Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,562 Bytes
9294bc7 7fafe48 9294bc7 7fafe48 9294bc7 7fafe48 9294bc7 7fafe48 9294bc7 7fafe48 9294bc7 |
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 |
import gradio as gr
import spaces
import torch
from diffusers import StableDiffusion3Pipeline
from adv_grpo.diffusers_patch.sd3_pipeline_with_logprob_fast import pipeline_with_logprob_random as pipeline_with_logprob
from adv_grpo.diffusers_patch.train_dreambooth_lora_sd3 import encode_prompt
from adv_grpo.ema import EMAModuleWrapper
from peft import PeftModel
from PIL import Image
import numpy as np
import os
from ml_collections import config_flags
from huggingface_hub import hf_hub_download
from huggingface_hub import login
login(os.environ["HF_TOKEN"])
# ---------------------------------------------------------
# GLOBAL VARIABLES
# ---------------------------------------------------------
pipeline = None
config = None
text_encoders = None
tokenizers = None
ema = None
transformer_trainable_parameters = None
def load_lora_from_subfolder():
repo_id = "benzweijia/Adv-GRPO"
subfolder = "DINO"
local_dir = "/tmp/DINO"
os.makedirs(local_dir, exist_ok=True)
for filename in ["adapter_config.json", "adapter_model.safetensors"]:
hf_hub_download(
repo_id=repo_id,
repo_type="model",
subfolder=subfolder,
filename=filename,
local_dir=local_dir,
force_download=False
)
# import pdb; pdb.set_trace()
return local_dir
# -------------- Load Config ------------------------------
def load_config():
"""
"""
import importlib.util
config_path = "config/base.py"
spec = importlib.util.spec_from_file_location("config", config_path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module.get_config()
# -------------- Embedding Function -----------------------
def compute_text_embeddings(prompt, text_encoders, tokenizers, max_sequence_length, device):
with torch.no_grad():
prompt_embeds, pooled_prompt_embeds = encode_prompt(
text_encoders, tokenizers, prompt, max_sequence_length
)
prompt_embeds = prompt_embeds.to(device)
pooled_prompt_embeds = pooled_prompt_embeds.to(device)
return prompt_embeds, pooled_prompt_embeds
# ---------------------------------------------------------
# GPU MODEL INITIALIZATION
# ---------------------------------------------------------
@spaces.GPU
def init_model():
global pipeline, config, text_encoders, tokenizers, ema, transformer_trainable_parameters
print("🔥 Loading config...")
config = load_config()
print("🔥 Loading SD3 base model on GPU...")
# import pdb; pdb.set_trace()
pipeline = StableDiffusion3Pipeline.from_pretrained(
"stabilityai/stable-diffusion-3.5-medium"
)
# freeze non-trainable params
pipeline.vae.requires_grad_(False)
pipeline.text_encoder.requires_grad_(False)
pipeline.text_encoder_2.requires_grad_(False)
pipeline.text_encoder_3.requires_grad_(False)
pipeline.transformer.requires_grad_(not config.use_lora)
text_encoders = [pipeline.text_encoder, pipeline.text_encoder_2, pipeline.text_encoder_3]
tokenizers = [pipeline.tokenizer, pipeline.tokenizer_2, pipeline.tokenizer_3]
pipeline.safety_checker = None
pipeline.set_progress_bar_config(disable=True)
# move to GPU
pipeline.vae.to("cuda")
pipeline.text_encoder.to("cuda")
pipeline.text_encoder_2.to("cuda")
pipeline.text_encoder_3.to("cuda")
pipeline.transformer.to("cuda")
config.train.lora_path = "benzweijia/Adv-GRPO/DINO"
config.use_lora = True
lora_dir = load_lora_from_subfolder()
if config.use_lora and config.train.lora_path:
print("🔥 Loading LoRA from:", config.train.lora_path)
pipeline.transformer = PeftModel.from_pretrained(
pipeline.transformer,
os.path.join(lora_dir,"DINO")
)
pipeline.transformer.set_adapter("default")
transformer_trainable_parameters = list(
filter(lambda p: p.requires_grad, pipeline.transformer.parameters())
)
# Setup EMA
ema = EMAModuleWrapper(
transformer_trainable_parameters,
decay=0.9,
update_step_interval=8,
device="cuda"
)
print("✅ Model initialized and ready.")
# ---------------------------------------------------------
# INFERENCE FUNCTION
# ---------------------------------------------------------
@spaces.GPU
def infer(prompt):
print("start infer")
global pipeline, config
print(pipeline)
if pipeline is None:
init_model()
print(pipeline)
prompts = [prompt]
# get prompt embedding
prompt_embeds, pooled_prompt_embeds = compute_text_embeddings(
prompts, text_encoders, tokenizers,
max_sequence_length=128,
device="cuda"
)
neg_embed, neg_pooled_embed = compute_text_embeddings(
[""], text_encoders, tokenizers,
max_sequence_length=128,
device="cuda"
)
neg_prompt_embeds = neg_embed.repeat(1, 1, 1)
neg_pooled_prompt_embeds = neg_pooled_embed.repeat(1, 1)
# generation seed
generator = torch.Generator().manual_seed(0)
with torch.no_grad():
images, _, _, _ = pipeline_with_logprob(
pipeline,
prompt_embeds=prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_prompt_embeds=neg_prompt_embeds,
negative_pooled_prompt_embeds=neg_pooled_prompt_embeds,
num_inference_steps=config.sample.eval_num_steps,
guidance_scale=config.sample.guidance_scale,
output_type="pt",
height=config.resolution,
width=config.resolution,
noise_level=0,
mini_num_image_per_prompt=1,
process_index=0,
sample_num_steps=config.sample.num_steps,
random_timestep=0,
generator=generator,
)
# Convert to PIL
pil = Image.fromarray(
(images[0].permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8)
)
# Fixed 512x512 for output
pil = pil.resize((512, 512))
return pil
# ---------------------------------------------------------
# GRADIO UI
# ---------------------------------------------------------
# init_model()
demo = gr.Interface(
fn=infer,
inputs=gr.Textbox(lines=2, label="Prompt"),
outputs=gr.Image(type="pil"),
title="Adv-GRPO(DINO)",
description="Enter a prompt and generate image using Adv-GRPO",
)
demo.launch()
|