Thin bash wrappers, one per pipeline stage, with the paper settings filled in. Each script's header has its usage and an example; arguments after the positionals are forwarded to the Python entry point (except run_humming_speedup.sh and push_to_hf.sh); to forward extra arguments, fill in the optional positionals first. Run the scripts from the repository root.
scripts/run_gptq.sh <model_id> <save_dir> [num_gpus]
scripts/run_build_database.sh <model_id> <layer_dir>
scripts/run_search_dl.sh <model_id> <layer_dir> [ear_target]
scripts/run_search_tl_calibrated.sh <model_id> <layer_dir> <bf16_results.json> <anchor_results.json> [min_recovery] [cal_bitwidth]
scripts/run_search_tl.sh <model_id> <layer_dir> [min_recovery] [num_gpus] [temperature] [top_p] [top_k]
scripts/run_build_checkpoint.sh <model_id> <layer_dir> <config | bits> <out_dir>
scripts/run_lmeval.sh <checkpoint> <out_dir> [num_gpus] [temperature] [top_p] [top_k]
scripts/run_ppl.sh <checkpoint> [datasets]
scripts/run_ear.sh <model_id> <layer_dir> <config.txt> ...
scripts/run_humming_speedup.sh <model_id> <layer_dir> [batch] [seq_len] [gpu] [bf16_gpus]
scripts/push_to_hf.sh <checkpoint> <hf_repo_id> [--private]
run_gptq.sh writes the per-layer weights to <save_dir>/<model_name>/4bit/ (4 is the calibration bitwidth); that directory is the <layer_dir> of every later stage.
run_search_dl.sh measures EAR with one forward pass per binary-search step; append --use-predicted-constraints for the purely predicted variant of paper Sec. 3.3.
run_search_tl_calibrated.sh is Algorithm 2 (single-point calibration + guardrail) and minimizes the KL Shapley costs. The anchor recovery is the mean per-benchmark score ratio between the two run_lmeval.sh results: one for the BF16 model and one for the uniform cal_bitwidth model from <layer_dir>, built with run_build_checkpoint.sh <model_id> <layer_dir> <cal_bitwidth> <out_dir> --quant-format fake (see the top-level README).
run_search_tl.sh is the benchmark-in-the-loop variant that runs the lm-eval suite at every binary-search step; use it for validation. It works in <layer_dir>/search_tl_work and keeps only the best candidate's BF16 fake-quant checkpoint (path in search_history.json under best.checkpoint).
Sampling defaults of run_lmeval.sh and run_search_tl.sh are Qwen3's (0.6 0.95 20). Use 1.0 0.95 20 for Qwen3.5 and 0.6 0.9 0 for Llama-3.3 (top_k 0 = off).
run_ear.sh reports EAR and top-K KL on the calibration set (512 x 2048 FineWeb-Edu tokens) for configs applied to the BF16 model; append e.g. --uniform-bitwidths 4,8 after the config files for uniform baselines.
run_humming_speedup.sh times model-level forward passes for uniform 2..8-bit Humming models against BF16 (no direct paper figure), with --single-gpu --compile on the quantized runs. gpu (default 0) is used for the Humming runs; bf16_gpus may be a comma list for models whose BF16 weights do not fit on one GPU. Humming is taken from $HUMMING_ROOT (default ../humming) if that directory exists, otherwise from the installed package.
push_to_hf.sh uploads a checkpoint built by slq_build_checkpoint.py, including its serve.sh launcher and model card, to the Hugging Face Hub.