Use SanaVideoPipeline to generate video latent, then use LTX2 Pipeline for Stage-2 refinement. Key technical points: 1. Manually pack vi

Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep. Ar

Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep. A

Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`)

setup new sequence

This is the shard function for T5 model.

Return frame-time positions belonging to size-1 (singleton) chunks. A size-1 chunk has no intra-chunk lookahead, so the anti-causal branch (b

Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the SA-Solver.

Performs a single optimization step. Args: closure (callable, optional): A closure that reevaluates the model and

Perform a single optimization step Main steps: 1. Determine if 8bit quantization is needed 2. Update first and second moment

Store rows of multiple numpy arrays in a single LMDB. Each row is stored separately with a naming convention.

Force the EMA parameters to be a direct copy of the given model parameters. This is used to create a snapshot for the rollout policy.

uniform: [low, high)

Compute training losses for a single timestep. :param model: the model to evaluate loss on. :param x_start: the [N x C x ...]

Compute training losses for a single timestep. :param model: the model to evaluate loss on. :param x_start: the [N x C x ...]

Input: - timestep: the timestep with shape [B*T] Output: the corresponding weighting [B*T]

Input: - timestep: the timestep with shape [B*T] Output: the corresponding weighting [B*T]

Update the reweighting using losses from a model. Call this method from each rank with a batch of timesteps and the correspon

Return tuple with min_len by repeating element at idx_repeat.

Input: grad_vk: (B, H, C+1, C), fp32 k: (B, N, H, C), fp16 v: (B, N, H, C), fp16 k_relu_mask: (B, N, H, C), bool

Input: grad_vk_q: (B, N, H, C+1), fp32 vk: (B, H, C+1, C), fp32 q: (B, N, H, C), fp16 q_relu_mask: (B, N, H, C),