ML Compiler Benchmarks: DVM & BladeDISC Reproduction

Reproduction of two ML compiler papers focused on dynamic shape compilation:

• DVM - Bytecode VM approach for dynamic tensor computation (arXiv 2603.24239v2)
• BladeDISC - Compiler-based symbolic shape optimization (SIGMOD 2024)

Results Summary

Paper	What was tested	Key result
DVM	Baselines (eager, JIT) on BERT-base	Recompilation costs 70-220x more than inference
BladeDISC	Full reproduction via Docker	Compile once, run 8 shapes: 1.09-1.56x speedup, numerically correct

Structure

├── reproduction_report.pdf    # Full report with embedded figures
├── code/                      # All benchmark scripts
├── results/                   # Raw CSV measurements (81 total)
└── plots/                     # 8 publication-quality figures

Hardware

• NVIDIA GeForce RTX 3050 Laptop GPU (4 GB VRAM)
• Windows 11 + WSL2 Ubuntu 22.04 + Docker (for BladeDISC)

How to reproduce

DVM baselines (Windows, needs PyTorch + transformers):

python code/dvm_baseline_benchmark.py
python code/plot_results.py

BladeDISC (needs WSL2 + Docker + NVIDIA Container Toolkit):

docker pull bladedisc/bladedisc:latest-runtime-torch-2.0.1-cu118
docker run --rm --gpus all -v $(pwd):/scripts:ro bladedisc/bladedisc:latest-runtime-torch-2.0.1-cu118 \
  bash -c "pip install transformers && python3 /scripts/code/bladedisc_reproduction.py"