dataset_code/sft_sftnews/offload/offoload_features_backup.py

import os
from tqdm import tqdm
from diffusers import AutoencoderKLHunyuanVideo
from transformers import (
    CLIPTextModel,
    CLIPTokenizer,
    LlamaModel,
    LlamaTokenizerFast,
    SiglipImageProcessor,
    SiglipVisionModel,
)
from diffusers.video_processor import VideoProcessor
from diffusers.utils import export_to_video, load_image

from dataset_tool import CollectionDataset, collate_fn_map
from omegaconf import OmegaConf
from torch.utils.data import DataLoader

import torch
import torch.distributed as dist
import torch.nn as nn
from torch.nn.parallel import DistributedDataParallel as DDP
import torchvision.transforms as transforms
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.animation import FuncAnimation
from IPython.display import HTML, display
from IPython.display import clear_output  # 用于清理历史输出

from accelerate import Accelerator, DistributedType
from accelerate.logging import get_logger
from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed

from utils_framepack import encode_image, encode_prompt

def main(rank, world_size):
    weight_dtype = torch.bfloat16
    batch_size = 2
    dataloader_num_workers = 0
    output_latent_folder = "/mnt/bn/yufan-dev-my/ysh/Datasets/fp_offload_latents"
    pretrained_model_name_or_path = "/mnt/bn/yufan-dev-my/ysh/Ckpts/hunyuanvideo-community/HunyuanVideo"
    siglip_model_name_or_path = "/mnt/bn/yufan-dev-my/ysh/Ckpts/lllyasviel/flux_redux_bfl"
    os.makedirs(output_latent_folder, exist_ok=True)

    device = "cuda"

    # Load the tokenizers
    tokenizer_one = LlamaTokenizerFast.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="tokenizer",
    )
    tokenizer_two = CLIPTokenizer.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="tokenizer_2",
    )
    feature_extractor = SiglipImageProcessor.from_pretrained(
        siglip_model_name_or_path,
        subfolder="feature_extractor",

    )

    vae = AutoencoderKLHunyuanVideo.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="vae",
        torch_dtype=torch.float32,
    )
    vae_scale_factor_spatial = vae.spatial_compression_ratio
    video_processor = VideoProcessor(vae_scale_factor=vae_scale_factor_spatial)

    text_encoder_one = LlamaModel.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="text_encoder",
        torch_dtype=weight_dtype,
    )
    text_encoder_two = CLIPTextModel.from_pretrained(
        pretrained_model_name_or_path,
        subfolder="text_encoder_2",
        torch_dtype=weight_dtype,
    )
    image_encoder = SiglipVisionModel.from_pretrained(
        siglip_model_name_or_path,
        subfolder="image_encoder",
        torch_dtype=weight_dtype,
    )

    vae.requires_grad_(False)
    text_encoder_one.requires_grad_(False)
    text_encoder_two.requires_grad_(False)
    image_encoder.requires_grad_(False)
    vae.eval()
    text_encoder_one.eval()
    text_encoder_two.eval()
    image_encoder.eval()

    vae = vae.to(device)
    text_encoder_one = text_encoder_one.to(device)
    text_encoder_two = text_encoder_two.to(device)
    image_encoder = image_encoder.to(device)

    configs = OmegaConf.load("512_collection_config_vae1011_aligned_full_dump.yaml")
    dataset = CollectionDataset.create_dataset_function(configs['train_data'],
                                                            configs['train_data_weights'],
                                                            **configs['data']['params'])
    dataloader = DataLoader(
        dataset,
        shuffle=False,
        batch_size=batch_size,
        num_workers=dataloader_num_workers,
        collate_fn=collate_fn_map,
        pin_memory=True,
        prefetch_factor=2 if dataloader_num_workers != 0 else None,
        persistent_workers=True if dataloader_num_workers != 0 else False,
    )

    for idx, batch in tqdm(enumerate(dataloader), total=len(dataloader), desc="Processing batches"):
        exis_flag = True
        num_frames = batch["video_metadata"]["num_frames"]
        for uttid, num_frame in batch["uttid"], num_frames:
            output_path = os.path.join(output_latent_folder, f"{uttid}_{num_frame}.pt")
            if not os.path.exists(output_path):
                exis_flag = False
                break
        if exis_flag:
            print("skipping!")
            continue

        with torch.no_grad():
            # Get Vae feature
            pixel_values = batch["videos"].permute(0, 2, 1, 3, 4).to(dtype=vae.dtype, device=device)
            vae_latents = vae.encode(pixel_values).latent_dist.sample()
            vae_latents = vae_latents * vae.config.scaling_factor

            # Encode prompts
            prompts = batch["prompts"]
            prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = encode_prompt(
                tokenizer=tokenizer_one,
                text_encoder=text_encoder_one,
                tokenizer_2=tokenizer_two,
                text_encoder_2=text_encoder_two,
                prompt=prompts,
                device=device,
            )

            # Prepare images
            image_tensor = batch["first_frames_images"]
            images = [transforms.ToPILImage()(x.to(torch.uint8)) for x in image_tensor]
            image = video_processor.preprocess(image=images, height=batch["videos"].shape[-2], width=batch["videos"].shape[-1])
            image_embeds = encode_image(
                feature_extractor,
                image_encoder,
                image,
                device=device,
                dtype=weight_dtype,
            )

        for uttid, cur_vae_latent, cur_prompt_embed, cur_pooled_prompt_embed, cur_prompt_attention_mask, cur_image_embed in zip(batch["uttid"], vae_latents, prompt_embeds, pooled_prompt_embeds, prompt_attention_mask, image_embeds):
            output_path = os.path.join(output_latent_folder, f"{uttid}_{pixel_values.shape[2]}.pt")
            torch.save(
                {
                    "vae_latent": cur_vae_latent.cpu().detach(),
                    "prompt_embed": cur_prompt_embed.cpu().detach(),
                    "pooled_prompt_embeds": cur_pooled_prompt_embed.cpu().detach(),
                    "prompt_attention_mask": cur_prompt_attention_mask.cpu().detach(),
                    "image_embeds": cur_image_embed.cpu().detach(),
                },
                output_path
            )
            print(f"save to: {output_path}")

def setup_distributed_env():
    dist.init_process_group(backend="nccl")
    torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))

def cleanup_distributed_env():
    dist.destroy_process_group()

if __name__ == "__main__":
    setup_distributed_env()

    global_rank = dist.get_rank()
    local_rank = int(os.environ["LOCAL_RANK"])
    device = torch.cuda.current_device()
    world_size = dist.get_world_size()

    main(world_size=world_size, rank = device)