PLENA Software

Quantization and evaluation toolkit for MX-quantized LLMs.

Documentation

Full documentation is hosted at https://aicrosssim.github.io/PLENA_Software/, including a getting-started guide, CLI reference for evaluation commands, and the quantization TOML config reference.

Setup

Installation

uv venv                          # Python >= 3.11.9
source .venv/bin/activate
uv sync                          # core deps (mase, fast-hadamard-transform, lm-eval, ...)
uv sync --all-extras             # add docs / evalplus / serve / bfcl / dse as needed

Per-table configs and run scripts for the paper live under plena_experiments/.

Key dependencies

• mase (mase[mx-ptq]) — quantization framework. Provides quantize_module_transform_pass, GPTQ, rotation search. Pinned via [tool.uv.sources] to the releases/plena-experiments branch of DeepWok/mase.
• fast-hadamard-transform — Hadamard kernels used by rotation search. Pulled transitively as a git dep; built with no-build-isolation.

Online DSE (hardware design-space exploration)

GP + Expected Improvement Bayesian optimization over the 9 PLENA HardwareConfig knobs (BLEN/MLEN/VLEN/HLEN, vector SRAM, HBM size/width/prefetch). Maximizes TPS for a given LLM workload by evaluating candidates in-process via PLENA_Simulator's LLaMAModel.

uv sync --extra dse                                   # adds botorch + gpytorch
git clone https://github.com/AICrossSim/PLENA_Simulator.git ../PLENA_Simulator

python plena_experiments/online_dse/scripts/online_dse_gp_ei.py \
    plena_experiments/online_dse/configs/dse_llama3_8b.json

Outputs land in results/online_dse/{cache.json, results.json} (override with --cache PATH / -o PATH). See plena_experiments/online_dse/README.md for the full config schema, output format, and known limitations.