Norm layer adaptive RMS normalization zero. Parameters: embedding_dim (`int`): The size of each embedding vector.

TeaCache parameters for `BooguImageTransformer2DModel` See https://github.com/ali-vilab/TeaCache/ for a more comprehensive understanding

Processor for implementing scaled dot-product attention with flash attention and variable length sequences. This processor is optimized for

Output class for BooguImagePipeline. Args: images (Union[List[PIL.Image.Image], np.ndarray]): List of denoised PIL image

r""" A feed-forward layer. Parameters: hidden_size (`int`): The dimensionality of the hidden layers in the model. This pa

Processor for implementing scaled dot-product attention with flash attention and variable length sequences. This processor implements: -

Double-stream self-attention processor without flash attention. This processor implements double-stream attention where: - Instruction a

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

Double-stream self-attention processor with flash attention and variable length sequences. This processor implements double-stream attention

Boogu-Image double-stream block. Here "double-stream" is the same idea as a "dual-stream" layer: instruction tokens and image tokens a

Boogu-Image image processor, with resize/crop behavior adapted from PixArt's image processor implementation. This class keeps a Diffuser

Rotary Position Embedding for Prompt Tuning tokens. This class generates rotary position embeddings specifically for prompt tuning tokens.

Basic Boogu-Image transformer block: attention + MLP + RMSNorm.

Output class for the scheduler's `step` function output. Args: prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels,

r""" Load LoRA layers into [`BooguImageTransformer2DModel`,`PromptEmbedding`]. Specific to [`BooguImagePipeline`,`BooguImageTurboPipeline`].

Base pipeline for Boogu text-to-image and image-editing inference. The pipeline coordinates the main components used by Boogu inference:

Boogu-Image pipeline variant with prompt-tuning support. This class keeps the generation behavior of `BooguImagePipeline` while adding a

Boogu-Image transformer with mixed stream topology. Early layers use double-stream (aka dual-stream) processing, then switch to single

`BooguImagePipeline` plus a DMD student few-step inference path. Enable it by passing `use_dmd_student_inference=True` to `__call__`. The DMD

Euler scheduler. This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic met

Simple RMS Normalization implementation using native PyTorch operations. This is a pure PyTorch implementation that matches the functionalit