acceptodds
Under review as a conference paper at ICLR 2027

SoloQ: Calibration-Free Quantization for Diffusion Language Models

Abstract

Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive language models through bidirectional diffusion-based token generation. However, their growing model sizes and high inference costs make efficient deployment challenging: full-sequence denoising repeatedly invokes compute-intensive forward passes, while block-diffusion models additionally introduce a memory-intensive KV-cache. Low-bit weight-activation quantization is therefore attractive, yet existing dLLM post-training quantization methods rely on calibration data despite activation distributions shifting across masking states and denoising steps. We present SoloQ, a calibration-free quantization framework that maps weights and activations into a normalized rotated basis with a predictable marginal distribution, enabling data-independent quantization. SoloQ combines a structured K-RPBH rotation with a lightweight rescaling correction for calibration-free quantization. Its predictable post-rotation distribution supports both distribution-matched codebooks and hardware-native NVFP4. For block-diffusion models, SoloQ further applies commit-time KV-cache quantization to compress persistent states without perturbing the actively denoised block. Across full-sequence dLLMs (LLaDA and Dream) and block-diffusion dLLMs (Fast-dLLM v2 and Nemotron-Labs-Diffusion), SoloQ retains accuracy under 4-bit quantization and outperforms calibration-based baselines on knowledge- and reasoning-intensive benchmarks. With NVFP4, SoloQ reduces peak memory by up to 2.61 and accelerates end-to-end inference by up to 2.24.

Primary area:
generative models
Keywords:
Quantization, Diffusion Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NMdTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SoloQ, a calibration-free quantization method for diffusion language models. The main motivation is that activation distributions in diffusion LLMs change across denoising steps and mask ratios, making it difficult for a small calibration set to cover all inference states. SoloQ addresses this issue through a structured random rotation, K-RPBH, which aims to make the activation distribution more suitable for low-bit quantization. The method also introduces a scale-based bias correction and a commit-time KV cache quantization strategy.

Authors evaluates SoloQ on several diffusion LLMs, including LLaDA, Dream, and Fast-dLLM v2, covering weight, activation, and KV cache quantization. The results show promising performance, with improvements in quantization accuracy and reductions in memory usage and inference latency.

Critical strengths
  1. The proposed K-RPBH rotation combines within-block mixing, random permutation, and cross-block mixing to improve the distribution of activation energy across channels. The ablation results show that adding permutation and cross-block mixing reduces KL divergence from 0.11396 to 0.00413 and energy imbalance from 67.89% to 22.12%. These results provide useful evidence that the proposed rotation makes activations more suitable for quantization.

  2. Activation distributions in diffusion LLMs vary across denoising steps and mask ratios, which makes conventional calibration-based quantization less reliable. SoloQ avoids the need for calibration data and is evaluated across several model families and quantization settings. This makes the method relevant to practical deployment, especially when representative calibration data are difficult to obtain.

Critical weaknesses
  1. The paper mainly shows that the proposed rotation improves the distribution of activation energy across coordinates. For a normalized vector x~=x/∣x∣2\tilde{x}=x/|x|_2 and a rotation RR, the transformed vector z=Rx~z=R\tilde{x} has a more balanced coordinate-wise energy distribution in expectation. However, this does not directly establish that quantization error after rotation will be smaller in terms of model output or downstream task accuracy. The current analysis does not provide a clear bound on the resulting output error or task loss. The paper would be stronger if it connected the rotation properties to weight-dependent sensitivity, error propagation, or an end-to-end quantization error bound.

  2. Although the paper explicitly identifies the limitations of OrbitQuant's RPBH rotation and introduces cross-block mixing to address them, the proposed K-RPBH remains a relatively incremental extension of an existing rotation scheme. The main modification is to add a Kronecker-structured cross-block mixing stage to improve block-wise energy balancing. While the paper provides theoretical analysis and empirical evidence for this design, it is less clear whether the modification constitutes a sufficiently substantial methodological advance beyond OrbitQuant, rather than an adaptation of an existing technique to diffusion language models. The authors should better clarify the fundamental novelty of K-RPBH and distinguish the contributions of the rotation itself from those arising from its application to diffusion LLMs and KV-cache quantization.

Questions for authors
  1. The paper claims that K-RPBH improves upon original RPBH by introducing cross-block mixing via Kronecker product Qk⊗IhQ_k\otimes I_h, and proves the expected coordinate variance equals 1/d1/d. However, this guarantee holds only in expectation over random QkQ_k. For a fixed pre-sampled K-RPBH matrix used during inference, there will be residual block-level energy imbalance. Could you quantify how the choice of fixed QkQ_k affects the KL divergence between the empirical coordinate distribution and the ideal spherical marginal distribution? And how sensitive is the final downstream generation accuracy to different random seeds of QkQ_k?

  2. In the ablation studies of K-RPBH, the paper evaluates PP and QkQ_k separately, but does not conduct a clean controlled experiment to disentangle the gain from cross-block mixing versus the codebook / bias correction. Could you provide an experiment that fixes the same quantization pipeline (same codebook, scale-only bias correction) and only swaps the rotation module between vanilla RPBH and K-RPBH on your dLLM benchmarks? This would help isolate how much performance gain purely comes from the cross-block mixing component of K-RPBH.

  3. OrbitQuant is the most closely related prior work that adopts RPBH rotation + Lloyd-Max codebook for diffusion models. However, the paper does not include a direct head-to-head comparison with OrbitQuant under identical dLLM setting (same model, same W4A4 bit-width, same evaluation tasks). Why is this baseline missing? What are the key differences in activation statistics between image DiT and dLLM that make OrbitQuant fail or underperform on diffusion language models?

2 comments

We greatly appreciate your efforts in reviewing our work and providing valuable feedback, which encourages us to improve our work.

W1

Thank you for the technical review. We agree that the paper argues only about the rotated activation distribution and gives no task-level bound. We connect the rotation to output error in three steps: a weight-dependent error model, measured error propagation, and task accuracy.

(1) Weight-dependent error model (new Prop. 2). For a rotated linear layer, we derive the expected output error under standard assumptions (uncorrelated coordinate errors, Gaussian coordinates within each block). It has three terms:

  • the error of a perfectly balanced rotation, set by the codebook MSE δ\delta (≈0.0095 for 4-bit Lloyd–Max), ∥x∥2\|x\|^2 and ∥W∥F2\|W\|_F^2.
  • an extra term proportional to the covariance between the activation block energies and the weight-column block energies, which grows when outlier activation blocks meet heavy weight columns.
  • a codebook-mismatch term, which grows when the block energies are unbalanced and is zero when they are equal.

RPBH applies the same permutation to xx and to the columns of WW, so an outlier channel and its weight column stay in one block and the covariance term does not shrink. K-RPBH's QkQ_k shrinks both the block-energy variance and this covariance by 2/(k+2)2/(k+2) in expectation (Prop. 3, stated in our answer to Q1 below), so both extra terms shrink. This is an error model, not a bound. The full statement and derivation will be in App. A.

(2) Measured error propagation. SoloQ-C at W4A4, mask ratios 0.1–0.9, 3 rotation seeds. "Layer error" is ∥Wx−y^∥/∥Wx∥\|Wx-\hat y\|/\|Wx\| of the FFN down-projection, averaged over layers and ratios. Logit KL and top-1 agreement are measured against the FP model.

Table R1. Error propagation.

Model (k)RotationLayer errorLogit KLTop-1 %
Dream (37)RPBH0.1680.151882.1
Dream (37)K-RPBH0.1060.108784.4
Dream (37)dense Haar0.1060.113384.1
Fast-dLLM v2 (37)RPBH0.1360.107083.6
Fast-dLLM v2 (37)K-RPBH0.1130.089485.2
Fast-dLLM v2 (37)dense Haar0.1130.086285.7
LLaDA-1.5 (3)RPBH0.1060.066788.1
LLaDA-1.5 (3)K-RPBH0.1060.068388.6

At k = 37, K-RPBH lowers the layer error and the logit KL of RPBH to the level of dense Haar. At k = 3 (LLaDA-1.5) the blocks are already nearly balanced, so the two rotations nearly coincide: the layer error is the same, and RPBH's logit KL is marginally lower (0.0667 vs. 0.0683). The gain of K-RPBH therefore grows with the block count k.

(3) Task accuracy. With only the rotation swapped, K-RPBH improves SoloQ-C over RPBH by +0.9 to +3.5 points on Fast-dLLM v2 and Dream (Table R3 of Q2).

Q1

Your point is right that Prop. 1 holds only in expectation over QkQ_k. We quantify (a) the distribution of one fixed draw and (b) the accuracy across draws. Each rotation seed redraws PP, DD and QkQ_k, so the seed variation includes the variation due to QkQ_k alone and does not understate it.

(a) One fixed draw. Five fixed draws, FFN down-projection, mask ratios 0.1–0.9 (k=37k = 37). "Worst block" is the per-token maximum block energy divided by the average, averaged over tokens and layers.

Table R2. One fixed draw (k=37k = 37).

RotationKL to fdf_d, DreamKL to fdf_d, Fast-dLLM v2Worst block
RPBH0.011–0.0180.009–0.0184.28–5.04×
K-RPBH1.7–4.7e-52.0–9.4e-51.49–1.69×
Dense Haar1.6–2.6e-51.7–5.2e-51.14×

Every fixed draw of K-RPBH stays close to dense Haar in KL. A residual block imbalance remains (1.5–1.7× vs. 1.14× for Haar), but it is about 3× smaller than RPBH's. Our new Prop. 3 quantifies this: for any orthogonal QkQ_k the block energies still average to 1, and for a Haar QkQ_k their expected variance is 2k+2\frac{2}{k+2} of RPBH's, i.e., 19.5× smaller at k=37k = 37 (a single realized draw adds a small O(1/h)O(1/h) term). At k=3k = 3 (LLaDA-1.5) the worst block drops from 1.22–1.28× to 1.08–1.20×, but the KL does not drop, so KL is only a proxy.

(b) Accuracy across seeds. Tables R3 and R4 (in our answer to W2/Q2/Q3) report means over 3 fixed random rotation seeds. Across these seeds, SoloQ accuracy varies by at most 2.7 points (std) on every task.

W2 / Q2 / Q3

K-RPBH builds on OrbitQuant's RPBH. Our contribution is diagnosing why RPBH fails on dLLMs and fixing it with a cross-block mixer. Below we separate the contributions (W2), swap only the rotation (Q2), and compare with the OrbitQuant recipe (Q3).

(1) Contributions (W2).

Rotation itself (model-agnostic):

  1. Diagnosis: a channel with energy fraction ff forces some RPBH block energy ≥kf\ge kf under every permutation, so RPBH's floor grows with the block count kk (k=37k=37 in Dream/Fast-dLLM v2 FFNs).
  2. Fix that brings the marginal to the Haar target, shown in theory and in measurement: a k×kk\times k Haar cross-block mixer for any kk (no Hadamard of order kk needed) at O(dk)O(dk) extra cost. In theory, every coordinate has variance exactly 1/d1/d in expectation (Prop. 1), and a single draw shrinks the block-energy variance to 2k+2\frac{2}{k+2} of RPBH's (Prop. 3, Q1). In measurement, one fixed K-RPBH matches the near-Gaussian target marginal fdf_d as closely as dense Haar at every mask ratio (Dream, k=37k = 37: KL 1.7–4.7e-5 vs. 1.6–2.6e-5 for Haar and 0.011–0.018 for RPBH, Q1).

Application to dLLMs and KV cache (not rotation novelty):

  1. Scale-only bias correction (BC), commit-time KV-cache quantization for block diffusion, the fold/unfold analysis, and calibration-free NVFP4 / E8P (QuIP#'s codebook) / INT4 paths.

(2) Experiments (Q2, Q3). "C/N, RPBH" is SoloQ-C/N with only the rotation swapped. "OrbitQuant recipe" is our implementation inside the SoloQ pipeline (RPBH, Lloyd–Max weights and activations, no BC). All rows use one harness and are means over the same 3 fixed random seeds. Fast-dLLM v2 IFEval is prompt-level strict. The Fast-dLLM v2 rows are re-runs with new seeds, so they are compared only with each other, not with Table 2.

Table R3. Fast-dLLM v2.

Fast-dLLM v2HEHE+IFEval
OrbitQuant recipe58.5354.7053.54
C, RPBH56.7052.6054.77
SoloQ-C60.1755.1057.18
N, RPBH60.2655.3856.67
SoloQ-N64.1659.1457.19

Table R4. Dream-7B.

Dream-7BARC-CWinoPIQA
OrbitQuant recipe53.2769.6972.25
C, RPBH55.5570.2273.56
SoloQ-C56.4671.7774.48
N, RPBH56.8071.5974.32
SoloQ-N56.3472.0974.61
FP1659.9874.4375.46

(3) Rotation only (Q2). The "C/N, RPBH" rows above are this experiment: same codebook and BC, with only the rotation swapped. K-RPBH improves SoloQ-C on all six tasks (+0.9 to +3.5). SoloQ-N gains less or not at all (Dream ARC-C −0.5), likely because its per-16 scales remove the codebook-mismatch term of Prop. 2 (W1). At k=3k = 3 (LLaDA-1.5, SoloQ-N HellaSwag, 3 seeds, not in the tables above), the gain is small: 72.99 → 73.54.

(4) OrbitQuant head-to-head (Q3). OrbitQuant targets image and video DiTs, and we did not port the recipe to dLLMs. Given the shared RPBH rotation, this was an omission, and the tables above add it. SoloQ-C beats the OrbitQuant recipe on all Dream tasks (+2.1 to +3.2) and on Fast-dLLM v2 IFEval (+3.6), and HE (+1.6) and HE+ (+0.4) are close. Codebook + BC alone (C, RPBH vs. the recipe) helps on Dream but not on Fast-dLLM v2 HE/HE+ (−1.8 / −2.1), where adding QkQ_k does (+3.5 / +2.5).

To compare activation statistics directly, we ran the same probe on FLUX.1-schnell, one of the DiTs evaluated in OrbitQuant, over all 4 sampling steps. As in both methods, hh is the largest power of two dividing dd, so k=d/hk = d/h.

Table R5. Activation statistics, image DiT vs. dLLM (5 rotation draws, ranges over sampling steps or mask ratios). ff is the energy share of the largest channel per token.

ModelLayerddkkffRPBH worst blockKL RPBHKL K-RPBHKL Haar
FLUX.1-schnellFFN down1228830.018–0.0211.08×2.4–3.2e-52.2–3.5e-52.3–3.1e-5
FLUX.1-schnellattn input307230.050–0.0791.14–1.19×1.0–1.3e-40.95–1.4e-40.98–1.6e-4
DreamFFN down18944370.093–0.1114.41–5.04×0.011–0.0181.7–4.7e-51.6–2.6e-5
Fast-dLLM v2FFN down18944370.089–0.1104.28–5.04×0.009–0.0182.0–9.4e-51.7–5.2e-5

The two differ in both structure and activations. FLUX has few large blocks (k=3k = 3) and a weak top channel (f≈0.02f \approx 0.02), so RPBH already reaches the Haar KL and QkQ_k adds nothing. The dLLMs have many small blocks (k=37k = 37) and a top channel carrying about 10% of the energy, so RPBH leaves a worst block of 4.3–5.0×, consistent with the floor kf≈3.3kf \approx 3.3–4.14.1 of contribution 1, and a KL more than 400× that of Haar. QkQ_k removes both.

We will add these results to the revised version.

Reviewer_dT4uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article proposes a calibration free quantization framework SoloQ for diffusion language models, aimed at addressing the dependency of traditional post training quantization methods on calibration data. The author believes that diffusion language models have dynamically changing activation distributions under different mask states and denoising time steps, so quantization methods based on fixed calibration data may lack robustness. Overall, the issues addressed in this article have practical significance, and the methods also involve certain technical designs. However, I have concerns about the necessity of the research objective of calibration free, the relationship between theoretical analysis and practical conclusions, and the adequacy of experimental comparisons.

Critical strengths

The research question has certain practical value. The diffusion language model requires multi-step denoising, and its activation distribution and computational cost have characteristics different from autoregressive models. Studying its low bit quantization has practical significance.

The experiment covers a wide range. The author evaluated various diffusion language model architectures, including both full sequence diffusion models and block diffusion models, and reported the inference acceleration results on actual hardware.

Critical weaknesses
  1. The necessity of calibration free quantification has not been fully demonstrated. The core motivation of the paper is that diffusion language models have dynamically changing activation distributions under different denoising time steps and mask states, so traditional quantization methods that rely on calibration data may have distribution mismatch problems. However, the dynamic changes in activation distribution do not necessarily mean that calibration should be completely avoided. In fact, different mask ratios and denoising states can be simulated by constructing corresponding calibration samples, and existing quantization methods for diffusion language models are also attempting to address this issue through state aware or time step aware strategies. Therefore, I believe that the paper has not fully explained why using non calibration quantification is a necessary choice to solve the above problems, rather than adopting a more reasonable calibration strategy.
  2. Theoretical analysis is insufficient to fully support the core proposition of predictable distributions independent of input states. The important theoretical basis of the paper is that random orthogonal rotation can make the normalized coordinate distribution approach a known distribution, thereby supporting quantization without calibration. However, the conclusion about the distribution of spherical edges in Section 3.2 is based on the condition of Haar random rotation, and the actual K-RPBH used is a rotation matrix with a specific structure, which is not equivalent to general Haar random rotation.
  3. Compared with competitive calibration based quantification methods, the experimental advantages are not yet sufficient. Firstly, not all models in Table 1 achieved optimal results. For example, on the Dream-7B, the average accuracy of STaR Quant is 63.59, while SoloQ-C and SoloQ-N are 62.06 and 62.53, respectively. This indicates that the proposed method is not consistently superior to calibrated quantization methods with state and time modeling capabilities on all architectures. Secondly, in the experiments of the block diffusion model, RTN and AWQ showed significant performance crashes, while QuaRot was marked as not applicable on Fast-dLLM v2. Therefore, some performance advantages are built on a baseline with relatively limited competitiveness.
  4. The partial results of FAIR Calib in Appendix Table 6 are from existing papers, while SoloQ uses the author's own evaluation results. Due to the fact that the two are not completely unified experimental setups, the relevant conclusions need to be interpreted more cautiously.
Questions for authors
  1. If representative calibration sets are constructed using different mask ratios and denoising time steps, do existing calibration based PTQ methods still have significant performance disadvantages? Can you provide a comparison curve between calibration data scale and quantification accuracy?
  2. What other quantifiable advantages does Calibration Free provide in real-world deployment besides avoiding collecting calibration samples? Can you compare the calibration preparation time and headquarters deployment cost of different methods? Proposition 1 mainly ensures the balance of coordinate variance in the expected sense. Can the author further explain why this conclusion is sufficient to support the predictability of the activated edge distribution after structured rotation, especially under different diffusion states?
  3. What are the differences in end-to-end inference latency and video memory usage between complete K-RPBH and RPBH under the same NVFP4 quantization kernel? Can the additional cross block hybrid computing overhead be offset by accuracy gains?
  4. Can you provide a more fair comparison with state aware calibration methods, especially for block diffusion models such as Fast-dLLM v2 and Nemotron Labs Diffusion?
2 comments

We thank the reviewer for the valuable feedback. We address each point below.

W1 / Q1

We agree that calibration sets covering different mask ratios and timesteps can reduce the distribution mismatch. Still, there are three reasons to avoid calibration for dLLMs.

(1) Calibration is a hidden hyperparameter. Accuracy varies with the calibration set [1], and dLLMs add mask ratios, timesteps and block states to the choice, so each dLLM PTQ method designs its own sampling scheme (DLLMQuant, Quant-dLLM).

(2) It is an extra step for every model and calibration set, taking 20 s (DuQuant) to 87–100 min (FlatQuant) in our runs (Table R4), while SoloQ needs no data.

(3) It is not needed. After K-RPBH, activations look nearly the same at every mask ratio (Table R3), and state-aware calibration does not exceed SoloQ (below).

Q1. We built the baseline you describe: K-RPBH + GPTQ (same rotation as SoloQ), calibrated on 32, 128 or 512 sequences that are clean or masked at ratios 0.1–0.9 (strat.). On Fast-dLLM v2 the masks follow the active block. DLLMQuant and STaR-Quant have no public code, so this is a generic state-aware baseline. We also ran DuQuant on the same sets. All rows are re-runs in one harness with fixed random seeds.

Table R1. Fast-dLLM v2.

MethodCalib.GSM8KHEHE+IFEval
GPTQclean 12873.4152.4349.4052.43
GPTQstrat. 3274.0054.351.848.98
GPTQstrat. 12875.7653.8750.6051.57
GPTQstrat. 51276.8052.450.048.80
DuQuantclean 12873.6251.848.852.87
DuQuantstrat. 3275.7456.751.855.82
DuQuantstrat. 12873.6954.949.451.39
DuQuantstrat. 51272.2550.045.753.97
SoloQ-Cnone76.9860.1755.1057.18
SoloQ-Nnone79.6064.1659.1457.19

Table R2. LLaDA-1.5.

MethodCalib.HellaSwagMMLU
GPTQclean 12872.8563.97
GPTQstrat. 3273.2063.83
GPTQstrat. 12872.5463.63
GPTQstrat. 51272.5163.78
DuQuantclean 12873.5463.72
DuQuantstrat. 3273.1363.64
DuQuantstrat. 12872.6064.30
DuQuantstrat. 51272.6564.39
SoloQ-Cnone73.0164.60
SoloQ-Nnone73.5464.70

On Fast-dLLM v2, SoloQ is above every calibrated row on HE, HE+ and IFEval, and on LLaDA-1.5 all rows are within 1.1 points. The 32 / 128 / 512 rows give the size curve: neither stratification nor more data closes the gap. We will add Tables R1/R2 and the protocol to the appendix.

W2 / Q2

You point it right that Eq. 3 is exact only for Haar rotations, and we will state this in §3.2.

(1) Theory. A block-rr output coordinate has variance 1h∑jQrj2Mj\frac{1}{h}\sum_j Q_{rj}^2 M_j (Eq. 5), where MjM_j is the energy of input block jj. RPBH has Qk=IQ_k = I, so an outlier-heavy block stays heavy. With a Haar QkQ_k, the variance is 1/d1/d in expectation (Prop. 1), which does not cover the one fixed rotation we deploy. Prop. 3 (our response to Reviewer NMdT, Q1) covers one fixed draw: the expected variance of block energies shrinks to 2k+2\frac{2}{k+2} of RPBH's, i.e., 19.5×19.5\times smaller at k=37k = 37.

(2) Measurement across diffusion states. One fixed rotation, mask ratios 0.1–0.9 (Dream FFN down-projection, k=37k = 37, five draws).

Table R3. Ranges over mask ratios.

RotationKL to fdf_dMax block imbalanceCodebook NMSE
None0.43–0.45––
RPBH0.011–0.0184.4–5.0×0.015–0.017
K-RPBH1.7–4.7e-51.5–1.7×0.0096
Dense Haar1.6–2.6e-51.14×0.0096

After K-RPBH the distribution barely changes with the diffusion state and matches dense Haar. Fast-dLLM v2 is similar (KL 2.0–9.4e-5 vs. 1.7–5.2e-5 for Haar).

(3) Deployment cost. Preparation on one H200 (calibration rows are per calibration set):

Table R4. Preparation cost.

MethodDataTimePeak extra memory
SoloQ-Nnone4 s1.7–2.3 GiB
SoloQ-Cnone40–46 s3.9–4.3 GiB
GPTQ, Fast-dLLM v232/128/5124/8/23 min5.8 GiB
GPTQ, LLaDA-1.532/128/5125/15/55 min2.6 GiB
DuQuant, Fast-dLLM v232/128/51220/21/28 s2.1 GiB
DuQuant, LLaDA-1.532/128/51220/22/30 s1.2 GiB
FlatQuant, Fast-dLLM v212887 min27.3 GiB
FlatQuant, Nemotron128100 min28.4 GiB

Calibration-based methods repeat this cost for every model and calibration set. At inference, Fig. 4 already includes the online K-RPBH rotation, and Q3 reports its overhead over RPBH.

[1] Williams, Miles, and Nikolaos Aletras. "On the impact of calibration data in post-training quantization and pruning." Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. 2024.

W3 / Q4

(1) Comparison with STaR-Quant. STaR-Quant, DLLMQuant and FAIR-Calib have no public code, so we could not reproduce them and instead quote the numbers reported in each paper. Before comparing, we checked our FP16 against STaR-Quant's FP16. The largest gaps are Dream TruthfulQA (ours 43.96 vs. 49.76) and LLaDA-1.5 C-EVAL (ours 66.05 vs. 70.05). On the other seven columns our FP16 is within 2.4 points of theirs, so Table R5 recomputes the averages over those seven columns.

Table R5. Averages without TruthfulQA and C-EVAL.

ModelFP16DLLMQuant++STaR-QuantSoloQ-CSoloQ-N
LLaDA-Base59.0254.8457.1458.1357.87
LLaDA-1.572.8367.8469.9071.0871.84
Dream-7B70.3464.7766.4065.7866.40

On Dream, SoloQ-N ties STaR-Quant (66.40 vs. 66.40) and SoloQ-C is 0.6 below. STaR-Quant's all-column lead on Dream (63.59 vs. 62.53 / 62.06 for SoloQ-N / C) comes from the two columns where the FP references differ. We therefore claim accuracy comparable to calibration-based methods, not superiority.

(2) Block-diffusion baselines (Q4). We agree that the collapsed RTN/AWQ rows and the N/A QuaRot row are uninformative. State-aware methods such as DLLMQuant and STaR-Quant have no public code, so we apply their key idea, calibration data sampled from the actual active-block state, to K-RPBH + GPTQ and DuQuant. We also compare with FlatQuant, which learns its transforms on calibration data.

Table R6. Fast-dLLM v2, calibrated PTQ vs. SoloQ (same harness as Table R1).

MethodCalibrationGSM8KHEHE+IFEval
K-RPBH+GPTQstrat. 12875.7653.8750.6051.57
DuQuantstrat. 12873.6954.949.451.39
FlatQuantWikiText-2 12878.9258.554.359.15
SoloQ-Cnone76.9860.1755.1057.18
SoloQ-Nnone79.6064.1659.1457.19

SoloQ is above GPTQ and DuQuant on HE, HE+ and IFEval, and on par with FlatQuant.

Table R7. Nemotron, calibrated PTQ vs. SoloQ (stratified WinoGrande-128 for GPTQ and DuQuant, one seed, same harness as Table 2).

MethodHEHE+MBPPMBPP+IFEval
K-RPBH+GPTQ66.4664.0275.4063.2358.08
DuQuant69.5165.8576.1962.4361.62
FlatQuant67.0764.0274.6062.9658.48
SoloQ-C75.6170.7375.6663.4962.35
SoloQ-N76.2272.5678.8467.4660.85

SoloQ-N leads all three calibrated runs by 6.7–9.8 points on HE and HE+, and on every task the better SoloQ variant is at or above every calibrated run.

We will add these results to the revised paper.

W4

We agree that Table 6 mixes two setups. FAIR-Calib has no public code and reports a different task set from Table 1, so we could not reproduce it or place it in Table 1. We therefore compared it separately on the five tasks both papers share, using its own FP16 row. In Table 6, only the SoloQ rows are ours. FP16/RTN/QuaRot/FlatQuant/FAIR-Calib are quoted from FAIR-Calib, and AWQ/DLLMQuant from STaR-Quant. Because the FP references differ, Table R8 compares retention rather than raw accuracy.

Table R8. FAIR-Calib vs. SoloQ on the five shared tasks, each against its own FP.

ModelMethodOwn-FP avgQuant. avgRetention
LLaDA-BaseFAIR-Calib65.7064.1297.6%
LLaDA-BaseSoloQ-C / N66.1565.16 / 65.2998.5% / 98.7%
Dream-7BFAIR-Calib68.0963.9193.9%
Dream-7BSoloQ-C / N68.7065.96 / 65.8196.0% / 95.8%

SoloQ retains more of its own FP accuracy on both models. We will mark Table 6 as a cross-paper comparison.

Q3

(1) Latency and memory. We measured both rotations with the same NVFP4 kernel and the Fig. 4 workload on RTX Pro 6000. Only the rotation differs: the RPBH arm sets Qk=IQ_k = I in both the offline weight rotation and the online activation rotation.

Table R9. End-to-end generation, same NVFP4 kernel.

ModelConfigGeneration time (ms)Peak memory (MiB)
Fast-dLLM v2BF168,05914,637
Fast-dLLM v2SoloQ-N, RPBH3,4956,228
Fast-dLLM v2SoloQ-N, K-RPBH3,6426,228
LLaDA-1.5BF162,95315,430
LLaDA-1.5SoloQ-N, RPBH1,4435,894
LLaDA-1.5SoloQ-N, K-RPBH1,4685,894

K-RPBH adds 4.2% (+147 ms) on Fast-dLLM v2 and 1.7% (+25 ms) on LLaDA-1.5, and peak memory is identical. QkQ_k itself takes only 0.18 MiB (Fast-dLLM v2) and 1.9 KB (LLaDA-1.5).

(2) Accuracy gain. For SoloQ-C, with only the rotation swapped, K-RPBH raises Fast-dLLM v2 HE, HE+ and IFEval by 2.4–3.5 points, for the same added rotation that costs 2–4% latency above. The full rotation-only comparison is in our response to Reviewer NMdT (W2 / Q2 / Q3).

Reviewer_jV9dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SOLOQ quantizes diffusion language models without calibration data. It combines K-RPBH rotations and scale correction with codebook or NVFP4 quantization. For block-diffusion models, committed KV states are quantized while the active block's KV states remain in full precision. Experiments cover five models.

Critical strengths
  1. K-RPBH improves downstream accuracy over RPBH on the tested LLaDA-1.5 and Fast-dLLM v2 tasks (Table 3). Figure 3 shows rotated marginals closer to the Gaussian reference and more uniform conditional coordinate variance across blocks.
  2. Table 4 isolates the contributions of permutation and cross-block mixing, while Figure 5 tests the effect of bias correction on downstream accuracy.
Critical weaknesses
  1. Section 5.2 emphasizes gains over DLLMQuant++ but does not discuss STaR-Quant's higher reported Dream mean (63.59 versus 62.53/62.06 for SOLOQ-N/C). This ranking is hard to interpret. Table 1's footnote acknowledges that some TruthfulQA baselines could not be reproduced and that the full-precision references differ. Although Appendix C.3 limits TruthfulQA and C-EVAL to controlled FP-to-quantized comparisons, both benchmarks are still included in the cross-method averages. Table 6 explicitly combines results reported by FAIR-Calib with results following the authors' evaluation protocol; its LLaDA-Base FP16 MMLU reference also differs from Table 1 (65.80 versus 64.15). Matched comparisons are needed to establish the claimed advantage. The abstract should specify which models and knowledge/reasoning benchmarks support superiority over calibration-based baselines.
  2. On Nemotron, W4A4 SOLOQ-N lowers average accuracy from 78.67 to 70.76 and IFEval from 97.27 to 60.85 (Table 2). The paper does not explain the much larger IFEval loss. This is a substantial instruction-following degradation, and I think it should be discussed more explicitly, especially given the retained-accuracy claim in the abstract.
Questions for authors
  1. Which Table 1 rows did you re-run, and which are quoted? Please compare SOLOQ, STaR-Quant and FAIR-Calib under matched settings on models covered by the compared methods, using the protocols in Appendix C and corresponding full-precision references. If some baselines cannot be reproduced, please explain the limitations and additionally report averages excluding non-comparable columns.
  2. Which IFEval checks account for the Nemotron drop? Please report the four component metrics and analyse representative failures.
  3. Does Figure 4 use folded or unfolded rotations? Please report accuracy for the timed native NVFP4 configuration; Fast-dLLM v2 SOLOQ-N scores 56.77 unfolded and 52.81 folded (Table 10).
  4. Table 12 reports 40.90 tokens/kJ, but its throughput and Figure 8's power (0.252 tokens/s at 7.48 W) imply 33.69 tokens/kJ. Were these measured under the same workload and power-accounting scope?
2 comments

We deeply appreciate your constructive reviews and comments, which have been invaluable in helping us strengthen our work. We address your concerns below and will integrate all feedback into the revised manuscript.

W1 / Q1

Thank you for checking our numbers so carefully. The mismatches come from one constraint: STaR-Quant, DLLMQuant and FAIR-Calib have no public code, so we could not reproduce them and instead quote the numbers reported in each paper.

In Table 1, the FP16 and SoloQ rows are our runs, and all baselines are quoted from STaR-Quant's tables. Before comparing, we checked our FP16 against STaR-Quant's FP16. The largest gaps are Dream TruthfulQA (ours 43.96 vs. 49.76) and LLaDA-1.5 C-EVAL (ours 66.05 vs. 70.05). On the other seven columns our FP16 is within 2.4 points of theirs. These two gaps are why C.3 restricts TruthfulQA and C-EVAL to FP-to-quantized comparisons, and including them in the cross-method average was inconsistent with that. Table R1 therefore recomputes the averages over the remaining seven columns.

Table R1. Averages without TruthfulQA and C-EVAL.

ModelFP16DLLMQuant++STaR-QuantSoloQ-CSoloQ-N
LLaDA-Base59.0254.8457.1458.1357.87
LLaDA-1.572.8367.8469.9071.0871.84
Dream-7B70.3464.7766.4065.7866.40

On these seven columns, SoloQ is above STaR-Quant on LLaDA-1.5 (+1.2 / +1.9 for C / N) and LLaDA-Base (+1.0 / +0.7, driven by GSM8K), and tied (SoloQ-N) or 0.6 below (SoloQ-C) on Dream. Over all nine columns STaR-Quant leads on Dream (63.59 vs. 62.53 / 62.06 for SoloQ-N / C), and that lead comes from the two non-comparable columns.

In Table 6, we compared FAIR-Calib separately because it has no public code to reproduce and reports a different task set. Only the SoloQ rows are ours. FP16/RTN/QuaRot/FlatQuant/FAIR-Calib are quoted from FAIR-Calib, and AWQ/DLLMQuant from STaR-Quant.

Table R2 compares retention on the five shared tasks, each against its own FP16 (hence LLaDA-Base MMLU 65.80, FAIR-Calib's FP16, vs. our 64.15). SoloQ retains more on both models.

Table R2. FAIR-Calib vs. SoloQ on the five shared tasks, each against its own FP.

ModelMethodOwn-FP avgQuant. avgRetention
LLaDA-BaseFAIR-Calib65.7064.1297.6%
LLaDA-BaseSoloQ-C / N66.1565.16 / 65.2998.5% / 98.7%
Dream-7BFAIR-Calib68.0963.9193.9%
Dream-7BSoloQ-C / N68.7065.96 / 65.8196.0% / 95.8%

As the closest matched comparison we can run, we calibrated GPTQ and DuQuant (public code) on LLaDA-1.5 in our own harness, with clean or mask-ratio-stratified data and the same rotation as SoloQ (Table R3). SoloQ-N matches or exceeds every calibrated row on both tasks without using data, and all rows are within 1.1 points. The full calibration sweep and the block-diffusion models are in our response to Reviewer dT4u.

Table R3. Matched comparison on LLaDA-1.5 (same harness, fixed random seeds).

MethodCalibrationHellaSwagMMLU
K-RPBH+GPTQclean 12872.8563.97
K-RPBH+GPTQstrat. 12872.5463.63
DuQuantclean 12873.5463.72
DuQuantstrat. 12872.6064.30
SoloQ-Cnone73.0164.60
SoloQ-Nnone73.5464.70

We will revise the abstract to claim accuracy comparable to calibration-based PTQ, not superiority, and §5.2 will report STaR-Quant's Dream lead and the seven-column averages.

W2 / Q2

Thank you for pointing this out. "97.27" is a typo for 67.27. With it, the FP average is 75.33, not 78.67, so SoloQ-N W4A4 loses 6.4 IFEval points, not 36, and 4.6 average points (75.33 → 70.76). Table R4 gives the four IFEval metrics.

Table R4. Four metrics (541 prompts. Cap: share of outputs hitting the 8,192-token limit. Loop: share with >30% repeated word 8-grams.)

ConfigP-strictI-strictP-looseI-looseAvg4Cap %Loop %
FP1660.6370.1464.7073.6267.274.8121.07
SoloQ-N W4A453.7965.1156.7567.7560.859.6130.13
SoloQ-N W4A4KV456.5667.2761.7470.8664.115.9128.28

The drop is spread over several instruction types (largest in startend, language and detectable_content/format), while length and punctuation checks are unchanged. It coincides with more outputs hitting the token cap (4.8% → 9.6%) or looping (21.1% → 30.1%). For example, an all-capitals song drifts into repeated lower-case lines, and a joke that must be wrapped in quotation marks loops until the token cap. This link is only correlational, and with one run per config, differences of about 3 points are within run-to-run noise.

We will correct Table 2, qualify the retained-accuracy claim.

Q3

Fig. 4 was measured in the unfolded topology. Folding appears only in App. D.3 (Tables 9/10). It measures speed with weights, activations and the KV cache all quantized (W4A4KV4), so its accuracy is the unfolded W4A4KV4 row of Table 2: 56.21 on Fast-dLLM v2 (56.77 with an FP KV cache).

To check that the native NVFP4 kernel gives the same accuracy, we ran both the kernel model and the fake-quantized model on Fast-dLLM v2, LLaDA-Base and Dream-7B in the same harness, with the same seed, on the same RTX Pro 6000 (Table R5). Both columns are these new runs, so they are compared with each other, not with Table 2.

Table R5. Fake quantization vs. native kernel (SoloQ-N W4A4KV4). Agreement is the share of problems where both models are right or both wrong. McNemar p tests the problems where they disagree.

ModelTaskFake quantizationNative kernelPer-problem agreementMcNemar p
Fast-dLLM v2HumanEval / HE+62.2 / 57.962.8 / 59.182–87%≥ 0.48
Fast-dLLM v2MBPP / MBPP+60.6 / 50.561.4 / 52.182–87%≥ 0.48
LLaDA-BaseC-EVAL68.2068.2091.8%1.0
Dream-7BARC-C57.1757.1794.7%1.0

The kernel gives the same accuracy on LLaDA and Dream. On Fast-dLLM v2 the HumanEval gaps (+0.6 / +1.2) are within the SoloQ-N std over 5 rotation seeds (2.7 / 1.8), and McNemar tests find no difference on any task (p ≥ 0.48).

Q4

Thank you for checking this. Yes, the workload (LLaDA-Base) and the power-accounting scope are the same. 40.90 is a typo, and it carried over into the 2.70× ratio. Using Fig. 8's total power (BF16: 0.13 tokens/s at 8.59 W, i.e., 15.13 tokens/kJ), NVFP4 gives 33.69 tokens/kJ, 2.23× BF16 (not 2.70×), with 12.9% less power (not 15%). Throughput (1.94×) is unchanged. We will correct Table 12 and App. E.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.