Resume training from a checkpoint. the checkpoint path looks like: exp_name/seed_200/checkpoint_model_000009/ the dcp_folder is in

x: [B, L, N, C]. grid_sizes: [B, 3]. freqs: [M, C // 2].

save_path: str, dir path to save the checkpoint step: int, step number of the checkpoint model: torch.nn.Module, model to save, might be

Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`. Args: seed (`int`): The seed to

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,

calculate the number of frames for video used for model inputs. Args: ele (dict): a dict contains the configuration of video.

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

Outs: Hidden: [B, M, N] in other words [B, Hidden, num_tokens] DY0_DY2: [B, 2M, N] in other words [B, 2 * Hidden, num_tokens]

Args: W0: [B, M, K] - [B, Hidden, D] W1: [B, K, M] - [B, D, Hidden] W2: [B, M, K] - [B, Hidden, D] X: [B, N, K] -

Fused Triton kernel for the 'ref_pytorch_swiglu_bwd_bwd_fused_cat_inp_out' computation. - Computes all pointwise outputs in one pass. - R

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

r""" Initialize the diffusion model backbone. Args: model_type (`str`, *optional*, defaults to 't2v'): Mo

Convert flow matching's prediction to x0 prediction. flow_pred: the prediction with shape [B, f, c, h, w] xt: the input noisy

Convert flow matching's prediction to x0 prediction. flow_pred: the prediction with shape [B, f, c, h, w] xt: the input noisy

Extract the denoised output from the interleaved sequence. Input: interleave_output: [b * ( (num_repeat + 1) * num_windo

Extract the denoised output from the interleaved sequence. Input: interleave_output: [b * ( (num_repeat + 1) * num_windo

Initialize the step_index counter for the scheduler.

Initialize the step_index counter for the scheduler.

Prepare the input for the AR video generation. input: noisy_input: [b * num_repeat, num_ar_chunks, ar_window_size, c, h,

Prepare the input for the AR video generation. input: noisy_input: [b * num_repeat, num_ar_chunks, ar_window_size, c, h,

Given a tensor containing the whole ODE sampling trajectories, randomly choose an intermediate timestep and return the latent as wel

Given a tensor containing the whole ODE sampling trajectories, randomly choose an intermediate timestep and return the latent as wel

Given a tensor containing the whole ODE sampling trajectories, randomly choose an intermediate timestep and return the latent as wel

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

Args: q: [b, s, d] k: [b, s, d] Returns: q: [b, s, d] k: [b, s, d]

Note: Forward: w1 @ (silu(w0 @ x) * (w2 @ x)) w0, w2: [b, dh, d] w1: [b, d, dh] x: [b, l, d] Fast linear layer

Note: Forward: w1 @ (silu(w0 @ x) * (w2 @ x)) w0, w2: [b, dh, d] w1: [b, d, dh] x: [b, l, d] Fast linear layer

Note: Forward: w1 @ (silu(w0 @ x) * (w2 @ x)) w0, w2: [b, dh, d] w1: [b, d, dh] x: [b, l, d] Fast linear layer

Note: Forward: w1 @ (silu(w0 @ x) * (w2 @ x)) w0, w2: [b, dh, d] w1: [b, d, dh] x: [b, l, d] Fast linear layer

q: [b, l, h, d] k: [b, l, h, d] v: [b, l, h, d] l = update_every - mini_batch_size + update_every * (num_chunks - 1)

q: [b, l, h, d] k: [b, l, h, d] v: [b, l, h, d] l = update_every - mini_batch_size + update_every * (num_chunks - 1)

q: [b, l, h, d] k: [b, l, h, d] v: [b, l, h, d] l = update_every - mini_batch_size + update_every * (num_chunks - 1) num_repeat

q: [b, l, h, d] k: [b, l, h, d] v: [b, l, h, d] l = update_every - mini_batch_size + update_every * (num_chunks - 1) num_repeat

Bidirectional LaCT with SwiGLU fast weight function. w0, w1, w2 are the fast weights. f(x) = w1 @ (silu(w0 @ x) * (w2 @ x)) About preci

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

When checkpoint the model, we need to first cache_model() copy_to_model() then save the model weights. resto

Validates fused_four_mm_same_out_interface against a float32 PyTorch reference.

Convert a single DL3DV folder to standard format. Args: folder_path: Path to the folder containing nerfstudio/ subfolder

Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is designed to discretize an

r""" Convert the model output to the corresponding type the UniPC algorithm needs. Args: model_output (`torch.Tensor`):

Validates fused_two_mm_same_out_wT_xT_triton against a float32 PyTorch reference.

Create dl3dv_sample_data_path.json file. Args: benchmark_dir: Path to the dl3dv_benchmark directory output_file: Path to

Download the benchmark based on the user inputs. 1. download the benchmark-meta.csv 2. based on the args, download the specific subs

Given a hash, download the relevant data from the huggingface repo :param filepath_dict: the cache dict that stores all the file relative paths

One step for the first-order DPMSolver (equivalent to DDIM). Args: model_output (`torch.Tensor`): The dir

TODO: rgb or bgr? in [0,1] or [-1,1]? videos: A list of videos each with shape [C, T, H, W].

Note: Forward: (silu(x @ w0) * (x @ w2)) @ w1 w0, w2: [b, d, dh] w1: [b, dh, d] q: [b, l, d] k: [b, l, d] v: [b,

Find the latest checkpoint in the output directory. Returns: str or None: Path to the latest checkpoint directory, or None if no

r""" Args: x(Tensor): Shape [B, L1, C] e(Tensor): Shape [B, C]

r""" Args: x(Tensor): Shape [B, L1, C] e(Tensor): Shape [B, C]

For each consecutive pair of cameras, add num_views linearly interpolated views.

Get a linear function that maps x to log(y), and returns exp(y'). f(4680) = 2.1135 f(10920) = 2.2846

Get a linear function that maps x to log(y), and returns exp(y'). f(4680) = 2.1135 f(10920) = 2.2846

Get a linear function that maps x to log(y), and returns exp(y'). f(4680) = 2.1135 f(10920) = 2.2846

Update the current scheduler with the interface's static method

Update the current scheduler with the interface's static method

Update the current scheduler with the interface's static method