debug

configs/vae_magvit_v2/inference/17x128x128_pixabay.py

@@ -11,24 +11,21 @@ dataset = dict(
 
 fps = 24 // 3
 is_vae = True
-use_pipeline = True
 
 # Define dataset
-root = None
-data_path = "CSV_PATH"
 max_test_samples = -1
-use_image_transform = False
-num_workers = 4
-video_contains_first_frame = True
-
 
 
 # Define acceleration
+num_workers = 4
 dtype = "bf16"
 grad_checkpoint = True
 plugin = "zero2"
 sp_size = 1
 
+use_pipeline = True
+video_contains_first_frame = True
+
 
 # Define model
 
@@ -52,7 +49,6 @@ model = dict(
     disable_space = True,
     custom_conv_padding = None,
     encoder_double_z = False,
-    custom_conv_padding=None,
 )
 
 discriminator = dict(

opensora/utils/config_utils.py

@@ -76,7 +76,7 @@ def merge_args(cfg, args, training=False):
         if cfg.get("discriminator") is not None:
             cfg.discriminator["from_pretrained"] = args.ckpt_path
         args.ckpt_path = None
-    if training and args.data_path is not None:
+    if (training or cfg.get("is_vae", False)) and args.data_path is not None:
         cfg.dataset["data_path"] = args.data_path
         args.data_path = None
     if not training and args.cfg_scale is not None:

scripts/inference-vae-v2.py

@@ -12,7 +12,7 @@ from opensora.registry import MODELS, SCHEDULERS, build_module
 from opensora.utils.config_utils import parse_configs
 from opensora.utils.misc import to_torch_dtype
 from opensora.datasets import prepare_dataloader, prepare_variable_dataloader
-from opensora.datasets import DATASETS, MODELS, build_module
+from opensora.registry import DATASETS, MODELS, build_module
 from opensora.acceleration.parallel_states import (
     get_data_parallel_group,
     set_data_parallel_group,