custom_pipeline.py

import numpy as np
import torch
from diffusers.pipelines.flux.pipeline_output import FluxPipeline, FluxPipelineOutput
from typing import List, Union, Optional, Dict, Any, Callable
from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps
from diffusers.utils import is_torch_xla_available

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

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False

# Constants for shift calculation
BASE_SEQ_LEN = 256
MAX_SEQ_LEN = 4096
BASE_SHIFT = 0.5
MAX_SHIFT = 1.2

# Helper functions
def calculate_timestep_shift(image_seq_len: int) -> float:
    """Calculates the timestep shift (mu) based on the image sequence length."""
    m = (MAX_SHIFT - BASE_SHIFT) / (MAX_SEQ_LEN - BASE_SEQ_LEN)
    b = BASE_SHIFT - m * BASE_SEQ_LEN
    mu = image_seq_len * m + b
    return mu

def prepare_timesteps(
    scheduler,
    num_inference_steps: Optional[int] = None,
    device: Optional[Union[str, torch.device]] = None,
    timesteps: Optional[List[int]] = None,
    sigmas: Optional[List[float]] = None,
    mu: Optional[float] = None,
) -> (torch.Tensor, int):
    """Prepares the timesteps for the diffusion process."""
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed.")

    if timesteps is not None:
        scheduler.set_timesteps(timesteps=timesteps, device=device)
    elif sigmas is not None:
        scheduler.set_timesteps(sigmas=sigmas, device=device)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, mu=mu)

    timesteps = scheduler.timesteps
    num_inference_steps = len(timesteps)
    return timesteps, num_inference_steps

# FLUX pipeline with CFG and intermediate outputs
class FluxWithCFGPipeline(FluxPipeline):
    """
    Flux pipeline with Classifier-Free Guidance and the ability to yield 
    intermediate images during the denoising process with progressively 
    increasing resolution for faster generation.
    """
    @torch.inference_mode()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
        timesteps: List[int] = None,
        guidance_scale: float = 3.5,
        num_images_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 512,
        yield_intermediates: bool = False,  # New parameter for yielding intermediates
    ):

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

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            height,
            width,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            max_sequence_length=max_sequence_length,
        )

        self._guidance_scale = guidance_scale
        self._joint_attention_kwargs = joint_attention_kwargs
        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

        lora_scale = (
            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            pooled_prompt_embeds,
            text_ids,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            lora_scale=lora_scale,
        )
        (
            negative_prompt_embeds,
            negative_pooled_prompt_embeds,
            negative_text_ids,
        ) = self.encode_prompt(
            prompt=negative_prompt,
            prompt_2=negative_prompt_2,
            prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=negative_pooled_prompt_embeds,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            lora_scale=lora_scale,
        )

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

        # 5. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
        image_seq_len = latents.shape[1]
        mu = calculate_timestep_shift(image_seq_len)
        timesteps, num_inference_steps = prepare_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            timesteps,
            sigmas,
            mu=mu,
        )
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

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

                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                # handle guidance
                if self.transformer.config.guidance_embeds:
                    guidance = torch.tensor([guidance_scale], device=device)
                    guidance = guidance.expand(latents.shape[0])
                else:
                    guidance = None

                noise_pred_text = self.transformer(
                    hidden_states=latents,
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
                )[0]

                noise_pred_uncond = self.transformer(
                    hidden_states=latents,
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=negative_pooled_prompt_embeds,
                    encoder_hidden_states=negative_prompt_embeds,
                    txt_ids=negative_text_ids,
                    img_ids=latent_image_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
                )[0]

                noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)

                # 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():
                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
                        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)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

                    # Yield intermediate images if requested
                    if yield_intermediates:
                        yield self._decode_latents_to_image(latents, height, width, output_type)

                if XLA_AVAILABLE:
                    xm.mark_step()

        # Final image decoding 
        if output_type == "latent":
            image = latents
        else:
            image = self._decode_latents_to_image(latents, height, width, output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return FluxPipelineOutput(images=image)

    def _decode_latents_to_image(self, latents, height, width, output_type):
        """Decodes the given latents into an image."""
        latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
        latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
        image = self.vae.decode(latents, return_dict=False)[0]
        return self.image_processor.postprocess(image, output_type=output_type)[0]