FastDeploy/examples/multimodal/stable_diffusion/export.md

English | 简体中文

Diffusion Model Export

The project supports two methods of model export, PPDiffusers model export and Diffusers model export. Here we introduce each of these two methods.

PPDiffusers Model Export

PPDiffusers is a Diffusion Model toolkit that supports cross-modal (e.g., image and speech) training and inference. It builds on the design of Diffusers by the 🤗 Huggingface team, and relies on PaddlePaddle framework and the PaddleNLP natural language processing library. The following describes how to use FastDeploy to deploy the Diffusion model provided by PPDiffusers for high performance.

Dependency Installation

The model export depends on paddlepaddle, paddlenlp and ppdiffusers, which can be installed quickly by running the following command using pip.

pip install -r requirements_paddle.txt

Model Export

Note: The StableDiffusion model needs to be downloaded during the model export process. In order to use the model and weights, you must accept the License required. Please visit HuggingFace's model card, to read the License carefully, and then sign the agreement.

Tips: Stable Diffusion is based on these Licenses: The CreativeML OpenRAIL M license is an Open RAIL M license, adapted from the work that BigScience and the RAIL Initiative are jointly carrying in the area of responsible AI licensing. See also the article about the BLOOM Open RAIL license on which this license is based.

You can run the following lines to export model.

python export_model.py --pretrained_model_name_or_path CompVis/stable-diffusion-v1-4 --output_path stable-diffusion-v1-4

The output model directory is as follows:

stable-diffusion-v1-4/
├── text_encoder
│   ├── inference.pdiparams
│   ├── inference.pdiparams.info
│   └── inference.pdmodel
├── unet
│   ├── inference.pdiparams
│   ├── inference.pdiparams.info
│   └── inference.pdmodel
└── vae_decoder
    ├── inference.pdiparams
    ├── inference.pdiparams.info
    └── inference.pdmodel

Parameters

Here is description of each command line parameter in export_model.py.

Parameter	Description
--pretrained_model_name_or_path	The diffusion pretrained model provided by ppdiffuers. Default is "CompVis/stable-diffusion-v1-4". For more diffusion pretrained models, please refer to ppdiffuser model list.
--output_path	Exported directory

Diffusers Model Export

Diffusers is a Diffusion Model toolkit built by HuggingFace to support cross-modal (e.g. image and speech) training and inference. The underlying model code is available in both a PyTorch implementation and a Flax implementation. This example shows how to use FastDeploy to deploy a PyTorch implementation of Diffusion Model for high performance.

Dependency Installation

The model export depends on onnx, torch, diffusers and transformers, which can be installed quickly by running the following command using pip.

pip install -r requirements_torch.txt

Model Export

Note: The StableDiffusion model needs to be downloaded during the model export process. In order to use the model and weights, you must accept the License required, and get the Token granted by HF Hub. Please visit HuggingFace's model card, to read the License carefully, and then sign the agreement.

Tips: Stable Diffusion is based on these Licenses: The CreativeML OpenRAIL M license is an Open RAIL M license, adapted from the work that BigScience and the RAIL Initiative are jointly carrying in the area of responsible AI licensing. See also the article about the BLOOM Open RAIL license on which this license is based.

If you are exporting a model for the first time, you need to log in to the HuggingFace client first. Run the following command to log in:

huggingface-cli login

After finishing the login, you can run the following lines to export model.

python export_torch_to_onnx_model.py --pretrained_model_name_or_path CompVis/stable-diffusion-v1-4 --output_path torch_diffusion_model

The output model directory is as follows:

torch_diffusion_model/
├── text_encoder
│   └── inference.onnx
├── unet
│   └── inference.onnx
└── vae_decoder
    └── inference.onnx

Parameters

Here is description of each command line parameter in export_torch_to_onnx_model.py.

Parameter	Description
--pretrained_model_name_or_path	The diffusion pretrained model provided by ppdiffuers, default is "CompVis/stable-diffusion-v1-4". For more diffusion pretrained models, please refer to HuggingFace model list.
--output_path	Exported directory