acceptodds
Under review as a conference paper at ICLR 2027

ScaleSweep: Accurate NVFP4 Post-Training Quantization of LLMs via Block Scale Initialization

Abstract

NVFP4 is a recently introduced hardware-supported FP4 format that improves the fidelity of 4-bit quantization through fine-grained block scales. However, existing NVFP4 scale initialization methods still rely on AbsMax initialization, which leaves a noticeable gap to the optimal solution. To address this, we propose ScaleSweep, a simple and efficient scale optimization method that sweeps over block scale candidates and selects the candidate that minimizes a target objective. We further provide a theoretical analysis of NVFP4 quantization and derive both lower and upper bounds for the required sweep range under mean square error (MSE) and weighted mean square error (WMSE) between the original tensor and the quantized reconstructed tensor. The proposed bounds substantially reduce the sweep space while preserving the optimal candidate and maintaining end-to-end inference throughput close to the baseline. Experiments on Llama and Qwen models demonstrate that ScaleSweep consistently improves quantization performance over existing initialization methods and further narrows the gap to full precision. In particular, under aggressive end-to-end quantization of weights, activations, KV cache, and query states, ScaleSweep preserves more than 93% of the full-precision performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
post-training quantization, nvfp4 quantization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JTVYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ScaleSweep searches neighboring FP8 E4M3 scales for NVFP4 blocks using MSE or importance-weighted MSE. It derives a block-size-16 MSE window of [-3, 7] and uses an empirical WMSE window of [-8, 7]. Experiments cover four Llama/Qwen models with RTN and GPTQ, from weight/activation quantization through additional KV-cache and query quantization, plus native Blackwell kernel and throughput measurements.

Critical strengths
  1. The MSE window has a concrete basis: a grid-containment upper bound and a lower bound backed by finite convex certificates (Section 4.2.2; Appendix B). This explains why a short search can replace full FP8 enumeration.

  2. The accuracy gains come with usable throughput. For Llama-3.1-8B with GPTQ, WAKVQ average accuracy rises from 66.09 with 4/6 to 67.35 with ScaleSweep (Table 3). Restricted WMSE and MSE retain about 93.9% and 99.5% of default throughput, well above full search (Table 4).

Critical weaknesses
  1. The closest methods are neither adequately distinguished nor compared. Section 2, lines 124-125, incorrectly limits ScaleSearch to attention inputs: ScaleSearch already searches neighboring FP8 bit patterns for weight/activation PTQ and selects scales by block MSE. H-Scale, Sections 3.2-3.3, uses the same diagonal activation-energy objective and a 16-candidate weight-scale search. Comparisons only with AbsMax and 4/6 leave the value of ScaleSweep's window and broader WMSE use unclear.

  2. The optimality claim is broader than the analysis supports. The abstract and conclusion claim preservation of the optimal candidate under both objectives, but Section 4.2.2 acknowledges that the optimal WMSE scale can be arbitrarily small. Thus [-8, 7] can exclude it. Figure 3's downstream comparison on one model does not quantify how often this happens or the objective gap. Since WMSE is the default, this distinction matters. The Q/K/V importance weights also need a clearer connection to attention-output error.

Questions for authors
  1. How does ScaleSweep compare with ScaleSearch and H-Scale under matched global scales, calibration, PTQ backend, and candidate budgets? Which gains come from the proved MSE window or from extending WMSE beyond weights?

  2. Is optimal-candidate preservation intended only for MSE? For WMSE, how often does [-8, 7] exclude the optimum, and what is the objective gap for weights, activations, and Q/K/V? What attention-output approximation justifies the Q/K/V weights, and is Q/K importance calibrated or computed online?

Reviewer_WeqxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates post-training quantization (PTQ) for large language models (LLMs) using the newly introduced NVFP4 format (combining an FP4 E2M1 core format with FP8 micro-block scales and an FP32 global scale). The authors observe that existing NVFP4 scale initialization strategies (such as AbsMax and the heuristic 4/64/6 method) rely heavily on simple bounds and leave a noticeable gap from optimal solutions.To bridge this gap efficiently, the paper proposes ScaleSweep, a block-scale optimization framework that searches for the optimal FP8 block scale minimizing a target objective. The core contributions include:

  1. Theoretical Bounds: Deriving mathematically justified lower and upper bounds for the optimal FP8 block scale under both Mean Square Error (MSE) and Weighted Mean Square Error (WMSE) objectives. These bounds shrink the search space to a tight local neighborhood in the FP8 bit-pattern space.

  2. ScaleSweep Framework: Restricting scale sweep to the derived compact range, which enables fast and practical calibration without sacrificing inference throughput.

  3. Comprehensive Evaluation: Validating ScaleSweep on Llama and Qwen model families across increasingly aggressive quantization settings—including weight-activation (WA), weight-activation-KV cache (WAKV), and weight-activation-KV cache-query states (WAKVQ) quantization—under both RTN and GPTQ pipelines. Results show consistent performance recovery improvements (recovering up to 93–95% of full-precision performance even under aggressive settings) with minimal hardware overhead.

Critical strengths
  1. First-of-its-Kind Theoretical Analysis for NVFP4 Scales: While prior works rely on heuristic initializations (like AbsMax or 4/64/6), this paper provides rigorous theoretical lower and upper bounds for FP8 block-scale optimization under both MSE and WMSE objectives, advancing the mathematical foundation of microscaling formats.

  2. Effective Space Reduction for Feasible Optimization: By translating global/local scale constraints into compact FP8 bit-pattern offsets (e.g., restricted offset ranges like [−3,7][-3, 7] or [−8,7][-8, 7]), ScaleSweep turns an otherwise expensive or unconstrained search into a highly efficient local sweep that preserves fast inference throughput.

  3. Extensive and Rigorous Evaluation: The empirical study is exceptionally thorough, testing multiple model sizes across Llama-3.1, Llama-3.2, and Qwen3, covering diverse reasoning benchmarks (MMLU Pro, GSM8K, IFEval, HellaSwag, WinoGrande), and examining progressive, multi-component quantization setups (WA, WAKV, WAKVQ) alongside Triton kernel latency benchmarks.

Critical weaknesses
  1. Dependence of WMSE on Diagonal Hessian Approximations: The WMSE objective relies on a diagonal approximation of the activation Gram matrix (XTXX^T X) to decouple channels. While standard in PTQ, ignoring off-diagonal cross-channel coupling terms may introduce sub-optimality, especially in layers with high feature correlation.

  2. Limited Scope of Downstream Task Alignment: As noted in the discussion of sweep ranges, local surrogate metrics (MSE/WMSE) do not always perfectly correlate with end-to-end downstream task performance. A more direct loss-aware or task-aware refinement could potentially close the remaining gap to full precision even further.

Questions for authors
  1. Generalization to Non-Transformer Architectures / Future Formats: How easily can the theoretical bounds derived for the E4M3 micro-block format generalize to other emerging micro-scaling formats (e.g., E5M2 or sub-byte integer-floating hybrids) or non-Transformer architectures (such as state-space models like Mamba)?

  2. Impact of Calibration Data Distribution: In Appendix E.3, the authors show that using task-specific data (like the GSM8K training split) as a calibration dataset yields better performance on math tasks compared to RedPajama. Could ScaleSweep benefit from a mixture-of-domains calibration strategy to prevent task-specific overfitting while maintaining general reasoning capabilities across diverse benchmarks?

Reviewer_Byw3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ScaleSweep, a method to improve NVFP4 post-training quantization by optimizing the FP8 scale shared by each 16-element FP4 block. With the tensor-level scale fixed, ScaleSweep searches a compact set of representable FP8 scales and selects the candidate minimizing either mean squared error (MSE) or importance-weighted MSE (WMSE). The authors derive format-specific search bounds for MSE and combine a theoretical upper bound with an empirical lower cutoff for WMSE, reducing the search to at most 11 and 16 candidates, respectively. The method integrates with RTN and GPTQ and supports quantization of weights, activations, KV caches, and query states. Experiments on four Llama and Qwen models across five benchmarks generally show improved accuracy over AbsMax and 4/6 initialization, with the WMSE variant retaining over 93% of the BF16 average score under the most aggressive quantization configuration. Accuracy is evaluated through simulated quantization, while separate Triton kernel and vLLM throughput measurements on Blackwell GPUs demonstrate substantially lower overhead than exhaustive scale search.

Critical strengths

The paper’s main technical strength is its analysis of the MSE search range for 16-element NVFP4 blocks, particularly the lower-bound argument and explicit convex-combination certificates in Appendix B. These provide a mathematical rationale for pruning candidate scales beyond the empirical window selection used in ScaleSearch. This analytical contribution is the main positive consideration in my assessment, although its limited scope does not, by itself, outweigh the concerns about incremental novelty.

Critical weaknesses
  1. Limited novelty and inaccurate positioning relative to ScaleSearch. Section 2 states that ScaleSearch applies only to attention inputs, whereas ScaleSearch explicitly covers PTQ of both weights and activations. Its Algorithm 1 already enumerates neighboring representable FP8 scales through integer offsets to their bit patterns and selects the candidate minimizing reconstruction MSE. ScaleSweep follows essentially the same core search procedure. Although the MSE search-range analysis and WMSE extension provide additional contributions, the paper does not adequately establish why these additions constitute a substantial methodological advance over ScaleSearch. The inaccurate characterization of this closely related work further weakens the paper’s novelty argument.

  2. Insufficient comparative evidence for the proposed refinements. The main comparisons for scale initialization and optimization are against AbsMax and 4/6, while ScaleSearch is omitted. This omission is particularly consequential because ScaleSearch already reports that the benefit of scale search is more pronounced for smaller blocks (§4.1, Figure 7). Improvements over AbsMax and 4/6 therefore do not establish an advantage over the closest existing scale-search method. A controlled comparison with ScaleSearch, measuring both downstream accuracy and computational overhead under matched quantization settings, is necessary to assess the practical value of the proposed search bounds and weighting scheme. Without this comparison, the empirical evidence is insufficient to substantiate a meaningful advance over existing work.

Questions for authors
  1. Comparison with ScaleSearch. ScaleSearch includes PTQ experiments through ModelOpt (§5.1, Table 1), in addition to its attention applications. Could the authors share the considerations behind not including it as a baseline for the corresponding PTQ experiments?

  2. Triton kernel implementation. Could the authors elaborate on how the Triton kernels described in §5.2 organize the evaluation of up to 11 MSE or 16 WMSE scale candidates on the GPU, particularly the parallelization strategy and reuse of input data across candidates?

  3. Interpretation of quantization overhead. Table 10 reports a relative latency of 1.13× for ScaleSweep_MSE at 512×8192, while ScaleSearch reports 1.74× for FP32 inputs of shape 2048×2048 in Table 6. Could the authors help us understand why the reported quantization overhead appears to differ so substantially between the two methods?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.