Spaces:
Sleeping
Sleeping
Update ltx_manager_helpers.py
Browse files- ltx_manager_helpers.py +171 -12
ltx_manager_helpers.py
CHANGED
|
@@ -1,8 +1,125 @@
|
|
| 1 |
# ltx_manager_helpers.py
|
| 2 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
|
| 4 |
class LtxPoolManager:
|
| 5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
def generate_video_fragment(
|
| 7 |
self,
|
| 8 |
motion_prompt: str, conditioning_items_data: list,
|
|
@@ -13,10 +130,44 @@ class LtxPoolManager:
|
|
| 13 |
):
|
| 14 |
worker_to_use = None
|
| 15 |
try:
|
| 16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
|
|
|
|
|
|
|
|
|
| 20 |
kwargs = {
|
| 21 |
"prompt": motion_prompt,
|
| 22 |
"negative_prompt": "blurry, distorted, bad quality, artifacts",
|
|
@@ -38,28 +189,36 @@ class LtxPoolManager:
|
|
| 38 |
"enhance_prompt": False,
|
| 39 |
}
|
| 40 |
|
| 41 |
-
# --- CORREÇÃO AQUI ---
|
| 42 |
# Verifica se o config do modelo especifica uma lista de timesteps.
|
| 43 |
# Se sim, usa essa lista. Se não, usa o num_inference_steps da UI.
|
| 44 |
first_pass_config = worker_to_use.config.get("first_pass", {})
|
| 45 |
if "timesteps" in first_pass_config:
|
| 46 |
print("Usando timesteps customizados do arquivo de configuração para o modelo distilled.")
|
| 47 |
kwargs["timesteps"] = first_pass_config["timesteps"]
|
| 48 |
-
# Quando usamos timesteps customizados, o num_inference_steps é inferido
|
| 49 |
-
# a partir do tamanho da lista de timesteps.
|
| 50 |
kwargs["num_inference_steps"] = len(first_pass_config["timesteps"])
|
|
|
|
|
|
|
| 51 |
else:
|
| 52 |
-
# Comportamento antigo para modelos não-destilados
|
| 53 |
print(f"Usando num_inference_steps da UI: {num_inference_steps}")
|
| 54 |
kwargs["num_inference_steps"] = int(num_inference_steps)
|
| 55 |
-
|
| 56 |
-
|
| 57 |
progress(0.1, desc=f"[Câmera LTX em {worker_to_use.device}] Filmando Cena {current_fragment_index}...")
|
| 58 |
result_tensor = worker_to_use.generate_video_fragment_internal(**kwargs).images
|
| 59 |
|
| 60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
return output_path, video_total_frames
|
| 63 |
|
| 64 |
finally:
|
| 65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# ltx_manager_helpers.py
|
| 2 |
+
# Gerente de Pool de Workers LTX para revezamento assíncrono em múltiplas GPUs.
|
| 3 |
+
# Este arquivo é parte do projeto Euia-AducSdr e está sob a licença AGPL v3.
|
| 4 |
+
# Copyright (C) 4 de Agosto de 2025 Carlos Rodrigues dos Santos
|
| 5 |
+
|
| 6 |
+
import torch
|
| 7 |
+
import gc
|
| 8 |
+
import os
|
| 9 |
+
import yaml
|
| 10 |
+
import numpy as np
|
| 11 |
+
import imageio
|
| 12 |
+
from pathlib import Path
|
| 13 |
+
import huggingface_hub
|
| 14 |
+
import threading
|
| 15 |
+
from PIL import Image
|
| 16 |
+
|
| 17 |
+
# Importa as funções e classes necessárias do inference.py
|
| 18 |
+
from inference import (
|
| 19 |
+
create_ltx_video_pipeline,
|
| 20 |
+
create_latent_upsampler,
|
| 21 |
+
ConditioningItem,
|
| 22 |
+
calculate_padding,
|
| 23 |
+
prepare_conditioning
|
| 24 |
+
)
|
| 25 |
+
from ltx_video.pipelines.pipeline_ltx_video import LTXMultiScalePipeline
|
| 26 |
+
|
| 27 |
+
class LtxWorker:
|
| 28 |
+
"""
|
| 29 |
+
Representa uma única instância do pipeline LTX, associada a uma GPU específica.
|
| 30 |
+
O pipeline é carregado na CPU por padrão e movido para a GPU sob demanda.
|
| 31 |
+
"""
|
| 32 |
+
def __init__(self, device_id='cuda:0'):
|
| 33 |
+
self.device = torch.device(device_id if torch.cuda.is_available() else 'cpu')
|
| 34 |
+
print(f"LTX Worker: Inicializando para o dispositivo {self.device} (carregando na CPU)...")
|
| 35 |
+
|
| 36 |
+
config_file_path = "configs/ltxv-13b-0.9.8-distilled.yaml"
|
| 37 |
+
with open(config_file_path, "r") as file:
|
| 38 |
+
self.config = yaml.safe_load(file)
|
| 39 |
+
|
| 40 |
+
LTX_REPO = "Lightricks/LTX-Video"
|
| 41 |
+
models_dir = "downloaded_models_gradio"
|
| 42 |
+
|
| 43 |
+
model_actual_path = huggingface_hub.hf_hub_download(
|
| 44 |
+
repo_id=LTX_REPO,
|
| 45 |
+
filename=self.config["checkpoint_path"],
|
| 46 |
+
local_dir=models_dir,
|
| 47 |
+
local_dir_use_symlinks=False
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
self.pipeline = create_ltx_video_pipeline(
|
| 51 |
+
ckpt_path=model_actual_path,
|
| 52 |
+
precision=self.config["precision"],
|
| 53 |
+
text_encoder_model_name_or_path=self.config["text_encoder_model_name_or_path"],
|
| 54 |
+
sampler=self.config["sampler"],
|
| 55 |
+
device='cpu'
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
print(f"LTX Worker para {self.device}: Compilando o transformer (isso pode levar um momento)...")
|
| 59 |
+
self.pipeline.transformer.to(memory_format=torch.channels_last)
|
| 60 |
+
try:
|
| 61 |
+
self.pipeline.transformer = torch.compile(self.pipeline.transformer, mode="reduce-overhead", fullgraph=True)
|
| 62 |
+
print(f"LTX Worker para {self.device}: Transformer compilado com sucesso.")
|
| 63 |
+
except Exception as e:
|
| 64 |
+
print(f"AVISO: A compilação do Transformer falhou em {self.device}: {e}. Continuando sem compilação.")
|
| 65 |
+
|
| 66 |
+
self.latent_upsampler = None
|
| 67 |
+
if self.config.get("pipeline_type") == "multi-scale":
|
| 68 |
+
print(f"LTX Worker para {self.device}: Carregando Latent Upsampler (Multi-Scale)...")
|
| 69 |
+
upscaler_path = huggingface_hub.hf_hub_download(
|
| 70 |
+
repo_id=LTX_REPO,
|
| 71 |
+
filename=self.config["spatial_upscaler_model_path"],
|
| 72 |
+
local_dir=models_dir,
|
| 73 |
+
local_dir_use_symlinks=False
|
| 74 |
+
)
|
| 75 |
+
self.latent_upsampler = create_latent_upsampler(upscaler_path, 'cpu')
|
| 76 |
+
|
| 77 |
+
print(f"LTX Worker para {self.device} pronto na CPU.")
|
| 78 |
+
|
| 79 |
+
def to_gpu(self):
|
| 80 |
+
"""Move o pipeline e o upsampler para a GPU designada."""
|
| 81 |
+
if self.device.type == 'cpu': return
|
| 82 |
+
print(f"LTX Worker: Movendo pipeline para {self.device}...")
|
| 83 |
+
self.pipeline.to(self.device)
|
| 84 |
+
if self.latent_upsampler:
|
| 85 |
+
print(f"LTX Worker: Movendo Latent Upsampler para {self.device}...")
|
| 86 |
+
self.latent_upsampler.to(self.device)
|
| 87 |
+
print(f"LTX Worker: Pipeline na GPU {self.device}.")
|
| 88 |
+
|
| 89 |
+
def to_cpu(self):
|
| 90 |
+
"""Move o pipeline de volta para a CPU e limpa a memória da GPU."""
|
| 91 |
+
if self.device.type == 'cpu': return
|
| 92 |
+
print(f"LTX Worker: Descarregando pipeline da GPU {self.device}...")
|
| 93 |
+
self.pipeline.to('cpu')
|
| 94 |
+
if self.latent_upsampler:
|
| 95 |
+
self.latent_upsampler.to('cpu')
|
| 96 |
+
gc.collect()
|
| 97 |
+
if torch.cuda.is_available():
|
| 98 |
+
torch.cuda.empty_cache()
|
| 99 |
+
print(f"LTX Worker: GPU {self.device} limpa.")
|
| 100 |
+
|
| 101 |
+
def generate_video_fragment_internal(self, **kwargs):
|
| 102 |
+
"""A lógica real da geração de vídeo, que espera estar na GPU."""
|
| 103 |
+
return self.pipeline(**kwargs)
|
| 104 |
|
| 105 |
class LtxPoolManager:
|
| 106 |
+
"""
|
| 107 |
+
Gerencia um pool de LtxWorkers, orquestrando um revezamento entre GPUs
|
| 108 |
+
para permitir que a limpeza de uma GPU ocorra em paralelo com a computação em outra.
|
| 109 |
+
"""
|
| 110 |
+
def __init__(self, device_ids=['cuda:2', 'cuda:3']):
|
| 111 |
+
print(f"LTX POOL MANAGER: Criando workers para os dispositivos: {device_ids}")
|
| 112 |
+
self.workers = [LtxWorker(device_id) for device_id in device_ids]
|
| 113 |
+
self.current_worker_index = 0
|
| 114 |
+
self.lock = threading.Lock()
|
| 115 |
+
self.last_cleanup_thread = None
|
| 116 |
+
|
| 117 |
+
def _cleanup_worker(self, worker):
|
| 118 |
+
"""Função alvo para a thread de limpeza."""
|
| 119 |
+
print(f"CLEANUP THREAD: Iniciando limpeza da GPU {worker.device} em background...")
|
| 120 |
+
worker.to_cpu()
|
| 121 |
+
print(f"CLEANUP THREAD: Limpeza da GPU {worker.device} concluída.")
|
| 122 |
+
|
| 123 |
def generate_video_fragment(
|
| 124 |
self,
|
| 125 |
motion_prompt: str, conditioning_items_data: list,
|
|
|
|
| 130 |
):
|
| 131 |
worker_to_use = None
|
| 132 |
try:
|
| 133 |
+
with self.lock:
|
| 134 |
+
if self.last_cleanup_thread and self.last_cleanup_thread.is_alive():
|
| 135 |
+
print("LTX POOL MANAGER: Aguardando limpeza da GPU anterior...")
|
| 136 |
+
self.last_cleanup_thread.join()
|
| 137 |
+
|
| 138 |
+
worker_to_use = self.workers[self.current_worker_index]
|
| 139 |
+
previous_worker_index = (self.current_worker_index - 1 + len(self.workers)) % len(self.workers)
|
| 140 |
+
worker_to_cleanup = self.workers[previous_worker_index]
|
| 141 |
+
|
| 142 |
+
cleanup_thread = threading.Thread(target=self._cleanup_worker, args=(worker_to_cleanup,))
|
| 143 |
+
cleanup_thread.start()
|
| 144 |
+
self.last_cleanup_thread = cleanup_thread
|
| 145 |
+
|
| 146 |
+
worker_to_use.to_gpu()
|
| 147 |
+
|
| 148 |
+
self.current_worker_index = (self.current_worker_index + 1) % len(self.workers)
|
| 149 |
+
|
| 150 |
+
target_device = worker_to_use.device
|
| 151 |
|
| 152 |
+
if use_attention_slicing:
|
| 153 |
+
worker_to_use.pipeline.enable_attention_slicing()
|
| 154 |
+
|
| 155 |
+
media_paths = [item[0] for item in conditioning_items_data]
|
| 156 |
+
start_frames = [item[1] for item in conditioning_items_data]
|
| 157 |
+
strengths = [item[2] for item in conditioning_items_data]
|
| 158 |
+
|
| 159 |
+
padded_h, padded_w = ((height - 1) // 32 + 1) * 32, ((width - 1) // 32 + 1) * 32
|
| 160 |
+
padding_vals = calculate_padding(height, width, padded_h, padded_w)
|
| 161 |
+
|
| 162 |
+
conditioning_items = prepare_conditioning(
|
| 163 |
+
conditioning_media_paths=media_paths, conditioning_strengths=strengths,
|
| 164 |
+
conditioning_start_frames=start_frames, height=height, width=width,
|
| 165 |
+
num_frames=video_total_frames, padding=padding_vals, pipeline=worker_to_use.pipeline,
|
| 166 |
+
)
|
| 167 |
|
| 168 |
+
for item in conditioning_items:
|
| 169 |
+
item.media_item = item.media_item.to(target_device)
|
| 170 |
+
|
| 171 |
kwargs = {
|
| 172 |
"prompt": motion_prompt,
|
| 173 |
"negative_prompt": "blurry, distorted, bad quality, artifacts",
|
|
|
|
| 189 |
"enhance_prompt": False,
|
| 190 |
}
|
| 191 |
|
|
|
|
| 192 |
# Verifica se o config do modelo especifica uma lista de timesteps.
|
| 193 |
# Se sim, usa essa lista. Se não, usa o num_inference_steps da UI.
|
| 194 |
first_pass_config = worker_to_use.config.get("first_pass", {})
|
| 195 |
if "timesteps" in first_pass_config:
|
| 196 |
print("Usando timesteps customizados do arquivo de configuração para o modelo distilled.")
|
| 197 |
kwargs["timesteps"] = first_pass_config["timesteps"]
|
|
|
|
|
|
|
| 198 |
kwargs["num_inference_steps"] = len(first_pass_config["timesteps"])
|
| 199 |
+
# Para modelos distilled, a UI de steps é ignorada, mas outros params do config são usados
|
| 200 |
+
kwargs.update({k: v for k, v in first_pass_config.items() if k != "timesteps"})
|
| 201 |
else:
|
|
|
|
| 202 |
print(f"Usando num_inference_steps da UI: {num_inference_steps}")
|
| 203 |
kwargs["num_inference_steps"] = int(num_inference_steps)
|
| 204 |
+
|
|
|
|
| 205 |
progress(0.1, desc=f"[Câmera LTX em {worker_to_use.device}] Filmando Cena {current_fragment_index}...")
|
| 206 |
result_tensor = worker_to_use.generate_video_fragment_internal(**kwargs).images
|
| 207 |
|
| 208 |
+
pad_l, pad_r, pad_t, pad_b = map(int, padding_vals)
|
| 209 |
+
slice_h = -pad_b if pad_b > 0 else None
|
| 210 |
+
slice_w = -pad_r if pad_r > 0 else None
|
| 211 |
+
cropped_tensor = result_tensor[:, :, :video_total_frames, pad_t:slice_h, pad_l:slice_w]
|
| 212 |
+
video_np = (cropped_tensor[0].permute(1, 2, 3, 0).cpu().float().numpy() * 255).astype(np.uint8)
|
| 213 |
+
|
| 214 |
+
with imageio.get_writer(output_path, fps=video_fps, codec='libx264', quality=8) as writer:
|
| 215 |
+
for frame in video_np:
|
| 216 |
+
writer.append_data(frame)
|
| 217 |
|
| 218 |
return output_path, video_total_frames
|
| 219 |
|
| 220 |
finally:
|
| 221 |
+
if use_attention_slicing and worker_to_use and worker_to_use.pipeline:
|
| 222 |
+
worker_to_use.pipeline.disable_attention_slicing()
|
| 223 |
+
|
| 224 |
+
ltx_manager_singleton = LtxPoolManager(device_ids=['cuda:2', 'cuda:3'])
|