JointQ: Joint Scale-and-Integer Quantization for Large Language Models
Organizations: Fujitsu Limited, Kanagawa, Japan
Abstract
GPTQ is the de facto standard solver for layer-wise post-training quantization (PTQ) of large language models, yet the solver itself has received less attention than weight preprocessing and cross-layer objectives. We target a bottleneck of GPTQ: its greedy procedure optimizes continuous group scales separately from the assignment of discrete integers. Evaluating each integer candidate with its closed-form optimal scale has been used for per-channel quantization; JointQ extends it to group-wise quantization, where a row's scales are coupled through the Hessian, and applies it to refine the GPTQ solution. Each local-search move changes one integer and recomputes its group's optimal scale in closed form, accounting for the other groups through the full Hessian; all scales are periodically re-solved jointly. The search starts from a row-wise selection among scale-refitted candidates including the GPTQ solution, so unregularized JointQ never has a larger layer-wise calibration error than GPTQ. JointQ preserves the GPTQ output format and is thus a drop-in replacement. Across four models and four settings, JointQ improves six-task average zero-shot accuracy over vanilla GPTQ in 14 of 16 comparisons and perplexity in 12 of 16. The mean accuracy gain increases from 0.0044 in W4G128 to 0.1559 in W3 per-channel, and JointQ also improves average accuracy over GPTQ with MSE-based scale search and activation ordering in all eight per-channel comparisons, though rarely in G128. An ablation shows that JointQ's diagonal Tikhonov regularization mitigates severe W3 per-channel perplexity degradation, so the largest gains are not due to the local search alone. Overall, JointQ's primary empirical benefit is downstream accuracy preservation in challenging low-bit and large-group regimes. Code is available in the OneComp library at https://github.com/FujitsuResearch/OneCompression.
Figures & tables
| W4G128 | W4 per-ch | W3G128 | W3 per-ch | ||||||
| Model | Method | PPL | Avg. | PPL | Avg. | PPL | Avg. | PPL | Avg. |
| Llama-2-7B | BF16 | PPL: 5.47 Avg.: 0.7038 | |||||||
| GPTQ | 5.89 | 0.6918 | 6.17 | 0.6816 | 6.80 | 0.6359 | 27.74 | 0.5020 | |
| JointQ | 5.67 | 0.6914 | 5.96 | 0.6836 | 6.78 | 0.6610 | 7.98 | 0.6032 | |
| Llama-3-8B | BF16 | PPL: 6.14 Avg.: 0.7442 | |||||||
| GPTQ | 14.09 | 0.7248 | 399.02 | 0.5960 | 57.69 | 0.5036 | 3381.46 | 0.3732 | |
| Setting | PPL W/T/L | Avg. W/L | Avg. |
|---|---|---|---|
| W4G128 | 3/0/1 | 3/1 | |
| W4 per-ch | 3/1/0 | 4/0 | |
| W3G128 | 3/0/1 | 4/0 | |
| W3 per-ch | 3/0/1 | 4/0 | |
| Overall | 12/1/3 | 15/1 | – |
| GPTQ | JointQ w/o stab. | JointQ | |||||
|---|---|---|---|---|---|---|---|
| Model | Setting | PPL | Avg. | PPL | Avg. | PPL | Avg. |
| Llama-2-7B | W4G128 | 5.89 | 0.6918 | 5.67 | 0.6896 | 5.67 | 0.6914 |
| W4 per-ch | 6.17 | 0.6816 | 6.27 | 0.6694 | 5.96 | 0.6836 | |
| W3G128 | 6.80 | 0.6359 | 6.89 | 0.6485 | 6.78 | 0.6610 | |
| W3 per-ch | 27.74 | 0.5020 | 23.02 | 0.4511 | 7.98 | 0.6032 | |
| Llama-3-8B | W4G128 | 14.09 | 0.7248 | 6.65 | 0.7348 | 6.65 | 0.7426 |
| Model | Method | PPL | Avg. | Time (s) |
|---|---|---|---|---|
| Llama-2-7B | BF16 | 5.47 | 0.7038 | – |
| GPTQ | 583.21 | 0.3607 | 266.6 | |
| JointQ | 26.33 | 0.4927 | 1164.1 | |
| Llama-3-8B | BF16 | 6.14 | 0.7442 | – |
| GPTQ | 195282.65 | 0.3682 | 292.1 | |
| JointQ | 108.77 | 0.4866 | 1330.4 |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting | Value | Role |
|---|---|---|
| Absolute tolerance. Algorithm 2 compares it with increases of , in the units of , and uses it as the guard in Eq. ( 23 ); Algorithms 1 and 3 compare it with decreases of . | ||
| Iteration cap | Maximum number of iterations of one call of Algorithm 2 . | |
| Algorithm 2 stops when more than this fraction of the rows has . | ||
| Algorithm 1 stops when fewer than this fraction of the rows is active. | ||
| Selection of the rows for the joint scale re-solve. | ||
| Diagonal damping of the fallback solve. |
| Layer type | W4G128 | W4 per-ch | W3G128 | W3 per-ch | W2G32 |
|---|---|---|---|---|---|
| q_proj | 0 | 0 | 0.01 | 0.15 | 0.1 |
| k_proj | 0 | 0 | 0.01 | 0.15 | 0.1 |
| v_proj | 0.01 | 0.15 | 0.01 | 0.15 | 0.1 |
| o_proj | 0 | 0 | 0 | 0.15 | 0 |
| gate_proj | 0.01 | 0.15 | 0.01 | 0.15 | 0 |
| up_proj | 0.01 | 0.15 | 0.01 | 0.15 | 0 |
| GPTQ-MSE | JointQ | ||||
|---|---|---|---|---|---|
| Model | Setting | PPL | Avg. | PPL | Avg. |
| Llama-2-7B | W4G128 | 5.97 | 0.6913 | 5.67 | 0.6914 |
| W4 per-ch | 27.21 | 0.5928 | 5.96 | 0.6836 | |
| W3G128 | 7.19 | 0.6568 | 6.78 | 0.6610 | |
| W3 per-ch | 855.39 | 0.4496 | 7.98 | 0.6032 | |
| Llama-3-8B | W4G128 | 8.84 | 0.7294 | 6.65 | 0.7426 |
| GPTQ ⋆ | JointQ | ||||||
|---|---|---|---|---|---|---|---|
| Model | Setting | PPL | Avg. | Time (s) | PPL | Avg. | Time (s) |
| Llama-2-7B | W4G128 | 5.67 | 0.6935 | 1301.1 | 5.67 | 0.6914 | 706.2 |
| W4 per-ch | 16.53 | 0.5753 | 362.9 | 5.96 | 0.6836 | 1906.5 | |
| W3G128 | 9.87 | 0.6077 | 1536.4 | 6.78 | 0.6610 | 965.9 | |
| W3 per-ch | 334.30 | 0.4512 | 363.8 | 7.98 | 0.6032 | 2161.8 | |
| Llama-3-8B | W4G128 | 6.55 | 0.7478 | 1454.3 | 6.65 | 0.7426 | 937.0 |
| GPTQ ⋆ | JointQ (GPTQ ⋆ cand.) | ||||||
|---|---|---|---|---|---|---|---|
| Model | Setting | PPL | Avg. | Time (s) | PPL | Avg. | Time (s) |
| Llama-2-7B | W4G128 | 5.67 | 0.6935 | 1301.1 | 5.66 | 0.6960 | 1805.4 |
| W3G128 | 9.87 | 0.6077 | 1536.4 | 6.56 | 0.6639 | 2226.3 | |
| Llama-3-8B | W4G128 | 6.55 | 0.7478 | 1454.3 | 6.55 | 0.7439 | 1913.8 |
| W3G128 | 8.37 | 0.7110 | 1705.5 | 8.41 | 0.7191 | 2341.4 | |
| Qwen3-8B | W4G128 | 9.93 | 0.7355 | 1546.3 | 9.88 | 0.7382 | 2089.2 |
| JointQ | GPTQ ⋆ | |||
|---|---|---|---|---|
| Sequences | PPL | Avg. | PPL | Avg. |
| 8 | ||||
| 16 | ||||
| 32 | ||||
| 64 | ||||
| 128 | ||||
| Bits | Sequences | JointQ PPL | JointQ w/o stab. PPL | PPL | PPL wins | Avg. (pt.) | Avg. wins |
|---|---|---|---|---|---|---|---|
| 4 | 8 | 20/20 | 16/20 | ||||
| 16 | 20/20 | 12/20 | |||||
| 32 | 20/20 | 16/20 | |||||
| 64 | 19/20 | 13/20 | |||||
| 128 | 20/20 | 14/20 | |||||
| 3 | 8 | 20/20 | 20/20 |
| Bits | Sequences | PPL (95% CI) | PPL wins | Avg. points (95% CI) | Avg. wins |
|---|---|---|---|---|---|
| 4 | 8 | [ ] | 18/20 | [ ] | 12/20 |
| 16 | [ ] | 20/20 | [ ] | 13/20 | |
| 32 | [ ] | 18/20 | [ ] | 10/20 | |
| 64 | [ ] | 19/20 | [ ] | 9/20 | |
| 128 | [ ] | 19/20 | [ ] | 14/20 | |
| 3 | 8 | [ ] | 2/20 | [ ] | 20/20 |
| Method | Setting | PPL | ARC-c | ARC-e | PIQA | WG | HS | BQ | Avg. | Time (s) |
|---|---|---|---|---|---|---|---|---|---|---|
| BF16 | – | 5.47 | 0.4480 | 0.7378 | 0.7845 | 0.6961 | 0.7617 | 0.7945 | 0.7038 | – |
| GPTQ | W4G128 | 5.89 | 0.4437 | 0.7176 | 0.7824 | 0.6867 | 0.7551 | 0.7654 | 0.6918 | 268.1 |
| JointQ | W4G128 | 5.67 | 0.4352 | 0.7176 | 0.7797 | 0.6961 | 0.7510 | 0.7688 | 0.6914 | 706.2 |
| GPTQ | W4 per-ch | 6.17 | 0.4369 | 0.6982 | 0.7824 | 0.6827 | 0.7401 | 0.7492 | 0.6816 | 260.3 |
| JointQ | W4 per-ch | 5.96 | 0.4480 | 0.7113 | 0.7791 | 0.6922 | 0.7336 | 0.7373 | 0.6836 | 1906.5 |
| GPTQ | W3G128 | 6.80 | 0.3857 | 0.6170 | 0.7563 | 0.6496 | 0.6906 | 0.7162 | 0.6359 | 263.1 |
| Method | Setting | PPL | ARC-c | ARC-e | PIQA | WG | HS | BQ | Avg. | Time (s) |
|---|---|---|---|---|---|---|---|---|---|---|
| BF16 | – | 6.14 | 0.5384 | 0.7757 | 0.8058 | 0.7332 | 0.7921 | 0.8199 | 0.7442 | – |
| GPTQ | W4G128 | 14.09 | 0.5060 | 0.7559 | 0.7824 | 0.7167 | 0.7810 | 0.8070 | 0.7248 | 277.9 |
| JointQ | W4G128 | 6.65 | 0.5401 | 0.7828 | 0.7965 | 0.7293 | 0.7843 | 0.8226 | 0.7426 | 937.0 |
| GPTQ | W4 per-ch | 399.02 | 0.3268 | 0.5206 | 0.6763 | 0.6464 | 0.6916 | 0.7144 | 0.5960 | 273.1 |
| JointQ | W4 per-ch | 7.41 | 0.4940 | 0.7580 | 0.7829 | 0.7301 | 0.7739 | 0.8092 | 0.7247 | 2564.0 |
| GPTQ | W3G128 | 57.69 | 0.2628 | 0.4482 | 0.6153 | 0.5943 | 0.4888 | 0.6122 | 0.5036 | 274.8 |
| Method | Setting | PPL | ARC-c | ARC-e | PIQA | WG | HS | BQ | Avg. | Time (s) |
|---|---|---|---|---|---|---|---|---|---|---|
| BF16 | – | 9.73 | 0.5640 | 0.8093 | 0.7758 | 0.6772 | 0.7489 | 0.8664 | 0.7403 | – |
| GPTQ | W4G128 | 10.33 | 0.5614 | 0.7879 | 0.7639 | 0.6693 | 0.7404 | 0.8709 | 0.7323 | 321.0 |
| JointQ | W4G128 | 10.17 | 0.5572 | 0.7950 | 0.7628 | 0.6756 | 0.7393 | 0.8719 | 0.7336 | 2019.8 |
| GPTQ | W4 per-ch | 11.04 | 0.5077 | 0.7441 | 0.7557 | 0.6835 | 0.7246 | 0.8529 | 0.7114 | 315.2 |
| JointQ | W4 per-ch | 10.71 | 0.5367 | 0.7740 | 0.7688 | 0.6859 | 0.7323 | 0.8465 | 0.7240 | 1631.7 |
| GPTQ | W3G128 | 11.75 | 0.4881 | 0.7193 | 0.7584 | 0.6527 | 0.6914 | 0.8413 | 0.6919 | 318.5 |
| Method | Setting | PPL | ARC-c | ARC-e | PIQA | WG | HS | BQ | Avg. | Time (s) |
|---|---|---|---|---|---|---|---|---|---|---|
| BF16 | – | 8.65 | 0.6041 | 0.8279 | 0.7976 | 0.7309 | 0.7877 | 0.8936 | 0.7736 | – |
| GPTQ | W4G128 | 8.85 | 0.6015 | 0.8224 | 0.8003 | 0.7301 | 0.7813 | 0.8896 | 0.7709 | 473.2 |
| JointQ | W4G128 | 8.89 | 0.5998 | 0.8245 | 0.7938 | 0.7269 | 0.7806 | 0.8939 | 0.7699 | 7435.3 |
| GPTQ | W4 per-ch | 9.11 | 0.5887 | 0.8001 | 0.7840 | 0.7040 | 0.7722 | 0.8657 | 0.7525 | 463.9 |
| JointQ | W4 per-ch | 9.40 | 0.5956 | 0.8266 | 0.8020 | 0.7324 | 0.7767 | 0.8930 | 0.7710 | 3923.5 |
| GPTQ | W3G128 | 10.07 | 0.5478 | 0.7862 | 0.7889 | 0.6914 | 0.7489 | 0.8508 | 0.7356 | 461.7 |
LLM Quantization
Large language model quantization reduces numerical precision to lower memory use and accelerate training or inference.
See all 438 papers