ONNX GenAI CogVideoX package

Public real-weight text-to-video derivative exported to ONNX by Mobius from zai-org/CogVideoX-2b revision 1137dacfc2c9c012bed6a0793f4ecf2ca8e7ba01 (apache-2.0). Canonical inference_metadata.yaml is hashless workflow IR with typed DDIM v_prediction, rank-5 transpose, latent unscale, chunked causal VAE decode, and temporal cache recurrence.

License, attribution, and modifications

CogVideoX-2B, including its transformer and VAE, is Apache-2.0 licensed. Section 4 permits public redistribution and derivatives when the license is provided, modified files carry prominent change notices, and applicable copyright, patent, trademark, and attribution notices are retained. This repository includes the upstream LICENSE; see NOTICE.md for attribution and modification details.

Download

hf download justinchuby/onnx-genai-cogvideox-2b --repo-type model --local-dir cogvideox-2b

Generic ONNX GenAI execution

git clone https://github.com/justinchuby/onnx-genai.git
cd onnx-genai && git checkout justinchuby/catalogue-diffusion-runtime
ORT=/path/to/onnxruntime-gpu/lib
VIDEO_WORKFLOW_PACKAGE_DIR=../cogvideox-2b VIDEO_WORKFLOW_INPUT_DIR=../cogvideox-2b/generic_runtime_inputs VIDEO_WORKFLOW_OUTPUT_DIR=../cogvideox-2b/generic_runtime_outputs VIDEO_WORKFLOW_STEPS=20 ONNX_GENAI_EP=cuda ONNX_GENAI_EP_FALLBACK=1 ONNX_GENAI_ORT_LIB=$ORT/libonnxruntime.so.1.28.0 ONNX_GENAI_ORT_LIB_DIR=$ORT cargo test -p onnx-genai-engine --features onnx-genai-ort/cuda --test video_workflow_e2e -- --ignored --nocapture

Evidence: generated.mp4 and frames/*.png from the Mobius runner; generic_runtime_outputs/generated.mp4 and its frame sequence from the ONNX GenAI workflow engine. Full versions, timings, and peak memory are in execution_evidence.json and generic_runtime_evidence.json.

license: apache-2.0 language:

• en tags:
• cogvideox
• video-generation
• thudm
• text-to-video inference: false

CogVideoX-2B

Demo Show

Model Introduction

CogVideoX is an open-source version of the video generation model originating from QingYing. The table below displays the list of video generation models we currently offer, along with their foundational information.

Data Explanation

• When testing using the diffusers library, all optimizations provided by the diffusers library were enabled. This solution has not been tested for actual VRAM/memory usage on devices other than NVIDIA A100 / H100. Generally, this solution can be adapted to all devices with NVIDIA Ampere architecture and above. If the optimizations are disabled, VRAM usage will increase significantly, with peak VRAM usage being about 3 times higher than the table shows. However, speed will increase by 3-4 times. You can selectively disable some optimizations, including:

pipe.enable_model_cpu_offload()
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()

• When performing multi-GPU inference, the enable_model_cpu_offload() optimization needs to be disabled.
• Using INT8 models will reduce inference speed. This is to ensure that GPUs with lower VRAM can perform inference normally while maintaining minimal video quality loss, though inference speed will decrease significantly.
• The 2B model is trained with FP16 precision, and the 5B model is trained with BF16 precision. We recommend using the precision the model was trained with for inference.
• PytorchAO and Optimum-quanto can be used to quantize the text encoder, Transformer, and VAE modules to reduce CogVideoX's memory requirements. This makes it possible to run the model on a free T4 Colab or GPUs with smaller VRAM! It is also worth noting that TorchAO quantization is fully compatible with torch.compile, which can significantly improve inference speed. FP8 precision must be used on devices with NVIDIA H100 or above, which requires installing the torch, torchao, diffusers, and accelerate Python packages from source. CUDA 12.4 is recommended.
• The inference speed test also used the above VRAM optimization scheme. Without VRAM optimization, inference speed increases by about 10%. Only the diffusers version of the model supports quantization.
• The model only supports English input; other languages can be translated into English during refinement by a large model.

Note

• Using SAT for inference and fine-tuning of SAT version models. Feel free to visit our GitHub for more information.

Quick Start 🤗

This model supports deployment using the huggingface diffusers library. You can deploy it by following these steps.

We recommend that you visit our GitHub and check out the relevant prompt optimizations and conversions to get a better experience.

1. Install the required dependencies

# diffusers>=0.30.1
# transformers>=0.44.0
# accelerate>=0.33.0 (suggest install from source)
# imageio-ffmpeg>=0.5.1
pip install --upgrade transformers accelerate diffusers imageio-ffmpeg 

2. Run the code

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

prompt = "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."

pipe = CogVideoXPipeline.from_pretrained(
    "THUDM/CogVideoX-2b",
    torch_dtype=torch.float16
)

pipe.enable_model_cpu_offload()
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()
video = pipe(
    prompt=prompt,
    num_videos_per_prompt=1,
    num_inference_steps=50,
    num_frames=49,
    guidance_scale=6,
    generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(video, "output.mp4", fps=8)

Quantized Inference

PytorchAO and Optimum-quanto can be used to quantize the Text Encoder, Transformer and VAE modules to lower the memory requirement of CogVideoX. This makes it possible to run the model on free-tier T4 Colab or smaller VRAM GPUs as well! It is also worth noting that TorchAO quantization is fully compatible with torch.compile, which allows for much faster inference speed.

# To get started, PytorchAO needs to be installed from the GitHub source and PyTorch Nightly.
# Source and nightly installation is only required until next release.

import torch
from diffusers import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel, CogVideoXPipeline
from diffusers.utils import export_to_video
+ from transformers import T5EncoderModel
+ from torchao.quantization import quantize_, int8_weight_only, int8_dynamic_activation_int8_weight

+ quantization = int8_weight_only

+ text_encoder = T5EncoderModel.from_pretrained("THUDM/CogVideoX-5b", subfolder="text_encoder", torch_dtype=torch.bfloat16)
+ quantize_(text_encoder, quantization())

+ transformer = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16)
+ quantize_(transformer, quantization())

+ vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-2b", subfolder="vae", torch_dtype=torch.bfloat16)
+ quantize_(vae, quantization())

# Create pipeline and run inference
pipe = CogVideoXPipeline.from_pretrained(
    "THUDM/CogVideoX-2b",
+    text_encoder=text_encoder,
+    transformer=transformer,
+    vae=vae,
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
pipe.vae.enable_tiling()

prompt = "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."

video = pipe(
    prompt=prompt,
    num_videos_per_prompt=1,
    num_inference_steps=50,
    num_frames=49,
    guidance_scale=6,
    generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(video, "output.mp4", fps=8)

Additionally, the models can be serialized and stored in a quantized datatype to save disk space when using PytorchAO. Find examples and benchmarks at these links:

• torchao
• quanto

Explore the Model

Welcome to our github, where you will find:

1. More detailed technical details and code explanation.
2. Optimization and conversion of prompt words.
3. Reasoning and fine-tuning of SAT version models, and even pre-release.
4. Project update log dynamics, more interactive opportunities.
5. CogVideoX toolchain to help you better use the model.
6. INT8 model inference code support.

Model License

The CogVideoX-2B model (including its corresponding Transformers module and VAE module) is released under the Apache 2.0 License.

The CogVideoX-5B model (Transformers module) is released under the CogVideoX LICENSE.

Citation

@article{yang2024cogvideox,
  title={CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer},
  author={Yang, Zhuoyi and Teng, Jiayan and Zheng, Wendi and Ding, Ming and Huang, Shiyu and Xu, Jiazheng and Yang, Yuanming and Hong, Wenyi and Zhang, Xiaohan and Feng, Guanyu and others},
  journal={arXiv preprint arXiv:2408.06072},
  year={2024}
}

Annotated inference metadata

Review inference_metadata.annotated.yaml for inline explanations of this package's workflow, tensor/state/cache contracts, and fail-closed omissions. inference_metadata.yaml remains the canonical machine-authored contract; automated validation confirms both files parse to the same metadata object.