Update app.py

app.py

@@ -5,7 +5,7 @@ import tempfile
 import os
 from torchvision import transforms
 
-from diffusers import LTXLatentUpsamplePipeline
+from diffusers import LTXLatentUpsamplePipeline, AutoModel
 #from pipeline_ltx_condition_control import LTXConditionPipeline, LTXVideoCondition
 from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXConditionPipeline, LTXVideoCondition
 from diffusers.utils import export_to_video, load_video
@@ -45,31 +45,72 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 #    token=os.getenv("HF_TOKEN"),
 #)
 
-from huggingface_hub import hf_hub_download
-from safetensors.torch import load_file as safe_load
+base_repo="Lightricks/LTX-Video"
+checkpoint_path="ltxv-13b-0.9.8-distilled.safetensors" 
+upscaler_repo="Lightricks/ltxv-spatial-upscaler-0.9.7"
+CACHE_DIR=os.getenv("HF_HOME_CACHE")
+FPS = 24
 
-# Baixa exatamente a variante desejada do repo oficial:
-weight_path = hf_hub_download(
-    repo_id="Lightricks/LTX-Video",
-    filename="ltxv-13b-0.9.8-distilled.safetensors",
-    #revision=os.getenv("LTXV_REVISION", "8984fa25007f376c1a299016d0957a37a2f797bb")
-)
+# 2. Baixar os arquivos do modelo base
+print(f"=== Baixando snapshot do repositório base: {base_repo} ===")
+ckpt_path_str = hf_hub_download(repo_id=base_repo, filename=checkpoint_path, cache_dir=CACHE_DIR)
+ckpt_path = Path(ckpt_path_str)
+    if not ckpt_path.is_file():
+        raise FileNotFoundError(f"Main checkpoint file not found: {ckpt_path}")
+
+# 3. Carregar cada componente da pipeline explicitamente
+print("=== Carregando componentes da pipeline... ===")
 
-pipeline = LTXConditionPipeline.from_pretrained(
+vae = AutoModel.from_pretrained(
     "Lightricks/LTX-Video",
-    #revision=os.getenv("LTXV_REVISION", "8984fa25007f376c1a299016d0957a37a2f797bb"),  # exemplo: SHA atual do repo oficial
-    offload_state_dict=False,
-    torch_dtype=torch.bfloat16,
-    cache_dir=os.getenv("HF_HOME_CACHE"),
-    token=os.getenv("HF_TOKEN"),
+    subfolder="vae",
+    dtype=torch_dtype,
+    cache_dir=CACHE_DIR
+)
+text_encoder = AutoModel.from_pretrained(
+    "Lightricks/LTX-Video",
+    subfolder="text_encoder",
+    dtype=torch_dtype,
+    cache_dir=CACHE_DIR
+)
+scheduler = AutoModel.from_pretrained(
+    "Lightricks/LTX-Video",
+    subfolder="scheduler",
+    dtype=torch_dtype,
+    cache_dir=CACHE_DIR
+)
+tokenizer = AutoModel.from_pretrained(
+    "Lightricks/LTX-Video",
+    subfolder="tokenizer",
+    dtype=torch_dtype,
+    cache_dir=CACHE_DIR
 )
 
+if hasattr(scheduler.config, 'use_dynamic_shifting') and scheduler.config.use_dynamic_shifting:
+    print("[Config] Desativando 'use_dynamic_shifting' no scheduler.")
+    scheduler.config.use_dynamic_shifting = False
+    
 
-# Carrega o state_dict e aplica no transformer já criado pelo model_index:
-state = safe_load(weight_path)
-pipeline.transformer.load_state_dict(state, strict=True)
+transformer = AutoModel.from_pretrained(
+    "Lightricks/LTX-Video",
+    subfolder="transformer",
+    dtype=torch.bfloat16,
+    cache_dir=CACHE_DIR
+)
+transformer.enable_layerwise_casting(
+    storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16
+)
 
 
+# 4. Montar a pipeline principal
+print("Montando a LTXConditionPipeline...")
+pipeline = LTXConditionPipeline(
+    vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, 
+    scheduler=scheduler, transformer=transformer, cache_dir=CACHE_DIR
+)
+pipeline.to(device)
+pipeline.vae.enable_tiling()
+
 
 
 pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained(