Close the dataset.

grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/

Generates embeddings and an attention mask for the given prompt(s), stripping the template part. Args: prompt (U

Take the mean over all non-batch dimensions.

Apply rotary embeddings (delegates to compiled ``_apply_rotary_emb``).

Pull dot_precision from `_resolve_launch_config` (honors PRECISION_OVERRIDE).

Flush stdin, wait for ffmpeg to finalize, and return the output path. Idempotent — calling twice is a no-op on the second call. Raises

Compute the density for sampling the timesteps when doing SD3 training. Courtesy: This was contributed by Rafie Walker in https://github.com/hugg

Finds a pre-trained G.pt model, downloading it if necessary. Alternatively, loads a model from a local path.

q: [B, Lq, Nq, C1]. k: [B, Lk, Nk, C1]. v: [B, Lk, Nk, C2]. Nq must be divisible by Nk. q_lens

Resolve a possibly ``hf://``-prefixed path to a local filesystem path. Accepts either: * a local path (returned unchanged if it exists), or

Update sampling progress Args: step: Current step number total_steps: Total number of steps

Compute per-frame gates shared across spatial positions. Delegates to the module-level compiled ``_compute_frame_gates``.

copy non-transformer params, skip specific params

Generate the next chunk, supporting overlap to ensure temporal continuity. Args: requires_grad: whether gradients are re

Update the current scheduler with the interface's static method

Compute the intermediate time steps for sampling. Args: skip_type: A `str`. The type for the spacing of the time steps. We suppor

Compute alpha_t of a given continuous-time label t in [0, T].

recalculate and accumulate KV cache, align with ar_flow_euler_sampler.accumulate_kv_cache - cur_kv_cache[block_id] structure is [cum_vk, k_sum

Reimplement ``LTX2VideoRotaryPosEmbed.prepare_video_coords`` with explicit per-frame positions. The default helper assumes contiguous ``torch.ara

Quantize a state tensor to 8bit Args: state_tensor: tensor to be quantized block_size: quantization block size

Get a term for the variational lower-bound. The resulting units are bits (rather than nats, as one might expect). This allows

Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.

Load state dict and convert relevant states to 8bit

Move only the modules and direct parameters used by the video-only forward.

Compute the mean and variance of the diffusion posterior: q(x_{t-1} | x_t, x_0)

Set a key to a value in a list of dictionaries.

Build pretrained InceptionV3 Parameters ---------- output_blocks : list of int Indices of blocks to return featur

Run :meth:`_apply_spatial`, chunking dim 0 to keep each call under PyTorch's 32-bit conv indexing limit. No-op for short inputs.

Compute RoPE frequencies for the local frame window.