Save captions to a file in the specified format. Args: captions: Dictionary mapping media paths to captions output_path: Path to s

Save the model weights.

Acquire a raw slot, carve it to block *idx*, enqueue a read, return the handle. The raw slot and its fetch status are returned together so the

Hand SDPA the full backend priority order; let torch's dispatcher pick at call time. ``sdpa_kernel(_SDPA_FULL_PRIORITY, set_priority=True)`` enabl

Round ``frames`` down to the nearest ``8k+1`` (the model's required frame count).

Advance one diffusion step for a single modality, or return ``None`` if absent.

Build a streaming wrapper, yield it, then tear down and free memory.

Number of generated tokens in *tile*.

Upcast the weight to the given dtype and optionally apply stochastic rounding. Input weight needs to have float8_e4m3fn or float8_e5m2 dtype.

Take a free raw slot, waiting any pending event before returning. Raises :class:`RuntimeError` if every slot is currently in use.

Accumulate ``sum((B * strength) @ A)`` across :class:`LoraProduct` items. The first product materializes a freshly-allocated aggregator via ``

Append :class:`VideoConditionByReferenceLatent` items for each reference path.

Best-effort human-readable backend name. Built-in callables expose ``.label`` (encoding the SDPA priority list for the Pytorch backends); fall

Expand the attention mask to include newly appended conditioning tokens. Each conditioning item appends M new reference tokens to the sequenc

Return a MultiModalGuider with params effective for the given sigma.

Build the middle block with two ResNet blocks and optional attention.

The guider calculates the guidance delta as (scale - 1) * (cond - uncond) for cfg and modality cfg, and as scale * (cond - uncond) fo

Process captions and save text embeddings. Under ``accelerate launch``, each process handles an interleaved shard of the dataset (rank/wo

Generate a 1D rectangular (pulse) mask. Args: length: Output length of the mask. left_ramp: Number of elements at the start o

Convert binary attention mask to additive form for transformer masking.

Create a caption projection for the transformer (V1/19B only).

Factory function to create a media captioner.

Derive a layout from a ``{name: tensor}`` dict. If ``dtype`` is given, non-FP8 dtypes are coerced to it (FP8 preserved). If ``None``, the sour

Returns True if the guider is doing isolated modality generation.

Returns True if the guider is doing perturbed generation.

Returns True if the guider is doing unconditional generation.

Build a factory from a dict of sigma_value -> MultiModalGuiderParams. Keys are sorted descending and used for bin lookup in params(si

Text-to-video training requires latents and text conditions. When ``with_audio`` is True, also requires audio latents.

Load the LTX transformer model. Args: checkpoint_path: Path to the safetensors checkpoint file device: Device to load model on

Log validation outputs (images, videos, or audio) to Weights & Biases.

Return the model configuration dictionary extracted from the checkpoint metadata.

Parse resolution buckets from string format to list of tuples (frames, height, width)

Parse a timecode string into a FrameTimecode object. Supports formats: - Frames: '123' - Seconds: '123s' or '123.45s' - Timecode: '00:

Read ``reference_downscale_factor`` from LoRA safetensors metadata (default 1).

Attach a compute-done event -- waited before this buffer is recycled.

Run denoising with a pre-built transformer. Same semantics as ``__call__`` but accepts a pre-built transformer so the model can be sha

Run features through the middle block.

Sample timesteps for a batch from a stretched shifted logit-normal distribution. Args: batch_size: Number of timesteps to sample

Returns True if the guider should skip the step.

Build the :class:`QuantizationPolicy` for this kind. ``checkpoint_path`` is required for both backends: ``FP8_SCALED_MM`` uses it to d

Convert the sample and its denoised version to velocity. Returns: Velocity

Rearrange channels back into spatial dimensions. Inverse of patchify - moves pixels from channels back into patch_size x patch_size blocks (d

Converts latent tensors between spatio-temporal formats and flattened sequence representations. Args: latents: Patch toke

Block until the read finishes; return its error, or ``None`` on success.

Convert waveform to log-mel spectrogram [batch, channels, time, n_mels].

Initialize the Encoder. Args: Arguments are configuration parameters, loaded from the audio VAE checkpoint config

Decode and accumulate all tiles of a temporal group into a local buffer. The buffer is local to the group and always starts at time 0

Adjust output shape to match target dimensions for variable-length audio. This function handles the common case where decoded audio s

Return the set of ``.weight`` keys touched by at least one LoRA in the list.

Vectorized (num_tiles, num_cond) bool: which tiles keep each conditioning token. A conditioning token is kept by a tile when its ``[start, end

Allocate exactly ``nbytes`` of pinned host memory via ``cudaHostRegister``. Bypasses PyTorch's ``CachingHostAllocator``, which rounds every ``

Apply all audio-targeting conditionings from the sample's conditions list.

Apply intrinsic conditioning using a binary mask. For each token, the mask value determines conditioning strength: - mask=1: condition

Concatenate reference latents to target sequence for reference conditioning (IC-LoRA style). The apply/skip decision is batch-wide (reference

Concatenate reference video pixels side-by-side with generated output if requested.

Apply all video-targeting conditionings from the sample's conditions list.

Save to ``out`` atomically via per-PID temp file + replace. Crash mid-write leaves an orphan ``.tmp.<pid>`` file that the skip logic ignores.

Convert an audio frame to a float32 ndarray with values in [-1, 1] and shape (channels, samples).

Per-block layout of the streamed tensors, taken from the meta model. Blocks may differ in shape and even in which tensors they have (e.g. Gemma4's

Build the argument parser for HDR IC-LoRA batch inference.

Build metadata dictionary for safetensors checkpoint. Delegates to the training strategy to get strategy-specific metadata that downst

Create a DiskWeightSource backed by a DiskBlockReader. Pool slots are sized to the largest block and carved per block on read, so hete