Spaces:
Paused
Paused
Update api/ltx/ltx_aduc_pipeline.py
Browse files- api/ltx/ltx_aduc_pipeline.py +155 -334
api/ltx/ltx_aduc_pipeline.py
CHANGED
|
@@ -1,375 +1,196 @@
|
|
| 1 |
# FILE: api/ltx/ltx_aduc_pipeline.py
|
| 2 |
-
# DESCRIPTION:
|
| 3 |
-
#
|
| 4 |
-
#
|
| 5 |
|
| 6 |
-
import gc
|
| 7 |
-
import json
|
| 8 |
import logging
|
| 9 |
-
import os
|
| 10 |
-
import shutil
|
| 11 |
-
import sys
|
| 12 |
-
import tempfile
|
| 13 |
import time
|
| 14 |
-
from pathlib import Path
|
| 15 |
-
from typing import Dict, List, Optional, Tuple, Union
|
| 16 |
-
|
| 17 |
import torch
|
| 18 |
-
import
|
| 19 |
-
import
|
| 20 |
-
from
|
| 21 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
# ==============================================================================
|
| 24 |
-
# ---
|
| 25 |
# ==============================================================================
|
| 26 |
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
|
|
|
|
|
|
| 42 |
|
| 43 |
-
# --- Módulos da nossa Arquitetura ---
|
| 44 |
-
try:
|
| 45 |
-
from managers.gpu_manager import gpu_manager
|
| 46 |
-
from api.ltx.vae_aduc_pipeline import vae_ltx_aduc_pipeline
|
| 47 |
-
from tools.video_encode_tool import video_encode_tool_singleton
|
| 48 |
-
from api.ltx.ltx_utils import build_ltx_pipeline_on_cpu, seed_everything
|
| 49 |
-
from api.ltx.ltx_aduc_manager import LatentConditioningItem, ltx_aduc_manager
|
| 50 |
-
from utils.debug_utils import log_function_io
|
| 51 |
-
except ImportError as e:
|
| 52 |
-
logging.critical(f"A crucial import from the local API/architecture failed. Error: {e}", exc_info=True)
|
| 53 |
-
sys.exit(1)
|
| 54 |
|
| 55 |
# ==============================================================================
|
| 56 |
-
# --- CLASSE
|
| 57 |
# ==============================================================================
|
| 58 |
|
| 59 |
-
|
| 60 |
class LtxAducPipeline:
|
| 61 |
"""
|
| 62 |
-
|
| 63 |
-
|
| 64 |
"""
|
| 65 |
-
|
| 66 |
-
@log_function_io
|
| 67 |
def __init__(self):
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
target_main_device_str = str(gpu_manager.get_ltx_device())
|
| 73 |
-
target_vae_device_str = str(gpu_manager.get_ltx_vae_device())
|
| 74 |
-
logging.info(f"LTX allocated to devices: Main='{target_main_device_str}', VAE='{target_vae_device_str}'")
|
| 75 |
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
self.main_device = torch.device("cpu")
|
| 80 |
-
self.vae_device = torch.device("cpu")
|
| 81 |
-
self.move_to_device(main_device_str=target_main_device_str, vae_device_str=target_vae_device_str)
|
| 82 |
-
|
| 83 |
-
self._apply_precision_policy()
|
| 84 |
-
logging.info(f"VideoService ready. Startup time: {time.time() - t0:.2f}s")
|
| 85 |
-
|
| 86 |
-
def _load_config(self) -> Dict:
|
| 87 |
-
"""Loads the YAML configuration file."""
|
| 88 |
-
config_path = LTX_VIDEO_REPO_DIR / "configs" / "ltxv-13b-0.9.8-distilled-fp8.yaml"
|
| 89 |
-
with open(config_path, "r") as file:
|
| 90 |
-
return yaml.safe_load(file)
|
| 91 |
-
|
| 92 |
-
def _resolve_model_paths_from_cache(self):
|
| 93 |
-
"""Finds the absolute paths to model files in the cache and updates the in-memory config."""
|
| 94 |
-
logging.info("Resolving model paths from Hugging Face cache...")
|
| 95 |
-
cache_dir = os.environ.get("HF_HOME")
|
| 96 |
-
try:
|
| 97 |
-
main_ckpt_path = hf_hub_download(repo_id=LTX_REPO_ID, filename=self.config["checkpoint_path"], cache_dir=cache_dir)
|
| 98 |
-
self.config["checkpoint_path"] = main_ckpt_path
|
| 99 |
-
if self.config.get("spatial_upscaler_model_path"):
|
| 100 |
-
upscaler_path = hf_hub_download(repo_id=LTX_REPO_ID, filename=self.config["spatial_upscaler_model_path"], cache_dir=cache_dir)
|
| 101 |
-
self.config["spatial_upscaler_model_path"] = upscaler_path
|
| 102 |
-
except Exception as e:
|
| 103 |
-
logging.critical(f"Failed to resolve model paths. Ensure setup.py ran correctly. Error: {e}", exc_info=True)
|
| 104 |
-
sys.exit(1)
|
| 105 |
-
|
| 106 |
-
@log_function_io
|
| 107 |
-
def move_to_device(self, main_device_str: str, vae_device_str: str):
|
| 108 |
-
"""Moves pipeline components to their designated target devices."""
|
| 109 |
-
target_main_device = torch.device(main_device_str)
|
| 110 |
-
target_vae_device = torch.device(vae_device_str)
|
| 111 |
-
self.main_device = target_main_device
|
| 112 |
-
self.vae_device = target_vae_device
|
| 113 |
-
self.pipeline.to(self.main_device)
|
| 114 |
-
self.pipeline.vae.to(self.vae_device)
|
| 115 |
-
if self.latent_upsampler: self.latent_upsampler.to(self.main_device)
|
| 116 |
-
logging.info("LTX models successfully moved to target devices.")
|
| 117 |
-
|
| 118 |
-
def move_to_cpu(self):
|
| 119 |
-
"""Moves all LTX components to CPU to free VRAM for other services."""
|
| 120 |
-
self.move_to_device(main_device_str="cpu", vae_device_str="cpu")
|
| 121 |
-
if torch.cuda.is_available(): torch.cuda.empty_cache()
|
| 122 |
-
|
| 123 |
-
def finalize(self):
|
| 124 |
-
"""Cleans up GPU memory after a generation task."""
|
| 125 |
-
gc.collect()
|
| 126 |
-
if torch.cuda.is_available(): torch.cuda.empty_cache()
|
| 127 |
-
try: torch.cuda.ipc_collect();
|
| 128 |
-
except Exception: pass
|
| 129 |
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
|
| 134 |
-
|
| 135 |
-
def generate_low_resolution(
|
| 136 |
self,
|
| 137 |
prompt_list: List[str],
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
|
|
|
|
|
|
|
|
|
| 141 |
"""
|
| 142 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
"""
|
| 144 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
used_seed = self._get_random_seed()
|
| 146 |
-
|
| 147 |
-
logging.info(f"Using randomly generated seed: {used_seed}")
|
| 148 |
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
is_narrative = len(prompt_list) > 1
|
| 152 |
num_chunks = len(prompt_list)
|
| 153 |
-
total_frames = self.
|
| 154 |
-
|
| 155 |
-
overlap_frames = 9 if
|
| 156 |
-
|
| 157 |
-
initial_conditions = []
|
| 158 |
-
if initial_media_items:
|
| 159 |
-
logging.info("Preparing initial conditioning items from raw media list...")
|
| 160 |
-
initial_conditions = vae_ltx_aduc_pipeline.generate_conditioning_items(
|
| 161 |
-
media_items=[item[0] for item in initial_media_items],
|
| 162 |
-
target_frames=[item[1] for item in initial_media_items],
|
| 163 |
-
strengths=[item[2] for item in initial_media_items],
|
| 164 |
-
target_resolution=(kwargs['height'], kwargs['width'])
|
| 165 |
-
)
|
| 166 |
-
|
| 167 |
-
temp_latent_paths = []
|
| 168 |
-
overlap_condition_item: Optional[LatentConditioningItem] = None
|
| 169 |
-
|
| 170 |
-
try:
|
| 171 |
-
for i, chunk_prompt in enumerate(prompt_list):
|
| 172 |
-
logging.info(f"Processing scene {i+1}/{num_chunks}: '{chunk_prompt[:50]}...'")
|
| 173 |
-
|
| 174 |
-
if i < num_chunks - 1:
|
| 175 |
-
current_frames_base = frames_per_chunk
|
| 176 |
-
else:
|
| 177 |
-
processed_frames_base = (num_chunks - 1) * frames_per_chunk
|
| 178 |
-
current_frames_base = total_frames - processed_frames_base
|
| 179 |
-
|
| 180 |
-
current_frames = current_frames_base + (overlap_frames if i > 0 else 0)
|
| 181 |
-
current_frames = self._align(current_frames, alignment_rule='n*8+1')
|
| 182 |
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
chunk_latents = self._generate_single_chunk_low(
|
| 187 |
-
prompt=chunk_prompt, num_frames=current_frames, seed=used_seed + i,
|
| 188 |
-
conditioning_items=current_conditions, **kwargs
|
| 189 |
-
)
|
| 190 |
-
if chunk_latents is None: raise RuntimeError(f"Failed to generate latents for scene {i+1}.")
|
| 191 |
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
)
|
| 199 |
-
|
| 200 |
-
if i > 0: chunk_latents = chunk_latents[:, :, overlap_frames:, :, :]
|
| 201 |
-
|
| 202 |
-
chunk_path = RESULTS_DIR / f"temp_chunk_{i}_{used_seed}.pt"
|
| 203 |
-
torch.save(chunk_latents.cpu(), chunk_path)
|
| 204 |
-
temp_latent_paths.append(chunk_path)
|
| 205 |
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
return video_path, latents_path, used_seed
|
| 212 |
-
except Exception as e:
|
| 213 |
-
logging.error(f"Error during unified generation: {e}", exc_info=True)
|
| 214 |
-
return None, None, None
|
| 215 |
-
finally:
|
| 216 |
-
for path in temp_latent_paths:
|
| 217 |
-
if path.exists(): path.unlink()
|
| 218 |
-
self.finalize()
|
| 219 |
-
|
| 220 |
-
# ==========================================================================
|
| 221 |
-
# --- UNIDADES DE TRABALHO E HELPERS INTERNOS ---
|
| 222 |
-
# ==========================================================================
|
| 223 |
-
|
| 224 |
-
def _log_conditioning_items(self, items: List[LatentConditioningItem]):
|
| 225 |
-
"""
|
| 226 |
-
Logs detailed information about a list of ConditioningItem objects.
|
| 227 |
-
This is a dedicated debug helper function.
|
| 228 |
-
"""
|
| 229 |
-
# Só imprime o log se o nível de logging for DEBUG
|
| 230 |
-
if logging.getLogger().isEnabledFor(logging.INFO):
|
| 231 |
-
log_str = ["\n" + "="*25 + " INFO: Conditioning Items " + "="*25]
|
| 232 |
-
if not items:
|
| 233 |
-
log_str.append(" -> Lista de conditioning_items está vazia.")
|
| 234 |
-
else:
|
| 235 |
-
for i, item in enumerate(items):
|
| 236 |
-
if hasattr(item, 'media_item') and isinstance(item.media_item, torch.Tensor):
|
| 237 |
-
t = item.media_item
|
| 238 |
-
log_str.append(
|
| 239 |
-
f" -> Item [{i}]: "
|
| 240 |
-
f"Tensor(shape={list(t.shape)}, "
|
| 241 |
-
f"device='{t.device}', "
|
| 242 |
-
f"dtype={t.dtype}), "
|
| 243 |
-
f"Target Frame = {item.media_frame_number}, "
|
| 244 |
-
f"Strength = {item.conditioning_strength:.2f}"
|
| 245 |
-
)
|
| 246 |
-
else:
|
| 247 |
-
log_str.append(f" -> Item [{i}]: Não contém um tensor válido.")
|
| 248 |
-
log_str.append("="*75 + "\n")
|
| 249 |
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
if kwargs.get("ltx_configs_override"):
|
| 267 |
-
self._apply_ui_overrides(first_pass_config, kwargs.get("ltx_configs_override"))
|
| 268 |
-
|
| 269 |
-
# 3. Monta o dicionário de argumentos SEM conditioning_items primeiro
|
| 270 |
-
pipeline_kwargs = {
|
| 271 |
-
"prompt": kwargs['prompt'],
|
| 272 |
-
"negative_prompt": kwargs['negative_prompt'],
|
| 273 |
-
"height": downscaled_height,
|
| 274 |
-
"width": downscaled_width,
|
| 275 |
-
"num_frames": kwargs['num_frames'],
|
| 276 |
-
"frame_rate": int(DEFAULT_FPS),
|
| 277 |
-
"generator": torch.Generator(device=self.main_device).manual_seed(kwargs['seed']),
|
| 278 |
-
"output_type": "latent",
|
| 279 |
-
#"conditioning_items": conditioning_items if conditioning_items else None,
|
| 280 |
-
"media_items": None,
|
| 281 |
-
"decode_timestep": self.config["decode_timestep"],
|
| 282 |
-
"decode_noise_scale": self.config["decode_noise_scale"],
|
| 283 |
-
"stochastic_sampling": self.config["stochastic_sampling"],
|
| 284 |
-
"image_cond_noise_scale": 0.01,
|
| 285 |
-
"is_video": True,
|
| 286 |
-
"vae_per_channel_normalize": True,
|
| 287 |
-
"mixed_precision": (self.config["precision"] == "mixed_precision"),
|
| 288 |
-
"offload_to_cpu": False,
|
| 289 |
-
"enhance_prompt": False,
|
| 290 |
-
#"skip_layer_strategy": SkipLayerStrategy.AttentionValues,
|
| 291 |
-
**first_pass_config
|
| 292 |
-
}
|
| 293 |
-
|
| 294 |
-
# --- Bloco de Logging para Depuração ---
|
| 295 |
-
# 4. Loga os argumentos do pipeline (sem os tensores de condição)
|
| 296 |
-
logging.info(f"\n[Info] Pipeline Arguments (BASE):\n {json.dumps(pipeline_kwargs, indent=2, default=str)}\n")
|
| 297 |
-
|
| 298 |
-
# Loga os conditioning_items separadamente com a nossa função helper
|
| 299 |
-
conditioning_items_list = kwargs.get('conditioning_items')
|
| 300 |
-
self._log_conditioning_items(conditioning_items_list)
|
| 301 |
-
# --- Fim do Bloco de Logging ---
|
| 302 |
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 309 |
|
| 310 |
-
|
| 311 |
-
|
| 312 |
-
@log_function_io
|
| 313 |
-
def _finalize_generation(self, final_latents: torch.Tensor, base_filename: str, seed: int) -> Tuple[str, str]:
|
| 314 |
-
"""Consolidates latents, decodes them to video, and saves final artifacts."""
|
| 315 |
-
logging.info("Finalizing generation: decoding latents to video.")
|
| 316 |
-
final_latents_path = RESULTS_DIR / f"latents_{base_filename}_{seed}.pt"
|
| 317 |
-
torch.save(final_latents, final_latents_path)
|
| 318 |
-
logging.info(f"Final latents saved to: {final_latents_path}")
|
| 319 |
|
| 320 |
-
|
| 321 |
-
final_latents, decode_timestep=float(self.config.get("decode_timestep", 0.05))
|
| 322 |
-
)
|
| 323 |
-
video_path = self._save_and_log_video(pixel_tensor, f"{base_filename}_{seed}")
|
| 324 |
-
return str(video_path), str(final_latents_path)
|
| 325 |
-
|
| 326 |
-
def _apply_ui_overrides(self, config_dict: Dict, overrides: Dict):
|
| 327 |
-
"""Applies advanced settings from the UI to a config dictionary."""
|
| 328 |
-
# Override step counts
|
| 329 |
-
for key in ["num_inference_steps", "skip_initial_inference_steps", "skip_final_inference_steps"]:
|
| 330 |
-
ui_value = overrides.get(key)
|
| 331 |
-
if ui_value and ui_value > 0:
|
| 332 |
-
config_dict[key] = ui_value
|
| 333 |
-
logging.info(f"Override: '{key}' set to {ui_value} by UI.")
|
| 334 |
|
| 335 |
-
|
| 336 |
-
with tempfile.TemporaryDirectory() as temp_dir:
|
| 337 |
-
temp_path = os.path.join(temp_dir, f"{base_filename}.mp4")
|
| 338 |
-
video_encode_tool_singleton.save_video_from_tensor(pixel_tensor, temp_path, fps=DEFAULT_FPS)
|
| 339 |
-
final_path = RESULTS_DIR / f"{base_filename}.mp4"
|
| 340 |
-
shutil.move(temp_path, final_path)
|
| 341 |
-
logging.info(f"Video saved successfully to: {final_path}")
|
| 342 |
-
return final_path
|
| 343 |
-
|
| 344 |
-
def _apply_precision_policy(self):
|
| 345 |
-
precision = str(self.config.get("precision", "bfloat16")).lower()
|
| 346 |
-
if precision in ["float8_e4m3fn", "bfloat16"]: self.runtime_autocast_dtype = torch.bfloat16
|
| 347 |
-
elif precision == "mixed_precision": self.runtime_autocast_dtype = torch.float16
|
| 348 |
-
else: self.runtime_autocast_dtype = torch.float32
|
| 349 |
-
logging.info(f"Runtime precision policy set for autocast: {self.runtime_autocast_dtype}")
|
| 350 |
-
|
| 351 |
-
def _align(self, dim: int, alignment: int = FRAMES_ALIGNMENT, alignment_rule: str = 'default') -> int:
|
| 352 |
-
"""Aligns a dimension to the nearest multiple of `alignment`."""
|
| 353 |
-
if alignment_rule == 'n*8+1':
|
| 354 |
-
return ((dim - 1) // alignment) * alignment + 1
|
| 355 |
-
return ((dim - 1) // alignment + 1) * alignment
|
| 356 |
-
|
| 357 |
-
def _calculate_aligned_frames(self, duration_s: float, min_frames: int = 1) -> int:
|
| 358 |
-
num_frames = int(round(duration_s * DEFAULT_FPS))
|
| 359 |
-
# Para a duração total, sempre arredondamos para cima para o múltiplo de 8 mais próximo
|
| 360 |
-
aligned_frames = self._align(num_frames, alignment=FRAMES_ALIGNMENT)
|
| 361 |
-
return max(aligned_frames, min_frames)
|
| 362 |
-
|
| 363 |
-
def _get_random_seed(self) -> int:
|
| 364 |
-
"""Always generates and returns a new random seed."""
|
| 365 |
-
return random.randint(0, 2**32 - 1)
|
| 366 |
|
| 367 |
-
#
|
| 368 |
-
# --- INSTANCIAÇÃO SINGLETON ---
|
| 369 |
-
# ==============================================================================
|
| 370 |
try:
|
| 371 |
ltx_aduc_pipeline = LtxAducPipeline()
|
| 372 |
-
logging.info("Global VideoService orchestrator instance created successfully.")
|
| 373 |
except Exception as e:
|
| 374 |
-
logging.critical(
|
| 375 |
-
|
|
|
|
| 1 |
# FILE: api/ltx/ltx_aduc_pipeline.py
|
| 2 |
+
# DESCRIPTION: A high-level client for submitting LTX video generation jobs to the pool manager.
|
| 3 |
+
# Its sole responsibility is to orchestrate the generation of a final LATENT tensor from prompts
|
| 4 |
+
# and initial conditions, without handling pixel decoding.
|
| 5 |
|
|
|
|
|
|
|
| 6 |
import logging
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
import time
|
|
|
|
|
|
|
|
|
|
| 8 |
import torch
|
| 9 |
+
import random
|
| 10 |
+
import json
|
| 11 |
+
from typing import List, Optional, Tuple, Union, Dict
|
| 12 |
+
|
| 13 |
+
# O cliente importa o MANAGER para submeter trabalhos
|
| 14 |
+
from api.ltx.ltx_aduc_manager import ltx_aduc_manager
|
| 15 |
+
|
| 16 |
+
# O cliente precisa da definição de LatentConditioningItem para os seus inputs
|
| 17 |
+
from api.ltx.vae_aduc_pipeline import LatentConditioningItem
|
| 18 |
+
|
| 19 |
+
# Importa o tipo do pipeline para anotações claras nas funções de trabalho
|
| 20 |
+
from ltx_video.pipelines.pipeline_ltx_video import LTXVideoPipeline
|
| 21 |
|
| 22 |
# ==============================================================================
|
| 23 |
+
# --- FUNÇÕES DE TRABALHO (Jobs a serem executados no Pool LTX) ---
|
| 24 |
# ==============================================================================
|
| 25 |
|
| 26 |
+
def _job_generate_latent_chunk(
|
| 27 |
+
pipeline: LTXVideoPipeline,
|
| 28 |
+
prompt: str,
|
| 29 |
+
negative_prompt: str,
|
| 30 |
+
height: int,
|
| 31 |
+
width: int,
|
| 32 |
+
num_frames: int,
|
| 33 |
+
seed: int,
|
| 34 |
+
conditioning_items: Optional[List[LatentConditioningItem]],
|
| 35 |
+
ltx_configs: Dict
|
| 36 |
+
) -> torch.Tensor:
|
| 37 |
+
"""
|
| 38 |
+
Função de trabalho que executa a geração de um único chunk (cena) de vídeo latente.
|
| 39 |
+
Esta função é executada DENTRO de um LTXMainWorker.
|
| 40 |
+
"""
|
| 41 |
+
generator = torch.Generator(device=pipeline.device).manual_seed(seed)
|
| 42 |
+
|
| 43 |
+
# Monta os argumentos para a chamada do pipeline
|
| 44 |
+
pipeline_kwargs = {
|
| 45 |
+
"prompt": prompt,
|
| 46 |
+
"negative_prompt": negative_prompt,
|
| 47 |
+
"height": height,
|
| 48 |
+
"width": width,
|
| 49 |
+
"num_frames": num_frames,
|
| 50 |
+
"frame_rate": 24, # Padrão, pode ser parametrizado se necessário
|
| 51 |
+
"generator": generator,
|
| 52 |
+
"output_type": "latent", # Ponto chave: sempre pedimos latentes
|
| 53 |
+
"conditioning_items": conditioning_items if conditioning_items else None,
|
| 54 |
+
**ltx_configs # Aplica configurações avançadas (guidance, steps, etc.)
|
| 55 |
+
}
|
| 56 |
+
|
| 57 |
+
logging.info(f"[LTX Job] Gerando chunk com {num_frames} frames para o prompt: '{prompt[:50]}...'")
|
| 58 |
+
|
| 59 |
+
# O pipeline já está na GPU correta dentro do worker
|
| 60 |
+
with torch.autocast(device_type=pipeline.device.type, dtype=torch.bfloat16):
|
| 61 |
+
latents_raw = pipeline(**pipeline_kwargs).images
|
| 62 |
|
| 63 |
+
# Retorna o tensor latente na CPU para liberar VRAM do worker para o próximo job
|
| 64 |
+
return latents_raw.cpu()
|
| 65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
# ==============================================================================
|
| 68 |
+
# --- A CLASSE CLIENTE (Interface Pública para Geração de Vídeo Latente) ---
|
| 69 |
# ==============================================================================
|
| 70 |
|
|
|
|
| 71 |
class LtxAducPipeline:
|
| 72 |
"""
|
| 73 |
+
Cliente de alto nível para orquestrar a geração de vídeo latente.
|
| 74 |
+
Submete trabalhos de geração de chunks de vídeo ao LTXAducManager.
|
| 75 |
"""
|
|
|
|
|
|
|
| 76 |
def __init__(self):
|
| 77 |
+
logging.info("✅ LTX ADUC Pipeline (Client) initialized and ready to submit jobs.")
|
| 78 |
+
# O __init__ é limpo, sem carregar modelos.
|
| 79 |
+
self.FRAMES_ALIGNMENT = 8
|
| 80 |
+
pass
|
|
|
|
|
|
|
|
|
|
| 81 |
|
| 82 |
+
def _get_random_seed(self) -> int:
|
| 83 |
+
"""Sempre gera e retorna uma nova semente aleatória."""
|
| 84 |
+
return random.randint(0, 2**32 - 1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
+
def _align(self, dim: int, alignment: int = 8) -> int:
|
| 87 |
+
"""Alinha uma dimensão para o múltiplo mais próximo."""
|
| 88 |
+
return ((dim + alignment - 1) // alignment) * alignment
|
| 89 |
|
| 90 |
+
def __call__(
|
|
|
|
| 91 |
self,
|
| 92 |
prompt_list: List[str],
|
| 93 |
+
initial_conditioning_items: Optional[List[LatentConditioningItem]] = None,
|
| 94 |
+
height: int = 432,
|
| 95 |
+
width: int = 768,
|
| 96 |
+
duration_in_seconds: float = 4.0,
|
| 97 |
+
ltx_configs: Optional[Dict] = None
|
| 98 |
+
) -> Tuple[Optional[torch.Tensor], Optional[int]]:
|
| 99 |
"""
|
| 100 |
+
Ponto de entrada principal para gerar um vídeo latente completo.
|
| 101 |
+
|
| 102 |
+
Args:
|
| 103 |
+
prompt_list: Lista de prompts, onde cada prompt é uma cena.
|
| 104 |
+
initial_conditioning_items: Lista de `LatentConditioningItem` para condicionar
|
| 105 |
+
a primeira cena.
|
| 106 |
+
height: Altura do vídeo.
|
| 107 |
+
width: Largura do vídeo.
|
| 108 |
+
duration_in_seconds: Duração total desejada do vídeo.
|
| 109 |
+
ltx_configs: Dicionário com configurações avançadas para o pipeline LTX
|
| 110 |
+
(guidance_scale, num_inference_steps, etc.).
|
| 111 |
+
|
| 112 |
+
Returns:
|
| 113 |
+
Uma tupla contendo:
|
| 114 |
+
- O tensor latente final completo (na CPU).
|
| 115 |
+
- A semente principal usada para a geração.
|
| 116 |
"""
|
| 117 |
+
t0 = time.time()
|
| 118 |
+
logging.info(f"LTX Client received a generation job with {len(prompt_list)} scenes.")
|
| 119 |
+
|
| 120 |
+
if not prompt_list:
|
| 121 |
+
raise ValueError("A lista de prompts não pode estar vazia.")
|
| 122 |
+
|
| 123 |
used_seed = self._get_random_seed()
|
| 124 |
+
logging.info(f"Generation seed set to: {used_seed}")
|
|
|
|
| 125 |
|
| 126 |
+
# --- Lógica de Divisão de Chunks e Sobreposição ---
|
|
|
|
|
|
|
| 127 |
num_chunks = len(prompt_list)
|
| 128 |
+
total_frames = self._align(int(duration_in_seconds * 24))
|
| 129 |
+
frames_per_chunk_base = total_frames // num_chunks
|
| 130 |
+
overlap_frames = self._align(9) if num_chunks > 1 else 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 131 |
|
| 132 |
+
final_latents_list = []
|
| 133 |
+
overlap_condition_item: Optional[LatentConditioningItem] = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
|
| 135 |
+
for i, chunk_prompt in enumerate(prompt_list):
|
| 136 |
+
current_conditions = []
|
| 137 |
+
if i == 0 and initial_conditioning_items:
|
| 138 |
+
current_conditions.extend(initial_conditioning_items)
|
| 139 |
+
if overlap_condition_item:
|
| 140 |
+
current_conditions.append(overlap_condition_item)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
+
# Calcula o número de frames para o chunk atual
|
| 143 |
+
num_frames_for_chunk = frames_per_chunk_base
|
| 144 |
+
if i == num_chunks - 1: # Último chunk pega o resto
|
| 145 |
+
processed_frames = sum(f.shape[2] for f in final_latents_list)
|
| 146 |
+
num_frames_for_chunk = total_frames - processed_frames
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
|
| 148 |
+
num_frames_for_chunk = self._align(num_frames_for_chunk)
|
| 149 |
+
|
| 150 |
+
# --- Submissão do Job para o Chunk Atual ---
|
| 151 |
+
chunk_latents = ltx_aduc_manager.submit_job(
|
| 152 |
+
job_type='ltx',
|
| 153 |
+
job_func=_job_generate_latent_chunk,
|
| 154 |
+
# Passa todos os argumentos necessários para a função de trabalho
|
| 155 |
+
prompt=chunk_prompt,
|
| 156 |
+
negative_prompt="blurry, low quality, bad anatomy, deformed", # Pode ser parametrizado
|
| 157 |
+
height=height,
|
| 158 |
+
width=width,
|
| 159 |
+
num_frames=num_frames_for_chunk,
|
| 160 |
+
seed=used_seed + i, # Semente diferente para cada chunk para variedade
|
| 161 |
+
conditioning_items=current_conditions,
|
| 162 |
+
ltx_configs=ltx_configs or {}
|
| 163 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
|
| 165 |
+
if chunk_latents is None:
|
| 166 |
+
logging.error(f"Failed to generate latents for scene {i+1}. Aborting generation.")
|
| 167 |
+
return None, used_seed
|
| 168 |
+
|
| 169 |
+
# --- Gerenciamento do "Eco Cinético" (Sobreposição) ---
|
| 170 |
+
if i < num_chunks - 1:
|
| 171 |
+
# Salva os últimos frames do chunk atual para condicionar o próximo
|
| 172 |
+
overlap_latents = chunk_latents[:, :, -overlap_frames:, :, :].clone()
|
| 173 |
+
overlap_condition_item = LatentConditioningItem(
|
| 174 |
+
latent_tensor=overlap_latents,
|
| 175 |
+
media_frame_number=0, # Sempre condiciona o início do próximo chunk
|
| 176 |
+
conditioning_strength=1.0 # Condicionamento forte
|
| 177 |
+
)
|
| 178 |
+
# Adiciona o chunk atual sem a sobreposição
|
| 179 |
+
final_latents_list.append(chunk_latents[:, :, :-overlap_frames, :, :])
|
| 180 |
+
else:
|
| 181 |
+
# Adiciona o último chunk completo
|
| 182 |
+
final_latents_list.append(chunk_latents)
|
| 183 |
|
| 184 |
+
# Concatena todos os chunks de latentes em um único tensor
|
| 185 |
+
final_latents = torch.cat(final_latents_list, dim=2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 186 |
|
| 187 |
+
logging.info(f"LTX Client job finished in {time.time() - t0:.2f}s. Final latent shape: {final_latents.shape}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
|
| 189 |
+
return final_latents, used_seed
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 190 |
|
| 191 |
+
# --- INSTÂNCIA SINGLETON DO CLIENTE ---
|
|
|
|
|
|
|
| 192 |
try:
|
| 193 |
ltx_aduc_pipeline = LtxAducPipeline()
|
|
|
|
| 194 |
except Exception as e:
|
| 195 |
+
logging.critical("CRITICAL: Failed to initialize the LtxAducPipeline client.", exc_info=True)
|
| 196 |
+
ltx_aduc_pipeline = None```
|