Open-Sora: Democratizing Efficient Video Production for All
Python
[exp]
Zangwei Zhengb26beabeb7b4
diff --git a/configs/opensora-v1-2/train/adapt-rf-gc.py b/configs/opensora-v1-2/train/adapt-rf-gc.pynew file mode 100644index 0000000..6729d66--- /dev/null+++ b/configs/opensora-v1-2/train/adapt-rf-gc.py@@ -0,0 +1,76 @@+# Dataset settings+dataset = dict(+ type="VariableVideoTextDataset",+ transform_name="resize_crop",+)+bucket_config = {+ # "144p": {1: (1.0, 200)},+ # # ---+ # "256": {1: (0.5, 200)},+ # "240p": {1: (0.5, 200)},+ # # ---+ # "360p": {1: (0.5, 120)},+ # "512": {1: (0.5, 120)},+ # # ---+ # "480p": {1: (0.5, 80)},+ # # ---+ # "720p": {1: (0.2, 40)},+ # "1024": {1: (0.2, 40)},+ # ---+ # "1080p": {1: (0.3, 20)},+ # ---+ "2048": {1: (1.0, 10)},+}++# Acceleration settings+num_workers = 4+num_bucket_build_workers = 16+dtype = "bf16"+grad_checkpoint = True+plugin = "zero2"+sp_size = 1++# Model settings+model = dict(+ type="STDiT3-XL/2",+ from_pretrained=None,+ qk_norm=True,+ enable_flashattn=True,+ enable_layernorm_kernel=True,+)+vae = dict(+ type="VideoAutoencoderKL",+ micro_batch_size=4,+ from_pretrained="PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers",+ subfolder="vae",+ local_files_only=True,+)+text_encoder = dict(+ type="t5",+ from_pretrained="DeepFloyd/t5-v1_1-xxl",+ model_max_length=300,+ shardformer=True,+ local_files_only=True,+)+scheduler = dict(+ type="rflow",+ use_discrete_timesteps=False,+ use_timestep_transform=True,+ sample_method="logit-normal",+)++# Log settings+seed = 42+outputs = "outputs"+wandb = False+epochs = 1000+log_every = 10+ckpt_every = 500++# optimization settings+load = None+batch_size = None+grad_clip = 1.0+lr = 1e-4+ema_decay = 0.99+adam_eps = 1e-15diff --git a/opensora/schedulers/rf/__init__.py b/opensora/schedulers/rf/__init__.pyindex 2ea0ee3..3e71d15 100644--- a/opensora/schedulers/rf/__init__.py+++ b/opensora/schedulers/rf/__init__.py@@ -1,6 +1,7 @@from functools import partialimport torch+from tqdm import tqdmfrom opensora.registry import SCHEDULERS@@ -67,7 +68,7 @@ class RFLOW:for t in timesteps]- for i, t in enumerate(timesteps):+ for i, t in tqdm(enumerate(timesteps)):z_in = torch.cat([z, z], 0)pred = model(z_in, torch.tensor([t] * z_in.shape[0], device=device), **model_args).chunk(2, dim=1)[0]pred_cond, pred_uncond = pred.chunk(2, dim=0)