Embed images with SigLIP and language tokens with embedding layer.

@func: get the loss and ema config of the model

Returns config for specified gemma variant.

Pad the last dimension of a vector to new_dim with zeros. Can be (batch_size x sequence_length x features_dimension) or (batch_size x feature

@func: fetch the models needed to reserve

Fix state dict keys to match current model architecture.

Build and cache orthonormal DCT-II matrix of size (L, L). For norm='ortho', inverse is its transpose (IDCT-III).

Return both video and image files (used for one-shot where a cell may be an image).

Resolve checkpoint and config paths from a checkpoint file or directory. Rules: - If input is a checkpoint file, load config from its

@func: returns a state dictionary for the optimizer and scaler

@func: initial the vae

@func: ema update in order to get a more stable version

quant actions with vqvae and prepare for Expert Gemma processing.

@func: filer the params for weight decay

Generate a manifest for one-shot transfer videos. Expected folder structure: static/videos/one_shot/<Category>/<Task>/<Method>/*.

Generate a manifest for real-world videos. Expected folder structure: static/videos/real_world/<Task>/{MINT,ACT,Pi0,Pi05*}.mp4 Outpu

Get a safe dtype for the given device type.

Enable gradient checkpointing for memory optimization.

@func :

@func :