K00B404's picture
Update app.py
7045c60 verified
raw
history blame
9.78 kB
import spaces
import argparse
import os
import time
from os import path
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
os.environ["TRANSFORMERS_CACHE"] = cache_path
os.environ["HF_HUB_CACHE"] = cache_path
os.environ["HF_HOME"] = cache_path
import gradio as gr
import torch
from diffusers import FluxPipeline
torch.backends.cuda.matmul.allow_tf32 = True
class timer:
def __init__(self, method_name="timed process"):
self.method = method_name
def __enter__(self):
self.start = time.time()
print(f"{self.method} starts")
def __exit__(self, exc_type, exc_val, exc_tb):
end = time.time()
print(f"{self.method} took {str(round(end - self.start, 2))}s")
if not path.exists(cache_path):
os.makedirs(cache_path, exist_ok=True)
from huggingface_hub import hf_hub_download
import torch
from diffusers import FluxPipeline, FluxTransformer2DModel, GGUFQuantizationConfig
ckpt_path = (
"https://huggingface.co/city96/flux.1-lite-8B-alpha-gguf/flux.1-lite-8B-alpha-Q3_K_S.gguf"
)
transformer = FluxTransformer2DModel.from_single_file(
ckpt_path,
quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
torch_dtype=torch.bfloat16,
)
pipe = FluxPipeline.from_pretrained(
"black-forest-labs/FLUX.1-dev",
transformer=transformer,
torch_dtype=torch.bfloat16,
)
# https://huggingface.co/martintomov/Hyper-FLUX.1-dev-gguf/resolve/main/hyper-flux-16step-Q3_K_M.gguf
#pipe = FluxPipeline.from_pretrained("flux1-schnell-Q3_K_S.gguf")
pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"))
pipe.fuse_lora(lora_scale=0.125)
pipe.enable_model_cpu_offload()
prompt = "A cat holding a sign that says hello world"
image = pipe(prompt, generator=torch.manual_seed(0)).images[0]
image.save("flux-gguf.png")
#pipe.to(device="cpu", dtype=torch.bfloat16)
#hf_hub_download(repo_id="city96/FLUX.1-schnell-gguf", filename="flux1-schnell-Q3_K_S.gguf")
with gr.Blocks(theme=gr.themes.Soft()) as demo:
gr.Markdown(
"""
<div style="text-align: center; max-width: 650px; margin: 0 auto;">
<h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem; display: contents;">Hyper-FLUX-8steps-LoRA</h1>
<p style="font-size: 1rem; margin-bottom: 1.5rem;">AutoML team from ByteDance</p>
</div>
"""
)
with gr.Row():
with gr.Column(scale=3):
with gr.Group():
prompt = gr.Textbox(
label="Your Image Description",
placeholder="E.g., A serene landscape with mountains and a lake at sunset",
lines=3
)
with gr.Accordion("Advanced Settings", open=False):
with gr.Group():
with gr.Row():
height = gr.Slider(label="Height", minimum=256, maximum=1152, step=64, value=1024)
width = gr.Slider(label="Width", minimum=256, maximum=1152, step=64, value=1024)
with gr.Row():
steps = gr.Slider(label="Inference Steps", minimum=6, maximum=25, step=1, value=16)
scales = gr.Slider(label="Guidance Scale", minimum=0.0, maximum=5.0, step=0.1, value=3.5)
seed = gr.Number(label="Seed (for reproducibility)", value=3413, precision=0)
generate_btn = gr.Button("Generate Image", variant="primary", scale=1)
with gr.Column(scale=4):
output = gr.Image(label="Your Generated Image")
gr.Markdown(
"""
<div style="max-width: 650px; margin: 2rem auto; padding: 1rem; border-radius: 10px; background-color: #f0f0f0;">
<h2 style="font-size: 1.5rem; margin-bottom: 1rem;">How to Use</h2>
<ol style="padding-left: 1.5rem;">
<li>Enter a detailed description of the image you want to create.</li>
<li>Adjust advanced settings if desired (tap to expand).</li>
<li>Tap "Generate Image" and wait for your creation!</li>
</ol>
<p style="margin-top: 1rem; font-style: italic;">Tip: Be specific in your description for best results!</p>
</div>
"""
)
@spaces.GPU
def process_image(height, width, steps, scales, prompt, seed):
global pipe
with torch.inference_mode(), torch.autocast("cpu", dtype=torch.bfloat16), timer("inference"):
return pipe(
prompt=[prompt],
generator=torch.Generator().manual_seed(int(seed)),
num_inference_steps=int(steps),
guidance_scale=float(scales),
height=int(height),
width=int(width),
max_sequence_length=256
).images[0]
generate_btn.click(
process_image,
inputs=[height, width, steps, scales, prompt, seed],
outputs=output
)
if __name__ == "__main__":
demo.launch()
'''
import spaces
import argparse
import os
import time
from os import path
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
# Setting up cache directories
cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
os.environ["TRANSFORMERS_CACHE"] = cache_path
os.environ["HF_HUB_CACHE"] = cache_path
os.environ["HF_HOME"] = cache_path
import gradio as gr
import torch
from diffusers import FluxPipeline
# Remove CUDA-specific settings since this will run on CPU
# torch.backends.cuda.matmul.allow_tf32 = True
class timer:
def __init__(self, method_name="timed process"):
self.method = method_name
def __enter__(self):
self.start = time.time()
print(f"{self.method} starts")
def __exit__(self, exc_type, exc_val, exc_tb):
end = time.time()
print(f"{self.method} took {str(round(end - self.start, 2))}s")
if not path.exists(cache_path):
os.makedirs(cache_path, exist_ok=True)
# Load the model in a CPU-friendly format (use float32 to save memory)
pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.float32)
pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"))
pipe.fuse_lora(lora_scale=0.125)
# Switch to CPU and use float32 for inference
pipe.to(device="cpu", dtype=torch.float32)
# Gradio UI setup
with gr.Blocks(theme=gr.themes.Soft()) as demo:
gr.Markdown(
"""
<div style="text-align: center; max-width: 650px; margin: 0 auto;">
<h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem; display: contents;">Hyper-FLUX-8steps-LoRA</h1>
<p style="font-size: 1rem; margin-bottom: 1.5rem;">AutoML team from ByteDance</p>
</div>
"""
)
with gr.Row():
with gr.Column(scale=3):
with gr.Group():
prompt = gr.Textbox(
label="Your Image Description",
placeholder="E.g., A serene landscape with mountains and a lake at sunset",
lines=3
)
with gr.Accordion("Advanced Settings", open=False):
with gr.Group():
with gr.Row():
height = gr.Slider(label="Height", minimum=256, maximum=1152, step=64, value=512)
width = gr.Slider(label="Width", minimum=256, maximum=1152, step=64, value=512)
with gr.Row():
steps = gr.Slider(label="Inference Steps", minimum=6, maximum=25, step=1, value=8)
scales = gr.Slider(label="Guidance Scale", minimum=0.0, maximum=5.0, step=0.1, value=3.5)
seed = gr.Number(label="Seed (for reproducibility)", value=3413, precision=0)
generate_btn = gr.Button("Generate Image", variant="primary", scale=1)
with gr.Column(scale=4):
output = gr.Image(label="Your Generated Image")
gr.Markdown(
"""
<div style="max-width: 650px; margin: 2rem auto; padding: 1rem; border-radius: 10px; background-color: #f0f0f0;">
<h2 style="font-size: 1.5rem; margin-bottom: 1rem;">How to Use</h2>
<ol style="padding-left: 1.5rem;">
<li>Enter a detailed description of the image you want to create.</li>
<li>Adjust advanced settings if desired (tap to expand).</li>
<li>Tap "Generate Image" and wait for your creation!</li>
</ol>
<p style="margin-top: 1rem; font-style: italic;">Tip: Be specific in your description for best results!</p>
</div>
"""
)
# Processing function for CPU execution
def process_image(height, width, steps, scales, prompt, seed):
global pipe
with torch.inference_mode(), timer("inference"):
return pipe(
prompt=[prompt],
generator=torch.Generator().manual_seed(int(seed)),
num_inference_steps=int(steps),
guidance_scale=float(scales),
height=int(height),
width=int(width),
max_sequence_length=256
).images[0]
generate_btn.click(
process_image,
inputs=[height, width, steps, scales, prompt, seed],
outputs=output
)
if __name__ == "__main__":
demo.launch()'''