initialize Sana models

initialize Sana models

Inject ``extra`` kwargs into ``kwargs``, slicing temporal dims. Tensors whose temporal axis is longer than ``num_chunk_frames`` are sliced to

Closed-form inverse of a 4x4 SE(3) batch.

Manage temporal grouping state for chunked processing.

Manage trim state for removing temporal padding.

Apply the RMSNorm normalization to the input tensor. Args: x (torch.Tensor): The input tensor. Returns:

Apply the RMSNorm normalization to the input tensor. Args: x (torch.Tensor): The input tensor.

Per-frame ``(yaw, pitch)`` in ``[-1, 1]`` with EMA smoothing.

normalize the prompts that may be nested (list/tuple) or contain single element tuples to List[str]. rules: - if it is a string, retur

normalize the prompts that may be nested (list/tuple) or contain single element tuples to List[str]. rules: - if it is a string, retur

Build raymap + chunk_plucker tensors the model consumes.

``"w-10,iw-5,none-3"`` → list of per-frame held-key lists.

Monkey-patch ``model.forward_long`` to inject sliced camera tensors.

Return ``(N-1, 3)`` per-frame translation and YXZ-euler rotation (deg).

Returns (d_splits, nw_override, ns_override, acc_override). `d_splits=1` → use baseline `_phase_b_kernel` with `_CHUNKWISE_TUNING` config. `d

Convert a decoded pixel chunk ``[-1, 1]`` to a CPU uint8 ``(B,T,H,W,3)``. The float->uint8 conversion and the channel-last permute/contiguous run

Pop all keys from ``model_kwargs`` except ``mask`` and ``data_info``. The popped entries are the "extras" (camera tensors, RoPE caches, etc.)

Forward through the transformer on the ACTIVE BLOCK ONLY and return x0. The active block's Q attends to ``[prefix, current]`` K/V via the

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

Get the prior KL term for the variational lower-bound, measured in bits-per-dim. This term can't be optimized, as it only dep

Apply special encoding to the first frame, following the logic in _run_generator. Args: frames: frame sequence [batch_si

Chunk-causal AR refinement — thin wrapper around ``RefinerChunkRunner``. Implements the canonical ``rf_shifted_sink`` KV-cache contract end-t

Returns (prec, dot_prec, state_fp32, num_warps). Uses ``PRECISION_OVERRIDE`` when set; otherwise falls back to ``_kcfg()`` (which picks ``STA

Materialise the five checkpoint paths, fetching from the Hub on first use. With ``--streaming_root`` set we read a LOCAL bundle (back-compat); ot

Memory-first path: finish stage-1, offload it, then refine/decode.

Initialize the inference pipeline for visualization on CPU, to be moved to GPU only when needed.

Determine if a parameter should be quantized to 8bit Rules: 1. linear layers: parameter size > min_8bit_size 2. 1x1 conv laye

Apply pixel shuffle and trim to output.

Temporarily remove TE init closures that are not used after construction.

Locate a CUDA toolkit with *usable* headers (cuda_runtime.h, nv/target, crt/host_config.h) and an nvcc, returning ``(home, include_dir)``. So

Decode with temporal tiling for long videos.

Encode with temporal tiling for long videos.