Test / api /ltx_server_refactored_complete.py
eeuuia's picture
Update api/ltx_server_refactored_complete.py
c4db7b7 verified
raw
history blame
22.7 kB
# FILE: api/ltx_server_refactored_complete.py
# DESCRIPTION: Final orchestrator for LTX-Video generation.
# This version includes the fix for the narrative generation overlap bug and
# consolidates all previous refactoring and debugging improvements.
import gc
import json
import logging
import os
import shutil
import sys
import tempfile
import time
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import random
import torch
import yaml
import numpy as np
from huggingface_hub import hf_hub_download
# ==============================================================================
# --- SETUP E IMPORTAÇÕES DO PROJETO ---
# ==============================================================================
# Configuração de logging e supressão de warnings
import warnings
warnings.filterwarnings("ignore")
logging.getLogger("huggingface_hub").setLevel(logging.ERROR)
log_level = os.environ.get("ADUC_LOG_LEVEL", "INFO").upper()
logging.basicConfig(level=log_level, format='[%(levelname)s] [%(name)s] %(message)s')
# --- Constantes de Configuração ---
DEPS_DIR = Path("/data")
LTX_VIDEO_REPO_DIR = DEPS_DIR / "LTX-Video"
RESULTS_DIR = Path("/app/output")
DEFAULT_FPS = 24.0
FRAMES_ALIGNMENT = 8
LTX_REPO_ID = "Lightricks/LTX-Video"
# Garante que a biblioteca LTX-Video seja importável
def add_deps_to_path():
repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
if repo_path not in sys.path:
sys.path.insert(0, repo_path)
logging.info(f"[ltx_server] LTX-Video repository added to sys.path: {repo_path}")
add_deps_to_path()
# --- Módulos da nossa Arquitetura ---
try:
from api.gpu_manager import gpu_manager
from managers.vae_manager import vae_manager_singleton
from tools.video_encode_tool import video_encode_tool_singleton
from api.ltx.ltx_utils import (
build_ltx_pipeline_on_cpu,
seed_everything,
load_image_to_tensor_with_resize_and_crop,
ConditioningItem,
)
from api.utils.debug_utils import log_function_io
except ImportError as e:
logging.critical(f"A crucial import from the local API/architecture failed. Error: {e}", exc_info=True)
sys.exit(1)
# ==============================================================================
# --- FUNÇÕES AUXILIARES DO ORQUESTRADOR ---
# ==============================================================================
@log_function_io
def calculate_padding(orig_h: int, orig_w: int, target_h: int, target_w: int) -> Tuple[int, int, int, int]:
"""Calculates symmetric padding required to meet target dimensions."""
pad_h = target_h - orig_h
pad_w = target_w - orig_w
pad_top = pad_h // 2
pad_bottom = pad_h - pad_top
pad_left = pad_w // 2
pad_right = pad_w - pad_left
return (pad_left, pad_right, pad_top, pad_bottom)
# ==============================================================================
# --- CLASSE DE SERVIÇO (O ORQUESTRADOR) ---
# ==============================================================================
class VideoService:
"""
Orchestrates the high-level logic of video generation, delegating low-level
tasks to specialized managers and utility modules.
"""
@log_function_io
def __init__(self):
t0 = time.perf_counter()
logging.info("Initializing VideoService Orchestrator...")
RESULTS_DIR.mkdir(parents=True, exist_ok=True)
target_main_device_str = str(gpu_manager.get_ltx_device())
target_vae_device_str = str(gpu_manager.get_ltx_vae_device())
logging.info(f"LTX allocated to devices: Main='{target_main_device_str}', VAE='{target_vae_device_str}'")
self.config = self._load_config()
self._resolve_model_paths_from_cache()
self.pipeline, self.latent_upsampler = build_ltx_pipeline_on_cpu(self.config)
self.main_device = torch.device("cpu")
self.vae_device = torch.device("cpu")
self.move_to_device(main_device_str=target_main_device_str, vae_device_str=target_vae_device_str)
self._apply_precision_policy()
vae_manager_singleton.attach_pipeline(self.pipeline, device=self.vae_device, autocast_dtype=self.runtime_autocast_dtype)
logging.info(f"VideoService ready. Startup time: {time.perf_counter()-t0:.2f}s")
def _load_config(self) -> Dict:
"""Loads the YAML configuration file."""
config_path = LTX_VIDEO_REPO_DIR / "configs" / "ltxv-13b-0.9.8-distilled-fp8.yaml"
logging.info(f"Loading config from: {config_path}")
with open(config_path, "r") as file:
return yaml.safe_load(file)
def _resolve_model_paths_from_cache(self):
"""Finds the absolute paths to model files in the cache and updates the in-memory config."""
logging.info("Resolving model paths from Hugging Face cache...")
cache_dir = os.environ.get("HF_HOME")
try:
main_ckpt_path = hf_hub_download(repo_id=LTX_REPO_ID, filename=self.config["checkpoint_path"], cache_dir=cache_dir)
self.config["checkpoint_path"] = main_ckpt_path
logging.info(f" -> Main checkpoint resolved to: {main_ckpt_path}")
if self.config.get("spatial_upscaler_model_path"):
upscaler_path = hf_hub_download(repo_id=LTX_REPO_ID, filename=self.config["spatial_upscaler_model_path"], cache_dir=cache_dir)
self.config["spatial_upscaler_model_path"] = upscaler_path
logging.info(f" -> Spatial upscaler resolved to: {upscaler_path}")
except Exception as e:
logging.critical(f"Failed to resolve model paths. Ensure setup.py ran correctly. Error: {e}", exc_info=True)
sys.exit(1)
@log_function_io
def move_to_device(self, main_device_str: str, vae_device_str: str):
"""Moves pipeline components to their designated target devices."""
target_main_device = torch.device(main_device_str)
target_vae_device = torch.device(vae_device_str)
logging.info(f"Moving LTX models -> Main Pipeline: {target_main_device}, VAE: {target_vae_device}")
self.main_device = target_main_device
self.pipeline.to(self.main_device)
self.vae_device = target_vae_device
self.pipeline.vae.to(self.vae_device)
if self.latent_upsampler: self.latent_upsampler.to(self.main_device)
logging.info("LTX models successfully moved to target devices.")
def move_to_cpu(self):
"""Moves all LTX components to CPU to free VRAM for other services."""
self.move_to_device(main_device_str="cpu", vae_device_str="cpu")
if torch.cuda.is_available(): torch.cuda.empty_cache()
def finalize(self):
"""Cleans up GPU memory after a generation task."""
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
try: torch.cuda.ipc_collect();
except Exception: pass
# ==========================================================================
# --- LÓGICA DE NEGÓCIO: ORQUESTRADOR PÚBLICO UNIFICADO ---
# ==========================================================================
@log_function_io
def generate_low_resolution(self, prompt: str, **kwargs) -> Tuple[Optional[str], Optional[str], Optional[int]]:
"""
[UNIFIED ORCHESTRATOR] Generates a low-resolution video from a prompt.
Handles both single-line and multi-line prompts transparently.
"""
logging.info("Starting unified low-resolution generation (random seed)...")
used_seed = self._get_random_seed()
seed_everything(used_seed)
logging.info(f"Using randomly generated seed: {used_seed}")
prompt_list = [p.strip() for p in prompt.splitlines() if p.strip()]
if not prompt_list: raise ValueError("Prompt is empty or contains no valid lines.")
is_narrative = len(prompt_list) > 1
logging.info(f"Generation mode detected: {'Narrative' if is_narrative else 'Simple'} ({len(prompt_list)} scene(s)).")
num_chunks = len(prompt_list)
total_frames = self._calculate_aligned_frames(kwargs.get("duration", 4.0))
frames_per_chunk = max(FRAMES_ALIGNMENT, (total_frames // num_chunks // FRAMES_ALIGNMENT) * FRAMES_ALIGNMENT)
# Overlap must be N*8+1 frames. 9 is the smallest practical value.
overlap_frames = 9 if is_narrative else 0
if is_narrative:
logging.info(f"Narrative mode: Using overlap of {overlap_frames} frames between chunks.")
temp_latent_paths = []
overlap_condition_item = None
try:
for i, chunk_prompt in enumerate(prompt_list):
logging.info(f"Processing scene {i+1}/{num_chunks}: '{chunk_prompt[:50]}...'")
if i < num_chunks - 1:
current_frames_base = frames_per_chunk
else: # Last chunk takes all remaining frames
processed_frames_base = (num_chunks - 1) * frames_per_chunk
current_frames_base = total_frames - processed_frames_base
current_frames = current_frames_base + (overlap_frames if i > 0 else 0)
# Ensure final frame count for generation is N*8+1
current_frames = self._align(current_frames, alignment_rule='n*8+1')
current_conditions = kwargs.get("initial_conditions", []) if i == 0 else []
if overlap_condition_item:
current_conditions.append(overlap_condition_item)
chunk_latents = self._generate_single_chunk_low(
prompt=chunk_prompt, num_frames=current_frames, seed=used_seed + i,
conditioning_items=current_conditions, **kwargs
)
if chunk_latents is None: raise RuntimeError(f"Failed to generate latents for scene {i+1}.")
if is_narrative and i < num_chunks - 1:
# 1. Criar tensor overlap latente
overlap_latents = chunk_latents[:, :, -overlap_frames:, :, :].clone()
logging.info(f"Criado overlap latente com shape: {list(overlap_latents.shape)}")
# 2. DECODIFICA o latente de volta para um tensor de PIXEL
logging.info("Decodificando latente de overlap para tensor de pixel...")
overlap_pixel_tensor = vae_manager_singleton.decode(
overlap_latents,
decode_timestep=float(self.config.get("decode_timestep", 0.05))
)
# O resultado de decode() está na CPU, no formato (B, C, F, H, W) e [0, 1]
# Precisamos normalizá-lo para [-1, 1] que é o que o pipeline espera.
overlap_pixel_tensor_normalized = (overlap_pixel_tensor * 2.0) - 1.0
logging.info(f"Tensor de pixel de overlap criado com shape: {list(overlap_pixel_tensor_normalized.shape)}")
# 3. Cria o ConditioningItem com o TENSOR DE PIXEL, não com o latente.
overlap_condition_item = ConditioningItem(
media_item=overlap_pixel_tensor_normalized,
media_frame_number=0,conditioning_strength=1.0
)
if i > 0:
chunk_latents = chunk_latents[:, :, overlap_frames:, :, :]
chunk_path = RESULTS_DIR / f"temp_chunk_{i}_{used_seed}.pt"
torch.save(chunk_latents.cpu(), chunk_path)
temp_latent_paths.append(chunk_path)
base_filename = "narrative_video" if is_narrative else "single_video"
return self._finalize_generation(temp_latent_paths, base_filename, used_seed)
except Exception as e:
logging.error(f"Error during unified generation: {e}", exc_info=True)
return None, None, None
finally:
for path in temp_latent_paths:
if path.exists(): path.unlink()
self.finalize()
# ==========================================================================
# --- UNIDADES DE TRABALHO E HELPERS INTERNOS ---
# ==========================================================================
# --- NOVA FUNÇÃO DE LOG DEDICADA ---
def _log_conditioning_items(self, items: List[ConditioningItem]):
"""
Logs detailed information about a list of ConditioningItem objects.
This is a dedicated debug helper function.
"""
# Só imprime o log se o nível de logging for DEBUG
if logging.getLogger().isEnabledFor(logging.INFO):
log_str = ["\n" + "="*25 + " INFO: Conditioning Items " + "="*25]
if not items:
log_str.append(" -> Lista de conditioning_items está vazia.")
else:
for i, item in enumerate(items):
if hasattr(item, 'media_item') and isinstance(item.media_item, torch.Tensor):
t = item.media_item
log_str.append(
f" -> Item [{i}]: "
f"Tensor(shape={list(t.shape)}, "
f"device='{t.device}', "
f"dtype={t.dtype}), "
f"Target Frame = {item.media_frame_number}, "
f"Strength = {item.conditioning_strength:.2f}"
)
else:
log_str.append(f" -> Item [{i}]: Não contém um tensor válido.")
log_str.append("="*75 + "\n")
# Usa o logger de debug para imprimir a mensagem completa
logging.info("\n".join(log_str))
@log_function_io
def _generate_single_chunk_low(self, **kwargs) -> Optional[torch.Tensor]:
"""[WORKER] Calls the pipeline to generate a single chunk of latents."""
height_padded, width_padded = (self._align(d) for d in (kwargs['height'], kwargs['width']))
downscale_factor = self.config.get("downscale_factor", 0.6666666)
vae_scale_factor = self.pipeline.vae_scale_factor
downscaled_height = self._align(int(height_padded * downscale_factor), vae_scale_factor)
downscaled_width = self._align(int(width_padded * downscale_factor), vae_scale_factor)
# 1. Começa com a configuração padrão
first_pass_config = self.config.get("first_pass", {}).copy()
# 2. Aplica os overrides da UI, se existirem
if kwargs.get("ltx_configs_override"):
self._apply_ui_overrides(first_pass_config, kwargs.get("ltx_configs_override"))
# 3. Monta o dicionário de argumentos SEM conditioning_items primeiro
pipeline_kwargs = {
"prompt": kwargs['prompt'],
"negative_prompt": kwargs['negative_prompt'],
"height": downscaled_height,
"width": downscaled_width,
"num_frames": kwargs['num_frames'],
"frame_rate": int(DEFAULT_FPS),
"generator": torch.Generator(device=self.main_device).manual_seed(kwargs['seed']),
"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,
**first_pass_config
}
# --- Bloco de Logging para Depuração ---
# 4. Loga os argumentos do pipeline (sem os tensores de condição)
logging.info(f"\n[Info] Pipeline Arguments (BASE):\n {json.dumps(pipeline_kwargs, indent=2, default=str)}\n")
# Loga os conditioning_items separadamente com a nossa função helper
conditioning_items_list = kwargs.get('conditioning_items')
self._log_conditioning_items(conditioning_items_list)
# --- Fim do Bloco de Logging ---
# 5. Adiciona os conditioning_items ao dicionário
pipeline_kwargs['conditioning_items'] = conditioning_items_list
# 6. Executa o pipeline com o dicionário completo
with torch.autocast(device_type=self.main_device.type, dtype=self.runtime_autocast_dtype, enabled="cuda" in self.main_device.type):
latents_raw = self.pipeline(**pipeline_kwargs).images
return latents_raw.to(self.main_device)
@log_function_io
def _finalize_generation(self, temp_latent_paths: List[Path], base_filename: str, seed: int) -> Tuple[str, str, int]:
"""Consolidates latents, decodes them to video, and saves final artifacts."""
logging.info("Finalizing generation: decoding latents to video.")
all_tensors_cpu = [torch.load(p) for p in temp_latent_paths]
final_latents = torch.cat(all_tensors_cpu, dim=2)
final_latents_path = RESULTS_DIR / f"latents_{base_filename}_{seed}.pt"
torch.save(final_latents, final_latents_path)
logging.info(f"Final latents saved to: {final_latents_path}")
pixel_tensor = vae_manager_singleton.decode(
final_latents, decode_timestep=float(self.config.get("decode_timestep", 0.05))
)
video_path = self._save_and_log_video(pixel_tensor, f"{base_filename}_{seed}")
return str(video_path), str(final_latents_path), seed
@log_function_io
def prepare_condition_items(self, items_list: List, height: int, width: int, num_frames: int) -> List[ConditioningItem]:
"""
[CORRIGIDO] Prepara ConditioningItems, garantindo que o tensor final
resida no dispositivo principal do pipeline (main_device).
"""
if not items_list: return []
height_padded, width_padded = self._align(height), self._align(width)
padding_values = calculate_padding(height, width, height_padded, width_padded)
conditioning_items = []
for media_item, frame, weight in items_list:
final_tensor = None
if isinstance(media_item, str):
# 1. Carrega a imagem. A função pode usar o VAE, então ela pode
# retornar um tensor em qualquer dispositivo.
tensor = load_image_to_tensor_with_resize_and_crop(media_item, height, width)
# 2. Aplica padding.
tensor = torch.nn.functional.pad(tensor, padding_values)
# 3. GARANTE que o tensor final esteja no dispositivo principal.
final_tensor = tensor.to(self.main_device, dtype=self.runtime_autocast_dtype)
elif isinstance(media_item, torch.Tensor):
# Se já for um tensor (ex: overlap), apenas garante que ele está no dispositivo principal.
final_tensor = media_item.to(self.main_device, dtype=self.runtime_autocast_dtype)
else:
logging.warning(f"Unknown conditioning media type: {type(media_item)}. Skipping.")
continue
safe_frame = max(0, min(int(frame), num_frames - 1))
conditioning_items.append(ConditioningItem(final_tensor, safe_frame, float(weight)))
self._log_conditioning_items(conditioning_items)
return conditioning_items
def _apply_ui_overrides(self, config_dict: Dict, overrides: Dict):
"""Applies advanced settings from the UI to a config dictionary."""
# Override step counts
for key in ["num_inference_steps", "skip_initial_inference_steps", "skip_final_inference_steps"]:
ui_value = overrides.get(key)
if ui_value and ui_value > 0:
config_dict[key] = ui_value
logging.info(f"Override: '{key}' set to {ui_value} by UI.")
def _save_and_log_video(self, pixel_tensor: torch.Tensor, base_filename: str) -> Path:
with tempfile.TemporaryDirectory() as temp_dir:
temp_path = os.path.join(temp_dir, f"{base_filename}.mp4")
video_encode_tool_singleton.save_video_from_tensor(pixel_tensor, temp_path, fps=DEFAULT_FPS)
final_path = RESULTS_DIR / f"{base_filename}.mp4"
shutil.move(temp_path, final_path)
logging.info(f"Video saved successfully to: {final_path}")
return final_path
def _apply_precision_policy(self):
precision = str(self.config.get("precision", "bfloat16")).lower()
if precision in ["float8_e4m3fn", "bfloat16"]: self.runtime_autocast_dtype = torch.bfloat16
elif precision == "mixed_precision": self.runtime_autocast_dtype = torch.float16
else: self.runtime_autocast_dtype = torch.float32
logging.info(f"Runtime precision policy set for autocast: {self.runtime_autocast_dtype}")
def _align(self, dim: int, alignment: int = FRAMES_ALIGNMENT, alignment_rule: str = 'default') -> int:
"""Aligns a dimension to the nearest multiple of `alignment`."""
if alignment_rule == 'n*8+1':
return ((dim - 1) // alignment) * alignment + 1
return ((dim - 1) // alignment + 1) * alignment
def _calculate_aligned_frames(self, duration_s: float, min_frames: int = 1) -> int:
num_frames = int(round(duration_s * DEFAULT_FPS))
# Para a duração total, sempre arredondamos para cima para o múltiplo de 8 mais próximo
aligned_frames = self._align(num_frames, alignment=FRAMES_ALIGNMENT)
return max(aligned_frames, min_frames)
def _get_random_seed(self) -> int:
"""Always generates and returns a new random seed."""
return random.randint(0, 2**32 - 1)
# ==============================================================================
# --- INSTANCIAÇÃO SINGLETON ---
# ==============================================================================
try:
video_generation_service = VideoService()
logging.info("Global VideoService orchestrator instance created successfully.")
except Exception as e:
logging.critical(f"Failed to initialize VideoService: {e}", exc_info=True)
sys.exit(1)