Add functionality to support the deep potential #827

Kick-H · 2024-12-15T13:49:26Z

Summary

I wrote a preliminary version and tested it with water system. After comparison, there was no problem with the force and energy. Here is the detailed installation and testing process.

The most complicated and time-consuming part of this version is probably the establishment of the neighbor list. I used the easiest method here, which requires completely mirroring the central cell to obtain the information of the 27 mirror boxes around it (boundary conditions p p p), and then completing the solution of the information of the atoms in the central box (the real box atoms).

This version was initially tested for speed and was several times slower than lammps, but it supports a larger simulation system size.

Modification

Added two files, force/deepmd.cu and force/deepmd.cuh, and added an example in examples/14_DP_water_msd, and the installation tutorial is also here (examples/14_DP_water_msd/Readme.md).

Others

Refer to the DP-lammps interface file pair_deepmd.cpp, which contains the dp.compute command, which can complete the solution of force and energy through coordinates and atomic types by calling dp's dependencies.

brucefan1983 · 2024-12-15T15:01:52Z

src/model/read_xyz.cu

@@ -548,8 +620,14 @@ void initialize_position(
  }

  std::vector<std::string> atom_symbols;
-  auto filename_potential = get_filename_potential();
+  auto filename_potential = get_out_potential();


This line should be separately writen in the #ifdef and #else parts

This line should be separately writen in the #ifdef and #else parts

OK, I've fixed it and it compiles.

brucefan1983 · 2024-12-15T15:04:57Z

Have you also tested virial stress and heat current?

Make output concise

Dankomaister · 2024-12-16T01:59:02Z

Nice work @Kick-H !
I curious why does deepmd need its own neighbour list implementation instead of just using the one thats already in GPUMD?

fix bug

Kick-H · 2024-12-16T11:09:47Z

Nice work @Kick-H ! I curious why does deepmd need its own neighbour list implementation instead of just using the one thats already in GPUMD?

This is related to the C++ interface of DP. If you are interested, welcome to check the source code.

Kick-H · 2024-12-16T11:11:44Z

Have you also tested virial stress and heat current?

I have tested stress. The heat flux test has not been done yet, but it is definitely worth doing.

Keep documents streamlined by removing unnecessary information

Dankomaister · 2024-12-17T03:52:34Z

Maybe worth replacing the cuda* specific calls with the macros we have (src/utilities/gpu_macro.cuh) ? Then this should work on AMD GPUs as well (assuming one has tensorflow and the deepmd lib)

Kick-H · 2024-12-17T08:50:47Z

Maybe worth replacing the cuda* specific calls with the macros we have (src/utilities/gpu_macro.cuh) ? Then this should work on AMD GPUs as well (assuming one has tensorflow and the deepmd lib)

OK, thank you very much for your suggestion.

replae the cuda* specific calls with the macros

brucefan1983 · 2024-12-19T18:02:26Z

General comments:

Do not modify read_xyz.cu, try to make a driver input file for DeePMD to be specified in run.in:

potential deepmd_driver_input_file.txt

The file deepmd_driver_input_file.txt can read as follows:

deepmd 2 H O
your_deepmd.pb

Move the special neighbor list code to deepmd.cu. Do not put special code in public file.
Use a separate makefile for deepmd. Do not touch the existing makefiles.
In summary, keep the other parts untouched. Check this PR for a very good model: A hybrid potential "NEP+ILP" for graphene and h-BN #835

QuantumMisaka · 2024-12-26T06:11:52Z

@Kick-H NICE WORK！But is this support for all backends and descriptors of DP model? or only support spefific (like TF backends) of DP model？

brucefan1983 · 2024-12-26T19:40:39Z

@Kick-H NICE WORK！But is this support for all backends and descriptors of DP model? or only support spefific (like TF backends) of DP model？

Currently only for the tensorflow backend, and the code is to be refactored (to make this part as isolated as possible) and optimized.

Delete the DUSE_tensorflow command of the Xu version

Kick-H · 2025-01-05T04:55:14Z

Hi, I tried to achieve it based on the code of @Kick-H these days and now there is some progress. 👼

Modification

Add two dp files: dp.cu and dp.cuh

Check dp in *force.cu"
  #ifdef DP_BHK
  } else if (strcmp(potential_name, "dp") == 0) {
    if (num_param != 3) {
      PRINT_INPUT_ERROR("potential should contain DP potential file behind setting file.\n");
    }
    potential.reset(new DP(param[2], number_of_atoms));
  #endif
Add two copy functions for GPU_Vector in gpu_vector.cuh to copy the GPU momery with an offset. I think it may be important so I didn't include them in macro definition.

Usage

I made a lot of mistakes when I built the environment before. So here, I want to show the usage method starting by setting up the environment. Maybe some experience could help you. My GPU is RTX4090 and system is Ubuntu 22.04. The environment controlling tool is conda

Create a new conda environment with python and activate it.
conda create -n tf-gpu2  python=3.9
conda activate tf-gpu2
Install CMake, CUDA-toolkit and Tensorflow. Please make sure the versions of CUDA-toolkit and Tensorflow is COMPATIBLE. My tensorflow version is 2.18.0.
pip install --upgrade cmake
conda install -c conda-forge cudatoolkit=11.8
pip install --upgrade tensorflow
Download DP source code and compile the source files following DP docs. Here is my cmake commands:
mkdir build
cd build
cmake -DENABLE_TENSORFLOW=TRUE -DUSE_CUDA_TOOLKIT=TRUE -DCMAKE_INSTALL_PREFIX=path_to_install -DUSE_TF_PYTHON_LIBS=TRUE ../
make -j
make install
We just need DP C++ interface, so we don't source all DP environment. The libraries will be installed in path_to_install.

Configure the makefile of GPUMD. The DP code is included by macro definition DP_BHK. So add it to CFLAGS
CFLAGS = -std=c++14 -O3 $(CUDA_ARCH) -DDP_BHK
Then we need to link the DP C++ libraries. Add this two lines to update the include and link paths and compile GPUMD.
INC += -Ipath_to_install/include/deepmd
LDFLAGS += -Lpath_to_install/lib -ldeepmd_cc
When run GPUMD, I get an error that could not find libraries of DP. So I need to add it to my library path. I choose a temporary method. Here is the run code:
LD_LIBRARY_PATH=path_to_install/lib:$LD_LIBRARY_PATH   gpumd
This DP interface need two files: setting file and DP potential file. The first file is very easy, used to make GPUMD know the atom number and types. For example:
dp 2 O H
! ! ! IMPORTANT ! ! ! The type list of setting file and potential file must be the same .

Test

The test files are in tests/dp/. I test on my computer and here are some results.

Energy conservation: run NVE in 1K with timestep 1fs

Speed: 4.7e4 atom*step/second for 1k5 atoms running 1e4 steps

DEBUG flag test:

Others

This interface may need more test 🚶 , such as virial. Welcome to test! ✨

TODO

Triclinic box is not supported.

Just support Tensorflow backend.

Thank you very much, I will check it as much as possiable.

Kick-H · 2025-01-05T04:59:28Z

I have change the #ifdef DP_BHK to #ifdef USE_TENSORFLOW, And the version is from @BBBuZHIDAO , which is faster. The old version has been backed up.

Kick-H · 2025-01-05T04:59:59Z

I will write a README and test it as soon as possible.

Kick-H · 2025-01-06T08:19:46Z

GPUMD supports DP potential project

Author: 徐克 (kickhsu[at]gmail.com)

Author: 卜河凯 (hekai_bu[at]whu.edu.cn)

WeChat Official Account: 微纳计算 (nanocomp)

0 Program Introduction

0.1 Necessary instructions

This is a test version.
Only potential function files ending with .pb in deepmd are supported, that is, the potential function files of the tensorflow version generated using dp --tf freeze.

0.2 Installation Dependencies

You must ensure that the new version of DP is installed and can run normally. This program contains DP-related dependencies.
The installation environment requirements of GPUMD itself must be met.

1 Installation details

Use the instance in AutoDL for testing。

If you need testing use AutoDL, please contact us.

And we have created an image in AutoDL that can run GPUMD-DP directly, which can be shared with the account that provides the user ID. Then, you will not require the following process and can be used directly.

2 GPUMD-DP installation (Offline version)

2.0 DP installation (Offline version)

Use the latest version of DP installation steps:

>> $ # Copy data and unzip files.
>> $ cd /root/autodl-tmp/
>> $ wget https://mirror.nju.edu.cn/github-release/deepmodeling/deepmd-kit/v3.0.0/deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.0 -O deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.0
>> $ wget https://mirror.nju.edu.cn/github-release/deepmodeling/deepmd-kit/v3.0.0/deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.1 -O deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.1
>> $ cat deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.0 deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.1 > deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh
>> $ # rm deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.0 deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh.1 # Please use with caution "rm"
>> $ sh deepmd-kit-3.0.0-cuda126-Linux-x86_64.sh -p /root/autodl-tmp/deepmd-kit -u # Just keep pressing Enter/yes.
>> $ source /root/autodl-tmp/deepmd-kit/bin/activate /root/autodl-tmp/deepmd-kit
>> $ dp -h

After running according to the above steps, using dp -h can successfully display no errors.

2.1 GPUMD-DP installation

The github link is Here.

>> $ wget https://codeload.github.com/Kick-H/GPUMD/zip/7af5267f4d8ba720830c154f11634a1942b66b08
>> $ cd ${GPUMD}/src-v0.1

Modify makefile as follows:

Line 19 is changed from CUDA_ARCH=-arch=sm_60 to CUDA_ARCH=-arch=sm_89 (for RTX 4090). Modify according to the corresponding graphics card model.
Line 25 is changed from INC = -I./ to INC = -I./ INC = -I./ -I/root/miniconda3/deepmd-kit/source/build/path_to_install/include/deepmd
Line 27 is changed fromLIBS = -lcublas -lcusolver to LIBS = -lcublas -lcusolver -L/root/miniconda3/deepmd-kit/source/build/path_to_install/lib -ldeepmd_cc

Then run the following installation command:

>> $ sudo echo "export LD_LIBRARY_PATH=/root/miniconda3/deepmd-kit/source/build/path_to_install/lib:$LD_LIBRARY_PATH" >> /root/.bashrc
>> $ source /root/.bashrc
>> $ make gpumd -j

2.2 Running Tests

>> $ cd /root/miniconda3/GPUMD-bu0/tests/dp
>> $ ../../src/gpumd

3 GPUMD-DP installation (Online version)

3.0 DP installation (Online version)

3.1 Conda environment

Create a new conda environment with python and activate it.

>> $ conda create -n tf-gpu2  python=3.9
>> $ conda activate tf-gpu2

3.2 Conda install some packages

Install CMake, CUDA-toolkit and Tensorflow. Please make sure the versions of CUDA-toolkit and Tensorflow is COMPATIBLE. My tensorflow version is 2.18.0.

>> $ pip install --upgrade cmake
>> $ conda install -c conda-forge cudatoolkit=11.8
>> $ pip install --upgrade tensorflow

3.3 download deep-kit and install

Download DP source code and compile the source files following DP docs. Here is cmake commands:

>> $ git clone https://github.com/deepmodeling/deepmd-kit.git
>> $ cd deepmd-kit/source
>> $ mkdir build
>> $ cd build
>> $ cmake -DENABLE_TENSORFLOW=TRUE -DUSE_CUDA_TOOLKIT=TRUE -DCMAKE_INSTALL_PREFIX=`path_to_install` -DUSE_TF_PYTHON_LIBS=TRUE ../
>> $ make -j
>> $ make install

We just need DP C++ interface, so we don't source all DP environment. The libraries will be installed in path_to_install.

3.4 Configure the makefile of `GPUMD`

The github link is Here.

>> $ wget https://codeload.github.com/Kick-H/GPUMD/zip/7af5267f4d8ba720830c154f11634a1942b66b08
>> $ cd ${GPUMD}/src
>> $ vi makefile

Configure the makefile of GPUMD. The DP code is included by macro definition USE_TENSORFLOW. So add it to CFLAGS

CFLAGS = -std=c++14 -O3 $(CUDA_ARCH) -DUSE_TENSORFLOW

Then we need to link the DP C++ libraries. Add this two lines to update the include and link paths and compile GPUMD.

INC += -Ipath_to_install/include/deepmd

LDFLAGS += -Lpath_to_install/lib -ldeepmd_cc

>> $ make gpumd -j

3.5 Run `GPUMD`

When run GPUMD, I get an error that could not find libraries of DP. So I need to add it to my library path. I choose a temporary method. Here is the run code:

LD_LIBRARY_PATH=path_to_install/lib:$LD_LIBRARY_PATH

Or you can add the environment to the ~/.bashrc

>> $ sudo echo "export LD_LIBRARY_PATH=/root/miniconda3/deepmd-kit/source/build/path_to_install/lib:$LD_LIBRARY_PATH" >> ~/.bashrc
>> $ source ~/.bashrc

3.6 Run Test

This DP interface need two files: setting file and DP potential file. The first file is very easy, used to make GPUMD know the atom number and types. For example:

dp 2 O H

Notice

The type list of setting file and potential file must be the same.

Kick-H

Tests that have been conducted include:

Basic information comparison with lammps, including force, energy, potential, etc.
Some dynamic processes.

The project has come to an end for the time being, and areas that need to be improved in the future:

Increase in calculation speed
Support for triclinic boxes

Kick-H added 5 commits December 15, 2024 20:13

Support GPUMD using DP potential function

c5fb40e

Support GPUMD using DP potential function

30e2f5d

Support GPUMD using DP potential function

7301398

Add a water example with GPUMD using DP potential

e6a5437

Delete the DP potential since too large

98d994d

brucefan1983 reviewed Dec 15, 2024

View reviewed changes

Update deepmd.cu

5485392

Make output concise

Update read_xyz.cu

89aba68

fix bug

Update model.xyz

bd6a853

Keep documents streamlined by removing unnecessary information

brucefan1983 changed the title ~~Add functionality to support Deepkit potential~~ Add functionality to support the deep potential Dec 16, 2024

Kick-H added 3 commits December 17, 2024 17:32

Update deepmd.cu

b971b15

Update deepmd.cu

982fd2b

replae the cuda* specific calls with the macros

Update deepmd.cu

e3618cc

FIX BUG: filename_potential is undefined

bb579eb

brucefan1983 marked this pull request as draft December 30, 2024 21:23

brucefan1983 requested a review from elindgren December 30, 2024 21:42

BBBuZHIDAO added 5 commits December 31, 2024 23:03

add dp files

005cc65

add func init dp file

29c658f

add dp nl init and delete code

a5ff51f

add a func to create dp position in gpu

49306ed

use stdvector to save dp nl

24b93af

Update read_xyz.cu

c907420

Delete the DUSE_tensorflow command of the Xu version

Kick-H marked this pull request as ready for review January 6, 2025 07:13

Kick-H marked this pull request as draft January 6, 2025 07:14

Kick-H added 21 commits January 8, 2025 18:55

Update dp.cu

ac26079

Update dp.cu

8a5f523

Update dp.cu

342d570

Merge branch 'brucefan1983:master' into master

c35a15c

Update dp.cu

59136db

Merge branch 'brucefan1983:master' into master

fc4c676

Update dp.cu

1571c32

Update neighbor.cu

5859e11

Update neighbor.cu

f2a7974

Update neighbor.cu

5278423

Update neighbor.cu

d55e46c

Update force.cu

900c6c6

Update force.cu

9652f64

Update force.cu

8522620

Update force.cu

5ea4270

Delete tests/dp directory

30233f8

Delete examples/14_DP_water_msd directory

cceed10

Update makefile

9a7584c

Create readme.md

9cf3db3

Update readme.md

5723df2

Add files via upload

69dc139

Kick-H commented Jan 11, 2025

View reviewed changes

Kick-H marked this pull request as ready for review January 11, 2025 03:37

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add functionality to support the deep potential #827

Add functionality to support the deep potential #827

Kick-H commented Dec 15, 2024

brucefan1983 Dec 15, 2024

Kick-H Dec 16, 2024

brucefan1983 commented Dec 15, 2024 •

edited

Loading

Dankomaister commented Dec 16, 2024

Kick-H commented Dec 16, 2024

Kick-H commented Dec 16, 2024

Dankomaister commented Dec 17, 2024

Kick-H commented Dec 17, 2024

brucefan1983 commented Dec 19, 2024 •

edited

Loading

QuantumMisaka commented Dec 26, 2024

brucefan1983 commented Dec 26, 2024

Kick-H commented Jan 5, 2025

Modification

Usage

Test

Others

TODO

Kick-H commented Jan 5, 2025

Kick-H commented Jan 5, 2025

Kick-H commented Jan 6, 2025 •

edited

Loading

Kick-H left a comment

Add functionality to support the deep potential #827

Are you sure you want to change the base?

Add functionality to support the deep potential #827

Conversation

Kick-H commented Dec 15, 2024

brucefan1983 Dec 15, 2024

Choose a reason for hiding this comment

Kick-H Dec 16, 2024

Choose a reason for hiding this comment

brucefan1983 commented Dec 15, 2024 • edited Loading

Dankomaister commented Dec 16, 2024

Kick-H commented Dec 16, 2024

Kick-H commented Dec 16, 2024

Dankomaister commented Dec 17, 2024

Kick-H commented Dec 17, 2024

brucefan1983 commented Dec 19, 2024 • edited Loading

QuantumMisaka commented Dec 26, 2024

brucefan1983 commented Dec 26, 2024

Kick-H commented Jan 5, 2025

Modification

Usage

Test

Others

TODO

Kick-H commented Jan 5, 2025

Kick-H commented Jan 5, 2025

Kick-H commented Jan 6, 2025 • edited Loading

GPUMD supports DP potential project

0 Program Introduction

0.1 Necessary instructions

0.2 Installation Dependencies

1 Installation details

2 GPUMD-DP installation (Offline version)

2.0 DP installation (Offline version)

2.1 GPUMD-DP installation

2.2 Running Tests

3 GPUMD-DP installation (Online version)

3.0 DP installation (Online version)

3.1 Conda environment

3.2 Conda install some packages

3.3 download deep-kit and install

3.4 Configure the makefile of GPUMD

3.5 Run GPUMD

3.6 Run Test

Notice

Kick-H left a comment

Choose a reason for hiding this comment

brucefan1983 commented Dec 15, 2024 •

edited

Loading

brucefan1983 commented Dec 19, 2024 •

edited

Loading

Kick-H commented Jan 6, 2025 •

edited

Loading

3.4 Configure the makefile of `GPUMD`

3.5 Run `GPUMD`