apuaachen/evt_ae

Artifacts of EVT ASPLOS'24

Artifact Evaluation of EVT ASPLOS'24

This repo is the EVT compiler targeting single-GPU & distributed training & inference of Deep Learning Models.

Compatibility

The following environment have been tested

Hardware	PyTorch Version	CUDA version
NVIDIA A100 Tensor Core GPU 40 GB	2.1.0a0+1767026	CUDA v12.1.66

Getting Started

We recommend using the docker image:

git clone https://github.com/apuaaChen/EVT_AE.git
cd EVT_AE
git submodule update --init --recursive
export MLCOMPILER_DIR=</path/to/your/clone>
bash build.sh evt_ae
docker run --gpus all --name evt_ae_test -v ${MLCOMPILER_DIR}:/workspace/EVT_AE -it evt_ae

Inside the docker container, to install gtl library:

cd /workspace/EVT_AE/python && bash install.sh

Reproduce the experimental results

To reproduce Figure 12, 13-17, run:

cd /workspace/EVT_AE
bash figure12.sh
bash figure13_17.sh

The reference results of the two figures are listed below:

evt_ae's Issues

3 evt_fuser UT failed

Hello, below 3 UT cases failed(accuracy checking failed): test_mm_partition10 / test_mm_partition11 / test_mm_partition12, other cases in the same .py are good, could you check if any bug in latest codes?