inference.py

import sys

sys.path.append(".")
import argparse
import os
import random
import numpy as np
import torch
import pickle
from configs.configuration_mmdit import MMDiTConfig
from models.modeling_motifimage import MotifImage

from safetensors.torch import load_file
# from tools.motif_api import PromptRewriter
# from tools.nsfw_filtering import ContentFilter

def load_sharded_model(model_index_path):
    """
    Loads a sharded model from a safetensors index file.

    Args:
        model_index_path (str): Path to the model.safetensors.index.json file.
    """
    with open(model_index_path, 'r') as f:
        index = json.load(f)

    sharded_state_dicts = {}
    folder = os.path.dirname(model_index_path)
    for weight_name, filename in index["weight_map"].items():
        if filename not in sharded_state_dicts:
            sharded_state_dicts[filename] = load_file(os.path.join(folder, filename), device="cpu")
    
    merged_state_dict = {}
    for weight_name, filename in index["weight_map"].items():
        merged_state_dict[weight_name] = sharded_state_dicts[filename][weight_name]
        
    merged_state_dict = {k: v for k, v in merged_state_dict.items() if 'dit' in k}
    return merged_state_dict


def main(args):

    # Check if the prompt file exists
    if not os.path.isfile(args.prompt_file):
        print(f"Error: The prompt file '{args.prompt_file}' does not exist.")
        sys.exit(1)

    # List of prompts
    with open(args.prompt_file) as f:
        prompts = [prompt.rstrip() for prompt in f.readlines()]

    # Load model configuration and model
    config = MMDiTConfig.from_json_file(args.model_config)
    config.height = args.resolution
    config.width = args.resolution

    model = MotifImage(config)

    # Load checkpoint
    try:
        ema_instance = torch.load(args.model_ckpt, weights_only=False)
        ema_instance = {k: v for k, v in ema_instance.items() if "dit" in k}
    except pickle.UnpicklingError as e:
        print(f"Error loading checkpoint: {e}")
        ema_instance = load_file(args.model_ckpt)
        ema_instance = {k: v for k, v in ema_instance.items() if "dit" in k}

    if "ema_model.bin" in args.model_ckpt:
        # EMA checkpoint loading
        for param, ema_param in zip(model.parameters(), ema_instance["shadow_params"]):
            param.data.copy_(ema_param.data)
    else:
        # Non-EMA checkpoint loading
        model.load_state_dict(ema_instance)

    model = model.cuda()
    model = model.to(dtype=torch.bfloat16)
    model.eval()

    # Use guidance scales from args or set default
    guidance_scales = args.guidance_scales if args.guidance_scales else [5.0]

    # If a single seed is passed without nargs, wrap it in a list
    if isinstance(args.seed, int):
        seeds = [args.seed]
    else:
        seeds = args.seed

    for seed in seeds:
        for guidance_scale in guidance_scales:
            # Output directory structure: base_dir/seed_xxx/guidance_yyy
            output_dir = os.path.join(args.output_dir, f"seed_{seed}", f"scale_{guidance_scale}")
            os.makedirs(output_dir, exist_ok=True)
            # Using for_loop when generating high-resolution images
            for i in range(0, len(prompts), args.batch_size):  # Process 1s prompts at a time
                # Set random seeds
                torch.manual_seed(seed)
                random.seed(seed)
                np.random.seed(seed)

                batch_prompts = prompts[i : i + args.batch_size]
                imgs = model.sample(
                    batch_prompts,
                    args.steps,
                    resolution=[args.resolution, args.resolution],
                    guidance_scale=guidance_scale,
                    step_scaling=1.0,
                    use_linear_quadratic_schedule=True,
                    linear_quadratic_emulating_steps=250,
                    get_intermediate_steps=args.streaming,
                    noisy_pad=args.noisy_pad,
                    zero_masking=args.zero_masking,
                )
                if args.streaming:
                    imgs, intermediate_imgs = imgs
                    if isinstance(intermediate_imgs, list):
                        for j, intermediate_img in enumerate(intermediate_imgs):
                            for k, img in enumerate(intermediate_img):
                                img.save(os.path.join(output_dir, f"{i + k:03d}_{j:03d}_intermediate.png"))
                    else:
                        # If intermediate_imgs is a single Image, save it directly
                        intermediate_imgs.save(os.path.join(output_dir, f"{i:03d}_0_intermediate.png"))
                for j, img in enumerate(imgs):
                    img.save(os.path.join(output_dir, f"{i + j:03d}_check.png"))


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Generate images with model")
    parser.add_argument("--model-config", type=str, required=True, help="Path to the model configuration file")
    parser.add_argument("--model-ckpt", type=str, required=True, help="Path to the model checkpoint file")
    parser.add_argument(
        "--seed", type=int, nargs="*", default=[7777], help="Random seed(s) for reproducibility (can provide multiple)"
    )
    # parser.add_argument("--slg", type=int, nargs="*", default=None, help="")
    parser.add_argument("--steps", type=int, default=50, help="Number of steps for image generation")
    parser.add_argument("--resolution", type=int, default=256, help="Resolution of output images")
    parser.add_argument("--batch-size", type=int, default=32,help="Batch size for image generation")
    parser.add_argument("--streaming", action="store_true", help="Enable streaming mode for intermediate steps")
    parser.add_argument("--noisy-pad", action="store_true")
    parser.add_argument("--zero-masking", action="store_true")
    parser.add_argument("--prompt-file", type=str, default="prompt_128.txt", help="Path to the prompt file")
    parser.add_argument("--guidance-scales", type=float, nargs="*", default=None, help="List of guidance scales")
    parser.add_argument("--output-dir", type=str, default="output", help="Base output directory for generated images")
    parser.add_argument("--lora-ckpt", action="store_true")
    args = parser.parse_args()

    main(args)