How GMM Algorithm Optimization Is Applied in the Framework

[SaturnWorld]

Dear developers and users:
In the paper 《Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning》,i find several optimizations for embedding net computations , such as
1. replace MATMUL and ADD operations with GEMM
2. replace CONCAT and ADD operations with GEMM
However, I do not find the relevant code implementation in the source code of the deepmd-kit framework. How does the framework apply this optimization solution? Thanks.

[tuoping]

@denghuilu

[denghuilu]

These optimizations did not apply to the deepmd-kit source code yet. The main reason is that we now have a better alternative, please see our introduction to model compression in the documentation. For the optimization of neural network structure, model compression can bring better speedups. You can enable this new feature in the deepmd-kit-v2.0 releases.

[denghuilu]

If you are interested in the paper's implementation, you can access the code by https://drive.google.com/file/d/1Rk42sTLHETIMD-qAdqTE0mQdM4vaVMJj/view

[SaturnWorld]

Thank you very much~