中文阅读

Fine-tune Cog family of video models for custom video generation under 24GB of GPU memory ⚡️📼

Clone the repository and make sure the requirements are installed: pip install -r requirements.txt and install diffusers from source by pip install git+https://github.com/huggingface/diffusers.

Then download a dataset:

# install `huggingface_hub`
huggingface-cli download \
  --repo-type dataset Wild-Heart/Disney-VideoGeneration-Dataset \
  --local-dir video-dataset-disney

Then launch LoRA fine-tuning for text-to-video (modify the different hyperparameters, dataset root, and other configuration options as per your choice):

# For LoRA finetuning of the text-to-video CogVideoX models
./train_text_to_video_lora.sh

# For full finetuning of the text-to-video CogVideoX models
./train_text_to_video_sft.sh

# For LoRA finetuning of the image-to-video CogVideoX models
./train_image_to_video_lora.sh

Assuming your LoRA is saved and pushed to the HF Hub, and named my-awesome-name/my-awesome-lora, we can now use the finetuned model for inference:

import torch
from diffusers import CogVideoXPipeline
from diffusers.utils import export_to_video

pipe = CogVideoXPipeline.from_pretrained(
    "THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16
).to("cuda")
+ pipe.load_lora_weights("my-awesome-name/my-awesome-lora", adapter_name="cogvideox-lora")
+ pipe.set_adapters(["cogvideox-lora"], [1.0])

video = pipe("<my-awesome-prompt>").frames[0]
export_to_video(video, "output.mp4", fps=8)

For Image-to-Video LoRAs trained with multiresolution videos, one must also add the following lines (see this Issue for more details):

from diffusers import CogVideoXImageToVideoPipeline

pipe = CogVideoXImageToVideoPipeline.from_pretrained(
    "THUDM/CogVideoX-5b-I2V", torch_dtype=torch.bfloat16
).to("cuda")

# ...

del pipe.transformer.patch_embed.pos_embedding
pipe.transformer.patch_embed.use_learned_positional_embeddings = False
pipe.transformer.config.use_learned_positional_embeddings = False

You can also check if your LoRA is correctly mounted here.

Below we provide additional sections detailing on more options explored in this repository. They all attempt to make fine-tuning for video models as accessible as possible by reducing memory requirements as much as possible.

Before starting the training, please check whether the dataset has been prepared according to the dataset specifications. We provide training scripts suitable for text-to-video and image-to-video generation, compatible with the CogVideoX model family. Training can be started using the train*.sh scripts, depending on the task you want to train. Let's take LoRA fine-tuning for text-to-video as an example.

• Configure environment variables as per your choice:

  export TORCH_LOGS="+dynamo,recompiles,graph_breaks"
  export TORCHDYNAMO_VERBOSE=1
  export WANDB_MODE="offline"
  export NCCL_P2P_DISABLE=1
  export TORCH_NCCL_ENABLE_MONITORING=0

• Configure which GPUs to use for training: GPU_IDS="0,1"

• Choose hyperparameters for training. Let's try to do a sweep on learning rate and optimizer type as an example:

  LEARNING_RATES=("1e-4" "1e-3")
  LR_SCHEDULES=("cosine_with_restarts")
  OPTIMIZERS=("adamw" "adam")
  MAX_TRAIN_STEPS=("3000")

• Select which Accelerate configuration you would like to train with: ACCELERATE_CONFIG_FILE="accelerate_configs/uncompiled_1.yaml". We provide some default configurations in the accelerate_configs/ directory - single GPU uncompiled/compiled, 2x GPU DDP, DeepSpeed, etc. You can create your own config files with custom settings using accelerate config --config_file my_config.yaml.

• Specify the absolute paths and columns/files for captions and videos.

  DATA_ROOT="/path/to/my/datasets/video-dataset-disney"
  CAPTION_COLUMN="prompt.txt"
  VIDEO_COLUMN="videos.txt"

• Launch experiments sweeping different hyperparameters:

  for learning_rate in "${LEARNING_RATES[@]}"; do
    for lr_schedule in "${LR_SCHEDULES[@]}"; do
      for optimizer in "${OPTIMIZERS[@]}"; do
        for steps in "${MAX_TRAIN_STEPS[@]}"; do
          output_dir="/path/to/my/models/cogvideox-lora__optimizer_${optimizer}__steps_${steps}__lr-schedule_${lr_schedule}__learning-rate_${learning_rate}/"
  
          cmd="accelerate launch --config_file $ACCELERATE_CONFIG_FILE --gpu_ids $GPU_IDS training/cogvideox_text_to_video_lora.py \
            --pretrained_model_name_or_path THUDM/CogVideoX-5b \
            --data_root $DATA_ROOT \
            --caption_column $CAPTION_COLUMN \
            --video_column $VIDEO_COLUMN \
            --id_token BW_STYLE \
            --height_buckets 480 \
            --width_buckets 720 \
            --frame_buckets 49 \
            --dataloader_num_workers 8 \
            --pin_memory \
            --validation_prompt \"BW_STYLE A black and white animated scene unfolds with an anthropomorphic goat surrounded by musical notes and symbols, suggesting a playful environment. Mickey Mouse appears, leaning forward in curiosity as the goat remains still. The goat then engages with Mickey, who bends down to converse or react. The dynamics shift as Mickey grabs the goat, potentially in surprise or playfulness, amidst a minimalistic background. The scene captures the evolving relationship between the two characters in a whimsical, animated setting, emphasizing their interactions and emotions:::BW_STYLE A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance\" \
            --validation_prompt_separator ::: \
            --num_validation_videos 1 \
            --validation_epochs 10 \
            --seed 42 \
            --rank 128 \
            --lora_alpha 128 \
            --mixed_precision bf16 \
            --output_dir $output_dir \
            --max_num_frames 49 \
            --train_batch_size 1 \
            --max_train_steps $steps \
            --checkpointing_steps 1000 \
            --gradient_accumulation_steps 1 \
            --gradient_checkpointing \
            --learning_rate $learning_rate \
            --lr_scheduler $lr_schedule \
            --lr_warmup_steps 400 \
            --lr_num_cycles 1 \
            --enable_slicing \
            --enable_tiling \
            --optimizer $optimizer \
            --beta1 0.9 \
            --beta2 0.95 \
            --weight_decay 0.001 \
            --max_grad_norm 1.0 \
            --allow_tf32 \
            --report_to wandb \
            --nccl_timeout 1800"
          
          echo "Running command: $cmd"
          eval $cmd
          echo -ne "-------------------- Finished executing script --------------------\n\n"
        done
      done
    done
  done
  

  To understand what the different parameters mean, you could either take a look at the args file or run the training script with --help.

Note: Training scripts are untested on MPS, so performance and memory requirements can differ widely compared to the CUDA reports below.

CogVideoX LoRA Finetuning
THUDM/CogVideoX-2b	THUDM/CogVideoX-5b
CogVideoX Full Finetuning
THUDM/CogVideoX-2b	THUDM/CogVideoX-5b

Supported and verified memory optimizations for training include:

• CPUOffloadOptimizer from torchao. You can read about its capabilities and limitations here. In short, it allows you to use the CPU for storing trainable parameters and gradients. This results in the optimizer step happening on the CPU, which requires a fast CPU optimizer, such as torch.optim.AdamW(fused=True) or applying torch.compile on the optimizer step. Additionally, it is recommended not to torch.compile your model for training. Gradient clipping and accumulation is not supported yet either.
• Low-bit optimizers from bitsandbytes. TODO: to test and make torchao ones work
• DeepSpeed Zero2: Since we rely on accelerate, follow this guide to configure your accelerate installation to enable training with DeepSpeed Zero2 optimizations.

Slaying OOMs with PyTorch

• Make scripts compatible with DDP
• Make scripts compatible with FSDP
• Make scripts compatible with DeepSpeed
• vLLM-powered captioning script
• Multi-resolution/frame support in prepare_dataset.py
• Analyzing traces for potential speedups and removing as many syncs as possible
• Support for QLoRA (priority), and other types of high usage LoRAs methods
• Test scripts with memory-efficient optimizer from bitsandbytes
• Test scripts with CPUOffloadOptimizer, etc.
• Test scripts with torchao quantization, and low bit memory optimizers (Currently errors with AdamW (8/4-bit torchao))
• Test scripts with AdamW (8-bit bitsandbytes) + CPUOffloadOptimizer (with gradient offloading) (Currently errors out)
• Sage Attention (work with the authors to support backward pass, and optimize for A100)