Spaces:
Sleeping
Sleeping
File size: 6,889 Bytes
f866cb0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 |
import torch
import numpy as np
from diffusers import FluxPipeline, FlowMatchEulerDiscreteScheduler
from typing import Any, Dict, List, Optional, Union
from PIL import Image
from torch.cuda import graphs
# Enable TF32 and memory format optimizations
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.benchmark = True
# Constants with optimized values
BASE_SEQ_LEN = 256
MAX_SEQ_LEN = 4096
BASE_SHIFT = 0.5
MAX_SHIFT = 1.2
BATCH_SIZE = 4 # Optimal batch size for A100
@torch.jit.script
def calculate_timestep_shift(image_seq_len: int) -> float:
"""Optimized timestep shift calculation using TorchScript"""
m = (MAX_SHIFT - BASE_SHIFT) / (MAX_SEQ_LEN - BASE_SEQ_LEN)
b = BASE_SHIFT - m * BASE_SEQ_LEN
return image_seq_len * m + b
def prepare_timesteps(
scheduler: FlowMatchEulerDiscreteScheduler,
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):
"""Optimized timestep preparation with CUDA graphs support"""
if device is None:
device = torch.device("cuda")
# Pre-calculate timesteps using CUDA graph
static_input = torch.tensor([], device=device)
g = torch.cuda.CUDAGraph()
with torch.cuda.graph(g):
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.to(memory_format=torch.channels_last)
num_inference_steps = len(timesteps)
return timesteps, num_inference_steps
# FLUX pipeline function
class FluxWithCFGPipeline(FluxPipeline):
"""
Extends the FluxPipeline to yield intermediate images during the denoising process
with progressively increasing resolution for faster generation.
"""
@torch.inference_mode()
def generate_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 = 4,
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 = 300,
):
"""Generates images and yields intermediate results during the denoising process."""
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_timestep_shift(image_seq_len)
timesteps, num_inference_steps = prepare_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.float16).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
# 6. Denoising loop
for i, t in enumerate(timesteps):
if self.interrupt:
continue
timestep = t.expand(latents.shape[0]).to(latents.dtype)
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]
# Yield intermediate result
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
torch.cuda.empty_cache()
# Final image
return self._decode_latents_to_image(latents, height, width, output_type)
self.maybe_free_model_hooks()
torch.cuda.empty_cache()
def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
"""Decodes the given latents into an image."""
vae = vae or self.vae
latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
image = vae.decode(latents, return_dict=False)[0]
return self.image_processor.postprocess(image, output_type=output_type)[0] |