- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper investigates an overlooked source of variability in Hessian-based post-training quantization: the order in which fixed calibration samples contribute to accumulated statistics. Across six Qwen and Llama models, four quantization methods, and 3-/4-bit weights, changing accumulation order produces downstream variation sometimes comparable to calibration resampling and can reverse method rankings. Fixed-activation diagnostics show that small Hessian differences are amplified during damped inversion. The authors introduce stabilization incorporating Neumaier compensated summation, substantially reducing order sensitivity with modest measured overhead.
Critical strengths
- Controlled experiments convincingly isolate accumulation-order sensitivity.
- Broad evaluation covers six models, four methods, and two bit widths, supported by numerical diagnostics.
- Practical mitigation achieves identical weights across orders in many settings, with approximately 5% measured quantization-time overhead.
Critical weaknesses
- Ordering variability is measured on one fixed calibration set per condition, limiting statistical generalization.
- Some stabilized implementations also modify normalization and reductions; missing ablations weaken attribution to compensated summation.
Questions for authors
- Do variability ratios and ranking reversals persist across multiple fixed calibration sets?
- What does each stabilization component contribute, and why does FP64 accumulation retain substantial variability?
4 comments
[W2 & Q2-1] The stabilized implementations also change normalization and reductions. What does each stabilization component contribute?
We thank the reviewer for asking for this breakdown. Besides Neumaier compensation, our stabilized implementations change the normalization, the scaling, and the reduction order, and all of these changes are needed. Leaving out any one of them leaves Ordering variability in at least one of our settings.
The three changes are as follows (App. D.3).
- Normalization (GPTAQ, BoA, TurboBoA): add up all samples and divide once, instead of rescaling a running mean after every batch. The GPTQ implementation we build on, from the open-source llm-compressor library (v0.10.0.1), already does this. For the ablation, we also ran GPTQ with a running mean.
- Scaling (GPTQ, GPTAQ): compute instead of .
- Reduction (BoA, TurboBoA): use a fixed summation tree when undoing the rotation.
We turned each change on and off, with and without compensation (Table R5-a; Qwen3-0.6B and Llama-3.2-1B, INT3, 10 orders).
- Compensation alone is not enough. Without compensation, no combination of the changes removes the variability ( 0.23–1.42). With compensation but with the running mean kept, it also remains ( 0.35–1.17).
- Scaling is needed as well. For GPTQ on Llama-3.2-1B, compensation with Normalization but without Scaling still gives 2 different models from the 10 orders ( 0.12). For GPTAQ on Llama-3.1-8B, the same configuration leaves 0.44 (Table R5-b). With Scaling added, all 10 orders give the same model in both cases.
- Reduction is needed for BoA on the 8B models. On the two small models, BoA and TurboBoA already give identical models with compensation and Normalization. For BoA on the 8B models, this configuration leaves 0.20 / 0.26, and adding Reduction lowers it to 0.02 / 0.13 (Table R5-b).
Why compensation needs Normalization. We attribute the variability that remains under a running mean to its rescaling step. Compensated summation corrects the rounding error of additions. A running mean also multiplies the partial sum by a ratio such as after every batch, and the rounding of that multiplication depends on which samples have been added so far. The correction term records only addition errors, so it does not capture this rounding. Normalizing once at the end turns the accumulation into a plain sum, to which compensation applies.
GPTQ. Our standard FP32 GPTQ already uses Normalization and Scaling, so its standard FP32 and stabilized versions differ only in compensation (App. D.3). There, compensation takes from 0.39–0.77 to 0.00 on all four models in Table 3 of the revised manuscript.
In the revised manuscript, we refer to the full configuration as "Stabilized" instead of "Neumaier" in Table 3 and §5.2, state in §5.1 that all of the changes are required, and add Table R5-a to App. D.4. The 8B-model ablation is Table 2 of the revised manuscript.
Table R5-a. Effect of each change with and without Neumaier compensation (, INT3, 10 orders). Bold: the full stabilized configuration.
| Method | Changes applied | Qwen3-0.6B, no compensation | Qwen3-0.6B, with compensation | Llama-3.2-1B, no compensation | Llama-3.2-1B, with compensation |
|---|---|---|---|---|---|
| GPTQ | None | 0.61 | 0.77 | 0.88 | 1.10 |
| Scaling | 0.80 | 1.02 | 1.20 | 1.15 | |
| Normalization | 0.88 | 0.00 | 1.11 | 0.12 | |
| Scaling + Normalization | 0.58 | 0.00 | 0.77 | 0.00 | |
| GPTAQ | None | 0.34 | 0.74 | 0.71 | 0.56 |
| Scaling | 0.50 | 0.55 | 0.80 | 0.71 | |
| Normalization | 0.35 | 0.00 | 0.86 | 0.00 | |
| Scaling + Normalization | 0.46 | 0.00 | 0.72 | 0.00 | |
| BoA | None | 0.27 | 0.83 | 0.29 | 0.47 |
| Normalization | 0.64 | 0.00 | 0.30 | 0.00 | |
| Reduction | 0.30 | 0.49 | 0.23 | 0.54 | |
| Normalization + Reduction | 0.42 | 0.00 | 0.39 | 0.00 | |
| TurboBoA | None | 1.42 | 0.59 | 0.26 | 0.50 |
| Normalization | 0.94 | 0.00 | 0.38 | 0.00 | |
| Reduction | 0.82 | 1.17 | 0.75 | 0.35 | |
| Normalization + Reduction | 0.95 | 0.00 | 0.35 | 0.00 |
Table R5-b. Effect of each change with compensation on the 8B models (, INT3, 10 orders). Bold: the full stabilized configuration.
| Method | Changes applied | Qwen3-8B, with compensation | Llama-3.1-8B, with compensation |
|---|---|---|---|
| GPTAQ | None | 0.59 | 0.43 |
| Scaling | 0.73 | 0.70 | |
| Normalization | 0.00 | 0.44 | |
| Scaling + Normalization | 0.00 | 0.00 | |
| BoA | None | 0.63 | 0.34 |
| Normalization | 0.20 | 0.26 | |
| Reduction | 0.57 | 0.36 | |
| Normalization + Reduction | 0.02 | 0.13 |
[Q2-2] Why does FP64 accumulation retain substantial variability?
We thank the reviewer for this question. FP64 accumulation is more precise, but its result is still cast to FP32, and tiny order-dependent differences in the FP64 Hessian can survive that cast. This is why FP64 retains variability. For BoA and TurboBoA, the FP64 control in the paper already matched the stabilized runs, so those results stand as reported. For GPTQ and GPTAQ it did not, and we have now matched it. The GPTQ results are essentially unchanged, and the GPTAQ variability disappears on the two models (Qwen3-0.6B and Llama-3.2-1B).
Precision of each operation. The FP64 control uses the following precision. Standard FP32 and the stabilized runs use FP32 in every step (App. A.2).
- Activation conversion: BF16 → FP64
- Matrix multiplication: FP64
- Scaling: FP64
- Accumulation: FP64
- Hessian cast: FP64 → FP32
- Normalization: FP32
- Solving: FP32
Implementation changes. Besides compensation, our stabilized implementations make method-specific changes that reduce rounding outside the summation itself (App. D.3). Two of them concern GPTQ and GPTAQ.
- Scaling: compute instead of , which avoids the rounding caused by the inexact value of .
- Normalization: add up all samples and divide once, instead of rescaling a running mean after every batch.
BoA and TurboBoA additionally use a fixed reduction order when undoing the rotation (Reduction).
What the FP64 control in the paper used. For BoA and TurboBoA, the FP64 control already used the same settings as the stabilized implementations, so their results in Table 3 are unchanged. For GPTQ, it lacked Scaling. For GPTAQ, it lacked both Scaling and Normalization.
Matching the control (Table R3). For GPTQ, adding Scaling gives of 0.00 / 0.29, against 0.01 / 0.33 in the paper (Qwen3-0.6B / Llama-3.2-1B). For GPTAQ, adding both changes gives 0.00 / 0.00, against 0.48 / 0.80.
FP64 with matched settings. There are 16 cases in which FP64 and the stabilized runs use the same settings: four GPTQ and GPTAQ cases on two models in Table R3, and 12 cases in Table 3 of the revised manuscript (GPTQ and GPTAQ on the other two models, and BoA and TurboBoA on all four models). FP64 leaves nonzero in 10 of them (three for GPTQ, two for GPTAQ, three for BoA, and two for TurboBoA), and the stabilized runs in 2 (none for TurboBoA). FP64 is the more precise accumulation, yet the variability remained, so the benefit of compensation is not explained by higher precision alone.
Where the FP64 variability comes from. We examined the remaining GPTQ case on Llama-3.2-1B, with both changes applied, for two orders of the same calibration set (seeds 777 and 5000). We compared the Hessians of the two orders just before and just after the cast to FP32. Layer 8 is the first layer in which the FP32 Hessians differ (Table R4). Before the cast, the largest difference in its down_proj Hessian is only ULP, where ULP is the gap between two neighboring FP32 numbers. After the cast, 2 of its entries differ by 1 ULP. This happens when a value lies right at the boundary between two FP32 numbers, so the two orders round it in different directions.
This is a concrete path by which order dependence reaches the FP32 computations that follow, and a change in one layer spreads to the later layers. In an additional experiment, an order difference in a single layer changes 14.1% / 13.2% of the integer weights over the whole model (Qwen3-0.6B / Llama-3.2-1B, new App. C.6). With the stabilized accumulation in the same setting, the quantized weights are bit-identical across all ten orders (§5.2).
Table R3. Ordering variability () of the FP64 control with and without the two changes (INT3, 10 orders).
| Method | Changes applied | Qwen3-0.6B | Llama-3.2-1B |
|---|---|---|---|
| GPTQ | Neither | 0.59 | 0.75 |
| Scaling only | 0.31 | 0.88 | |
| Normalization only (paper) | 0.01 | 0.33 | |
| Both (same as stabilized) | 0.00 | 0.29 | |
| GPTAQ | Neither (paper) | 0.48 | 0.80 |
| Scaling only | 0.70 | 0.74 | |
| Normalization only | 0.21 | 0.89 | |
| Both (same as stabilized) | 0.00 | 0.00 |
Table R4. Hessian differences between order seeds 777 and 5000 in layer 8 (Llama-3.2-1B, GPTQ INT3, FP64 with both changes).
| Module | Before cast: largest difference (ULP) | After cast: entries that differ |
|---|---|---|
q/k/v (shared) | 0 of | |
o_proj | 0 of | |
gate/up (shared) | 0 of | |
down_proj | 2 of (both by 1 ULP) |
Gentle Reminder: Follow-up on our response
Dear Reviewer,
Thank you again for your time and thoughtful feedback on our submission.
As the discussion period is approaching its end on November 18 (AoE), we wanted to kindly follow up to see if our previous responses and the revised manuscript have adequately addressed your concerns.
If any points remain unclear or if you have any further questions, please let us know. We would be more than happy to clarify them promptly before the deadline.
Best regards,
Authors
Official Response to Reviewer ykNV
Following the reviewer's suggestion, we repeated the ordering experiment on three more calibration sets. Ordering variability appears on all four sets, and a ranking reversal between GPTQ and GPTAQ also occurs on one of the additional sets.
Setup. Besides the calibration set of the paper (Base set), we used three other fixed sets of random WikiText-2 samples (Sets 1–3). On each set, we quantized the model 10 times, accumulating the same samples in a different order each time, for GPTQ and GPTAQ at INT3 on Qwen3-0.6B and Llama-3.2-1B.
Variability. σord is nonzero on all four sets in every case, between 0.32 and 1.00 (Table R13). Pooled over the four sets, it is 0.66–1.17 times σsamp, the standard deviation across independently sampled sets. The paper also reports six different fixed sets on Llama-3.1-8B (three corpora, two sizes), all with σord between 0.33 and 0.61 (Table 1).
Ranking reversals. Pairing the two methods by order seed as in §3.2, GPTQ scores higher than GPTAQ in 2 of the 10 pairs on Set 3 of Llama-3.2-1B, although GPTAQ is ahead on average. A reversal is therefore not limited to the calibration set of the paper. Reversals are less frequent here than in §3.2, because the two methods are further apart on these two models: 3–5 pp on the Base set (Table 3 of the revised manuscript), against 0.3 pp on Qwen3-8B in §3.2.
We have added Table R13 to App. B.2 in the revised manuscript.
Table R13. Ordering variability on four fixed calibration sets (INT3). σord (in percentage points) over 10 accumulation orders on each set.
Pooled over sets: the square root of the mean of the four variances.