File size: 7,189 Bytes
3d1c0e1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 | import os
import sys
sys.path.append("./")
import torch
from src.transformer import Transformer2DModel
from src.pipeline import Pipeline
from src.scheduler import Scheduler
from transformers import (
CLIPTextModelWithProjection,
CLIPTokenizer,
)
from diffusers import VQModel
import gradio as gr
import time
from torchao.quantization.quant_api import (
quantize_,
float8_weight_only,
)
device = 'cuda'
def get_quantization_method(method):
quantization_methods = {
'fp8': lambda: float8_weight_only(),
'none': None
}
return quantization_methods.get(method, None)
def load_models(quantization_method='none'):
model_path = "MeissonFlow/Meissonic"
dtype = torch.float16
model = Transformer2DModel.from_pretrained(model_path, subfolder="transformer", torch_dtype=dtype)
vq_model = VQModel.from_pretrained(model_path, subfolder="vqvae", torch_dtype=dtype)
text_encoder = CLIPTextModelWithProjection.from_pretrained(
"laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
torch_dtype=dtype
)
tokenizer = CLIPTokenizer.from_pretrained(model_path, subfolder="tokenizer")
scheduler = Scheduler.from_pretrained(model_path, subfolder="scheduler")
if quantization_method != 'none':
quant_method = get_quantization_method(quantization_method)
if quant_method:
quantize_(model, quant_method())
pipe = Pipeline(vq_model, tokenizer=tokenizer, text_encoder=text_encoder, transformer=model, scheduler=scheduler)
return pipe.to(device)
# Global variable to store the pipeline
global_pipe = None
current_quantization = 'none'
def initialize_pipeline(quantization):
global global_pipe, current_quantization
if global_pipe is None or current_quantization != quantization:
global_pipe = load_models(quantization)
current_quantization = quantization
return global_pipe
def generate_images(prompt, negative_prompt, seed, randomize_seed, width, height,
guidance_scale, num_inference_steps, quantization_method, batch_size=1,
progress=gr.Progress(track_tqdm=True)):
if randomize_seed or seed == 0:
seed = torch.randint(0, MAX_SEED, (1,)).item()
torch.manual_seed(seed)
# Initialize or update pipeline if needed
pipe = initialize_pipeline(quantization_method)
# Reset CUDA memory stats
torch.cuda.reset_peak_memory_stats()
start_time = time.time()
# Handle batch generation
if isinstance(prompt, str):
prompts = [prompt] * batch_size
else:
prompts = prompt[:batch_size]
images = pipe(
prompt=prompts,
negative_prompt=[negative_prompt] * batch_size,
height=height,
width=width,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps
).images
# Calculate performance metrics
inference_time = time.time() - start_time
memory_used = torch.cuda.max_memory_reserved() / (1024 ** 3) # Convert to GB
performance_info = f"""
Inference Time: {inference_time:.2f} seconds
Memory Used: {memory_used:.2f} GB
Quantization: {quantization_method}
"""
return images[0] if batch_size == 1 else images, seed, performance_info
MAX_SEED = 2**32 - 1
MAX_IMAGE_SIZE = 1024
default_negative_prompt = "worst quality, low quality, low res, blurry, distortion, watermark, logo, signature, text, jpeg artifacts, signature, sketch, duplicate, ugly, identifying mark"
examples = [
"Two actors are posing for a pictur with one wearing a black and white face paint.",
"A large body of water with a rock in the middle and mountains in the background.",
"A white and blue coffee mug with a picture of a man on it.",
"The sun is setting over a city skyline with a river in the foreground.",
"A black and white cat with blue eyes.",
"Three boats in the ocean with a rainbow in the sky.",
"A robot playing the piano.",
"A cat wearing a hat.",
"A dog in a jungle."
]
css = """
#col-container {
margin: 0 auto;
max-width: 640px;
}
"""
with gr.Blocks(css=css) as demo:
with gr.Column(elem_id="col-container"):
gr.Markdown("# Meissonic Text-to-Image Generator (with FP8 Support)")
with gr.Row():
prompt = gr.Text(
label="Prompt",
show_label=False,
max_lines=1,
placeholder="Enter your prompt",
container=False,
)
run_button = gr.Button("Run", scale=0, variant="primary")
result = gr.Image(label="Result", show_label=False)
performance_info = gr.Textbox(label="Performance Metrics", lines=4)
with gr.Accordion("Advanced Settings", open=False):
quantization = gr.Radio(
choices=['none', 'fp8'],
value='none',
label="Quantization Method",
)
negative_prompt = gr.Text(
label="Negative prompt",
max_lines=1,
placeholder="Enter a negative prompt",
value=default_negative_prompt,
)
seed = gr.Slider(
label="Seed",
minimum=0,
maximum=MAX_SEED,
step=1,
value=0,
)
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
with gr.Row():
width = gr.Slider(
label="Width",
minimum=256,
maximum=MAX_IMAGE_SIZE,
step=32,
value=1024,
)
height = gr.Slider(
label="Height",
minimum=256,
maximum=MAX_IMAGE_SIZE,
step=32,
value=1024,
)
with gr.Row():
guidance_scale = gr.Slider(
label="Guidance scale",
minimum=0.0,
maximum=20.0,
step=0.1,
value=9.0,
)
num_inference_steps = gr.Slider(
label="Number of inference steps",
minimum=1,
maximum=100,
step=1,
value=64,
)
batch_size = gr.Slider(
label="Batch Size",
minimum=1,
maximum=8,
step=1,
value=1,
)
gr.Examples(examples=examples, inputs=[prompt])
gr.on(
triggers=[run_button.click, prompt.submit],
fn=generate_images,
inputs=[
prompt,
negative_prompt,
seed,
randomize_seed,
width,
height,
guidance_scale,
num_inference_steps,
quantization,
batch_size,
],
outputs=[result, seed, performance_info],
)
demo.launch()
|