requirements-paper.txt

# Provenance only -- do NOT `pip install -r` this to use the tool. Install the
# package instead: `pip install -e ".[qe,serve,eval]"` (floors-only, tracks
# current releases). This file records the exact environment behind the paper's
# reported numbers.
#
# Hardware: 2x NVIDIA A100 SXM 80 GB, Python 3.11, CUDA 12.x. TPOT is specific
# to this setup; newer vLLM or different GPUs (e.g. H100 with W8A8 FP8) shift
# TPOT and may change the selected lambda*.
#
# Install vllm first (it pulls torch 2.10.0+cu128), then upgrade transformers:
# vLLM 0.19.0 defaults to transformers 4.57.6, which lacks the `gemma4` arch;
# 5.5.3 is needed for the Gemma models and runs fine despite vLLM's `<5` pin.
vllm==0.19.0
torch==2.10.0
transformers==5.5.3
tokenizers==0.22.2

# Clustering and QE pipeline
sentence-transformers
scikit-learn
datasets
kneed
matplotlib

# QE (ModernBERT) training only; the wheel must match your installed torch.
flash-attn==2.8.3