acceptodds
Under review as a conference paper at ICLR 2027

Prefix-Point: An Adaptive Non-Uniform Format for Diffusion Model Quantization

Abstract

Post-training quantization of diffusion models is challenging because activation distributions vary substantially across layers and denoising steps. Prior work mitigates this with step-aware scaling for INT quantization, but scaling adjusts only the range of a distribution and not its shape, so it cannot follow activations that shift between heavy-tailed shapes, with most values near zero, and light-tailed shapes spread evenly across the range. Non-uniform formats represent such shapes better, yet existing ones fix a single codebook at design time from an assumed distribution and reuse it throughout the network and therefore degrade generation quality sharply at low precision ( bits). We propose Prefix-Point (PXP), a format that divides the value range into exponent levels and sets the resolution of each level independently, allocating bit-width where it reduces quantization error most, and fitting this allocation separately to the distribution of each layer and denoising step. Since the resulting design space grows exponentially with bit-width, we cast level allocation as a multiple-choice knapsack problem and solve it exactly by dynamic programming. Its floating-point-like structure keeps decoding and the multiplier cheaper than in existing non-uniform formats. Across class-conditional, text-to-image and video diffusion, PXP lowers FID by up to over the strongest existing non-uniform format while being up to faster.

Primary area:
generative models
Keywords:
Quantization, Diffusion Model, DiTs, Non-Uniform Formats

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_sDX5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Prefix-Point (PXP), an adaptive non-uniform numerical format for post-training quantization of diffusion models. The key observation is that activation distributions vary not only in scale but also substantially in shape across transformer layers and denoising timesteps, making a single fixed non-uniform codebook suboptimal. PXP retains an exponent-mantissa representation but encodes exponent levels using variable-length prefix codes, such that exponent regions contributing larger quantization errors receive shorter prefixes and consequently more mantissa bits.

Critical strengths
  1. The paper identifies a meaningful and underexplored failure mode of existing diffusion quantization methods.
  2. PXP is conceptually simple but non-trivial.
  3. Results at very low precision are unusually strong.
Critical weaknesses
  1. The strongest concern is that the baseline comparison does not fully disentangle format design from calibration flexibility. PXP uses a separate calibrated prefix table for each layer and timestep, while the reported NF/AF/SF/BOF baselines use a single codebook shared across layers and denoising steps. I would like to see the results that use a single fixed PXP format shared across layers and timesteps.
  2. The evaluation deliberately removes advanced diffusion PTQ techniques, making the practical state-of-the-art comparison incomplete. The paper explicitly evaluates all formats using absolute-max scaling without equivalent transformations, rotations, learned clipping, or advanced calibration, in order to isolate the format effect. However, the Related Work simultaneously argues that PXP is orthogonal to existing timestep-aware calibration and transformation methods and can be combined with them. That important claim is not empirically verified.
  3. Hardware efficiency claims are compelling but should be interpreted more narrowly than some of the presentation suggests. The reported speedups come from a custom RTL design synthesized with Yosys/ABC using the NanGate 45 nm library, not from wall-clock execution on an existing GPU/NPU/accelerator.
  4. The Related Work section would be more complete by including additional representative PTQ methods for diffusion models, such as SVDQuant, EDA-DM, TFMQ-DM, and DilateQuant.
Questions for authors
  1. How much of PXP's gain comes from the prefix-point representation itself versus the fact that PXP is calibrated separately for every layer and timestep? Could the authors provide an equally fine-grained adaptive baseline under a comparable metadata/storage budget?
  2. The paper argues that PXP is orthogonal to existing diffusion PTQ calibration/transformation methods. Could the authors integrate PXP into one representative strong diffusion PTQ pipeline and report whether the gain remains after such transformations?
8 comments
AuthorsOpenReview ↗

General Response to Reviewer sDX5

We sincerely thank you for your positive assessment of our work and for recognizing the significance of the identified quantization challenge and PXP's strong low-precision performance. We also appreciate your constructive feedback and insightful questions. We address each comment in detail below.

AuthorsOpenReview ↗

(Question 1 - part 1/2) Global, layer-only, timestep-only and per layer-and-timestep PXP codebooks.

Thank you for highlighting the ambiguity in attributing quality gains to the format itself vs. layer- and timestep-specific codebooks. The table below compares four codebook granularities: global, per-layer, per-timestep, and joint per-layer-and-timestep (as reported in the paper) for 256×256 ImageNet generation using DiT-XL/2 (CFG = 1.5) over 50 denoising steps at W4A6. For ease of comparison, we also include the formats evaluated in Table 1 of the main paper.

MethodBit-widthGranularityIS ↑FID ↓sFID ↓Precision ↑Recall ↑
FP16 (Baseline)236.115.4017.570.810.67
INTW4A6Global2.58220.06422.060.030.08
NFW4A6Global53.6244.36162.970.440.56
AFW4A6Global72.8739.5643.640.450.70
SFW4A6Global34.4177.1062.210.270.62
BOFW4A6Global7.53160.2774.080.110.27
PXP (Ours)W4A6Global121.9619.0059.600.620.64
PXP (Ours)W4A6Per-layer179.9010.0523.820.720.67
PXP (Ours)W4A6Per-timestep158.9112.6235.520.690.65
PXP (Ours)W4A6Per-layer and timestep182.509.7023.500.730.67

The results show that both per-layer and per-timestep codebooks provide substantial improvements over a global codebook, with per-layer codebooks offering the greatest benefit and achieving performance close to that of per-layer-and-timestep codebooks. Notably, even with a single global codebook, PXP substantially outperforms all existing formats in IS, FID and precision. These results highlight two complementary advantages of PXP: its flexible quantization levels provide a more effective format than existing alternatives, while its adaptivity enables codebooks to be optimized for individual layers and timesteps. Unlike existing formats, which are fixed based on assumed prior distributions, PXP naturally supports such fine-grained adaptation.

AuthorsOpenReview ↗

(Question 1 - part 2/2) Comparison with equally fine-grained adaptive baseline.

For comparisons against equally fine-grained adaptive baselines, we compare PXP with AFP [1] and FloatMax [2], both of which independently determine the floating-point format for each layer and timestep, offering the same adaptation granularity as PXP.

The table below compares the generation results of AFP, FloatMax, and PXP on DiT-XL/2 (CFG=1.5) over 50 denoising steps at W4A6 quantization precision. All methods use absolute-maximum scaling with a group size of 128 and are calibrated using 64 samples.

MethodBit-widthIS ↑FID ↓sFID ↓Prec ↑Rec ↑
FP16 (Baseline)236.115.4017.570.810.67
AFPW4A690.1027.8132.750.520.70
FloatMaxW4A632.3062.48206.600.340.54
PXP (Ours)W4A6182.509.7023.500.730.67

The results show that AFP and FloatMax still perform significantly worse than PXP. AFP performs better due to its greater flexibility in jointly optimizing the exponent width and exponent bias, whereas FloatMax restricts the exponent width to 2 or 3 for hardware efficiency and uses a fixed exponent bias. Notably, although AFP and FloatMax employ the same granularity as PXP (i.e., per-layer-and-timestep formats), they still fall short of PXP in generation quality. We also refer the reviewer to Appendix I, where we compare PXP with FP-Cal, which uses an identical format to AFP but selects the optimal format based on minimum MSE, whereas AFP uses minimum KL divergence. These results highlight the advantages of PXP's additional flexibility in adapting mantissa widths across exponent levels.

References:

[1] Liu, Fangxin, et al. "Improving neural network efficiency via post-training quantization with adaptive floating-point." 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2021.

[2] Chung, Seoho, et al. "FloatMax: An Efficient Accelerator for Transformer-based Models Exploiting Tensor-wise Adaptive Floating-Point Quantization." 2024 IEEE 42nd International Conference on Computer Design (ICCD). IEEE, 2024.

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

The paper proposes Prefix-Point (PXP), a non-uniform format for diffusion-model post-training quantization. It uses variable-length exponent prefixes to assign different mantissa widths across magnitude ranges while keeping the total word length fixed. Dynamic programming finds the optimal allocation under the specified quantizer and prefix-code constraints. Tables are calibrated per layer and timestep, then fixed at inference. Experiments on DiT-XL/2, PixArt-Sigma, and Open-Sora show clear low-bit gains, while synthesis and accelerator modeling suggest potential hardware savings.

Critical strengths
  1. The format design makes sense: prefix lengths control where mantissa precision is spent, and the optimization finds the best allocation within the stated constraints. This gives the paper a concrete contribution at the representation level while retaining structured arithmetic.
  2. The low-bit gains are substantial. On DiT-XL/2 with 100 steps and CFG=1.5, W4A6 PXP reaches FID 8.62 versus 39.02 for AF (Table 1). Under the same sampling setup, W16A6 PXP reaches 5.65, compared with 16.27 for AF and 5.42 for FP16 (Table A1). The latter comparison is particularly useful for showing the benefit for activation quantization. Results across tasks and the additional ablations provide further support.
Critical weaknesses
  1. It is still hard to separate the format’s benefit from its adaptation granularity. PXP adapts tables per layer and timestep, while the main fixed-codebook baselines share tables across both. The calibrated-minifloat and Lloyd-Max comparisons help, but do not fully resolve this issue. Comparisons with stronger methods, especially FP4DiT and VETA-DiT, would also make the competitiveness claim more convincing.
  2. The hardware claims need a clearer scope. Appendix J uses logic synthesis and accelerator modeling, assumes a fully utilized array, and uses FP16 arithmetic for LUT baselines. These results suggest potential savings, but leave open how much survives under different arithmetic precision, utilization, or bandwidth limits. The main speedup claims should make these assumptions explicit.

I find the contribution promising, and these concerns seem addressable. A concrete, feasible commitment to stronger comparisons under a clear protocol, together with clearer hardware claims, would make me consider a modest upward revision. Supporting results would strengthen the case for acceptance.

Questions for authors
  1. Could you provide, or commit to adding, comparisons with FP4DiT and representative methods such as VETA-DiT or Q-DiT? Please match the sampling and evaluation setup, report calibration budgets and quantization settings, and account for resource differences from mixed precision or additional branches.
  2. Could you compare globally shared, layer-only, timestep-only, and layer-and-timestep PXP tables? Please keep calibration data, scaling, and weight quantization fixed, and optimize each variant under its own sharing constraint. Both MSE and generation metrics would help separate the effects.
  3. Applying Equations (4)–(7) to Figure 3’s 4-bit codebook maps normalized input 0.49 to magnitude 0.375, although 0.5 is representable and closer. Is this within-interval clamping intentional and used in both calibration and inference? I understand the DP guarantee to cover the specified quantizer. If the implementation allows cross-interval rounding, how does the additive cost account for it?
  4. How sensitive are the hardware estimates to lower utilization or limited bandwidth? What motivates FP16 decoded precision for LUT baselines, and would lower decoded precision preserve their generation quality while changing PXP’s estimated advantage?
9 comments
AuthorsOpenReview ↗

General Response to Reviewer QmcC

We sincerely thank you for your positive assessment of our work and for recognizing the contributions and promising results of PXP. We also appreciate your constructive feedback and insightful questions. We address each comment in detail below.

AuthorsOpenReview ↗

(Question 1 - part 1/2) Comparison with FP4DiT.

Thank you for suggesting these additional comparisons to strengthen the paper. For comparison with FP4DiT [1], the table below compares the generation results of FP4DiT and PXP on DiT-XL/2 (CFG=1.5) over 50 denoising steps at W4A6 quantization precision. All methods use absolute-maximum scaling with a group size of 128 and are calibrated using 64 samples. Since FP4DiT relies on manually selected FP formats for weights and activations and does not report results for DiT-XL/2 in the paper, we evaluate two configurations: (1) the official configuration, which uses E2M1 for weights (except for the first linear layer, which uses E3M0) and E2M3 for activations; and (2) a configuration that searches across all FP formats for activations and selects the best-performing format, E3M2. Since FP4DiT uses adaptive rounding [2], we also evaluate a variant using round-to-nearest (RTN). Unless otherwise specified, FP4DiT uses adaptive rounding.

MethodBit-widthIS ↑FID ↓sFID ↓Prec ↑Rec ↑
FP16 (Baseline)-236.115.4017.570.810.67
FP4DiT (paper setting)W4A622.3076.56261.000.280.48
FP4DiT (best format)W4A6151.4111.8925.430.680.72
FP4DiT (best format) + RTNW4A694.9626.5138.830.520.73
PXP (Ours)W4A6182.509.7023.500.730.67

The results show that FP4DiT still performs significantly worse than PXP. Notably, even when searching across all FP formats to identify its best activation format, FP4DiT fails to match PXP. Moreover, adaptive rounding substantially improves FP4DiT over round-to-nearest, highlighting its importance for generation quality. In contrast, PXP achieves better generation quality through its adaptability to the input distribution, despite relying on round-to-nearest rather than adaptive rounding.

References:

[1] Chen, Ruichen, Keith G. Mills, and Di Niu. "FP4DiT: Towards Effective Floating Point Quantization for Diffusion Transformers." Transactions on Machine Learning Research (2025).

[2] Nagel, Markus, et al. "Up or down? adaptive rounding for post-training quantization." International conference on machine learning. PMLR, 2020.

AuthorsOpenReview ↗

(Question 1 - part 2/2) Comparison with VETA-DiT.

As suggested, we also compare PXP against VETA-DiT, which primarily proposes a calibration and transformation pipeline to make input distributions more quantization-friendly. Since these techniques are orthogonal to the underlying quantization format and can be applied to any format, we adopt the same pipeline for PXP to ensure a fair comparison, including GPTQ for weights, zero-point for both weight and activation quantization, SmoothQuant (α=0.5\alpha=0.5), and KLT and Hadamard transformations. We then compare the default INT format against PXP within the VETA-DiT pipeline. We also compare other non-uniform formats in Table 1 using the VETA-DiT pipeline. Following VETA-DiT's experimental setup, we evaluate DiT-XL/2 with CFG=1.5 over 50 denoising steps using 10,000 generated ImageNet images under W4A4 quantization in the table below. Each method is calibrated using 64 calibration samples.

MethodBit-WidthIS ↑FID ↓sFID ↓Precision ↑Recall ↑
FP16168.064.7017.680.800.66
VETA-DiT (INT)W4A4135.667.9019.110.730.69
NF (VETA-DiT)W4A4140.637.2018.700.750.69
AF (VETA-DiT)W4A4130.568.4119.240.730.70
SF (VETA-DiT)W4A4123.699.6219.940.720.70
BOF (VETA-DiT)W4A4126.259.6123.050.720.67
PXP (VETA-DiT)W4A4144.096.7319.120.750.69

The results show that VETA-DiT transformations substantially reduce the performance gap between INT, other non-uniform formats, and PXP by making the distributions more quantization-friendly. Nevertheless, PXP still achieves significantly better generation quality than vanilla VETA-DiT using INT and all other non-uniform formats in IS and FID. Notably, among the non-uniform formats, only NF and PXP achieve better generation quality than INT. This is likely because VETA-DiT's transformations make the distributions more closely resemble a standard normal, making them more suitable for INT quantization. NF can better exploit this distribution than other fixed formats, while PXP's adaptive quantization levels provide further gains over both INT and NF.

Reviewer_QB9DOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Prefix-Point (PXP), an adaptive non-uniform numerical format for post-training quantization of diffusion models whose activation distributions vary across layers and timesteps. PXP uses variable-length exponent prefixes to allocate mantissa precision across exponent levels, decomposes quantization MSE into level-wise terms, and obtains optimal prefix-length allocation within the proposed format through dynamic programming. Experiments on DiT-XL/2, PixArt-Σ, and Open-Sora evaluate low-bit image and video generation, supplemented by RTL implementations and synthesis-based hardware evaluations.

Critical strengths
  1. A technically meaningful adaptive numerical format design. PXP combines prefix-based exponent encoding with variable mantissa precision, allowing quantization codewords to be allocated adaptively according to the activation distribution. This design provides greater representational flexibility than conventional fixed floating-point formats while maintaining a fixed total bit-width.

  2. Substantial improvements in generation quality at very low precision. The paper demonstrates strong performance under aggressive quantization settings, particularly W4A5 and W4A6. For example, on DiT-XL/2 with W4A6 quantization, 100 denoising steps, and CFG=1.5, PXP reduces FID from 39.02 for the strongest non-uniform baseline to 8.62, approaching the FP16 result of 5.42.

Critical weaknesses
  1. It is unclear how much of the gain comes from the prefix format itself and how much comes from choosing different tables for different layers and timesteps. The FP-Cal and Lloyd-Max controls help, but the paper does not compare PXP with a global table, layer-wise tables, timestep-wise tables, and joint tables.

  2. The paper omits FP4DiT, which already studies floating-point PTQ for diffusion transformers. The claim that previous diffusion quantization remains within INT is therefore too broad, and the comparison with existing floating-point PTQ is incomplete.

[1] FP4DiT: Towards Effective Floating Point Quantization for Diffusion Transformers.

Questions for authors
  1. What does PXP add over FP4DiT and adaptive floating-point formats such as AFP and FloatMax? Please clarify the differences and compare with FP4DiT at matched bit widths and sampling settings.

  2. How much improvement comes from adapting tables across layers and timesteps? Please compare global, layer-wise, timestep-wise, and joint PXP tables.

[2] Improving Neural Network Efficiency via Post-Training Quantization With Adaptive Floating-Point.

[3] FloatMax: An Efficient Accelerator for Transformer-Based Models Exploiting Tensor-Wise Adaptive Floating-Point Quantization.

6 comments
AuthorsOpenReview ↗

General Response to Reviewer QB9D

We sincerely thank you for your positive assessment of our work and for recognizing the technical contributions and performance improvements of PXP, as well as for your constructive feedback and insightful questions. We address each comment in detail below.

AuthorsOpenReview ↗

(Question 1) Comparison with FP4DiT, AFP and FloatMax.

Thank you for pointing out FP4DiT [1], AFP [2], and FloatMax [3], which, like PXP, employ non-uniform floating-point formats. Although AFP and FloatMax optimize the exponent width for each tensor, allowing the number of exponent levels to adapt to the input distribution, all three methods assign a uniform mantissa width across all exponent levels within a tensor. PXP removes this constraint by allowing each exponent level to have its own mantissa width, introducing an additional degree of flexibility to distribute the available quantization levels according to the input distribution.

The table below compares the generation results of FP4DiT, AFP, FloatMax, and PXP on DiT-XL/2 (CFG=1.5) over 50 denoising steps at W4A6 quantization precision. AFP originally also employs mixed-precision optimization to allocate bit-widths across layers. Since bit-width allocation is orthogonal to format design, we evaluate AFP at the same precision as PXP, FloatMax, and FP4DiT to ensure a fair comparison. All methods use absolute-maximum scaling with a group size of 128 and are calibrated using 64 samples. Since FP4DiT relies on manually selected FP formats for weights and activations and does not report results for DiT-XL/2 in the paper, we evaluate two configurations: (1) the official configuration, which uses E2M1 for weights (except for the first linear layer, which uses E3M0) and E2M3 for activations; and (2) a configuration that searches across all FP formats for activations and selects the best-performing format, E3M2. Since FP4DiT uses adaptive rounding [4], we also evaluate a variant using round-to-nearest (RTN). Unless otherwise specified, FP4DiT uses adaptive rounding.

MethodBit-widthIS ↑FID ↓sFID ↓Prec ↑Rec ↑
FP16 (Baseline)236.115.4017.570.810.67
FP4DiT (paper setting)W4A622.3076.56261.000.280.48
FP4DiT (best format)W4A6151.4111.8925.430.680.72
FP4DiT (best format) + RTNW4A694.9626.5138.830.520.73
AFPW4A690.1027.8132.750.520.70
FloatMaxW4A632.3062.48206.600.340.54
PXP (Ours)W4A6182.509.7023.500.730.67

The results show that FP4DiT, AFP, and FloatMax still perform significantly worse than PXP. Notably, even when searching across all FP formats to identify its best activation format, FP4DiT fails to match PXP. Moreover, adaptive rounding substantially improves FP4DiT over round-to-nearest, highlighting its importance for generation quality. Among AFP and FloatMax, AFP achieves better performance due to its greater flexibility in jointly optimizing the exponent width and exponent bias, whereas FloatMax restricts the exponent width to 2 or 3 for hardware efficiency and uses a fixed exponent bias. Notably, although AFP and FloatMax employ the same granularity as PXP (i.e., per-layer-and-timestep formats), they still fall short of PXP in generation quality. These results highlight the advantages of PXP's additional flexibility in adapting mantissa widths across exponent levels.

References:

[1] Chen, Ruichen, Keith G. Mills, and Di Niu. "FP4DiT: Towards Effective Floating Point Quantization for Diffusion Transformers." Transactions on Machine Learning Research (2025).

[2] Liu, Fangxin, et al. "Improving neural network efficiency via post-training quantization with adaptive floating-point." 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2021.

[3] Chung, Seoho, et al. "FloatMax: An Efficient Accelerator for Transformer-based Models Exploiting Tensor-wise Adaptive Floating-Point Quantization." 2024 IEEE 42nd International Conference on Computer Design (ICCD). IEEE, 2024.

[4] Nagel, Markus, et al. "Up or down? adaptive rounding for post-training quantization." International conference on machine learning. PMLR, 2020.

AuthorsOpenReview ↗

(Question 2) Comparing global, layer-wise, timestep-wise and joint layer-and-timestep-wise PXP Codebooks.

Thank you for highlighting the ambiguity in attributing quality gains to the format itself vs. layer- and timestep-specific codebooks. The table below compares four codebook granularities: global, per-layer, per-timestep, and joint per-layer-and-timestep (as reported in the paper) for 256×256 ImageNet generation using DiT-XL/2 (CFG = 1.5) over 50 denoising steps at W4A6. For ease of comparison, we also include the formats evaluated in Table 1 of the main paper.

MethodBit-widthGranularityIS ↑FID ↓sFID ↓Precision ↑Recall ↑
FP16 (Baseline)--236.115.4017.570.810.67
INTW4A6Global2.58220.06422.060.030.08
NFW4A6Global53.6244.36162.970.440.56
AFW4A6Global72.8739.5643.640.450.70
SFW4A6Global34.4177.1062.210.270.62
BOFW4A6Global7.53160.2774.080.110.27
PXP (Ours)W4A6Global121.9619.0059.600.620.64
PXP (Ours)W4A6Per-layer179.9010.0523.820.720.67
PXP (Ours)W4A6Per-timestep158.9112.6235.520.690.65
PXP (Ours)W4A6Per-layer and timestep182.509.7023.500.730.67

The results show that both per-layer and per-timestep codebooks provide substantial improvements over a global codebook, with per-layer codebooks offering the greatest benefit and achieving performance close to that of per-layer-and-timestep codebooks. Notably, even with a single global codebook, PXP substantially outperforms all existing formats in IS, FID and precision. These results highlight two complementary advantages of PXP: its flexible quantization levels provide a more effective format than existing alternatives, while its adaptivity enables codebooks to be optimized for individual layers and timesteps. Unlike existing formats, which are fixed based on assumed prior distributions, PXP naturally supports such fine-grained adaptation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback and positive assessment of our work, which has substantially improved our paper. For the convenience of the reviewers and AC, we summarize the key strengths identified by the reviewers, the concerns raised, and our corresponding responses below. All revisions are highlighted in blue in the updated PDF.

Strengths Noted by Reviewers

Key strengthsReviewers
A technically meaningful adaptive numerical format designQB9D\color{steelblue}{\mathbf{QB9D}}
The format design makes senseQmcC\color{indianred}{\mathbf{QmcC}}
Concrete contribution at the representation levelQmcC\color{indianred}{\mathbf{QmcC}}
Identifies a meaningful and underexplored failure mode in diffusion model quantizationsDX5\color{goldenrod}{\mathbf{sDX5}}
Conceptually simple but non-trivial approachsDX5\color{goldenrod}{\mathbf{sDX5}}
Substantial improvements in generation quality at very low precisionQB9D\color{steelblue}{\mathbf{QB9D}}, QmcC\color{indianred}{\mathbf{QmcC}}, sDX5\color{goldenrod}{\mathbf{sDX5}}
Useful activation quantization results in Table A1QmcC\color{indianred}{\mathbf{QmcC}}
Results across tasks and additional ablations provide further supportQmcC\color{indianred}{\mathbf{QmcC}}

Key Concerns and Our Responses

Key concernsReviewersOur responsesRevision
Unclear contribution of format design vs. codebook granularityQB9D\color{steelblue}{\mathbf{QB9D}}, QmcC\color{indianred}{\mathbf{QmcC}}, sDX5\color{goldenrod}{\mathbf{sDX5}}Compared global, layer-wise, timestep-wise, and joint layer-and-timestep-wise PXP codebooks.App. L, Tab. A4
Comparison with FP4DiTQB9D\color{steelblue}{\mathbf{QB9D}}, QmcC\color{indianred}{\mathbf{QmcC}}Provided a detailed comparison with FP4DiT under matched settings.App. J, Tab. A2
Comparison with AFP and FloatMaxQB9D\color{steelblue}{\mathbf{QB9D}}Provided comparisons with AFP and FloatMax under matched settings.App. J, Tab. A2
Comparison with VETA-DiTQmcC\color{indianred}{\mathbf{QmcC}}, sDX5\color{goldenrod}{\mathbf{sDX5}}Compared PXP and existing formats integrated into the VETA-DiT pipeline.App. K, Tab. A3
Ambiguity between cross-interval rounding and within-interval clampingQmcC\color{indianred}{\mathbf{QmcC}}Corrected the presentation in the revised manuscript and compared the generation quality of cross-interval rounding and within-interval clamping.Sec. 4, Eqs. 6 and 8
Sensitivity to hardware utilization and memory bandwidthQmcC\color{indianred}{\mathbf{QmcC}}Provided hardware synthesis results across different hardware utilization levels and bandwidth limits.App. P, Tabs. A7 and A8
Effects of lower decoding precision for LUT-based formatsQmcC\color{indianred}{\mathbf{QmcC}}Decoded LUT-based formats to FP8 and provided corresponding hardware synthesis and generation quality results.App. O, Tabs. A5 and A6
Comparison with equally fine-grained adaptive baselinessDX5\color{goldenrod}{\mathbf{sDX5}}Compared PXP with AFP and FloatMax using per-layer-and-timestep-wise formats, matching PXP's granularity.App. J, Tab. A2
Hardware results do not reflect actual wall-clock executionsDX5\color{goldenrod}{\mathbf{sDX5}}Implemented PXP and existing formats on an A100 GPU and provided actual inference latency comparisons.App. Q, Tab. A9
Missing related work referencessDX5\color{goldenrod}{\mathbf{sDX5}}Included the suggested references in the revised manuscript.Sec. 2

Reviewer Responses Following Rebuttal

ReviewerResponse statusSummary
QB9D\color{steelblue}{\mathbf{QB9D}}✓\color{green}{\checkmark}Remains Positive
QmcC\color{indianred}{\mathbf{QmcC}}✓\color{green}{\checkmark}Concerns addressed
sDX5\color{goldenrod}{\mathbf{sDX5}}✓\color{green}{\checkmark}Concerns addressed

Revisions in the Manuscript (marked in blue)

Summary of the additionRevision
Fixed the presentation error in Eq. 6 and added Eq. 8 to eliminate the ambiguity between cross-interval rounding and within-interval clamping, as pointed out by reviewer QmcC\color{indianred}{\mathbf{QmcC}}Sec. 4, Eqs. 6 and 8
Included the missing references suggested by reviewer sDX5\color{goldenrod}{\mathbf{sDX5}}Sec. 2
Comparison with floating-point format-based methods (FP4DiT, AFP, and FloatMax)App. J, Tab. A2
Integration of PXP and existing formats with VETA-DiTApp. K, Tab. A3
Effect of codebook granularityApp. L, Tab. A4
Effect of reduced decoding precision for LUT-based formatsApp. O, Tabs. A5 and A6
Sensitivity to utilization and bandwidth limitsApp. P, Tabs. A7 and A8
Inference latency on an A100 GPUApp. Q, Tab. A9

We thank the reviewers again for their valuable feedback and would be happy to provide any further clarification .

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.