app.py

from PIL import Image
import gradio as gr
from huggingface_hub import hf_hub_download, snapshot_download
import torch
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.set_float32_matmul_precision('high')
setattr(torch.nn.Linear, 'reset_parameters', lambda self: None)
setattr(torch.nn.LayerNorm, 'reset_parameters', lambda self: None)
import os
import time
import argparse
from tokenizer_image.vq_model import VQ_models
from models.gpt import GPT_models
from models.generate import generate
from t5 import T5Embedder
os.environ["TOKENIZERS_PARALLELISM"] = "false"

device = "cuda"

model2ckpt = {
    "GPT-XL": ("vq_ds16_t2i.pt", "t2i_XL_stage2_512.pt", 512),
}

def load_model(args):
    ckpt_folder = './'
    t5_folder = os.path.join(ckpt_folder, "flan-t5-xl")
    if not os.path.exists(t5_folder):
        os.makedirs(t5_folder, exist_ok=True)
    vq_ckpt, gpt_ckpt, image_size = model2ckpt[args.gpt_model]
    hf_hub_download(repo_id="peizesun/llamagen_t2i", filename=vq_ckpt, local_dir=ckpt_folder)
    hf_hub_download(repo_id="peizesun/llamagen_t2i", filename=gpt_ckpt, local_dir=ckpt_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="config.json", local_dir=t5_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="pytorch_model-00001-of-00002.bin", local_dir=t5_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="pytorch_model-00002-of-00002.bin", local_dir=t5_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="pytorch_model.bin.index.json", local_dir=t5_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="special_tokens_map.json", local_dir=t5_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="spiece.model", local_dir=t5_folder)
    hf_hub_download(repo_id="google/flan-t5-xl", filename="tokenizer_config.json", local_dir=t5_folder)
    # create and load model
    vq_model = VQ_models[args.vq_model](
        codebook_size=args.codebook_size,
        codebook_embed_dim=args.codebook_embed_dim)
    vq_model.to(device)
    vq_model.eval()
    checkpoint = torch.load(f"{ckpt_folder}{vq_ckpt}", map_location="cpu")
    vq_model.load_state_dict(checkpoint["model"])
    del checkpoint
    print(f"image tokenizer is loaded")

    # create and load gpt model
    precision = {'none': torch.float32, 'bf16': torch.bfloat16, 'fp16': torch.float16}[args.precision]
    latent_size = image_size // args.downsample_size
    gpt_model = GPT_models[args.gpt_model](
        vocab_size=args.codebook_size,
        block_size=latent_size ** 2,
        num_classes=args.num_classes,
        cls_token_num=args.cls_token_num,
        model_type=args.gpt_type,
    ).to(device=device, dtype=precision)
    
    checkpoint = torch.load(f"{ckpt_folder}{gpt_ckpt}", map_location="cpu")
    if args.from_fsdp: # fspd
        model_weight = checkpoint
    elif "model" in checkpoint:  # ddp
        model_weight = checkpoint["model"]
    elif "module" in checkpoint: # deepspeed
        model_weight = checkpoint["module"]
    elif "state_dict" in checkpoint:
        model_weight = checkpoint["state_dict"]
    else:
        raise Exception("please check model weight")
    # if 'freqs_cis' in model_weight:
    #     model_weight.pop('freqs_cis')
    gpt_model.load_state_dict(model_weight, strict=False)
    gpt_model.eval()
    del checkpoint
    print(f"gpt model is loaded")

    if args.compile:
        print(f"compiling the model...")
        gpt_model = torch.compile(
            gpt_model,
            mode="reduce-overhead",
            fullgraph=True
        ) # requires PyTorch 2.0 (optional)
    else:
        print(f"no need to compile model in demo") 

    t5_model = T5Embedder(
        device=device, 
        local_cache=True, 
        cache_dir=ckpt_folder, 
        dir_or_name="flan-t5-xl",
        torch_dtype=precision,
        model_max_length=args.t5_feature_max_len,
    )

    return t5_model, vq_model, gpt_model, image_size


def infer(cfg_scale, top_k, top_p, temperature, prompt, seed):
    prompts = [prompt for _ in range(4)]
    caption_embs, emb_masks = t5_model.get_text_embeddings(prompts)

    if not args.no_left_padding:
        print(f"processing left-padding...")    
        # a naive way to implement left-padding
        new_emb_masks = torch.flip(emb_masks, dims=[-1])
        new_caption_embs = []
        for idx, (caption_emb, emb_mask) in enumerate(zip(caption_embs, emb_masks)):
            valid_num = int(emb_mask.sum().item())
            print(f'  prompt {idx} token len: {valid_num}')
            new_caption_emb = torch.cat([caption_emb[valid_num:], caption_emb[:valid_num]])
            new_caption_embs.append(new_caption_emb)
        new_caption_embs = torch.stack(new_caption_embs)
    else:
        new_caption_embs, new_emb_masks = caption_embs, emb_masks
    c_indices = new_caption_embs * new_emb_masks[:,:, None]
    c_emb_masks = new_emb_masks
    qzshape = [len(c_indices), args.codebook_embed_dim, latent_size, latent_size]

    t1 = time.time()
    torch.manual_seed(seed)
    index_sample = generate(
        gpt_model, c_indices, latent_size ** 2,
        c_emb_masks,
        cfg_scale=cfg_scale, cfg_interval=args.cfg_interval,
        temperature=temperature, top_k=top_k,
        top_p=top_p, sample_logits=True, 
        )
    sampling_time = time.time() - t1
    print(f"gpt sampling takes about {sampling_time:.2f} seconds.")    

    t2 = time.time()
    samples = vq_model.decode_code(index_sample, qzshape) # output value is between [-1, 1]
    decoder_time = time.time() - t2
    print(f"decoder takes about {decoder_time:.2f} seconds.")
    # Convert to PIL.Image format:
    samples = samples.mul(127.5).add_(128.0).clamp_(0, 255).permute(0, 2, 3, 1).to("cpu", torch.uint8).numpy()
    samples = [Image.fromarray(sample) for sample in samples]
    return samples


parser = argparse.ArgumentParser()
parser.add_argument("--t5-path", type=str, default='.')
parser.add_argument("--t5-feature-max-len", type=int, default=120)
parser.add_argument("--t5-feature-dim", type=int, default=2048)
parser.add_argument("--no-left-padding", action='store_true', default=False)
parser.add_argument("--gpt-model", type=str, choices=list(GPT_models.keys()), default="GPT-XL")
parser.add_argument("--gpt-type", type=str, choices=['c2i', 't2i'], default="t2i", help="class-conditional or text-conditional")
parser.add_argument("--from-fsdp", action='store_true')
parser.add_argument("--cls-token-num", type=int, default=120, help="max token number of condition input")
parser.add_argument("--precision", type=str, default='bf16', choices=["none", "fp16", "bf16"]) 
parser.add_argument("--compile", action='store_true', default=False)
parser.add_argument("--vq-model", type=str, choices=list(VQ_models.keys()), default="VQ-16")
parser.add_argument("--codebook-size", type=int, default=16384, help="codebook size for vector quantization")
parser.add_argument("--codebook-embed-dim", type=int, default=8, help="codebook dimension for vector quantization")
parser.add_argument("--downsample-size", type=int, choices=[8, 16], default=16)
parser.add_argument("--num-classes", type=int, default=1000)
parser.add_argument("--cfg-scale", type=float, default=7.5)
parser.add_argument("--cfg-interval", type=float, default=-1)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--top-k", type=int, default=2000,help="top-k value to sample with")
parser.add_argument("--temperature", type=float, default=1.0, help="temperature value to sample with")
parser.add_argument("--top-p", type=float, default=1.0, help="top-p value to sample with")
args = parser.parse_args()

t5_model, vq_model, gpt_model, image_size = load_model(args)
latent_size = image_size // args.downsample_size

examples = [
    "A fluffy golden retriever puppy with big, soulful eyes sits in a sunlit garden, surrounded by colorful flowers and butterflies fluttering around its wagging tail.",
    "A steaming bowl of Pho, filled with translucent rice noodles and thin slices of savory beef, topped with a heaping of fresh bean sprouts, a wedge of lime on the side, and a sprinkle of chopped green onions and cilantro.",
    "An ethereal black and white landscape, where a solitary, sinuous black tree stands stark against a stark white snowy backdrop. Its branches twist intricately towards the sky, casting dramatic shadows on the untouched snow below.",        
]

with gr.Blocks() as demo:
    gr.Markdown("<h1 style='text-align: center'>Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation</h1>")

    with gr.Tabs():
        with gr.TabItem('Generate'):
            with gr.Row():
                with gr.Column():
                    cfg_scale = gr.Slider(minimum=1, maximum=25, step=0.1, value=7.5, label='Classifier-free Guidance Scale')
                    top_k = gr.Slider(minimum=1, maximum=16384, step=1, value=4000, label='Top-K')
                    top_p = gr.Slider(minimum=0., maximum=1.0, step=0.1, value=1.0, label="Top-P")
                    temperature = gr.Slider(minimum=0., maximum=1.0, step=0.1, value=1.0, label='Temperature')
                    seed = gr.Slider(minimum=0, maximum=1000, step=1, value=0, label='Seed')
                    with gr.Row():
                        text_prompt = gr.Textbox(
                            label="Enter your prompt",
                            show_label=False,
                            max_lines=1,
                            placeholder="Enter your prompt",
                        )
                    button = gr.Button("Generate", variant="primary")
                    gr.Examples(
                        label="Examples (select one example, and click Generate button)",
                        examples=examples,
                        inputs=text_prompt,
                        # outputs=[result],
                        # fn=generate,
                    )
                with gr.Column():
                    output = gr.Gallery(label='Generated Images', height=700)
                    button.click(infer, inputs=[cfg_scale, top_k, top_p, temperature, text_prompt, seed], outputs=[output])
    demo.queue()
    demo.launch(debug=True, share=True)