How Much Does Random Rotation Help GPTQ? Sharp Worst-Case Rate Gaps in the High-Rate Regime
Organizations: University of Illinois Urbana-Champaign (UIUC)
Abstract
GPTQ is Babai's nearest-plane algorithm on the lattice of the Cholesky factor of the layer's input second moment . In the high-rate, entropy-coded regime it needs of its Cholesky pivots more bits than WaterSIC, which is within bit of the rate-distortion limit for every ; random rotations close most of this gap in practice, and Ordentlich and Polyanskiy asked for the worst-case gap between rotated GPTQ and WaterSIC. We show that after a Haar rotation the -th pivot converges almost surely, uniformly in and over spectra of bounded condition number, to , a classical formula from the theory of Lyapunov exponents, and use it to answer the question. Over spectra with condition number the worst rotated gap is bits, attained only by two-point spectra, just bits below the worst case without rotation, so rotation cannot replace waterfilling in general. It nearly can on real layers: rotation cuts the cost of an -fraction of outlier eigenvalues of size from about to to leading order, but cannot remove spectral spread on a logarithmic scale; for non-Gaussian weights it also adds a weight-side rate cost, which we characterize exactly under a standard high-rate assumption. On Hessians of Qwen3 and Llama-3.1 the rotated gap has median bit and maximum bit, against -- bits in the worst case for the same condition numbers, and the limit matches it to within bit.
Figures & tables
| asympt. (ii) | saved | ||||
|---|---|---|---|---|---|
| 1.53 | 0.46 | 0.18 | 0.39 | 1.06 | |
| 4.32 | 2.56 | 2.50 | 0.225 | 1.76 | |
| 7.35 | 5.27 | 5.24 | 0.149 | 2.08 | |
| 16.82 | 14.22 | 14.20 | 0.073 | 2.60 | |
| 75.61 | 72.00 | 72.00 | 0.018 | 3.61 |
| model | bits | GPTQ | GPTQ+Had. | PWSIC |
|---|---|---|---|---|
| Qwen3-1.7B | 2 | 152.95 | 26.93 | 26.06 |
| 3 | 23.53 | 18.80 1.14 | 17.33 0.02 | |
| Llama-3.1-8B-I. | 2 | 41.37 | 11.18 | 10.37 |
| 3 | 12.81 | 7.85 | 7.76 |
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| asympt. | no rotation | |||||
|---|---|---|---|---|---|---|
| 0.950 | 0.179 | 0.82 | 0.999 | 3.322 | 0.463 | |
| 3.680 | 0.098 | 0.90 | 3.821 | 6.644 | 2.564 | |
| 6.734 | 0.066 | 0.91 | 6.850 | 9.966 | 5.272 | |
| 16.240 | 0.034 | 0.93 | 16.316 | 19.932 | 14.219 | |
| 75.083 | 0.009 | 0.97 | 75.111 | 79.726 | 72.000 |
| numerical | exact formula | expansion to | expansion to | |
|---|---|---|---|---|
| 0.634 | 0.6338 | 0.5922 | 0.6166 | |
| 1.646 | 1.6457 | 1.6300 | 1.6409 | |
| 2.865 | 2.8647 | 2.8568 | 2.8630 | |
| 5.578 | 5.5779 | 5.5747 | 5.5775 |
| spectrum | no rotation | Haar | had2 | had1 | ||||
|---|---|---|---|---|---|---|---|---|
| two-point , | 0.890 | 2.830 | +0.0167 | 0.0001 | +0.0305 | 0.0000 | 0.0221 | 0.0751 |
| two-point , | 2.563 | 4.090 | +0.0008 | 0.0004 | +0.0044 | 0.0004 | 0.0385 | 0.1469 |
| power law , | 0.397 | 1.043 | +0.0040 | +0.0002 | 0.0023 | +0.0002 | 0.0072 | 0.0239 |
| log-uniform, | 0.634 | 1.720 | +0.0093 | +0.0003 | +0.0082 | +0.0003 | 0.0186 | 0.0394 |
| Marchenko–Pastur, ratio | 0.0025 | 0.099 | +0.0008 | +0.0000 | +0.0008 | +0.0000 | 0.0000 | 0.0000 |
| model | input | Haar | Hadamard | natural | act-order | PCA | |||
|---|---|---|---|---|---|---|---|---|---|
| Qwen3-1.7B | q/k/v | 5.1e4 | 0.039 | 0.041 | 0.039 | 0.238 | 0.076 | 1.15 | 3.48 |
| o_proj | 2.9e4 | 0.081 | 0.082 | 0.081 | 0.271 | 0.192 | 1.07 | 3.17 | |
| gate/up | 5.7e4 | 0.026 | 0.026 | 0.026 | 0.130 | 0.023 | 0.85 | 3.55 | |
| down_proj | 2.7e4 | 0.057 | 0.057 | 0.056 | 0.341 | 0.229 | 0.78 | 3.11 | |
| Qwen3-8B-Base | q/k/v | 6.8e4 | 0.040 | 0.040 | 0.040 | 0.261 | 0.077 | 1.08 | 3.65 |
| o_proj | 3.7e4 | 0.070 | 0.071 | 0.075 | 0.212 | 0.146 | 1.00 | 3.28 |
| model | input | no rotation | act-order | Hadamard (same basis) | weight-side cost |
|---|---|---|---|---|---|
| Qwen3-1.7B | q/k/v | 0.238 / 0.236 | 0.076 / 0.075 | 0.039 / 0.039 | 0.019 |
| o_proj | 0.271 / 0.265 | 0.192 / 0.189 | 0.080 / 0.079 | 0.036 | |
| gate/up | 0.130 / 0.128 | 0.023 / 0.022 | 0.026 / 0.026 | 0.012 | |
| down_proj | 0.341 / 0.338 | 0.229 / 0.228 | 0.056 / 0.055 | 0.034 | |
| Llama-3.1-8B-I. | q/k/v | 0.180 / 0.179 | 0.085 / 0.085 | 0.030 / 0.029 | 0.010 |
| o_proj | 0.247 / 0.244 | 0.147 / 0.145 | 0.094 / 0.092 | 0.032 |
| configuration | bits | Qwen3-1.7B | Llama-3.1-8B-I. |
|---|---|---|---|
| entropy-coded GPTQ, no rotation | 2.01 / 2.50 / 3.00 | 152.95 / 30.26 / 23.53 | 41.37 / 30.13 / 12.81 |
| entropy-coded GPTQ + Hadamard | 2.01 / 2.50 / 3.00 | 26.93 / 19.60 / 18.56 | 11.18 / 8.55 / 7.85 |
| PlainWaterSIC, no rotation | 2.01 / 2.51 / 3.00 | 26.06 / 19.44 / 17.31 | 10.37 / 8.45 / 7.76 |
| PlainWaterSIC + Hadamard | 2.51 | 19.69 | 8.61 |
| INT3 g128 sym, no rotation / + Hadamard | 3.125 | 62.32 / 23.44 | 21.15 / 9.64 |
| INT3 g128 asym, no rotation / + Hadamard | 3.125 | 49.87 / 20.12 | 21.26 / 8.45 |
Rotation-Based Quantization
Rotation-based quantization uses orthogonal transforms to redistribute values before low-bit encoding, reducing outlier-driven error in models and data.
See all 57 papers