Read a grouped config value, falling back to legacy flat names.

Return the block size if this a block-scaled format, or `None` otherwise.

Quantize a tensor to FP4. ## Sample Code ### With Four Over Six ```python x = torch.tensor(1024, 1024, dtype=torch.bfloat16, d

Return True if the backend is available on the current machine.

Perform a matrix multiplication (`a @ b.T`) between two quantized tensors. ## Sample Code Each tensor may be provided in either high or

Get video information without loading frames (cached).

Return the maximum allowed E2M1 value for the rule.

Return the scale dtype if this a block-scaled format, or `None` otherwise.

Diffusion forward corruption process. Input: - clean_latent: the clean latent with shape [B, C, H, W] - noise

Return the maximum allowed E4M3 value for the rule.

Quantize model with FourOverSix and optionally skip selected modules. `filtered_modules` supports: - Exact module names, e.g. "head.head

Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`):

Convert [B, S_local, N, D] to [B, S_total, N_local, D].

Return the quantization configuration for the weight tensors.

Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`. Args: seed (`int`): The seed to

In-place copy a QuantizedTensor's data into a pre-allocated slot. Keeps the slot's `values`/`scale_factors`/`amax` buffers persistent (their

Get the path where this submodule will be installed in the Modal image.

Get an SQLAlchemy session for the SQLite database.

Return True if the backend supports the given inputs and output data type.

Expand grouped release configs into the flat runtime schema. The training and inference code historically reads fields such as ``config.batch

x: [B, L, N, C]. grid_sizes: [B, 3]. freqs: [M, C // 2].

Check if we can sample num_frames starting from start_frame. Returns True if we can sample num_frames without exceeding video bounds.

Remove FSDP wrapper prefixes from parameter names.

Same item-list format as ``_gather_errors_for_buffer`` but with no collective — used outside the warmup window or when distributed is off.

Remove FSDP wrapper prefixes used by some EMA checkpoints.

Configure LoRA for a WanDiffusionWrapper model Args: transformer: The transformer model to apply LoRA to model_name: 'generat

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

Return the quantization configuration for the activation tensors.

Get the evaluator class for the given PTQ method.

Replace eligible modules with FourOverSix NVFP4 modules using the runtime config.

Extract the generator state dict from common LongLive checkpoint layouts.

Randomly generate a timestep tensor based on the generator's task type. It uniformly samples a timestep from the range [min_timestep,

Store a single block error into the matching bucket. Args: error_block: (block_size, C, H, W) tensor timestep_index:

Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'.

Clone a QuantizedTensor by cloning its internal tensors.