live_preview_helpers.py

import torch
import numpy as np
from diffusers import FluxPipeline, AutoencoderTiny, FlowMatchEulerDiscreteScheduler
from typing import Any, Dict, List, Optional, Union

# Helper functions
def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
    max_shift: float = 1.16,
):
    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 retrieve_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,
    **kwargs,
):
    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:
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    elif sigmas is not None:
        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

# FLUX pipeline function
@torch.inference_mode()
def flux_pipe_call_that_returns_an_iterable_of_images(
    self,
    prompt: Union[str, List[str]] = None,
    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,
    output_type: Optional[str] = "pil",
    return_dict: bool = True,
    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
    max_sequence_length: int = 512,
    good_vae: Optional[Any] = None,
):
    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
    self.check_inputs(
        prompt,
        prompt_2,
        height,
        width,
        prompt_embeds=prompt_embeds,
        pooled_prompt_embeds=pooled_prompt_embeds,
        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
    batch_size = 1 if isinstance(prompt, str) else len(prompt)
    device = self._execution_device

    # 3. Encode prompt
    lora_scale = joint_attention_kwargs.get("scale", None) if 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,
    )
    # 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_shift(
        image_seq_len,
        self.scheduler.config.base_image_seq_len,
        self.scheduler.config.max_image_seq_len,
        self.scheduler.config.base_shift,
        self.scheduler.config.max_shift,
    )
    timesteps, num_inference_steps = retrieve_timesteps(
        self.scheduler,
        num_inference_steps,
        device,
        timesteps,
        sigmas,
        mu=mu,
    )
    self._num_timesteps = len(timesteps)

    # Handle guidance
    guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None

    # 6. Denoising loop with optimizations
    skip_preview_steps = max(1, num_inference_steps // 8)  # Only preview every 8th step for speed
    
    for i, t in enumerate(timesteps):
        if self.interrupt:
            continue

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

        # Use mixed precision for transformer call
        with torch.autocast(device_type=device.type, dtype=torch.bfloat16):
            noise_pred = 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]
        
        # Only yield preview for certain steps to reduce overhead
        if i % skip_preview_steps == 0 or i == len(timesteps) - 1:
            latents_for_image = self._unpack_latents(latents, height, width, self.vae_scale_factor)
            latents_for_image = (latents_for_image / self.vae.config.scaling_factor) + self.vae.config.shift_factor
            
            # Use fast VAE decode with minimal memory allocation
            with torch.no_grad():
                image = self.vae.decode(latents_for_image, return_dict=False)[0]
            yield self.image_processor.postprocess(image, output_type=output_type)[0]
        
        # Scheduler step with memory optimization
        latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
        
        # Only clear cache every few steps, not every step
        if i % 4 == 0:
            torch.cuda.empty_cache()

    # Final image using good_vae
    latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
    latents = (latents / good_vae.config.scaling_factor) + good_vae.config.shift_factor
    image = good_vae.decode(latents, return_dict=False)[0]
    self.maybe_free_model_hooks()
    torch.cuda.empty_cache()
    yield self.image_processor.postprocess(image, output_type=output_type)[0]