rotate the camera around the object, and change the radius and elevation periodically

Initialize the logging folder and wandb run Given: args: output_path: path to save the output wandb_host: wan

r""" Initialize model parameters using Xavier initialization.

r""" Initialize model parameters using Xavier initialization.

Load the model using torch.distributed.checkpoint ckpt_path: str, dir path where the checkpoint is saved shard_strategy: str, either "ful

Main function to convert all DL3DV benchmark data.

W0, W1: [B, K, M] X0, X1: [B, K, N] W2: [B, M, N] X2: [B, K, M]

One step for the second-order multistep DPMSolver. Args: model_output_list (`List[torch.Tensor]`): The di

One step for the third-order multistep DPMSolver. Args: model_output_list (`List[torch.Tensor]`): The dir

One step for the UniC (B(h) version). Args: this_model_output (`torch.Tensor`): The model outputs at `x_

One step for the UniP (B(h) version). Alternatively, `self.solver_p` is used if is specified. Args: model_output (`torch

x of shape [B, C, F, H, W] pad zeros to the begging frames, pad self.patch_size[0] -1 frames with all zeros in the beginning

Plot logit normal distribution weighting function Args: num_train_timesteps: Number of training timesteps mu: Mean in lo

A few custom initialization steps that should be called after the object is created. Currently, the only one we have is to bind a few

Block causal LaCT with SwiGLU fast weight function. Apply then Update => Shifted Block Causal LaCT w0, w1, w2 are the fast weights. f

Block causal LaCT with SwiGLU fast weight function. Apply then Update => Shifted Block Causal LaCT w0, w1, w2 are the fast weights. f

Block causal LaCT with SwiGLU fast weight function. Apply then Update => Shifted Block Causal LaCT w0, w1, w2 are the fast weights. f

Args: x: [B, n, D2], fp32 dx: [B, n, D2], bf16 dx_momentum: [B, n, D2], fp32 momentum_mu: [B], fp32 tgt_s

Process one scene following the rules from process_dl3dv.py This is copied from Ziwen Chen's code here: https://github.com/arthurhero/Long-L

Shapes: W0: [B, M, K] W1: [B, K, M] W2: [B, M, K] X : [B, N, K] dOut (a.k.a. V): [B, N, K] lr0, lr1, lr2: [B,

In previous fwd pass: dx0 = lr0 * dh * x2 * sigma * (1 + x0 * (1 - sigma)) dx2 = lr2 * dh * silu(x0) hidden_lr1 = lr1 * x2 * silu(x0)

q: [b, s, n_h, d] k: [b, s, n_h, d]

q: [b, s, n_h, d] k: [b, s, n_h, d]

Resize and crop image to target_size, adjusting camera parameters accordingly. Args: image: PIL Image target_size: (heig

Will be used to apply activation checkpointing to the WanAttentionBlock. Check if the submodule is a WanAttentionBlock.

Save the model using torch.distributed.checkpoint ckp_path: str, dir path to save the checkpoint

Save tensor as RGB image.

config: config object should have the following keys: - train: train config. - train.fsdp_modules: list of modules to wrap with FSDP.

Args: dy: [b, d, l], gradient of the outer loss wrt the y x: [b, d, l], input of the silu activation outs: dx: [b, d,

Args: dy: [b, d, l], gradient of the outer loss wrt the y x: [b, d, l], input of the silu activation outs: dx: [b, d,

Args: dy: [b, d, l], gradient of the outer loss wrt the y x: [b, d, l], input of the silu activation outs: dx: [b, d,

q: [b, l, h, d] k: [b, l, h, d] v: [b, l, h, d] kv_cache: [2, b, l, h, d] Three cases: 1. first ar noise chunk, where kv_ca

q: [b, l, h, d] k: [b, l, h, d] v: [b, l, h, d] kv_cache: [2, b, l, h, d] Three cases: 1. first ar noise chunk, where kv_ca

Note: 1. need all rank have the same size and type of x. 2. it will auto-pad and shard among the scatter_dim given the sp world s

Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the multistep UniPC.

Reference implementation in PyTorch (for correctness checks). Shapes / layouts (contiguous, unless noted): W0: [B, M, K] W1: [B,

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features

image_size is (h, w) or just a int (as size).