leocyusa/ra-quant-vulnerability-maps

Reasoning-Aware Quantization: Vulnerability Maps for DeepSeek-R1 Distilled Models

Pre-computed, per-module INT4 vulnerability maps for DeepSeek-R1 distilled reasoning models, plus accuracy–energy results for selective mixed-precision compression. Use the maps to decide which layers to keep in higher precision without repeating the profiling (336 single-module evaluations per benchmark for a 14B model).

Produced with compute provided by OpenToken, at Carnegie Mellon University Africa (Leonard Twagirayezu; supervisor Prof. Prasenjit Mitra). Code: https://github.com/Leonard2019-samantha/ra_quant

Contents

Path	What it holds
vulnerability_maps/qwen14b_<dataset>.csv / .json	R1-Distill-Qwen-14B, 336 (layer, projection) pairs, one file per benchmark (MuSiQue: 334)
vulnerability_maps/qwen7b_math500.csv / .json	R1-Distill-Qwen-7B, 196 pairs, MATH-500
vulnerability_maps/all_vulnerability_maps.csv	All of the above in one long table (1,874 rows)
selective_results/selective_compression_results.csv	Accuracy, energy, output tokens and power for FP16, full INT4 and top-10/20/30/40% protection
figures/	Vulnerability heatmaps and the 14B accuracy–energy frontier (PNG, PDF)

Benchmarks: GSM8K, MATH-500, ProofWriter (depth-5), FOLIO, MuSiQue (token F1).

Vulnerability map columns

Column	Meaning
model	Hugging Face model ID
dataset	Benchmark
rank	1 = most vulnerable. Ties broken by layer, then projection
layer	Transformer layer index (0-based)
projection	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
group	attention or mlp
full_name	Module path, e.g. model.layers.23.self_attn.q_proj
accuracy	Calibration accuracy with only this module quantized to INT4
accuracy_drop	baseline_accuracy − accuracy; > 0 means the module is vulnerable
baseline_accuracy	FP16 calibration accuracy

Method

• Quantization: BitsAndBytes NF4, weight-only, round-to-nearest, float16 compute.
• Sweep: with the FP16 model resident, each (layer, projection) pair is quantized alone, the calibration split (first 50% of the dataset) is evaluated, and the pair is restored. 14B: 15 calibration questions per benchmark; 7B: 30.
• Selective results: evaluated on the held-out second 50%; 14B: 30 questions × 3 runs.
• Energy: NVML power sampled every 100 ms; energy = mean power × wall-clock time.
• Hardware: 1× Tesla V100-SXM3 32 GB (350 W limit), OpenToken.

Usage

import pandas as pd
m = pd.read_csv("vulnerability_maps/qwen14b_gsm8k.csv")
k = int(0.10 * len(m))                       # protect the top 10%
keep_fp16 = m.nsmallest(k, "rank")["full_name"].tolist()
# Pass keep_fp16 as llm_int8_skip_modules in BitsAndBytesConfig,
# or restore those modules from an FP16 copy after loading in 4-bit.

Caveats

• Many modules tie on accuracy_drop because calibration sets are small; treat ranks within a tie as equal.
• MuSiQue calibration accuracy is near zero (14B: 1 of 15), so its map is not informative.
• In the 14B selective runs, non-quantized layers and restored modules ran in bfloat16 (FP16 baseline in float16); V100s lack bfloat16 tensor cores, which may raise the measured energy of INT4 and top-k conditions.
• Energy values are specific to the V100-SXM3 and are not comparable across GPUs.

Citation

Twagirayezu, L. and Mitra, P. (2026). Reasoning-Aware Compression: Identifying and
Protecting Vulnerable Reasoning Circuits for Energy-Efficient LLM Deployment.
Under review, ACL Rolling Review.

Acknowledgement and license

Compute for these maps was provided by OpenToken. Released under CC BY 4.0: free to use and adapt with attribution.