Configuration to use when quantizing a tensor. Args: backend (QuantizeBackend): The backend to use for quantization. If no backend i

A quantized tensor.

Causal 3d convolusion.

Causal 3d convolusion.

Configuration for quantizing a model with Four Over Six. Args: activation_scale_rule (ScaleRule | None): The scaling rule to use for

`UniPCMultistepScheduler` is a training-free framework designed for the fast sampling of diffusion models. This model inherits from [`Schedu

Block scale selection rules for NVFP4 quantization. - `abs_max`: Between 4 and 6, select the block scale that minimizes the maximum

Dataset that concatenates multiple videos from a folder into a fixed-length video. Each item consists of multiple video segments concatenated

Data types.

`FlowDPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs. This model inherits from [`SchedulerMixin`] and [

Text-only dataset for multi-shot training and inference. Supports two input modes: **txt file** — each line is one caption. Each sample uses

Causal self-attention with a single head.

Backends for quantizing a tensor to NVFP4 or MXFP4. - `cuda`: CUDA implementation. Requires a Blackwell GPU, and currently only supports

Rounding styles for quantization. - `nearest`: Round to the nearest FP4 value. - `stochastic`: Round to the nearest FP4 value after appl

Causal self-attention with a single head.

Fouroversix deserializer for gpt oss model.

LongLive2.0 diffusion inference pipeline using Ulysses sequence parallelism.

Bucketed ring buffer for storing prediction errors on CPU. Two layouts are supported: * **1D (timestep-only)** — when ``num_blocks <= 0``.

A PTQ experiment with results.

Drop-in replacement for GptOssExperts layer that uses FP4 quantization.

Representation for a kernel that quantizes a tensor to FP4.

Evaluation coordinator for running PTQ experiments locally.

Backends for matrix multiplication with FP4. - `cutlass`: CUTLASS implementation. This requires a Blackwell GPU. - `pytorch`: PyTorch im

Evaluation coordinator for running PTQ experiments on Modal.

Configuration for quantizing modules with Four Over Six. Args: activation_scale_rule (ScaleRule | None): The scaling rule to use for

Wan 5B diffusion wrapper backed by the Ulysses SP causal model.

Lazy torch.compile wrapper that falls back to eager on compile/runtime errors.

Optimize a model with SpinQuant.

A lightweight wrapper that routes a linear layer through TransformerEngine.

Causal self-attention with Ulysses sequence/head exchange.

Evaluate a model using AWQ.

Base class for evaluation coordinators.