Add project code

README.md

@@ -1,3 +1,36 @@
+# 15-745 Project - MagPy with TVM
+
+This branch is the integration of MagPy with TVM, for eager-mode DNN
+program compilation in TVM.
+
+## Project Reports
+
+* [Project proposal](reports/15745_Project_Proposal.pdf)
+* [Project milestone report](reports/15745_Project_Milestone_Report.pdf)
+* Project final report (TBA)
+
+## Reproduce the results
+
+Please follow the steps below to reproduce the evaluation results.
+
+**Step 1.** Install TVM via
+
+```bash
+python -m pip install --pre -U -f https://mlc.ai/wheels mlc-ai-nightly-cu123
+```
+
+**Step 2.** Install MagPy by following [the MagPy installation](#installation).
+
+**Step 3.** Run evaluation via
+
+```bash
+./eval_script/run_15.sh
+```
+
+
+
+---
+
 # MagPy
 MagPy is a JIT compiler for PyTorch programs. It can extract the operator graph from PyTorch programs and optimize the graph with a wide range of deep learning graph compilers.
 

eval_script/run_15.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+
+set -euxo pipefail
+
+TIME_TAG=`date +%y%m%d-%H%M%S`
+
+export GPU_NUM_DEVICES=1
+export PJRT_DEVICE=GPU
+
+LOG_DIR=logs/e2e
+mkdir -p $LOG_DIR
+
+for bs in 1 16
+do
+    for model in align bert deberta densenet monodepth quantized resnet tridentnet
+    for model in resnet
+    do
+        for compile in eager dynamo sys script sys-torchscript sys-tvm
+        do
+            rm -rf $LOG_DIR/$model.$bs.$compile.log
+            LD_PRELOAD=build/ldlong.v3.9.12.so python3 run.py --bs $bs --model $model --compile $compile 2>&1 | tee $LOG_DIR/$model.$bs.$compile.log
+        done
+    done
+done
+
+wait