Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings to the given query or key

Factory function that returns a validator for device arguments. Base supported formats: 'cpu', 'cuda', or 'cuda:x' (where x is an integer).

Normalize input_images into a two-level batch structure with per-sample lists: - List[List[PIL.Image.Image]] or - Li

Initialization for cache.

r""" Returns target `(height, width)` after optional downscaling and rounding to `vae_scale_factor` multiples. Args:

Format the instruction with the system prompt. If `return_str` is True, it will call `self.instruction_rewriter_processor.tokenizer.a

Get interleaved instruction embeddings from VLM (self.mllm), aligned with training: - Build VLM inputs via processor.apply_chat_templ

Get system text prompts for rewriting text instructions. Returns a tuple of lists: (rewrite_text_prompts, magic_prompts)

To duplicate text embeddings and attention mask for each generation per instruction, using mps friendly method

Resize input PIL images for VLM encoding, matching dataset behavior exactly as in BOOGUTrainTorchIterableTI2IDataset.preprocess_vlm_i

Initialize the attention processor.

Build chat template content with interleaved text and images. If `system_prompt_follows_task_type` is True, the system prompt will be

Concatenate instruction (text & image) and reference image features (instruction first, then image). Args: img_hidden_st

Concatenate instruction (text & image) and reference image features (instruction first, then image). Args: img_hidden_st

Project m0 onto m1 by treating each [H, W] slice as a matrix. Args: m0: Input tensor to be decomposed, shape [B, C, H, W]

"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at t

Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.prompt_embedding`. All kwargs are forwarded to `sel

Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and `self.text_encoder`. All kwargs ar

r""" Return state dict for lora weights and the network alphas. <Tip warning={true}> We support loading A1111 formatted LoRA

Normalize platform tags the same way as the mjunya wheel browser.

Preprocess the input images.

Preprocess the input images.

Set the processor used by the local instruction rewriter.

mllm's setter

processor's setter

Set or clear the prompt-tuning embedding module.

scheduler's setter

transformer's setter

Initialize Taylor storage for the first step/module access. The target location is `cache_dic['cache'][-1][stream][layer][module]`.

Validate whether the device and offload strategy are compatible.

Format the instruction with the system prompt. `input_images` could be List[str] or List[PIL.Image.Image]. `List[str]` means a list o

Initialize the step_index counter for the scheduler.

Batched Newton-Schulz iteration. Accepts: - (H, W) -> returns (H, W) - (N, H, W) -> returns (N, H, W)

Rewrite input text instructions with input images using self.text_instruction_rewriter. Supports batch inputs (list[str]). Returns a

Rewrite input text instructions using self.text_instruction_rewriter. Supports batch inputs (list[str]). Returns a list[str] where ea

x: [N] return: kernel of x

Simple RMS normalization implementation using native PyTorch. Args: x: Input tensor [..., hidden_size] weigh

x: [N] or [B, N] return: lambda * softmax(x / tau)

Split concatenated features back to instruction and image features. Inverse operation of _concat_instruction_image_features.

Split concatenated features back to instruction and image features. Inverse operation of _concat_instruction_image_features.

Unpad the input tensors for flash attention. Same implementation as BooguImageAttnProcessorFlash2Varlen.

Unpad the input tensors for flash attention. Args: query_layer: Query tensor of shape (batch_size, seq_len, num_heads, h

Apply Cache-DiT caching to the pipeline. Args: pipeline: The pipeline to apply Cache-DiT caching to. all_layers: Whether