Multi-Device Support

The prosperity of the PaddlePaddle ecosystem depends on the contributions of developers and users. We warmly welcome more models contributed to PaddlePaddle's multi-hardware adaptation.

1. Hardware Support List

The list of models adapted for various hardware in PaddleScience is summarized as follows (excluding cases only on AIStudio):

Mathematics (AI for Math)Technology Science (AI for Technology)Material Science (AI for Material)Earth Science (AI for Earth Science)Chemical Science (AI for Chemistry)

Problem Type	Case Name	NVIDIA	Hygon	Taichu	MetaX
Helmholtz Equation	SPINN(Helmholtz3D)	✅		✅	✅
Phase Field Equation	Allen-Cahn	✅			✅
Differential Equation	Laplace Equation	✅		✅	✅
Differential Equation	Burgers Equation	✅			✅
Differential Equation	Non-linear Partial Differential Equation	✅			✅
Differential Equation	Lorenz Equation	✅		✅	✅
Differential Equation	Rossler Equation	✅		✅	✅
Operator Learning	DeepONet	✅		✅	✅
Integral Equation	Volterra Integral Equation	✅		✅	✅
Optical Rogue Wave	Optical rogue wave	✅		✅	✅
Domain Decomposition	XPINN	✅		✅	✅
Brusselator Diffusion System	3D-Brusselator	✅
Symbolic Regression	Transformer4SR	✅

Problem Type	Case Name	NVIDIA	Hygon	Taichu	MetaX
Car Surface Drag Prediction	DrivAerNet	✅			✅
1D Linear Convection Problem	1D Linear Convection	✅			✅
Unsteady Incompressible Fluid	2D Lid-Driven Cavity Buoyancy Flow	✅			✅
Steady Incompressible Fluid	Re3200 2D Steady Lid-Driven Cavity Flow	✅			✅
Steady Incompressible Fluid	2D Darcy Flow	✅		✅	✅
Steady Incompressible Fluid	2D Pipe Flow	✅			✅
Steady Incompressible Fluid	3D Intracranial Aneurysm	✅			✅
Steady Incompressible Fluid	Flow Around Arbitrary 2D Geometry	✅		✅	✅
Unsteady Incompressible Fluid	2D Unsteady Lid-Driven Cavity Flow	✅		✅	✅
Unsteady Incompressible Fluid	Re100 2D Flow Around Cylinder	✅			✅
Unsteady Incompressible Fluid	Re100~750 2D Flow Around Cylinder	✅			✅
Compressible Fluid	2D Air Shock Wave	✅			✅
General Flow Field Simulation	Aerodynamic Shape Design	✅
Fluid-Structure Interaction	Vortex-Induced Vibration	✅			✅
Multiphase Flow	Gas-Liquid Two-Phase Flow	✅			✅
High-Resolution Flow Field Reconstruction	2D Turbulent Flow Field Reconstruction	✅			✅
Solver Coupling	CFD-GCN	✅			✅
Stress Analysis	1D Euler Beam Deformation	✅		✅	✅
Stress Analysis	2D Plate Deformation	✅			✅
Stress Analysis	3D Bracket Deformation	✅			✅
Stress Analysis	Structural Vibration Simulation	✅			✅
Stress Analysis	2D Elastic-Plastic Structure	✅			✅
Stress Analysis and Inverse Problem	3D Car Control Arm Deformation	✅			✅
Stress Analysis and Inverse Problem	3D Heart Simulation	✅			✅
Topology Optimization	2D Topology Optimization	✅			✅
Thermal Simulation	1D Heat Exchanger Thermal Simulation	✅			✅
Thermal Simulation	2D Thermal Simulation	✅			✅
Thermal Simulation	2D Chip Thermal Simulation	✅
Operator Learning	NeuralOperator	✅
Car Surface Drag Prediction	DrivAerNetPlusPlus	✅
Solver Coupling	NSFNets	✅			✅
High-Resolution Flow Field Reconstruction	PhyCRNet	✅
Solver Coupling	NSFNet4	✅

Problem Type	Case Name	NVIDIA	Hygon	Taichu	MetaX
Material Design	Diffuser Design (Inverse Problem)	✅		✅	✅
Material Design	CGCNN	✅		✅

Problem Type	Case Name	NVIDIA	Hygon	Taichu	MetaX
Weather Forecasting	Extformer-MoE Weather Forecasting	✅		✅	✅
Weather Forecasting	FourCastNet Weather Forecasting	✅		✅
Weather Forecasting	NowCastNet Weather Forecasting	✅			✅
Weather Forecasting	GraphCast Weather Forecasting	✅		✅	✅
Weather Forecasting	DGMR Weather Forecasting	✅
Seismic Waveform Inversion	VelocityGAN Seismic Waveform Inversion	✅			✅
Traffic Prediction	TGCN Traffic Flow Prediction	✅			✅
Weather Forecasting	EarthFormer Weather Forecasting	✅			✅
Traffic Prediction	IOPS Traffic Flow Prediction	✅			✅
Weather Forecasting	Pangu-Weather Weather Forecasting	✅			✅
Weather Forecasting	FengWu Weather Forecasting	✅			✅

Problem Type	Case Name	NVIDIA	Hygon	Taichu	MetaX
Chemical Molecule Generation	Moflow	✅			✅
Chemical Reaction Prediction	IFM	✅			✅

2. Running Guide

For hardware already supported by PaddleScience, we provide a running example for each hardware, taking 1D Euler Beam Deformation as an example.

Note

Please ensure that you have correctly installed PaddlePaddle corresponding to the computing hardware in your environment, otherwise please refer to PaddleCustomDevice to connect your hardware code to PaddlePaddle.

NVIDIAHygonTaichuMetaX

# Install PaddleScience
git clone -b develop https://github.com/PaddlePaddle/PaddleScience.git
# If github clone is slow, you can use gitee clone
# git clone -b develop https://gitee.com/paddlepaddle/PaddleScience.git

cd PaddleScience

# install paddlesci with editable mode
python -m pip install -e . -i https://pypi.tuna.tsinghua.edu.cn/simple
cd examples/euler_beam

Model Training CommandModel Evaluation CommandModel Export CommandModel Inference Command

python euler_beam.py

python euler_beam.py mode=eval EVAL.pretrained_model_path=https://paddle-org.bj.bcebos.com/paddlescience/models/euler_beam/euler_beam_pretrained.pdparams

python euler_beam.py mode=export

python euler_beam.py mode=infer

# Install PaddleScience
git clone -b develop https://github.com/PaddlePaddle/PaddleScience.git
# If github clone is slow, you can use gitee clone
# git clone -b develop https://gitee.com/paddlepaddle/PaddleScience.git

cd PaddleScience

# install paddlesci with editable mode
python -m pip install -e . -i https://pypi.tuna.tsinghua.edu.cn/simple
cd examples/euler_beam

Model Training CommandModel Evaluation CommandModel Export CommandModel Inference Command

python euler_beam.py

# Test your trained model
python euler_beam.py mode=eval EVAL.pretrained_model_path=$YOUR_MODEL_PATH
# Test officially provided pre-trained model
python euler_beam.py mode=eval EVAL.pretrained_model_path=https://paddle-org.bj.bcebos.com/paddlescience/models/euler_beam/euler_beam_pretrained.pdparams

python euler_beam.py mode=export

python euler_beam.py mode=infer

# Install PaddleScience
git clone -b develop https://github.com/PaddlePaddle/PaddleScience.git
# If github clone is slow, you can use gitee clone
# git clone -b develop https://gitee.com/paddlepaddle/PaddleScience.git

cd PaddleScience

# install paddlesci with editable mode
python -m pip install -e . -i https://pypi.tuna.tsinghua.edu.cn/simple
cd examples/euler_beam

Model Training CommandModel Evaluation CommandModel Export CommandModel Inference Command

python euler_beam.py

# Test your trained model
python euler_beam.py mode=eval EVAL.pretrained_model_path=$YOUR_MODEL_PATH
# Test officially provided pre-trained model
python euler_beam.py mode=eval EVAL.pretrained_model_path=https://paddle-org.bj.bcebos.com/paddlescience/models/euler_beam/euler_beam_pretrained.pdparams

python euler_beam.py mode=export

python euler_beam.py mode=infer INFER.device=sdaa

# Install PaddleScience
git clone -b develop https://github.com/PaddlePaddle/PaddleScience.git
# If github clone is slow, you can use gitee clone
# git clone -b develop https://gitee.com/paddlepaddle/PaddleScience.git

cd PaddleScience

# install paddlesci with editable mode
python -m pip install -e . -i https://pypi.tuna.tsinghua.edu.cn/simple
cd examples/euler_beam

Model Training CommandModel Evaluation CommandModel Export CommandModel Inference Command

python euler_beam.py

# Test your trained model
python euler_beam.py mode=eval EVAL.pretrained_model_path=$YOUR_MODEL_PATH
# Test officially provided pre-trained model
python euler_beam.py mode=eval EVAL.pretrained_model_path=https://paddle-org.bj.bcebos.com/paddlescience/models/euler_beam/euler_beam_pretrained.pdparams

python euler_beam.py mode=export

python euler_beam.py mode=infer

3. Contribution Guide

We provide reference accuracy and corresponding pre-trained model weights based on NVIDIA CUDA training at the beginning of the public case documentation. If you need to run on specified hardware, you can refer to the following steps:

1. If your hardware type has not been connected to PaddlePaddle, you can refer to the PaddleCustomDevice official documentation to connect to the PaddlePaddle framework. If your hardware type has been connected to PaddlePaddle but has not been added to the hardware support list of PaddleScience, please add your hardware type in ppsci/utils/config.py and deploy/python_infer/base.py.

2. Prepare necessary datasets according to the steps given in the case documentation.

3. If the model documentation provides a model training command, you need to perform full training on your hardware, save training logs, record the best model accuracy and the best model weights. These contents are generally automatically saved in the case folder during the training process.

4. If the model documentation provides a model evaluation command, you need to evaluate the accuracy of the best model saved in the third step on your hardware, save evaluation logs, and record evaluation accuracy. These contents are generally automatically saved in the case folder during the evaluation process.

   Note

   For full model training accuracy, the default requirement is that the best accuracy aligns with NVIDIA CUDA accuracy. Specifically, if the case accuracy indicator is relative error (such as L2 relative error), the indicator cannot exceed the reference value by ± 0.5%. If the case accuracy indicator is an error like MSE/MAE, it should be in the same order of magnitude as the reference value.

5. If the model documentation provides model export and inference commands, please verify whether model export and inference can be executed normally on the new hardware and align with CUDA inference results according to the model export and inference commands.

6. After completing the above steps, you can add your hardware support information (✅) to the corresponding model in the table of 1. Hardware Support List, and then submit a PR to PaddleScience. Your PR should include at least the following contents:

   • Add running instructions for using the model based on your hardware environment in 2. Running Guide
   • Best model weight file saved during training (.pdparams file)
   • Running logs such as training/evaluation (.log file)
   • Software versions used to verify model accuracy, including but not limited to:
     • PaddlePaddle version
     • PaddleCustomDevice version (if any)
   • Machine environment used to verify model accuracy, including but not limited to:
     • Chip model
     • System version
     • Hardware driver version
     • Operator library version, etc.

4. More Documentation

For more documents on PaddlePaddle multi-hardware adaptation and usage, please refer to:

• PaddlePaddle User Guide
• PaddlePaddle Hardware Support
• PaddleCustomDevice