support want2v lora train

[gushiqiao]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces support for WanT2V training and inference, adding new datasets (WanT2VVideoDataset, WanT2VCachedDataset, PromptDataset), a WanT2VInferencer, a WanT2VModel, and dynamic time-shifting scheduling. Feedback focuses on improving robustness and performance: handling empty video paths and directory checks in dataset resolution, wrapping metadata parsing in try-except blocks, allowing configurable batch sizes, caching VAE latent statistics to avoid redundant tensor creation, and preventing a potential division-by-zero error in the dynamic time-shift scheduler.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Wrap the integer parsing of height and width in a try-except block to prevent the dataset loading from crashing due to malformed metadata.

Suggested change

	def _maybe_add_target_hw(self, sample, height, width):
	if height in (None, "") or width in (None, ""):
	return
	sample["target_height"] = int(height)
	sample["target_width"] = int(width)
	def _maybe_add_target_hw(self, sample, height, width):
	if height in (None, "") or width in (None, ""):
	return
	try:
	sample["target_height"] = int(height)
	sample["target_width"] = int(width)
	except ValueError:
	logger.warning("Invalid height/width values: height={}, width={}", height, width)

[gemini-code-assist]

Allow the batch_size to be configured from data_config instead of hardcoding it to 1.

Suggested change

	return DataLoader(
	dataset,
	batch_size=1,
	shuffle=shuffle if sampler is None else False,
	sampler=sampler,
	num_workers=data_config.get("num_workers", 8),
	pin_memory=data_config.get("pin_memory", True),
	)
	return DataLoader(
	dataset,
	batch_size=data_config.get("batch_size", 1),
	shuffle=shuffle if sampler is None else False,
	sampler=sampler,
	num_workers=data_config.get("num_workers", 8),
	pin_memory=data_config.get("pin_memory", True),
	)

[gemini-code-assist]

Pre-compute and cache latent_mean and latent_std on the target device during initialization to avoid recreating these tensors on every training step.

Suggested change

	if self.load_vae:
	self.vae = AutoencoderKLWan.from_pretrained(
	model_path,
	subfolder="vae",
	torch_dtype=self.vae_dtype,
	).to(self.device)
	self.vae.requires_grad_(False)
	if model_config.get("enable_vae_tiling", False):
	self.vae.enable_tiling()
	if self.load_vae:
	self.vae = AutoencoderKLWan.from_pretrained(
	model_path,
	subfolder="vae",
	torch_dtype=self.vae_dtype,
	).to(self.device)
	self.vae.requires_grad_(False)
	if model_config.get("enable_vae_tiling", False):
	self.vae.enable_tiling()
	self.latent_mean = torch.tensor(self.vae.config.latents_mean, device=self.device, dtype=self.vae_dtype).view(1, self.vae.config.z_dim, 1, 1, 1)
	self.latent_std = 1.0 / torch.tensor(self.vae.config.latents_std, device=self.device, dtype=self.vae_dtype).view(1, self.vae.config.z_dim, 1, 1, 1)

[gemini-code-assist]

Use the cached self.latent_mean and self.latent_std attributes instead of recreating them on every call.

Suggested change

	latent_mean = torch.tensor(self.vae.config.latents_mean, device=self.device, dtype=self.vae_dtype).view(1, self.vae.config.z_dim, 1, 1, 1)
	latent_std = 1.0 / torch.tensor(self.vae.config.latents_std, device=self.device, dtype=self.vae_dtype).view(1, self.vae.config.z_dim, 1, 1, 1)
	latent = (latent - latent_mean) * latent_std
	latent = (latent - self.latent_mean) * self.latent_std

[gemini-code-assist]

Use the cached self.latent_mean and self.latent_std attributes instead of recreating them on every call.

Suggested change

	latents_mean = torch.tensor(self.vae.config.latents_mean, device=self.device, dtype=self.vae_dtype).view(1, self.vae.config.z_dim, 1, 1, 1)
	latents_std = 1.0 / torch.tensor(self.vae.config.latents_std, device=self.device, dtype=self.vae_dtype).view(1, self.vae.config.z_dim, 1, 1, 1)
	latent = latent.to(dtype=self.vae_dtype) / latents_std + latents_mean
	latent = latent.to(dtype=self.vae_dtype) / self.latent_std + self.latent_mean