# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
from typing import Any, Callable

import numpy as np
import PIL
import torch
from transformers import Qwen2TokenizerFast, Qwen3ForCausalLM

from ...loaders import Flux2LoraLoaderMixin
from ...models import AutoencoderKLFlux2, Flux2Transformer2DModel
from ...models.transformers.transformer_flux2 import Flux2KVAttnProcessor, Flux2KVParallelSelfAttnProcessor
from ...schedulers import FlowMatchEulerDiscreteScheduler
from ...utils import is_torch_xla_available, logging, replace_example_docstring
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline
from .image_processor import Flux2ImageProcessor
from .pipeline_output import Flux2PipelineOutput


if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from PIL import Image
        >>> from diffusers import Flux2KleinKVPipeline

        >>> pipe = Flux2KleinKVPipeline.from_pretrained(
        ...     "black-forest-labs/FLUX.2-klein-9b-kv", torch_dtype=torch.bfloat16
        ... )
        >>> pipe.to("cuda")
        >>> ref_image = Image.open("reference.png")
        >>> image = pipe("A cat dressed like a wizard", image=ref_image, num_inference_steps=4).images[0]
        >>> image.save("flux2_kv_output.png")
        ```
"""


# Copied from diffusers.pipelines.flux2.pipeline_flux2.compute_empirical_mu
def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float:
    a1, b1 = 8.73809524e-05, 1.89833333
    a2, b2 = 0.00016927, 0.45666666

    if image_seq_len > 4300:
        mu = a2 * image_seq_len + b2
        return float(mu)

    m_200 = a2 * image_seq_len + b2
    m_10 = a1 * image_seq_len + b1

    a = (m_200 - m_10) / 190.0
    b = m_200 - 200.0 * a
    mu = a * num_steps + b

    return float(mu)


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
    scheduler,
    num_inference_steps: int | None = None,
    device: str | torch.device | None = None,
    timesteps: list[int] | None = None,
    sigmas: list[float] | None = None,
    **kwargs,
):
    r"""
    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.

    Args:
        scheduler (`SchedulerMixin`):
            The scheduler to get timesteps from.
        num_inference_steps (`int`):
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
        device (`str` or `torch.device`, *optional*):
            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        timesteps (`list[int]`, *optional*):
            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
            `num_inference_steps` and `sigmas` must be `None`.
        sigmas (`list[float]`, *optional*):
            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
            `num_inference_steps` and `timesteps` must be `None`.

    Returns:
        `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
        second element is the number of inference steps.
    """
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
    if timesteps is not None:
        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accepts_timesteps:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" timestep schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    elif sigmas is not None:
        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accept_sigmas:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" sigmas schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
def retrieve_latents(
    encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample"
):
    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
        return encoder_output.latent_dist.sample(generator)
    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
        return encoder_output.latent_dist.mode()
    elif hasattr(encoder_output, "latents"):
        return encoder_output.latents
    else:
        raise AttributeError("Could not access latents of provided encoder_output")


class Flux2KleinKVPipeline(DiffusionPipeline, Flux2LoraLoaderMixin):
    r"""
    The Flux2 Klein KV pipeline for text-to-image generation with KV-cached reference image conditioning.

    On the first denoising step, reference image tokens are included in the forward pass and their attention K/V
    projections are cached. On subsequent steps, the cached K/V are reused without recomputing, providing faster
    inference when using reference images.

    Reference:
    [https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence)

    Args:
        transformer ([`Flux2Transformer2DModel`]):
            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKLFlux2`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`Qwen3ForCausalLM`]):
            [Qwen3ForCausalLM](https://huggingface.co/docs/transformers/en/model_doc/qwen3#transformers.Qwen3ForCausalLM)
        tokenizer (`Qwen2TokenizerFast`):
            Tokenizer of class
            [Qwen2TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2TokenizerFast).
    """

    model_cpu_offload_seq = "text_encoder->transformer->vae"
    _callback_tensor_inputs = ["latents", "prompt_embeds"]

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKLFlux2,
        text_encoder: Qwen3ForCausalLM,
        tokenizer: Qwen2TokenizerFast,
        transformer: Flux2Transformer2DModel,
        is_distilled: bool = True,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            scheduler=scheduler,
            transformer=transformer,
        )

        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
        self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
        self.tokenizer_max_length = 512
        self.default_sample_size = 128

        # Set KV-cache-aware attention processors
        self._set_kv_attn_processors()

    @staticmethod
    def _get_qwen3_prompt_embeds(
        text_encoder: Qwen3ForCausalLM,
        tokenizer: Qwen2TokenizerFast,
        prompt: str | list[str],
        dtype: torch.dtype | None = None,
        device: torch.device | None = None,
        max_sequence_length: int = 512,
        hidden_states_layers: list[int] = (9, 18, 27),
    ):
        dtype = text_encoder.dtype if dtype is None else dtype
        device = text_encoder.device if device is None else device

        prompt = [prompt] if isinstance(prompt, str) else prompt

        all_input_ids = []
        all_attention_masks = []

        for single_prompt in prompt:
            messages = [{"role": "user", "content": single_prompt}]
            text = tokenizer.apply_chat_template(
                messages,
                tokenize=False,
                add_generation_prompt=True,
                enable_thinking=False,
            )
            inputs = tokenizer(
                text,
                return_tensors="pt",
                padding="max_length",
                truncation=True,
                max_length=max_sequence_length,
            )

            all_input_ids.append(inputs["input_ids"])
            all_attention_masks.append(inputs["attention_mask"])

        input_ids = torch.cat(all_input_ids, dim=0).to(device)
        attention_mask = torch.cat(all_attention_masks, dim=0).to(device)

        # Forward pass through the model
        output = text_encoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            output_hidden_states=True,
            use_cache=False,
        )

        # Only use outputs from intermediate layers and stack them
        out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1)
        out = out.to(dtype=dtype, device=device)

        batch_size, num_channels, seq_len, hidden_dim = out.shape
        prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim)

        return prompt_embeds

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_text_ids
    def _prepare_text_ids(
        x: torch.Tensor,  # (B, L, D) or (L, D)
        t_coord: torch.Tensor | None = None,
    ):
        B, L, _ = x.shape
        out_ids = []

        for i in range(B):
            t = torch.arange(1) if t_coord is None else t_coord[i]
            h = torch.arange(1)
            w = torch.arange(1)
            l = torch.arange(L)

            coords = torch.cartesian_prod(t, h, w, l)
            out_ids.append(coords)

        return torch.stack(out_ids)

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_latent_ids
    def _prepare_latent_ids(
        latents: torch.Tensor,  # (B, C, H, W)
    ):
        r"""
        Generates 4D position coordinates (T, H, W, L) for latent tensors.

        Args:
            latents (torch.Tensor):
                Latent tensor of shape (B, C, H, W)

        Returns:
            torch.Tensor:
                Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0,
                H=[0..H-1], W=[0..W-1], L=0
        """

        batch_size, _, height, width = latents.shape

        t = torch.arange(1)  # [0] - time dimension
        h = torch.arange(height)
        w = torch.arange(width)
        l = torch.arange(1)  # [0] - layer dimension

        # Create position IDs: (H*W, 4)
        latent_ids = torch.cartesian_prod(t, h, w, l)

        # Expand to batch: (B, H*W, 4)
        latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1)

        return latent_ids

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_image_ids
    def _prepare_image_ids(
        image_latents: list[torch.Tensor],  # [(1, C, H, W), (1, C, H, W), ...]
        scale: int = 10,
    ):
        r"""
        Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents.

        This function creates a unique coordinate for every pixel/patch across all input latent with different
        dimensions.

        Args:
            image_latents (list[torch.Tensor]):
                A list of image latent feature tensors, typically of shape (C, H, W).
            scale (int, optional):
                A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th
                latent is: 'scale + scale * i'. Defaults to 10.

        Returns:
            torch.Tensor:
                The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all
                input latents.

        Coordinate Components (Dimension 4):
            - T (Time): The unique index indicating which latent image the coordinate belongs to.
            - H (Height): The row index within that latent image.
            - W (Width): The column index within that latent image.
            - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1)
        """

        if not isinstance(image_latents, list):
            raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.")

        # create time offset for each reference image
        t_coords = [scale + scale * t for t in torch.arange(0, len(image_latents))]
        t_coords = [t.view(-1) for t in t_coords]

        image_latent_ids = []
        for x, t in zip(image_latents, t_coords):
            x = x.squeeze(0)
            _, height, width = x.shape

            x_ids = torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1))
            image_latent_ids.append(x_ids)

        image_latent_ids = torch.cat(image_latent_ids, dim=0)
        image_latent_ids = image_latent_ids.unsqueeze(0)

        return image_latent_ids

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._patchify_latents
    def _patchify_latents(latents):
        batch_size, num_channels_latents, height, width = latents.shape
        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
        latents = latents.permute(0, 1, 3, 5, 2, 4)
        latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2)
        return latents

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpatchify_latents
    def _unpatchify_latents(latents):
        batch_size, num_channels_latents, height, width = latents.shape
        latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width)
        latents = latents.permute(0, 1, 4, 2, 5, 3)
        latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2)
        return latents

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._pack_latents
    def _pack_latents(latents):
        """
        pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels)
        """

        batch_size, num_channels, height, width = latents.shape
        latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1)

        return latents

    @staticmethod
    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpack_latents_with_ids
    def _unpack_latents_with_ids(x: torch.Tensor, x_ids: torch.Tensor) -> list[torch.Tensor]:
        """
        using position ids to scatter tokens into place
        """
        x_list = []
        for data, pos in zip(x, x_ids):
            _, ch = data.shape  # noqa: F841
            h_ids = pos[:, 1].to(torch.int64)
            w_ids = pos[:, 2].to(torch.int64)

            h = torch.max(h_ids) + 1
            w = torch.max(w_ids) + 1

            flat_ids = h_ids * w + w_ids

            out = torch.zeros((h * w, ch), device=data.device, dtype=data.dtype)
            out.scatter_(0, flat_ids.unsqueeze(1).expand(-1, ch), data)

            # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W)

            out = out.view(h, w, ch).permute(2, 0, 1)
            x_list.append(out)

        return torch.stack(x_list, dim=0)

    def _set_kv_attn_processors(self):
        """Replace default attention processors with KV-cache-aware variants."""
        for block in self.transformer.transformer_blocks:
            block.attn.set_processor(Flux2KVAttnProcessor())
        for block in self.transformer.single_transformer_blocks:
            block.attn.set_processor(Flux2KVParallelSelfAttnProcessor())

    def encode_prompt(
        self,
        prompt: str | list[str],
        device: torch.device | None = None,
        num_images_per_prompt: int = 1,
        prompt_embeds: torch.Tensor | None = None,
        max_sequence_length: int = 512,
        text_encoder_out_layers: tuple[int] = (9, 18, 27),
    ):
        device = device or self._execution_device

        if prompt is None:
            prompt = ""

        prompt = [prompt] if isinstance(prompt, str) else prompt

        if prompt_embeds is None:
            prompt_embeds = self._get_qwen3_prompt_embeds(
                text_encoder=self.text_encoder,
                tokenizer=self.tokenizer,
                prompt=prompt,
                device=device,
                max_sequence_length=max_sequence_length,
                hidden_states_layers=text_encoder_out_layers,
            )

        batch_size, seq_len, _ = prompt_embeds.shape
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        text_ids = self._prepare_text_ids(prompt_embeds)
        text_ids = text_ids.to(device)
        return prompt_embeds, text_ids

    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._encode_vae_image
    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
        if image.ndim != 4:
            raise ValueError(f"Expected image dims 4, got {image.ndim}.")

        image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax")
        image_latents = self._patchify_latents(image_latents)

        latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype)
        latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(
            image_latents.device, image_latents.dtype
        )
        image_latents = (image_latents - latents_bn_mean) / latents_bn_std

        return image_latents

    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline.prepare_latents
    def prepare_latents(
        self,
        batch_size,
        num_latents_channels,
        height,
        width,
        dtype,
        device,
        generator: torch.Generator,
        latents: torch.Tensor | None = None,
    ):
        # VAE applies 8x compression on images but we must also account for packing which requires
        # latent height and width to be divisible by 2.
        height = 2 * (int(height) // (self.vae_scale_factor * 2))
        width = 2 * (int(width) // (self.vae_scale_factor * 2))

        shape = (batch_size, num_latents_channels * 4, height // 2, width // 2)
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )
        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device=device, dtype=dtype)

        latent_ids = self._prepare_latent_ids(latents)
        latent_ids = latent_ids.to(device)

        latents = self._pack_latents(latents)  # [B, C, H, W] -> [B, H*W, C]
        return latents, latent_ids

    # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline.prepare_image_latents
    def prepare_image_latents(
        self,
        images: list[torch.Tensor],
        batch_size,
        generator: torch.Generator,
        device,
        dtype,
    ):
        image_latents = []
        for image in images:
            image = image.to(device=device, dtype=dtype)
            imagge_latent = self._encode_vae_image(image=image, generator=generator)
            image_latents.append(imagge_latent)  # (1, 128, 32, 32)

        image_latent_ids = self._prepare_image_ids(image_latents)

        # Pack each latent and concatenate
        packed_latents = []
        for latent in image_latents:
            # latent: (1, 128, 32, 32)
            packed = self._pack_latents(latent)  # (1, 1024, 128)
            packed = packed.squeeze(0)  # (1024, 128) - remove batch dim
            packed_latents.append(packed)

        # Concatenate all reference tokens along sequence dimension
        image_latents = torch.cat(packed_latents, dim=0)  # (N*1024, 128)
        image_latents = image_latents.unsqueeze(0)  # (1, N*1024, 128)

        image_latents = image_latents.repeat(batch_size, 1, 1)
        image_latent_ids = image_latent_ids.repeat(batch_size, 1, 1)
        image_latent_ids = image_latent_ids.to(device)

        return image_latents, image_latent_ids

    def check_inputs(
        self,
        prompt,
        height,
        width,
        prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
    ):
        if (
            height is not None
            and height % (self.vae_scale_factor * 2) != 0
            or width is not None
            and width % (self.vae_scale_factor * 2) != 0
        ):
            logger.warning(
                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
            )

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

    @property
    def attention_kwargs(self):
        return self._attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def current_timestep(self):
        return self._current_timestep

    @property
    def interrupt(self):
        return self._interrupt

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        image: list[PIL.Image.Image] | PIL.Image.Image | None = None,
        prompt: str | list[str] = None,
        height: int | None = None,
        width: int | None = None,
        num_inference_steps: int = 4,
        sigmas: list[float] | None = None,
        num_images_per_prompt: int = 1,
        generator: torch.Generator | list[torch.Generator] | None = None,
        latents: torch.Tensor | None = None,
        prompt_embeds: torch.Tensor | None = None,
        output_type: str = "pil",
        return_dict: bool = True,
        attention_kwargs: dict[str, Any] | None = None,
        callback_on_step_end: Callable[[int, int, dict], None] | None = None,
        callback_on_step_end_tensor_inputs: list[str] = ["latents"],
        max_sequence_length: int = 512,
        text_encoder_out_layers: tuple[int] = (9, 18, 27),
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            image (`PIL.Image.Image` or `List[PIL.Image.Image]`, *optional*):
                Reference image(s) for conditioning. On the first denoising step, reference tokens are included in the
                forward pass and their attention K/V are cached. On subsequent steps, the cached K/V are reused without
                recomputing.
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation.
            height (`int`, *optional*):
                The height in pixels of the generated image.
            width (`int`, *optional*):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 4):
                The number of denoising steps.
            sigmas (`List[float]`, *optional*):
                Custom sigmas for the denoising schedule.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                Generator(s) for deterministic generation.
            latents (`torch.Tensor`, *optional*):
                Pre-generated noisy latents.
            prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated text embeddings.
            output_type (`str`, *optional*, defaults to `"pil"`):
                Output format: `"pil"` or `"np"`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether to return a `Flux2PipelineOutput` or a plain tuple.
            attention_kwargs (`dict`, *optional*):
                Extra kwargs passed to attention processors.
            callback_on_step_end (`Callable`, *optional*):
                Callback function called at the end of each denoising step.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                Tensor inputs for the callback function.
            max_sequence_length (`int`, defaults to 512):
                Maximum sequence length for the prompt.
            text_encoder_out_layers (`tuple[int]`):
                Layer indices for text encoder hidden state extraction.

        Examples:

        Returns:
            [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`.
        """

        # 1. Check inputs
        self.check_inputs(
            prompt=prompt,
            height=height,
            width=width,
            prompt_embeds=prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
        )

        self._attention_kwargs = attention_kwargs
        self._current_timestep = None
        self._interrupt = False

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        device = self._execution_device

        # 3. prepare text embeddings
        prompt_embeds, text_ids = self.encode_prompt(
            prompt=prompt,
            prompt_embeds=prompt_embeds,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            text_encoder_out_layers=text_encoder_out_layers,
        )

        # 4. process images
        if image is not None and not isinstance(image, list):
            image = [image]

        condition_images = None
        if image is not None:
            for img in image:
                self.image_processor.check_image_input(img)

            condition_images = []
            for img in image:
                image_width, image_height = img.size
                if image_width * image_height > 1024 * 1024:
                    img = self.image_processor._resize_to_target_area(img, 1024 * 1024)
                    image_width, image_height = img.size

                multiple_of = self.vae_scale_factor * 2
                image_width = (image_width // multiple_of) * multiple_of
                image_height = (image_height // multiple_of) * multiple_of
                img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop")
                condition_images.append(img)
                height = height or image_height
                width = width or image_width

        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

        # 5. prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_ids = self.prepare_latents(
            batch_size=batch_size * num_images_per_prompt,
            num_latents_channels=num_channels_latents,
            height=height,
            width=width,
            dtype=prompt_embeds.dtype,
            device=device,
            generator=generator,
            latents=latents,
        )

        image_latents = None
        image_latent_ids = None
        if condition_images is not None:
            image_latents, image_latent_ids = self.prepare_image_latents(
                images=condition_images,
                batch_size=batch_size * num_images_per_prompt,
                generator=generator,
                device=device,
                dtype=self.vae.dtype,
            )

        # 6. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
        if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas:
            sigmas = None
        image_seq_len = latents.shape[1]
        mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps)
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            sigmas=sigmas,
            mu=mu,
        )
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

        # 7. Denoising loop with KV caching
        # Step 0 with ref images: forward_kv_extract (full pass, cache ref K/V)
        # Steps 1+: forward_kv_cached (reuse cached ref K/V)
        # No ref images: standard forward
        self.scheduler.set_begin_index(0)
        kv_cache = None

        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                self._current_timestep = t
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                if i == 0 and image_latents is not None:
                    # Step 0: include ref tokens, extract KV cache
                    latent_model_input = torch.cat([image_latents, latents], dim=1).to(self.transformer.dtype)
                    latent_image_ids = torch.cat([image_latent_ids, latent_ids], dim=1)

                    noise_pred, kv_cache = self.transformer(
                        hidden_states=latent_model_input,
                        timestep=timestep / 1000,
                        guidance=None,
                        encoder_hidden_states=prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_image_ids,
                        joint_attention_kwargs=self.attention_kwargs,
                        return_dict=False,
                        kv_cache_mode="extract",
                        num_ref_tokens=image_latents.shape[1],
                    )

                elif kv_cache is not None:
                    # Steps 1+: use cached ref KV, no ref tokens in input
                    noise_pred = self.transformer(
                        hidden_states=latents.to(self.transformer.dtype),
                        timestep=timestep / 1000,
                        guidance=None,
                        encoder_hidden_states=prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_ids,
                        joint_attention_kwargs=self.attention_kwargs,
                        return_dict=False,
                        kv_cache=kv_cache,
                        kv_cache_mode="cached",
                    )[0]

                else:
                    # No reference images: standard forward
                    noise_pred = self.transformer(
                        hidden_states=latents.to(self.transformer.dtype),
                        timestep=timestep / 1000,
                        guidance=None,
                        encoder_hidden_states=prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_ids,
                        joint_attention_kwargs=self.attention_kwargs,
                        return_dict=False,
                    )[0]

                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]

                if latents.dtype != latents_dtype:
                    if torch.backends.mps.is_available():
                        latents = latents.to(latents_dtype)

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)

                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

                if XLA_AVAILABLE:
                    xm.mark_step()

        # Clean up KV cache
        if kv_cache is not None:
            kv_cache.clear()

        self._current_timestep = None

        latents = self._unpack_latents_with_ids(latents, latent_ids)

        latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype)
        latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(
            latents.device, latents.dtype
        )
        latents = latents * latents_bn_std + latents_bn_mean
        latents = self._unpatchify_latents(latents)
        if output_type == "latent":
            image = latents
        else:
            image = self.vae.decode(latents, return_dict=False)[0]
            image = self.image_processor.postprocess(image, output_type=output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return Flux2PipelineOutput(images=image)