FlexPosit MICRO 2026 Artifact

See the paper's Artifact Appendix for the full description. This README is a quick reference.

Reproduce

export HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   # required for gated Llama-2

bash 01_install.sh              # ~40 min
bash 02_headline_ppl.sh         # Table 2       13-17 h
bash 03_ablation.sh             # Table 3       1 h
bash 04_mpq_granularity.sh      # Table 4       1 min
bash 05_hardware.sh             # Figures 11-12, Table 10  ~2-3 h
bash 06_act_quant.sh            # Table 6       0.5-1 h

Each script writes its CSV to results/ and ends with a verify.py diff against expected/ (±0.1 PPL tolerance).

Requirements

• 1x NVIDIA GPU with ≥40 GB VRAM and compute capability 8.0–9.0 (A40, A6000, L40S, A100, RTX 6000 Ada, H100, H200). Blackwell (compute 10.0+) requires PyTorch 2.5+ with CUDA 12.4+; swap requirements.txt accordingly.
• ~600 GB free disk
• Miniforge/Miniconda; NVIDIA driver ≥520
• HuggingFace token with access to meta-llama/Llama-2-7b-hf

If Meta approval is pending, pass SKIP_LLAMA2=1 to reproduce the other 8 models. On a 24 GB card, pass SKIP_QWEN14B=1 to skip the 14B model.

Contact

Yimin Gao yg9bq@virginia.edu