app.py

import subprocess
import os

# 필수 패키지 설치 (이미 설치되어 있다면 무시됩니다)
subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
)
subprocess.run("pip install huggingface_hub==0.25.0", shell=True)
subprocess.run("pip install numpy==1.26.4 sentencepiece sacremoses transformers gradio safetensors torchvision diffusers", shell=True)

# 체크포인트 폴더 생성 및 모델 스냅샷 다운로드
os.makedirs("/home/user/app/checkpoints", exist_ok=True)
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Alpha-VLLM/Lumina-Image-2.0", local_dir="/home/user/app/checkpoints")

hf_token = os.environ["HF_TOKEN"]

# ★ 중요: CUDA 초기화 전에 spaces 패키지를 임포트합니다.
import spaces

# 이제 CUDA와 관련된 라이브러리들을 임포트합니다.
import argparse
import builtins
import json
import math
import multiprocessing as mp
import random
import socket
import traceback

import torch
import gradio as gr
import numpy as np
from safetensors.torch import load_file
from torchvision.transforms.functional import to_pil_image

# 번역 파이프라인 (한글 프롬프트를 영어로 번역)
from transformers import pipeline
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")

from imgproc import generate_crop_size_list
import models
from transport import Sampler, create_transport

from multiprocessing import Process, Queue, set_start_method, get_context

class ModelFailure:
    pass

# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(prompt_batch, text_encoder, tokenizer, proportion_empty_prompts, is_train=True):
    captions = []
    for caption in prompt_batch:
        if random.random() < proportion_empty_prompts:
            captions.append("")
        elif isinstance(caption, str):
            captions.append(caption)
        elif isinstance(caption, (list, np.ndarray)):
            captions.append(random.choice(caption) if is_train else caption[0])

    with torch.no_grad():
        text_inputs = tokenizer(
            captions,
            padding=True,
            pad_to_multiple_of=8,
            max_length=256,
            truncation=True,
            return_tensors="pt",
        )

        print(f"Text Encoder Device: {text_encoder.device}")
        text_input_ids = text_inputs.input_ids.cuda()
        prompt_masks = text_inputs.attention_mask.cuda()
        print(f"Text Input Ids Device: {text_input_ids.device}")
        print(f"Prompt Masks Device: {prompt_masks.device}")

        prompt_embeds = text_encoder(
            input_ids=text_input_ids,
            attention_mask=prompt_masks,
            output_hidden_states=True,
        ).hidden_states[-2]
        text_encoder.cpu()

    return prompt_embeds, prompt_masks

@torch.no_grad()
def model_main(args, master_port, rank):
    # diffusers, transformers 등의 내부 임포트를 위해 함수 내부에서 임포트합니다.
    from diffusers.models import AutoencoderKL
    from transformers import AutoModel, AutoTokenizer

    # 기본 print 함수를 오버라이드하여 출력 지연을 최소화합니다.
    original_print = builtins.print
    def print(*args, **kwargs):
        kwargs.setdefault("flush", True)
        original_print(*args, **kwargs)
    builtins.print = print

    train_args = torch.load(os.path.join(args.ckpt, "model_args.pth"))
    print("Loaded model arguments:", json.dumps(train_args.__dict__, indent=2))

    print(f"Creating lm: Gemma-2-2B")
    dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[args.precision]

    text_encoder = AutoModel.from_pretrained("google/gemma-2-2b", torch_dtype=dtype, token=hf_token).eval().to("cuda")
    cap_feat_dim = text_encoder.config.hidden_size
    if args.num_gpus > 1:
        raise NotImplementedError("Inference with >1 GPUs not yet supported")

    tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b", token=hf_token)
    tokenizer.padding_side = "right"

    vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", token=hf_token).cuda()

    print(f"Creating DiT: {train_args.model}")
    model = models.__dict__[train_args.model](
        in_channels=16,
        qk_norm=train_args.qk_norm,
        cap_feat_dim=cap_feat_dim,
    )
    model.eval().to("cuda", dtype=dtype)

    assert train_args.model_parallel_size == args.num_gpus
    if args.ema:
        print("Loading EMA model.")
    print('Loading model weights...')
    ckpt_path = os.path.join(
        args.ckpt,
        f"consolidated{'_ema' if args.ema else ''}.{rank:02d}-of-{args.num_gpus:02d}.safetensors",
    )
    if os.path.exists(ckpt_path):
        ckpt = load_file(ckpt_path)
    else:
        ckpt_path = os.path.join(
            args.ckpt,
            f"consolidated{'_ema' if args.ema else ''}.{rank:02d}-of-{args.num_gpus:02d}.pth",
        )
        assert os.path.exists(ckpt_path)
        ckpt = torch.load(ckpt_path, map_location="cuda")
    model.load_state_dict(ckpt, strict=True)
    print('Model weights loaded.')
    
    return text_encoder, tokenizer, vae, model

@torch.no_grad()
def inference(args, infer_args, text_encoder, tokenizer, vae, model):
    dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[args.precision]
    train_args = torch.load(os.path.join(args.ckpt, "model_args.pth"))
    torch.cuda.set_device(0)
    with torch.autocast("cuda", dtype):
        (
            cap,
            neg_cap,
            system_type,
            resolution,
            num_sampling_steps,
            cfg_scale,
            cfg_trunc,
            renorm_cfg,
            solver,
            t_shift,
            seed,
            scaling_method,
            scaling_watershed,
            proportional_attn,
        ) = infer_args

        system_prompt = system_type
        cap = system_prompt + cap
        if neg_cap != "":
            neg_cap = system_prompt + neg_cap

        metadata = dict(
            real_cap=cap,
            real_neg_cap=neg_cap,
            system_type=system_type,
            resolution=resolution,
            num_sampling_steps=num_sampling_steps,
            cfg_scale=cfg_scale,
            cfg_trunc=cfg_trunc,
            renorm_cfg=renorm_cfg,
            solver=solver,
            t_shift=t_shift,
            seed=seed,
            scaling_method=scaling_method,
            scaling_watershed=scaling_watershed,
            proportional_attn=proportional_attn,
        )
        print("> Parameters:", json.dumps(metadata, indent=2))

        try:
            # 샘플러 설정
            if solver == "dpm":
                transport = create_transport("Linear", "velocity")
                sampler = Sampler(transport)
                sample_fn = sampler.sample_dpm(
                    model.forward_with_cfg,
                    model_kwargs=model_kwargs,
                )
            else:
                transport = create_transport(
                    args.path_type,
                    args.prediction,
                    args.loss_weight,
                    args.train_eps,
                    args.sample_eps,
                )
                sampler = Sampler(transport)
                sample_fn = sampler.sample_ode(
                    sampling_method=solver,
                    num_steps=num_sampling_steps,
                    atol=args.atol,
                    rtol=args.rtol,
                    reverse=args.reverse,
                    time_shifting_factor=t_shift,
                )
            # 해상도 및 latent 공간 크기 계산
            resolution = resolution.split(" ")[-1]
            w, h = resolution.split("x")
            w, h = int(w), int(h)
            latent_w, latent_h = w // 8, h // 8
            if int(seed) != 0:
                torch.random.manual_seed(int(seed))
            z = torch.randn([1, 16, latent_h, latent_w], device="cuda").to(dtype)
            z = z.repeat(2, 1, 1, 1)

            with torch.no_grad():
                if neg_cap != "":
                    cap_feats, cap_mask = encode_prompt([cap] + [neg_cap], text_encoder, tokenizer, 0.0)
                else:
                    cap_feats, cap_mask = encode_prompt([cap] + [""], text_encoder, tokenizer, 0.0)
               
            cap_mask = cap_mask.to(cap_feats.device)

            model_kwargs = dict(
                cap_feats=cap_feats,
                cap_mask=cap_mask,
                cfg_scale=cfg_scale,
                cfg_trunc=1 - cfg_trunc,
                renorm_cfg=renorm_cfg,
            )

            print(f"> Caption: {cap}")
            print(f"> Number of sampling steps: {num_sampling_steps}")
            print(f"> CFG scale: {cfg_scale}")
            print("> Starting sampling...")
            if solver == "dpm":
                samples = sample_fn(z, steps=num_sampling_steps, order=2, skip_type="time_uniform_flow", method="multistep", flow_shift=t_shift)
            else:
                samples = sample_fn(z, model.forward_with_cfg, **model_kwargs)[-1]
            samples = samples[:1]
            print("Sample dtype:", samples.dtype)

            vae_scale = {
                "sdxl": 0.13025,
                "sd3": 1.5305,
                "ema": 0.18215,
                "mse": 0.18215,
                "cogvideox": 1.15258426,
                "flux": 0.3611,
            }["flux"]
            vae_shift = {
                "sdxl": 0.0,
                "sd3": 0.0609,
                "ema": 0.0,
                "mse": 0.0,
                "cogvideox": 0.0,
                "flux": 0.1159,
            }["flux"]
            print(f"> VAE scale: {vae_scale}, shift: {vae_shift}")
            print("Samples shape:", samples.shape)
            samples = vae.decode(samples / vae_scale + vae_shift).sample
            samples = (samples + 1.0) / 2.0
            samples.clamp_(0.0, 1.0)

            img = to_pil_image(samples[0].float())
            print("> Generated image successfully.")

            return img, metadata
        except Exception:
            print(traceback.format_exc())
            return ModelFailure()

def none_or_str(value):
    if value == "None":
        return None
    return value

def parse_transport_args(parser):
    group = parser.add_argument_group("Transport arguments")
    group.add_argument(
        "--path-type",
        type=str,
        default="Linear",
        choices=["Linear", "GVP", "VP"],
        help="Type of path for transport: 'Linear', 'GVP' (Geodesic Vector Pursuit), or 'VP' (Vector Pursuit).",
    )
    group.add_argument(
        "--prediction",
        type=str,
        default="velocity",
        choices=["velocity", "score", "noise"],
        help="Prediction model for the transport dynamics.",
    )
    group.add_argument(
        "--loss-weight",
        type=none_or_str,
        default=None,
        choices=[None, "velocity", "likelihood"],
        help="Weighting of different loss components: 'velocity', 'likelihood', or None.",
    )
    group.add_argument("--sample-eps", type=float, help="Sampling parameter in the transport model.")
    group.add_argument("--train-eps", type=float, help="Training epsilon to stabilize learning.")

def parse_ode_args(parser):
    group = parser.add_argument_group("ODE arguments")
    group.add_argument(
        "--atol",
        type=float,
        default=1e-6,
        help="Absolute tolerance for the ODE solver.",
    )
    group.add_argument(
        "--rtol",
        type=float,
        default=1e-3,
        help="Relative tolerance for the ODE solver.",
    )
    group.add_argument("--reverse", action="store_true", help="Run the ODE solver in reverse.")
    group.add_argument(
        "--likelihood",
        action="store_true",
        help="Enable likelihood calculation during the ODE solving process.",
    )

def find_free_port() -> int:
    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
    sock.bind(("", 0))
    port = sock.getsockname()[1]
    sock.close()
    return port

# 한글 프롬프트가 감지되면 영어로 번역하는 함수
def translate_if_korean(text: str) -> str:
    import re
    if re.search(r"[ㄱ-ㅎㅏ-ㅣ가-힣]", text):
        print("Translating Korean prompt to English...")
        translation = translator(text)
        return translation[0]["translation_text"]
    return text

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--num_gpus", type=int, default=1)
    parser.add_argument("--ckpt", type=str, default='/home/user/app/checkpoints', required=False)
    parser.add_argument("--ema", action="store_true")
    parser.add_argument("--precision", default="bf16", choices=["bf16", "fp32"])
    parser.add_argument("--hf_token", type=str, default=None, help="Hugging Face read token for accessing gated repo.")
    parser.add_argument("--res", type=int, default=1024, choices=[256, 512, 1024])
    parser.add_argument("--port", type=int, default=12123)

    parse_transport_args(parser)
    parse_ode_args(parser)
    args = parser.parse_known_args()[0]

    if args.num_gpus != 1:
        raise NotImplementedError("Multi-GPU Inference is not yet supported")

    master_port = find_free_port()
    text_encoder, tokenizer, vae, model = model_main(args, master_port, 0)
    description = "Lumina-Image 2.0 ([Github](https://github.com/Alpha-VLLM/Lumina-Image-2.0/tree/main))"

    # 커스텀 CSS: 메뉴 컨테이너의 너비를 줄이고, 배경이 잘 보이도록 반투명 배경을 적용합니다.
    custom_css = """
    body {
        background: linear-gradient(135deg, #1a2a6c, #b21f1f, #fdbb2d);
        font-family: 'Helvetica', sans-serif;
        color: #333;
    }
    .gradio-container {
        background: rgba(255, 255, 255, 0.85); /* 반투명 배경 */
        max-width: 800px;               /* 컨테이너 최대 너비 조정 */
        margin: 20px auto;              /* 중앙 정렬 */
        border-radius: 15px;
        box-shadow: 0 8px 16px rgba(0, 0, 0, 0.25);
        padding: 20px;
    }
    .gradio-title {
        font-weight: bold;
        font-size: 1.5em;
        text-align: center;
        margin-bottom: 10px;
    }
    """

    with gr.Blocks(css=custom_css) as demo:
        with gr.Row():
            gr.Markdown(f"<div class='gradio-title'>{description}</div>")
        with gr.Row():
            with gr.Column():
                cap = gr.Textbox(
                    lines=2,
                    label="Caption",
                    interactive=True,
                    value="Majestic landscape photograph of snow-capped mountains under a dramatic sky at sunset. The mountains dominate the lower half of the image, with rugged peaks and deep crevasses visible. A glacier flows down the right side, partially illuminated by the warm light. The sky is filled with fiery orange and golden clouds, contrasting with the cool tones of the snow. The central peak is partially obscured by clouds, adding a sense of mystery. The foreground features dark, shadowed forested areas, enhancing the depth. High contrast, natural lighting, warm color palette, photorealistic, expansive, awe-inspiring, serene, visually balanced, dynamic composition.",
                    placeholder="Enter a caption."
                )
                neg_cap = gr.Textbox(
                    lines=2,
                    label="Negative Caption",
                    interactive=True,
                    value="",
                    placeholder="Enter a negative caption."
                )
                default_value = "You are an assistant designed to generate high-quality images with the highest degree of image-text alignment based on textual prompts."
                system_type = gr.Dropdown(
                    value=default_value,
                    choices=[
                        "You are an assistant designed to generate high-quality images with the highest degree of image-text alignment based on textual prompts.",
                        "You are an assistant designed to generate superior images with the superior degree of image-text alignment based on textual prompts or user prompts.",
                        ""
                    ],
                    label="System Type"
                )
                with gr.Row():
                    res_choices = [f"{w}x{h}" for w, h in generate_crop_size_list((args.res // 64) ** 2, 64)]
                    default_value = "1024x1024"
                    resolution = gr.Dropdown(value=default_value, choices=res_choices, label="Resolution")
                with gr.Row():
                    num_sampling_steps = gr.Slider(minimum=1, maximum=70, value=40, step=1, interactive=True, label="Sampling Steps")
                    seed = gr.Slider(minimum=0, maximum=int(1e5), value=0, step=1, interactive=True, label="Seed (0 for random)")
                    cfg_trunc = gr.Slider(minimum=0, maximum=1, value=0, step=0.01, interactive=True, label="CFG Truncation")
                with gr.Row():
                    solver = gr.Dropdown(value="euler", choices=["euler", "midpoint", "rk4"], label="Solver")
                    t_shift = gr.Slider(minimum=1, maximum=20, value=6, step=1, interactive=True, label="Time Shift")
                    cfg_scale = gr.Slider(minimum=1.0, maximum=20.0, value=4.0, interactive=True, label="CFG Scale")
                with gr.Row():
                    renorm_cfg = gr.Dropdown(value=True, choices=[True, False, 2.0], label="CFG Renorm")
                with gr.Accordion("Advanced Settings for Resolution Extrapolation", open=False):
                    with gr.Row():
                        scaling_method = gr.Dropdown(value="Time-aware", choices=["Time-aware", "None"], label="RoPE Scaling Method")
                        scaling_watershed = gr.Slider(minimum=0.0, maximum=1.0, value=0.3, interactive=True, label="Linear/NTK Watershed")
                    with gr.Row():
                        proportional_attn = gr.Checkbox(value=True, interactive=True, label="Proportional Attention")
                with gr.Row():
                    submit_btn = gr.Button("Submit", variant="primary")
            with gr.Column():
                output_img = gr.Image(label="Generated Image", interactive=False)
                with gr.Accordion(label="Generation Parameters", open=True):
                    gr_metadata = gr.JSON(label="Metadata", show_label=False)
        with gr.Row():
            prompts = [
                "Close-up portrait of a young woman with light brown hair, looking to the right, illuminated by warm, golden sunlight. Her hair is gently tousled, catching the light and creating a halo effect around her head. She wears a white garment with a V-neck, visible in the lower left of the frame. The background is dark and out of focus, enhancing the contrast between her illuminated face and the shadows. Soft, ethereal lighting, high contrast, warm color palette, shallow depth of field, natural backlighting, serene and contemplative mood, cinematic quality, intimate and visually striking composition.",
                "하늘을 나는 용, 신비로운 분위기, 구름 위를 날며 빛나는 비늘을 가진, 전설 속의 존재, 강렬한 색채와 디테일한 묘사.",
                "Aesthetic photograph of a bouquet of pink and white ranunculus flowers in a clear glass vase, centrally positioned on a wooden surface. The flowers are in full bloom, displaying intricate layers of petals with a soft gradient from pale pink to white. The vase is filled with water, visible through the clear glass, and the stems are submerged. In the background, a blurred vase with green stems is partially visible, adding depth to the composition. The lighting is warm and natural, casting soft shadows and highlighting the delicate textures of the petals. The scene is serene and intimate, with a focus on the organic beauty of the flowers. Photorealistic, shallow depth of field, soft natural lighting, warm color palette, high contrast, glossy texture, tranquil, visually balanced.",
                "한只优雅的白猫穿着一件紫色的旗袍，旗袍上绣有精致的牡丹花图案，显得高贵典雅。它头上戴着一朵金色的发饰，嘴里叼着一根象征好运的红色丝带。周围环绕着许多飘动的纸鹤和金色的光点，营造出一种祥瑞和梦幻的氛围。超写实风格。"
            ]
            prompts = [[p] for p in prompts]
            gr.Examples(prompts, [cap], label="Examples")
        @spaces.GPU(duration=200)
        def on_submit(cap, neg_cap, system_type, resolution, num_sampling_steps, cfg_scale, cfg_trunc, renorm_cfg, solver, t_shift, seed, scaling_method, scaling_watershed, proportional_attn, progress=gr.Progress(track_tqdm=True)):
            # 한글 프롬프트가 감지되면 영어로 번역
            cap = translate_if_korean(cap)
            if neg_cap and neg_cap.strip():
                neg_cap = translate_if_korean(neg_cap)
            infer_args = (cap, neg_cap, system_type, resolution, num_sampling_steps, cfg_scale, cfg_trunc, renorm_cfg, solver, t_shift, seed, scaling_method, scaling_watershed, proportional_attn)
            result = inference(args, infer_args, text_encoder, tokenizer, vae, model)
            if isinstance(result, ModelFailure):
                raise RuntimeError("Model failed to generate the image.")
            return result
        submit_btn.click(
            on_submit,
            [cap, neg_cap, system_type, resolution, num_sampling_steps, cfg_scale, cfg_trunc, renorm_cfg, solver, t_shift, seed, scaling_method, scaling_watershed, proportional_attn],
            [output_img, gr_metadata],
        )
        def show_scaling_watershed(scaling_m):
            return gr.update(visible=scaling_m == "Time-aware")
        scaling_method.change(show_scaling_watershed, scaling_method, scaling_watershed)
    demo.queue().launch(server_name="0.0.0.0")

if __name__ == "__main__":
    main()