pipeline.py

from diffusers import DiffusionPipeline
import torch
import numpy as np
import importlib.util
import sys
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import os
from safetensors.torch import load_file
from .vae import AutoencoderKL
from .mar import mar_base, mar_large, mar_huge

# inheriting from DiffusionPipeline for HF
class MARModel(DiffusionPipeline):

    def __init__(self):
        super().__init__()

    @torch.no_grad()
    def __call__(self, *args, **kwargs):
        """
        This method downloads the model and VAE components,
        then executes the forward pass based on the user's input.
        """
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")



        # init the mar model architecture
        buffer_size = kwargs.get("buffer_size", 64)
        diffloss_d = kwargs.get("diffloss_d", 3)
        diffloss_w = kwargs.get("diffloss_w", 1024)
        num_sampling_steps = kwargs.get("num_sampling_steps", 100)
        model_type = kwargs.get("model_type", "mar_base")

        model_mapping = {
            "mar_base": mar_base,
            "mar_large": mar_large,
            "mar_huge": mar_huge
        }

        num_sampling_steps_diffloss = 100  # Example number of sampling steps   

        # download the pretrained model and set diffloss parameters
        if model_type == "mar_base":
            diffloss_d = 6
            diffloss_w = 1024
            model_path = "mar-base.safetensors"
        elif model_type == "mar_large":
            diffloss_d = 8
            diffloss_w = 1280
            model_path = "mar-large.safetensors"
        elif model_type == "mar_huge":
            diffloss_d = 12
            diffloss_w = 1536
            model_path = "mar-huge.safetensors"
        else:
            raise NotImplementedError
        # download and load the model weights (.safetensors or .pth)
        model_checkpoint_path = hf_hub_download(
            repo_id=kwargs.get("repo_id", "jadechoghari/mar"),
            filename=kwargs.get("model_filename", model_path)
        )

        model_fn = model_mapping[model_type]

        model = model_fn(
            buffer_size=64,
            diffloss_d=diffloss_d,
            diffloss_w=diffloss_w,
            num_sampling_steps=str(num_sampling_steps_diffloss)
        ).cuda()

        # use safetensors
        state_dict = load_file(safetensors_path)
        model.load_state_dict(state_dict)
        model.eval()

        # download and load the vae
        vae_checkpoint_path = hf_hub_download(
            repo_id=kwargs.get("repo_id", "jadechoghari/mar"),
            filename=kwargs.get("vae_filename", "kl16.ckpt")
        )
        vae_checkpoint_path = kwargs.get("vae_checkpoint_path", vae_checkpoint_path)

        vae = AutoencoderKL(embed_dim=16, ch_mult=(1, 1, 2, 2, 4), ckpt_path=vae_checkpoint_path)
        vae = vae.to(device).eval()

        # set up user-specified or default values for generation
        seed = kwargs.get("seed", 6)
        torch.manual_seed(seed)
        np.random.seed(seed)

        num_ar_steps = kwargs.get("num_ar_steps", 64)
        cfg_scale = kwargs.get("cfg_scale", 4)
        cfg_schedule = kwargs.get("cfg_schedule", "constant")
        temperature = kwargs.get("temperature", 1.0)
        class_labels = kwargs.get("class_labels", 207, 360, 388, 113, 355, 980, 323, 979)

        # generate the tokens and images
        with torch.cuda.amp.autocast():
            sampled_tokens = model.sample_tokens(
                bsz=len(class_labels), num_iter=num_ar_steps,
                cfg=cfg_scale, cfg_schedule=cfg_schedule,
                labels=torch.Tensor(class_labels).long().cuda(),
                temperature=temperature, progress=True
            )

            sampled_images = vae.decode(sampled_tokens / 0.2325)

        output_dir = kwargs.get("output_dir", "./")
        os.makedirs(output_dir, exist_ok=True)
    
        # save the images
        image_path = os.path.join(output_dir, "sampled_image.png")
        samples_per_row = kwargs.get("samples_per_row", 4)
    
        save_image(
            sampled_images, image_path, nrow=int(samples_per_row), normalize=True, value_range=(-1, 1)
        )
    
        # return as a pil image
        image = Image.open(image_path)
    
        return image