"),
token=os.getenv("HF_TOKEN"),
)
self.config["spatial_upscaler_model_path"] = spatial_upscaler_path
print(f"[DEBUG] Upscaler em: {spatial_upscaler_path}")
print("[DEBUG] Construindo pipeline...")
pipeline = create_ltx_video_pipeline(
ckpt_path=self.config["checkpoint_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"],
)
print("[DEBUG] Pipeline pronto.")
latent_upsampler = None
if self.config.get("spatial_upscaler_model_path"):
print("[DEBUG] Construindo latent_upsampler...")
latent_upsampler = create_latent_upsampler(self.config["spatial_upscaler_model_path"], device="cpu")
print("[DEBUG] Upsampler pronto.")
print(f"[DEBUG] _load_models() tempo total={time.perf_counter()-t0:.3f}s")
return pipeline, latent_upsampler
def _promote_fp8_weights_to_bf16(self, module):
if not isinstance(module, torch.nn.Module):
print("[DEBUG] Promoção FP8→BF16 ignorada: alvo não é nn.Module.")
return
f8 = getattr(torch, "float8_e4m3fn", None)
if f8 is None:
print("[DEBUG] torch.float8_e4m3fn indisponível.")
return
p_cnt = b_cnt = 0
for _, p in module.named_parameters(recurse=True):
try:
if p.dtype == f8:
with torch.no_grad():
p.data = p.data.to(torch.bfloat16); p_cnt += 1
except Exception:
pass
for _, b in module.named_buffers(recurse=True):
try:
if hasattr(b, "dtype") and b.dtype == f8:
b.data = b.data.to(torch.bfloat16); b_cnt += 1
except Exception:
pass
print(f"[DEBUG] FP8→BF16: params_promoted={p_cnt}, buffers_promoted={b_cnt}")
def _apply_precision_policy(self):
prec = str(self.config.get("precision", "")).lower()
self.runtime_autocast_dtype = torch.float32
print(f"[DEBUG] Aplicando política de precisão: {prec}")
if prec == "float8_e4m3fn":
self.runtime_autocast_dtype = torch.bfloat16
force_promote = os.getenv("LTXV_FORCE_BF16_ON_FP8", "0") == "1"
print(f"[DEBUG] FP8 detectado. force_promote={force_promote}")
if force_promote and hasattr(torch, "float8_e4m3fn"):
try:
self._promote_fp8_weights_to_bf16(self.pipeline)
except Exception as e:
print(f"[DEBUG] Promoção FP8→BF16 na pipeline falhou: {e}")
try:
if self.latent_upsampler:
self._promote_fp8_weights_to_bf16(self.latent_upsampler)
except Exception as e:
print(f"[DEBUG] Promoção FP8→BF16 no upsampler falhou: {e}")
elif prec == "bfloat16":
self.runtime_autocast_dtype = torch.bfloat16
elif prec == "mixed_precision":
self.runtime_autocast_dtype = torch.float16
else:
self.runtime_autocast_dtype = torch.float32
def _prepare_conditioning_tensor(self, filepath, height, width, padding_values):
print(f"[DEBUG] Carregando condicionamento: {filepath}")
tensor = load_image_to_tensor_with_resize_and_crop(filepath, height, width)
tensor = torch.nn.functional.pad(tensor, padding_values)
out = tensor.to(self.device, dtype=self.runtime_autocast_dtype) if self.device == "cuda" else tensor.to(self.device)
print(f"[DEBUG] Cond shape={tuple(out.shape)} dtype={out.dtype} device={out.device}")
return out
def _dividir_latentes_em_partes(self, latents_brutos, quantidade: int):
"""
Divide um tensor de latentes em `quantidade` partes e retorna uma lista de clones.
Args:
latents_brutos: tensor [B, C, T, H, W]
quantidade: número de partes que queremos dividir
Returns:
List[Tensor]: lista de `quantidade` partes, cada uma cloneada
"""
total = latents_brutos.shape[2] # dimensão temporal
partes = []
if quantidade <= 1 or quantidade > total:
return [latents_brutos.clone()]
# calcular tamanho aproximado de cada parte
step = total // quantidade
overlap = 0 # sobreposição mínima de 1 frame entre partes
for i in range(quantidade):
start = i * step
end = start + step
if i == quantidade - 1:
end = total # última parte vai até o final
else:
end += overlap # sobreposição
parte = latents_brutos[:, :, start-1:end+1, :, :].clone()
partes.append(parte)
return partes
def dividir_latentes(latents_brutos):
total = latents_brutos.shape[2] # dimensão temporal (latentes)
if total % 2 == 1: # ÍMPAR
cut = total // 2
primeira = latents_brutos[:, :, :cut+1, :, :].clone()
segunda = latents_brutos[:, :, cut:, :, :].clone()
else: # PAR
cut = total // 2
# primeira parte até o meio, mas o último frame deve ser ajustado
primeira = latents_brutos[:, :, :cut+1, :, :].clone()
segunda = latents_brutos[:, :, cut:, :, :].clone()
return primeira, segunda
def _concat_mp4s_no_reencode(self, mp4_list: List[str], out_path: str):
"""
Concatena múltiplos MP4s sem reencode usando o demuxer do ffmpeg.
ATENÇÃO: todos os arquivos precisam ter mesmo codec, fps, resolução etc.
"""
if not mp4_list or len(mp4_list) < 2:
raise ValueError("Forneça pelo menos dois arquivos MP4 para concatenar.")
# Cria lista temporária para o ffmpeg
with tempfile.NamedTemporaryFile("w", delete=False, suffix=".txt") as f:
for mp4 in mp4_list:
f.write(f"file '{os.path.abspath(mp4)}'\n")
list_path = f.name
cmd = f"ffmpeg -y -f concat -safe 0 -i {list_path} -c copy {out_path}"
print(f"[DEBUG] Concat: {cmd}")
try:
subprocess.check_call(shlex.split(cmd))
finally:
try:
os.remove(list_path)
except Exception:
pass
def generate(
self,
prompt,
negative_prompt,
mode="text-to-video",
start_image_filepath=None,
middle_image_filepath=None,
middle_frame_number=None,
middle_image_weight=1.0,
end_image_filepath=None,
end_image_weight=1.0,
input_video_filepath=None,
height=512,
width=704,
duration=2.0,
frames_to_use=9,
seed=42,
randomize_seed=True,
guidance_scale=3.0,
improve_texture=True,
progress_callback=None,
# Sempre latent → VAE → MP4 (simples)
external_decode=True,
):
t_all = time.perf_counter()
print(f"[DEBUG] generate() begin mode={mode} external_decode={external_decode} improve_texture={improve_texture}")
if self.device == "cuda":
torch.cuda.empty_cache(); torch.cuda.reset_peak_memory_stats()
self._log_gpu_memory("Início da Geração")
if mode == "image-to-video" and not start_image_filepath:
raise ValueError("A imagem de início é obrigatória para o modo image-to-video")
if mode == "video-to-video" and not input_video_filepath:
raise ValueError("O vídeo de entrada 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); print(f"[DEBUG] Seed usado: {used_seed}")
FPS = 24.0; MAX_NUM_FRAMES = 2570
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)))
print(f"[DEBUG] Frames alvo: {actual_num_frames} (dur={duration}s @ {FPS}fps)")
height_padded = ((height - 1) // 32 + 1) * 32
width_padded = ((width - 1) // 32 + 1) * 32
padding_values = calculate_padding(height, width, height_padded, width_padded)
print(f"[DEBUG] Dimensões: ({height},{width}) -> pad ({height_padded},{width_padded}); padding={padding_values}")
generator = torch.Generator(device=self.device).manual_seed(used_seed)
conditioning_items = []
if mode == "image-to-video":
start_tensor = self._prepare_conditioning_tensor(start_image_filepath, height, width, padding_values)
conditioning_items.append(ConditioningItem(start_tensor, 0, 1.0))
if middle_image_filepath and middle_frame_number is not None:
middle_tensor = self._prepare_conditioning_tensor(middle_image_filepath, height, width, padding_values)
safe_middle_frame = max(0, min(int(middle_frame_number), actual_num_frames - 1))
conditioning_items.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight)))
if end_image_filepath:
end_tensor = self._prepare_conditioning_tensor(end_image_filepath, height, width, padding_values)
last_frame_index = actual_num_frames - 1
conditioning_items.append(ConditioningItem(end_tensor, last_frame_index, float(end_image_weight)))
print(f"[DEBUG] Conditioning items: {len(conditioning_items)}")
# Sempre pedimos latentes (simples)
call_kwargs = {
"prompt": prompt,
"negative_prompt": negative_prompt,
"height": height_padded,
"width": width_padded,
"num_frames": actual_num_frames,
"frame_rate": int(FPS),
"generator": generator,
"output_type": "latent",
"conditioning_items": conditioning_items if conditioning_items else None,
"media_items": None,
"decode_timestep": self.config["decode_timestep"],
"decode_noise_scale": self.config["decode_noise_scale"],
"stochastic_sampling": self.config["stochastic_sampling"],
"image_cond_noise_scale": 0.01,
"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,
}
print(f"[DEBUG] output_type={call_kwargs['output_type']} skip_layer_strategy={call_kwargs['skip_layer_strategy']}")
if mode == "video-to-video":
media = load_media_file(
media_path=input_video_filepath,
height=height,
width=width,
max_frames=int(frames_to_use),
padding=padding_values,
).to(self.device)
call_kwargs["media_items"] = media
print(f"[DEBUG] media_items shape={tuple(media.shape)}")
latents = None
try:
ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
if improve_texture:
if not self.latent_upsampler:
raise ValueError("Upscaler espacial não carregado.")
# --- PASSO 1: GERAÇÃO DE LATENTES EM BAIXA RESOLUÇÃO ---
print("[DEBUG] Multi-escala: Iniciando Passo 1 (geração de latentes base).")
first_pass_args = self.config.get("first_pass", {}).copy()
first_pass_kwargs = call_kwargs.copy()
first_pass_kwargs.update({
"guidance_scale": float(guidance_scale),
"stg_scale": first_pass_args.get("stg_scale"),
"rescaling_scale": first_pass_args.get("rescaling_scale"),
"skip_block_list": first_pass_args.get("skip_block_list"),
})
schedule = first_pass_args.get("timesteps") or first_pass_args.get("guidance_timesteps")
if schedule:
first_pass_kwargs["timesteps"] = schedule
first_pass_kwargs["guidance_timesteps"] = schedule
downscale_factor = self.config.get("downscale_factor", 2)
original_height = first_pass_kwargs["height"]
original_width = first_pass_kwargs["width"]
divisor = 24
target_height_p1 = original_height // downscale_factor
height_p1 = round(target_height_p1 / divisor) * divisor
if height_p1 == 0: height_p1 = divisor
first_pass_kwargs["height"] = height_p1
target_width_p1 = original_width // downscale_factor
width_p1 = round(target_width_p1 / divisor) * divisor
if width_p1 == 0: width_p1 = divisor
first_pass_kwargs["width"] = width_p1
print(f"[DEBUG] Passo 1: Dimensões reduzidas e ajustadas para {height_p1}x{width_p1}")
with ctx:
first_pass_result = self.pipeline(**first_pass_kwargs)
latents_low_res = first_pass_result.images
log_tensor_info(latents_low_res, "Latentes (Passo 1)")
del first_pass_result
gc.collect()
if self.device == "cuda": torch.cuda.empty_cache()
# --- PASSO INTERMEDIÁRIO: UPSCALE DOS LATENTES ---
print("[DEBUG] Multi-escala: Fazendo upscale dos latentes com latent_upsampler.")
with ctx:
latents_high_res = self.latent_upsampler(latents_low_res)
log_tensor_info(latents_high_res, "Latentes (Pós-Upscale)")
del latents_low_res
gc.collect()
if self.device == "cuda": torch.cuda.empty_cache()
# --- PASSO 2: REFINAMENTO EM ALTA RESOLUÇÃO ---
print("[DEBUG] Multi-escala: Iniciando Passo 2 (refinamento em alta resolução).")
second_pass_args = self.config.get("second_pass", {}).copy()
second_pass_kwargs = call_kwargs.copy()
height_p2 = height_p1 * 2
width_p2 = width_p1 * 2
second_pass_kwargs["height"] = height_p2
second_pass_kwargs["width"] = width_p2
print(f"[DEBUG] Passo 2: Dimensões definidas para {height_p2}x{width_p2} para corresponder ao upscale.")
second_pass_kwargs.update({
"guidance_scale": float(guidance_scale),
"stg_scale": second_pass_args.get("stg_scale"),
"rescaling_scale": second_pass_args.get("rescaling_scale"),
"skip_block_list": second_pass_args.get("skip_block_list"),
})
schedule_p2 = second_pass_args.get("timesteps") or second_pass_args.get("guidance_timesteps")
if schedule_p2:
timesteps_para_refinamento = schedule_p2
print(f"[DEBUG] Passo 2: Usando {len(timesteps_para_refinamento)} timesteps pré-definidos do config para refinamento.")
else:
strength_p2 = second_pass_args.get("strength", second_pass_args.get("denoising_strength", 0.4))
num_steps_passo2_total = second_pass_args.get("num_inference_steps", 20)
self.pipeline.scheduler.set_timesteps(num_steps_passo2_total, device=self.device)
todos_os_timesteps_p2 = self.pipeline.scheduler.timesteps
ponto_de_corte = int(len(todos_os_timesteps_p2) * (1.0 - strength_p2))
timesteps_para_refinamento = todos_os_timesteps_p2[ponto_de_corte:]
print(f"[DEBUG] Passo 2: Calculando {len(timesteps_para_refinamento)} timesteps manuais (strength ≈ {strength_p2})")
second_pass_kwargs["timesteps"] = timesteps_para_refinamento
if "strength" in second_pass_kwargs: del second_pass_kwargs["strength"]
second_pass_kwargs["latents"] = latents_high_res
with ctx:
second_pass_result = self.pipeline(**second_pass_kwargs)
latents = second_pass_result.images
log_tensor_info(latents, "Latentes Finais (Passo 2)")
else:
# --- PASSO ÚNICO (SINGLE-PASS) ---
single_pass_kwargs = call_kwargs.copy()
first_pass_config = self.config.get("first_pass", {})
single_pass_kwargs.update(
{
"guidance_scale": float(guidance_scale),
"stg_scale": first_pass_config.get("stg_scale"),
"rescaling_scale": first_pass_config.get("rescaling_scale"),
"skip_block_list": first_pass_config.get("skip_block_list"),
}
)
schedule = first_pass_config.get("timesteps") or first_pass_config.get("guidance_timesteps")
if mode == "video-to-video":
schedule = [0.7]; print("[INFO] Modo video-to-video (etapa única): timesteps=[0.7]")
if isinstance(schedule, (list, tuple)) and len(schedule) > 0:
single_pass_kwargs["timesteps"] = schedule
single_pass_kwargs["guidance_timesteps"] = schedule
print(f"[DEBUG] Single-pass: timesteps_len={len(schedule) if schedule else 0}")
print("\n[INFO] Executando pipeline de etapa única...")
with ctx:
result = self.pipeline(**single_pass_kwargs)
latents = result.images
print(f"[DEBUG] Latentes (single-pass): shape={tuple(latents.shape)}")
# --- DECODIFICAÇÃO E CODIFICAÇÃO DE VÍDEO FINAL ---
latents_cpu = latents.detach().to("cpu", non_blocking=True)
if self.device == "cuda":
torch.cuda.empty_cache()
try:
torch.cuda.ipc_collect()
except Exception:
pass
lat_a, lat_b = self._dividir_latentes(latents_cpu)
lat_a1, lat_a2 = self._dividir_latentes(lat_a)
lat_b1, lat_b2 = self._dividir_latentes(lat_b)
latents_parts = [lat_a1, lat_a2, lat_b1, lat_b2]
temp_dir = tempfile.mkdtemp(prefix="ltxv_"); self._register_tmp_dir(temp_dir)
results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
partes_mp4 = []
par = 0
for part in latents_parts:
print(f"[DEBUG] Partição {par}: {tuple(part.shape)}")
par = par + 1
output_video_path = os.path.join(temp_dir, f"output_{used_seed}_{par}.mp4")
print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
pixel_tensor = vae_manager_singleton.decode(
part.to(self.device, non_blocking=True),
decode_timestep=float(self.config.get("decode_timestep", 0.05))
)
log_tensor_info(pixel_tensor, "Pixel tensor (VAE saída)")
print("[DEBUG] Codificando MP4 a partir do tensor de pixels...")
video_encode_tool_singleton.save_video_from_tensor(
pixel_tensor,
output_video_path,
fps=call_kwargs["frame_rate"],
progress_callback=progress_callback
)
candidate = os.path.join(results_dir, f"output_par_{par}.mp4")
try:
shutil.move(output_video_path, candidate)
final_output_path = candidate
print(f"[DEBUG] MP4 parte {par} movido para {final_output_path}")
partes_mp4.append(final_output_path)
except Exception as e:
print(f"[DEBUG] Falha no move; usando tmp como final: {e}")
final_concat = os.path.join(results_dir, f"concat_fim_{used_seed}.mp4")
self._concat_mp4s_no_reencode(partes_mp4, final_concat)
self._log_gpu_memory("Fim da Geração")
return final_concat, used_seed
except Exception as e:
print("[DEBUG] EXCEÇÃO NA GERAÇÃO:")
print("".join(traceback.format_exception(type(e), e, e.__traceback__)))
raise
# ltx_server.py
def generate(
self,
prompt,
negative_prompt,
mode="text-to-video",
start_image_filepath=None,
middle_image_filepath=None,
middle_frame_number=None,
middle_image_weight=1.0,
end_image_filepath=None,
end_image_weight=1.0,
input_video_filepath=None,
height=512,
width=704,
duration=2.0,
frames_to_use=9,
seed=42,
randomize_seed=True,
guidance_scale=3.0, # Valor de referência/fallback
improve_texture=True,
progress_callback=None,
external_decode=True,
):
t_all = time.perf_counter()
print(f"[DEBUG] generate() begin mode={mode} external_decode={external_decode} improve_texture={improve_texture}")
if self.device == "cuda":
torch.cuda.empty_cache(); torch.cuda.reset_peak_memory_stats()
self._log_gpu_memory("Início da Geração")
if mode == "image-to-video" and not start_image_filepath:
raise ValueError("A imagem de início é obrigatória para o modo image-to-video")
if mode == "video-to-video" and not input_video_filepath:
raise ValueError("O vídeo de entrada 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); print(f"[DEBUG] Seed usado: {used_seed}")
FPS = 24.0; MAX_NUM_FRAMES = 2570
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)))
print(f"[DEBUG] Frames alvo: {actual_num_frames} (dur={duration}s @ {FPS}fps)")
height_padded = ((height - 1) // 32 + 1) * 32
width_padded = ((width - 1) // 32 + 1) * 32
padding_values = calculate_padding(height, width, height_padded, width_padded)
print(f"[DEBUG] Dimensões: ({height},{width}) -> pad ({height_padded},{width_padded}); padding={padding_values}")
generator = torch.Generator(device=self.device).manual_seed(used_seed)
conditioning_items = []
if mode == "image-to-video":
start_tensor = self._prepare_conditioning_tensor(start_image_filepath, height, width, padding_values)
conditioning_items.append(ConditioningItem(start_tensor, 0, 1.0))
if middle_image_filepath and middle_frame_number is not None:
middle_tensor = self._prepare_conditioning_tensor(middle_image_filepath, height, width, padding_values)
safe_middle_frame = max(0, min(int(middle_frame_number), actual_num_frames - 1))
conditioning_items.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight)))
if end_image_filepath:
end_tensor = self._prepare_conditioning_tensor(end_image_filepath, height, width, padding_values)
last_frame_index = actual_num_frames - 1
conditioning_items.append(ConditioningItem(end_tensor, last_frame_index, float(end_image_weight)))
print(f"[DEBUG] Conditioning items: {len(conditioning_items)}")
call_kwargs = {
"prompt": prompt,
"negative_prompt": negative_prompt,
"height": height_padded,
"width": width_padded,
"num_frames": actual_num_frames,
"frame_rate": int(FPS),
"generator": generator,
"output_type": "latent",
"conditioning_items": conditioning_items if conditioning_items else None,
"media_items": None,
"decode_timestep": self.config.get("decode_timestep"),
"decode_noise_scale": self.config.get("decode_noise_scale"),
"stochastic_sampling": self.config.get("stochastic_sampling"),
"image_cond_noise_scale": self.config.get("image_cond_noise_scale", 0.01),
"is_video": True,
"vae_per_channel_normalize": self.config.get("vae_per_channel_normalize", True),
"mixed_precision": (self.config.get("precision") == "mixed_precision"),
"offload_to_cpu": False,
"enhance_prompt": False,
"skip_layer_strategy": SkipLayerStrategy[self.config.get("stg_mode", "AttentionValues")],
}
print(f"[DEBUG] output_type={call_kwargs['output_type']} skip_layer_strategy={call_kwargs['skip_layer_strategy']}")
if mode == "video-to-video":
media = load_media_file(
media_path=input_video_filepath,
height=height,
width=width,
max_frames=int(frames_to_use),
padding=padding_values,
).to(self.device)
call_kwargs["media_items"] = media
print(f"[DEBUG] media_items shape={tuple(media.shape)}")
latents = None
try:
ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
if improve_texture:
if not self.latent_upsampler:
raise ValueError("Upscaler espacial não carregado.")
# --- PASSO 1: GERAÇÃO DE LATENTES EM BAIXA RESOLUÇÃO ---
print("[DEBUG] Multi-escala: Iniciando Passo 1 (geração de latentes base).")
first_pass_args = self.config.get("first_pass", {}).copy()
first_pass_kwargs = call_kwargs.copy()
first_pass_kwargs.update({
"guidance_scale": first_pass_args.get("guidance_scale", guidance_scale),
"stg_scale": first_pass_args.get("stg_scale"),
"rescaling_scale": first_pass_args.get("rescaling_scale"),
"skip_block_list": first_pass_args.get("skip_block_list"),
"guidance_timesteps": first_pass_args.get("guidance_timesteps"),
"timesteps": first_pass_args.get("timesteps")
})
print(f"[DEBUG] Passo 1: Parâmetros do config carregados: guidance_scale={first_pass_kwargs['guidance_scale']}, stg_scale={first_pass_kwargs['stg_scale']}")
downscale_factor = self.config.get("downscale_factor", 2)
original_height = first_pass_kwargs["height"]
original_width = first_pass_kwargs["width"]
divisor = 24
target_height_p1 = original_height // downscale_factor
height_p1 = round(target_height_p1 / divisor) * divisor
if height_p1 == 0: height_p1 = divisor
first_pass_kwargs["height"] = height_p1
target_width_p1 = original_width // downscale_factor
width_p1 = round(target_width_p1 / divisor) * divisor
if width_p1 == 0: width_p1 = divisor
first_pass_kwargs["width"] = width_p1
print(f"[DEBUG] Passo 1: Dimensões reduzidas e ajustadas para {height_p1}x{width_p1}")
with ctx:
first_pass_result = self.pipeline(**first_pass_kwargs)
latents_low_res = first_pass_result.images
log_tensor_info(latents_low_res, "Latentes (Passo 1)")
del first_pass_result
gc.collect()
if self.device == "cuda": torch.cuda.empty_cache()
# --- PASSO INTERMEDIÁRIO: UPSCALE DOS LATENTES ---
print("[DEBUG] Multi-escala: Fazendo upscale dos latentes com latent_upsampler.")
with ctx:
latents_high_res = self.latent_upsampler(latents_low_res)
log_tensor_info(latents_high_res, "Latentes (Pós-Upscale)")
del latents_low_res
gc.collect()
if self.device == "cuda": torch.cuda.empty_cache()
# --- PASSO 2: REFINAMENTO EM ALTA RESOLUÇÃO ---
print("[DEBUG] Multi-escala: Iniciando Passo 2 (refinamento em alta resolução).")
second_pass_args = self.config.get("second_pass", {}).copy()
second_pass_kwargs = call_kwargs.copy()
second_pass_kwargs.update({
"guidance_scale": second_pass_args.get("guidance_scale", guidance_scale),
"stg_scale": second_pass_args.get("stg_scale"),
"rescaling_scale": second_pass_args.get("rescaling_scale"),
"skip_block_list": second_pass_args.get("skip_block_list"),
"guidance_timesteps": second_pass_args.get("guidance_timesteps"),
"timesteps": second_pass_args.get("timesteps")
})
print(f"[DEBUG] Passo 2: Parâmetros do config carregados: guidance_scale={second_pass_kwargs['guidance_scale']}, stg_scale={second_pass_kwargs['stg_scale']}")
height_p2 = height_p1 * 2
width_p2 = width_p1 * 2
second_pass_kwargs["height"] = height_p2
second_pass_kwargs["width"] = width_p2
print(f"[DEBUG] Passo 2: Dimensões definidas para {height_p2}x{width_p2}")
second_pass_kwargs["latents"] = latents_high_res
with ctx:
second_pass_result = self.pipeline(**second_pass_kwargs)
latents = second_pass_result.images
log_tensor_info(latents, "Latentes Finais (Passo 2)")
else:
# --- PASSO ÚNICO (SINGLE-PASS) ---
single_pass_kwargs = call_kwargs.copy()
single_pass_kwargs.update({
"guidance_scale": self.config.get("guidance_scale", guidance_scale),
"stg_scale": self.config.get("stg_scale"),
"rescaling_scale": self.config.get("rescaling_scale"),
"skip_block_list": self.config.get("skip_block_list"),
"guidance_timesteps": self.config.get("guidance_timesteps"),
"timesteps": self.config.get("timesteps"),
"num_inference_steps": self.config.get("num_inference_steps", 20)
})
print("\n[INFO] Executando pipeline de etapa única...")
with ctx:
result = self.pipeline(**single_pass_kwargs)
latents = result.images
print(f"[DEBUG] Latentes (single-pass): shape={tuple(latents.shape)}")
# --- DECODIFICAÇÃO E CODIFICAÇÃO DE VÍDEO FINAL ---
latents_cpu = latents.detach().to("cpu", non_blocking=True)
if self.device == "cuda":
torch.cuda.empty_cache()
try: torch.cuda.ipc_collect()
except Exception: pass
lat_a, lat_b = self._dividir_latentes(latents_cpu)
if lat_a is not None:
lat_a1, lat_a2 = self._dividir_latentes(lat_a)
else:
lat_a1, lat_a2 = None, None
if lat_b is not None:
lat_b1, lat_b2 = self._dividir_latentes(lat_b)
else:
lat_b1, lat_b2 = None, None
latents_parts = [p for p in [lat_a1, lat_a2, lat_b1, lat_b2] if p is not None]
if not latents_parts:
latents_parts = [latents_cpu]
temp_dir = tempfile.mkdtemp(prefix="ltxv_"); self._register_tmp_dir(temp_dir)
results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
partes_mp4 = []
par = 0
for part in latents_parts:
par += 1
print(f"[DEBUG] Partição {par}: {tuple(part.shape)}")
output_video_path = os.path.join(temp_dir, f"output_{used_seed}_{par}.mp4")
print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
pixel_tensor = vae_manager_singleton.decode(
part.to(self.device, non_blocking=True),
decode_timestep=float(self.config.get("decode_timestep", 0.05))
)
log_tensor_info(pixel_tensor, "Pixel tensor (VAE saída)")
print("[DEBUG] Codificando MP4 a partir do tensor de pixels...")
video_encode_tool_singleton.save_video_from_tensor(
pixel_tensor,
output_video_path,
fps=call_kwargs["frame_rate"],
progress_callback=progress_callback
)
candidate = os.path.join(results_dir, f"output_par_{par}.mp4")
try:
shutil.move(output_video_path, candidate)
print(f"[DEBUG] MP4 parte {par} movido para {candidate}")
partes_mp4.append(candidate)
except Exception as e:
print(f"[DEBUG] Falha no move; usando tmp como final: {e}")
partes_mp4.append(output_video_path)
final_concat = os.path.join(results_dir, f"concat_fim_{used_seed}.mp4")
if partes_mp4:
if len(partes_mp4) == 1:
shutil.move(partes_mp4[0], final_concat)
print(f"[DEBUG] Apenas uma parte, movida para {final_concat}")
else:
self._concat_mp4s_no_reencode(partes_mp4, final_concat)
else:
print("[WARN] Nenhuma parte de vídeo foi gerada para concatenar.")
return None, used_seed
self._log_gpu_memory("Fim da Geração")
return final_concat, used_seed
except Exception as e:
print("[DEBUG] EXCEÇÃO NA GERAÇÃO:")
print("".join(traceback.format_exception(type(e), e, e.__traceback__)))
raise
finally:
try:
del latents, latents_low_res, latents_high_res, second_pass_result, first_pass_result, result
except NameError:
pass
except Exception as e:
print(f"[DEBUG] Erro na limpeza de variáveis: {e}")
gc.collect()
if self.device == "cuda":
try:
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
except Exception as e:
print(f"[DEBUG] Limpeza GPU no finally falhou: {e}")
try:
self.finalize(keep_paths=[])
except Exception as e:
print(f"[DEBUG] finalize() no finally falhou: {e}")
print("Criando instância do VideoService. O carregamento do modelo começará agora...")
video_generation_service = VideoService() |