FlashTP [paper]

Overview

FlashTP is a high-performance Tensor-Product library for Machine Learning Interatomic Potentials (MLIPs). It is a drop-in replacement for channelwise tensorproduct ("uvu") that uses e3nn.o3.TensorProduct.

FlashTP implements fused, sparsity-aware CUDA kernels to accelerate the Tensor-Product layer in MLIPs, leveraging optimizations such as:

• Kernel fusion
• Sparse computation
• Input reuse

For full details on the optimizations, please see our ICML ’25 paper.

TODO

• Performance Improvements
• JIT implemenation

Prerequisites

• Python 3.11 (Not a strict requirement)
• CUDA Toolkit 12 or higher
• PyTorch 2.4.1 or higher
• e3nn
• torch_scatter (for accurate e3nn performance benchmarks)

Installation and features

1. Clone this repository:

   git clone https://github.com/SNU-ARC/flashTP.git
   cd flashTP

2. Ensure CUDA toolkit 12 or higher is installed on your system.

3. Install the Python dependencies and the package itself: (Compiling the CUDA kernels can take up to 10 minutes)

   By default, builds optimized kernels for NVIDIA A100 and H100 GPUs.

   You can customize CUDA_ARCH_LIST to match the Compute Capability of your NVIDIA GPU. Please view this link for Compute Capability.

   pip install -r requirements.txt
   CUDA_ARCH_LIST="80;90" pip install . --no-build-isolation 

   We recommend using Virtualenv or Conda to manage the dependencies.

How to use FlashTP

Standard e3nn TensorProduct

import torch
import e3nn.o3
from torch_scatter import scatter

# Initalization
tp = e3nn.o3.TensorProduct(i_in1,i_in2,i_out,inst_tuple,shared_weights=False, internal_weights=False)

# Execution
in1 = in1_node[edge_src]
out_large_e3nn = tp(in1,in2,weight)
out_e3nn = scatter(out_large_e3nn, edge_dst.to(torch.int64), dim=0, dim_size=total_node, reduce="sum")

FlashTP replacement

import torch
import flashTP_e3nn

# Initalization
flashtp = flashTP_e3nn.uvu_TP(i_in1,i_in2,i_out,inst_tuple, device="cuda", dtype=used_dtype)

# Execution
out_ours = flashtp(in1,in2,weight, edge_src, edge_dst)

Evaluation

Run ./example/test.sh to run microbenchmarks in ./ir_config. For detailed information on the input parameters, please refer to ./example/flashTP_test.py.

cd example
bash ./test.sh

For SevenNet integration please checkout flash branch of SevenNet.

Citation

Please cite our paper if you find our work useful.

@inproceedings{leeflashtp,
  title={FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials},
  author={Lee, Seung Yul and Kim, Hojoon and Park, Yutack and Jeong, Dawoon and Han, Seungwu and Park, Yeonhong and Lee, Jae W},
  booktitle={Forty-second International Conference on Machine Learning}
}