Build the embeddings processor on the target device.

Build the checkpoint-key remapping for the Gemma multimodal encoder. The vision-tower mapping differs between transformers <5 and >=5 because

Pre-load each block into its own contiguous pinned CPU buffer with LoRA fusion.

Build sd-ops that fold prequant ``*_scale`` siblings into their parent tensor at load time. *scales* is keyed by the **post-rename** param key

Return a DataLoader over this rank's interleaved shard of ``dataset``. When ``overwrite`` is False, items whose outputs already exist (per ``i

Return a DataLoader over this rank's interleaved shard of ``dataset``. When ``overwrite`` is False, items whose outputs already exist (per ``i

Build a binary mask from a pixel-space spatial region (y1, x1, y2, x2).

Build STG perturbation config that skips self-attention in the specified blocks.

Build the Gemma text encoder (non-streaming path).

Build the sd_ops downcast map from the same suffix registry as the matcher.

Construct the `vllm serve ...` argv.

Load text encoder, encode all validation prompts, cache on CPU, then unload.

Caption a video, sending its audio track as a separate modality. vLLM does not extract a video's audio on its own (and its ``use_audio

Causal-pad *x*, free it, then run the raw ``nn.Conv3d``. This avoids the peak where both the original and padded tensors are live simultaneous

Return names of any parameters/buffers still on meta device.

Clamp frame, height, and width tilings to the latent's extents. ``latent_shape`` is ``(F, H, W)`` in latent units.

Remove common LLM-generated prefixes from captions.

Clean up old checkpoints.

Wrap an iterator to clean up *model* memory once it is exhausted or abandoned.

Clean up old training state files, using the same keep_last_n as checkpoints.

Collect trainable parameters based on training mode.

Return a list of .mp4/.mov files from *input_path* (file or directory).

Builds a `(B, 1, T, 2)` tensor containing timestamps for each latent frame. This helper method underpins `get_patch_grid_bounds` for

Compute log-mel spectrogram from waveform using causal STFT bases. Args: audio: Waveform tensor of shape (B, C, T). Return

Compute the index permutation that places valid tokens before pads in each row. Stable sort: valid tokens keep their relative order. Idempotent fo

Compute float mask for spatial crop region (y1, x1, y2, x2) in pixel coords. Returns [B, seq_len] in {0, 1}.

Compute VAE-aligned temporal subsample indices. Keeps frame 0 (the VAE's standalone first-frame latent), then takes every ``factor``-th frame

Concatenate two [C, F, H, W] videos horizontally, matching height and padding frames.

Check if a validation condition targets the audio modality.

Check if a validation condition targets the video modality.

Enqueue H2D copy + LoRA fusion on the copy stream and wait on compute. *cpu_buffer* is one contiguous source buffer carved by the same layout

Create AudioLatentTools for the given video duration.

Create conditioning items for video generation. Args: conditioning_attention_strength: Scalar attention weight in [0, 1].

Select and create the appropriate feature extractor based on config. Detection logic: - V1: V2 config keys absent → projection lives in transf

Generate and save a model card for the trained model.

Create learning rate scheduler based on config.

Create VideoLatentTools for the given output dimensions.

Decode latent to HDR video, optionally cropping to target size. Args: crop_size: ``(width, height)`` to crop decoded frames to, or

Decode latents to video frames.

Discover all valid sample files across all data sources. Uses a fast two-pass approach: first globs all sources in parallel to build f

Run libx264 frame.encode + container.mux on a background thread while the caller produces numpy chunks on the current thread. The 1-slot queue

Load VAE encoders, encode all conditioning media, cache on CPU, then unload.

Encode a conditioning image to latent space. Returns [B, C, 1, H', W'] on CPU.

Encode all conditioning media for a single validation sample.

Encode media for prefix/suffix/mask/spatial_crop/frozen-video conditions.

Encode video into non-patchified latent representation. Args: vae: Video VAE encoder model video: Input tensor of shape [B, C, F,

Schedule a read for *idx* if one is not already pending.

Extract audio track from a video file, trimmed/padded to match video duration.

Extract the audio track to a 16 kHz mono PCM WAV (matches pretraining). Raises ``CalledProcessError`` when the video has no audio stream.

Add a valid sample to the sample_files tracking.

Filter out videos with insufficient frames.

Workspace-based final norm + [ada] + SiLU + conv_out + unpatchify.

Find the checkpoint file to load, handling both file and directory paths.

Find chunk size for in-place temporal convolution. The chunk size ensures the last chunk has at least 3 frames (the temporal kernel size), avo

Fuse LoRA deltas directly into GPU block weights via ``fuse_rule``.

Generate one sample: build conditioned states, denoise, decode.

Compute ``(sigma_down, sigma_up)`` for one DDIM ancestral sampling step. Both inputs are in the rescaled parameterization ``sigma / alpha``. R

Map sigma value to bucket index.

Get all media files from the input path.

Get set of existing reference video paths without loading the actual files. Args: output_path: Path to the reference video paths file

Move cached prompt embeddings to device, returning (v_pos, a_pos, v_neg, a_neg).

Map precision string to quanto dtype.

Group tiles by their temporal output slice.

Infer the reference downscale factor from target and reference dimensions.

Infer spatial and temporal scale factors by peeking at one sample pair.

Infer spatial scale factor between condition and target resolutions.

Infer temporal scale factor between condition and target latent frame counts. The first latent frame encodes a single pixel frame (the VAE's c

Initialize audio-specific components.

Initialize audio-video cross-attention components.

Initialize the training data loader using the strategy's data sources.

Initialize the optimizer and learning rate scheduler.

Initialize preprocessors for LTX.

Initialize the timestep sampler based on the config.

Initialize transformer blocks for LTX.

Initialize video-specific components.

Initialize Weights & Biases run, resuming an existing run if its id is provided.

Add noise to latents and create training targets. Returns (noisy, targets, timesteps, mask, sigmas).

Return True if *module_name* names a Linear that should be fp8-cast.

Create an ``FP8Linear`` matching the shape/bias of *layer*.

Linear -> sRGB OETF per IEC 61966-2-1. Input assumed in [0, 1].

Load a mask image/video and downsample to latent-space dimensions. Returns a binary float tensor of shape [1, F', H', W'] where F', H', W' are

Load an audio mask (.pt or .wav) and resample to target audio latent length. Returns a float tensor of shape [1, num_tokens] in patchified tok

Load captions and compute their output embedding paths.

Load captions from a CSV file and compute output embedding paths.

Load captions from a JSON file and compute output embedding paths.

Load captions from a JSONL file and compute output embedding paths.

Load checkpoint if specified in config, then resolve resume state.

Load VAE decoder, audio decoder, and vocoder. Kept on CPU until generation.