Source-linked AI summary
Benchmarking Post-Training Quantization of Large Language Models under Microscaling Floating Point Formats
Manyi Zhang, Ji-Fu Li, Zhongao Sun, Haoli Bai, Hui-Ling Zhen, Zhenhua Dong, Xianzhi Yu
TL;DR
Existing PTQ methods largely target integer formats, leaving their behavior under MXFP insufficiently understood. This work systematically evaluates PTQ across MXFP formats, methods, models, modalities, and benchmarks. MXFP8 is near-lossless, whereas MXFP4 is sensitive to format compatibility and scaling-factor error, which pre-scale optimization can mitigate.
Problem
Existing PTQ research predominantly targets integer formats, leaving its effectiveness and failure modes under MXFP largely underexplored.
Method
The study evaluates over seven PTQ methods across MXFP8 and MXFP4, multiple quantization settings, benchmarks, LLM families, and multimodal models.
Results
MXFP8 supports near-lossless quantization, while MXFP4 degrades accuracy; error compensation and affine methods are more compatible, and trends remain consistent across models and modalities.
Takeaways & Limitations
MXFP should be treated as a distinct numerical regime requiring format-aware PTQ design, with pre-scale optimization recommended for MXFP4.
Takeaways & Limitations
The study covers 7B/8B-scale models and MXFP formats but does not evaluate substantially larger models or NVIDIA-specific NVFP formats.
Abstract
from arXiv · showhide
Microscaling Floating-Point (MXFP) has emerged as a promising low-precision format for large language models (LLMs). Despite various post-training quantization (PTQ) algorithms being proposed, they mostly focus on integer quantization, while their applicability and behavior under MXFP formats remain largely unexplored. To address this gap, this work conducts a systematic investigation of PTQ under MXFP formats, encompassing over 7 PTQ algorithms, 15 evaluation benchmarks, and 3 LLM families. The key findings include: 1) MXFP8 consistently achieves near-lossless performance, while MXFP4 introduces substantial accuracy degradation and remains challenging; 2) PTQ effectiveness under MXFP depends strongly on format compatibility, with some algorithmic paradigms being consistently more effective than others; 3) PTQ performance exhibits highly consistent trends across model families and modalities, in particular, quantization sensitivity is dominated by the language model rather than the vision encoder in multimodal LLMs; 4) The scaling factor of quantization is a critical error source in MXFP4, and a simple pre-scale optimization strategy can significantly mitigate its impact. Together, these results provide practical guidance on adapting existing PTQ methods to MXFP quantization.
1 Introduction
This paper studies how post-training quantization methods behave under MXFP, addressing limited evidence because existing PTQ research mainly targets integer formats. Across models and modalities, MXFP8 is near-lossless, while MXFP4 remains challenging and requires format-aware methods.
- The study evaluates PTQ under MXFP8 and MXFP4 across multiple method classes, quantization settings, benchmarks, LLM families, and multimodal models.
- MXFP8 provides near-lossless 8-bit weight-activation quantization, whereas MXFP4 causes non-negligible accuracy degradation and remains an open challenge.
- Error compensation and affine transformation methods are more compatible with MXFP, especially at low bit-widths, and combining them improves performance; RTN remains competitive.
- PTQ effectiveness remains consistent across model families and modalities, with multimodal sensitivity dominated by the language model rather than the vision encoder.
- Scaling-factor quantization introduces noticeable MXFP4 error, while pre-scale operation is a practically effective mitigation strategy.
2 Preliminary
This section introduces quantization and the MXFP numerical regime, then organizes PTQ methods by how they transform or compensate for weights and activations. MXFP uses block quantization with shared scaling, making format-specific behavior central to method design.
- 2.1 Low-Bit Integer (INT) and Floating-Point (FP) Quantization: Quantization maps high-precision tensors to lower-bit representations; integer quantization clips normalized values, while floating-point quantization selects the nearest representable value.
- 2.2 Microscaling Floating Point (MXFP) Quantization: MXFP uses block quantization with block size 32 and a shared UE8M0 scaling type; MXFP8 uses E4M3 or E5M2, while MXFP4 uses E2M1.
- The study categorizes PTQ into channel-wise transformation, error compensation, rotational transformation, and affine transformation, evaluating representative algorithms from each class.
- Channel-wise methods adjust activation and weight ranges, including SmoothQuant and AWQ, while error-compensation methods model quantization discrepancies using GPTQ or MR-GPTQ.
- Rotational methods use orthogonal transformations, whereas affine methods learn rescaling transformations to redistribute numerical magnitudes before quantization.
3 Evaluations
Across MXFP evaluations, 8-bit settings are generally lossless, while 4-bit settings expose substantial accuracy and method-compatibility challenges. PTQ trends remain consistent across models and modalities, with multimodal sensitivity concentrated in the LLM and scaling-factor error materially affecting MXFP4.
- 3.2 Performance on Different MXFP Quantization Settings (RQ1): 8-bit weight-activation MXFP quantization is generally lossless, whereas W4A4 causes severe degradation across LLMs and MLLMs.W4A4 recovery falls to 87.25%–96.79% for Llama3.1-8B-Instruct, 86.37%–95.28% for openPangu-Embedded-7B-V1.1, and 92.72%–97.36% for Qwen2.5-VL-7B.
- 3.3 Comparison of PTQ Methods (RQ2): Affine transformation is most robust under W4A4, with FlatQuant reaching 96.57% recovery and 8.03 perplexity on Llama-3.1-8B-Instruct.FlatQuant surpasses RTN, SpinQuant, and MR-GPTQ in the cited W4A4 comparison.
- 3.3 Comparison of PTQ Methods (RQ2): Error compensation and affine transformations are better aligned with MXFP, especially at low bit-widths, while RTN remains a strong baseline.Error compensation methods recover 97.03%–97.36% under W4A8 versus 96.25%–96.33% for channel-wise transformations on openPangu-Embedded-7B-V1.1.
- 3.4 Impact of Model Families and Modalities (RQ3): PTQ effectiveness is highly consistent across models and modalities, with an average pairwise Pearson correlation of 0.917 across performance curves.The comparison covers four models and multiple MXFP quantization settings.
- 3.4 Impact of Model Families and Modalities (RQ3): In multimodal LLMs, quantization sensitivity is dominated by the LLM rather than the ViT, favoring higher LLM precision and more aggressive ViT quantization.On Qwen2.5-VL-7B, W4A4 causes about 3% recovery loss in the LLM versus approximately 1% in the ViT; visual tokens are comparatively robust.
- 3.5 Quantization Components: Scaling-factor error is a significant MXFP4 failure source, and pre-scale optimization is recommended to reduce its impact.Using high-precision scales with low-precision values clearly reduces relative perplexity compared with low-precision values and scales.
4 Conclusion
This work identifies consistent PTQ behavior under MXFP across model families and modalities, while showing that MXFP4 scaling-factor quantization is a key error source mitigated by pre-scale optimization.
- PTQ performance exhibits highly aligned trends across model families and modalities, with quantization sensitivity in multimodal models dominated by the language component.This favors preserving higher precision in the language model when designing mixed-precision multimodal systems.
- MXFP should be treated as a distinct numerical regime rather than a drop-in replacement for integer formats.The findings provide practical guidance for developing MXFP-aware quantization methods.
- MXFP4 scaling-factor quantization is a key error source, and simple pre-scale optimization can effectively mitigate its impact.The study therefore treats scaling-factor handling as central to MXFP4 quantization design.
Limitation
The study evaluates 7B/8B-scale LLMs and MLLMs using MXFP formats, leaving the generality of its conclusions at larger scales and for alternative microscaling formats unresolved.
- The study covers 7B/8B-scale LLMs and MLLMs with MXFP, but does not evaluate substantially larger models or NVIDIA-specific NVFP formats.The untested alternatives include 30B-scale models and NVFP4/NVFP8.
- How well the conclusions extend to larger model scales or microscaling formats with different exponent and scaling designs remains unclear.The authors leave systematic investigation of these settings for future work.
A.1 Benchmarks and Evaluation Details
The evaluation spans non-reasoning, reasoning, and image-text benchmarks, using standard benchmark-specific protocols to assess language and multimodal model performance.
- Non-Reasoning Benchmarks: The benchmark suite covers non-reasoning tasks including PIQA, Winogrande, Hellaswag, ARC-Easy, and ARC-Challenge.These tasks are evaluated with lm-evaluationharness and the vllm backend.
- Reasoning Benchmarks: Reasoning evaluation includes MATH500 with Avg@1 and AIME24/AIME25 with Avg@16.The AIME scores average accuracy over 16 independent generations.
- Image-Text Benchmarks: Image-text evaluation uses OCRBench, MMBench, and TextVQA to assess optical character recognition and multimodal understanding.OCRBench includes text recognition, scene-text VQA, document-oriented VQA, key information extraction, and handwritten mathematical expression recognition.
A.2.1 Results on Reasoning Benchmarks
The supplied passages identify Table 7 as the source of detailed reasoning-benchmark results, but do not report the table’s individual comparisons or values.
- Detailed results on the reasoning benchmarks are provided in Table 7.
A.2.2 Results on openPangu-VL-7B
The section points readers to detailed openPangu-VL-7B evaluation results in Table 8.
- Table 8 provides detailed openPangu-VL-7B results for reference.
A.2.3 Results under More Quantization Scenarios
Across additional bit-width settings, 4-bit weight quantization is the primary performance bottleneck, while 8-bit activation or KV-cache quantization often causes smaller changes. The section also examines activation distributions under different quantization methods.
- Experimental scope: The evaluation extends PTQ comparisons across openPangu-Embedded-7B-V1.1 and Llama-3.1-8B-Instruct under additional W4A16 and W4A8KV8 configurations.The broader comparison is presented in Tables 9 and 10.
- Results on W4A16: 4-bit weight quantization is the primary bottleneck: most methods show noticeable performance drops under W4A16, even with 16-bit activations.QuaRot achieves a 95.27% recovery rate on openPangu-Embedded-7B-V1.1, below the risky-regime threshold of 97%.
- Results on W4A16: W4A8 usually causes only a slight additional decrease relative to W4A16, reinforcing that 4-bit weights alone are a significant challenge.SpinQuant declines from 95.10% to 94.46%, while FlatQuant improves from 96.87% to 97.12% on openPangu-Embedded-7B-V1.1.
- Results on W4A8KV8: Under W4A8KV8, RTN, SpinQuant, and FlatQuant maintain near-stable recovery rates compared with W4A8 on openPangu-Embedded-7B-V1.1.Their recovery rates change from 95.44%, 94.46%, and 97.12% to 95.14%, 94.33%, and 97.42%, respectively.
- Activation distributions: Figure 5 compares q_proj activation distributions in layer 8 of openPangu-Embedded-7B-V1.1 across quantization methods, emphasizing changes in dynamic range and distribution.The figure uses inference-time activations under various quantization methods.