Skip to content

Scaling Laws for Mixed Quantization

Conference: EMNLP 2026
Paper: ACL Anthology
Area: Model Compression
Keywords: mixed-precision quantization, scaling laws, block granularity, post-training quantization, deployment optimization

TL;DR

Presents the first unified scaling law for post-training mixed-precision quantization in large language models, modeling perplexity degradation as a separable multiplicative function of model parameter count, low-precision parameter ratio, and block size, establishing that larger models tolerate higher low-precision ratios while the absolute loss benefit of fine-grained block scaling diminishes with scale.

Background & Motivation

The rapid expansion of large language models to hundreds of billions of parameters has made efficient inference deployment increasingly challenging. Post-training quantization (PTQ) serves as the primary technique to curb memory footprint and memory-bandwidth bottlenecks. In practical production deployments, engineering workflows frequently adopt mixed-precision quantization schemes rather than uniform low-bit arithmetic. This is largely driven by the empirical reality that pre-trained Transformer weights and activations naturally develop extreme magnitude outliers; forcing all tensors uniformly into very low bit-widths triggers severe perplexity spikes. Consequently, practitioners keep outlier-heavy sensitive weights or critical layers at high precision while casting the vast majority of parameters to low precision. In parallel, modern hardware architectures—most notably NVIDIA's Blackwell platform with MXFP4 and MXFP6 formats—incorporate native support for block-wise microscaling, where groups of \(Q_b\) elements share a common exponent scale factor, delivering high dynamic range and silicon compute efficiency simultaneously.

However, while training-time compute-optimal scaling laws (e.g., Kaplan and Chinchilla) are thoroughly characterized, and recent work has initiated the study of precision scaling under uniform bit-widths, the interaction between mixed-precision PTQ, model scale, and hardware block granularity remains poorly understood. When deploying models of varying capacities, practitioners currently rely on ad-hoc, expensive trial-and-error searches across combinatorial layer assignments, lacking any analytical formulation to anticipate how much loss degradation a target compression budget will induce or how block size trade-offs shift as models grow.

To bridge this foundational gap on the inference and deployment side, this paper establishes a sample-efficient, extrapolatable analytical scaling framework for mixed-precision quantization. Core Idea: model post-training mixed-precision degradation through a separable multiplicative scaling law \(\delta(N, Q_r, Q_b) = C e^{A Q_r} N^{-\gamma_N} (Q_b + d)^{\gamma_c}\) parameterized by low-precision parameter ratio \(Q_r\), block granularity \(Q_b\), and model size \(N\), enabling zero-shot cross-scale extrapolation from small checkpoints to tens of billions of parameters.

Method

The proposed formulation provides a tractable continuous surrogate over the combinatorial space of mixed-precision configurations, establishing predictable power-law and exponential trends that connect micro-level quantization settings to macro-level task loss degradation.

Overall Architecture

The framework investigates autoregressive causal language models under post-training mixed-precision quantization. For a model with \(N\) parameters pre-trained on \(D\) tokens, mixed-precision quantization partitions parameter matrices into a low-precision subset \(W_l\) (e.g., 4-bit or 2-bit) and a high-precision subset \(W_h\) (e.g., BF16). The total evaluation loss is modeled as the standard Chinchilla power-law pre-training loss plus an additive quantization-induced degradation term \(\delta(N, Q_r, Q_b)\). Because the combinatorial number of layer or operator precision assignments satisfying a given global budget is astronomically large, the partition is treated as a random variable over the configuration space. The framework models both the lower envelope (the sample minimum \(\delta_{opt}\), representing achievable optimal performance) and the center of mass (the expectation \(\mathbb{E}[\delta]\), representing average performance). The end-to-end pipeline spans configuration space parameterization, random partition sampling, small-model parameter fitting, and cross-scale extrapolation.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Pre-trained LLM and Compression Budget<br/>Model size N, training tokens D, target hardware constraints"] --> B["Dual-Parameter Search Space & Partition Sampling<br/>Define low-precision ratio Qr and block size Qb, sample component allocations"]
    B --> C["Unified Scaling Law Construction<br/>Fit separable multiplicative weak and strong laws for optimal and expected loss"]
    C --> D["Sample-Efficient Calibration & Scale Extrapolation<br/>Calibrate power-law and exponential constants on small checkpoints (N <= 8B)"]
    D --> E["Output: Optimal PTQ Deployment Decisions<br/>Predict loss degradation and identify hardware block overhead tipping points"]

Key Designs

1. Dual-Parameter Search Space & Partition Sampling: Formalizing Low-Precision Ratio and Block Size To address the combinatorial intractability of heterogeneous bit allocation across deep networks, the method parameterizes the mixed-precision space using two orthogonal physical variables: the quantization ratio \(Q_r\) and the block granularity \(Q_b\). The quantization ratio represents the global budget of low-precision parameters: $\(Q_r(W_l, W_h) := \frac{\|W_l\|_0}{\|W_l\|_0 + \|W_h\|_0}\)$ while \(Q_b\) denotes the number of tensor values sharing a single scale factor (e.g., 16, 32, 64, 128, 256 in MX formats). Because exhaustive combinatorial optimization is infeasible, the authors evaluate configurations via uniform random sampling over layer-wise or matmul-wise assignments. Drawing 100 random partition allocations per setting yields sample mean \(\delta_\mu\) and sample minimum \(\delta_{min}\), which serve as consistent estimators for \(\mathbb{E}[\delta]\) and \(\delta_{opt}\). Crucially, cross-model Spearman rank correlations of layer-wise sensitivity were found to be modest (averaging only 0.41–0.60), demonstrating that fine-grained layer sensitivity does not reliably transfer across model scales. This validates maintaining \(Q_r\) as a compact, global scalar budget rather than introducing brittle per-layer variables.

2. Unified Scaling Law Construction: Separable Multiplicative Formulation Across Regimes To capture how quantization errors compound across cascaded layers, the paper models the degradation surface through a separable multiplicative hypothesis. When block granularity \(Q_b\) is held constant, the formulation obeys the Weak Scaling Law: $\(\delta_{opt}(N, Q_r) = C e^{A Q_r} N^{-\gamma_N}, \quad \mathbb{E}[\delta(N, Q_r)] = C' e^{A' Q_r} N^{-\gamma'_N}\)$ Here, \(e^{A Q_r}\) reflects the compounding exponential growth of quantization noise as more components are compressed, while \(N^{-\gamma_N}\) captures the capacity of larger parameter spaces to absorb perturbations. At \(Q_r = 0\), this smoothly reduces to the uniform Precision Scaling Law of Kumar et al. (2024). When explicitly modeling varying block sizes, the formulation expands into the Strong Scaling Law: $\(\delta_{opt}(N, Q_r, Q_b) = C e^{A Q_r} N^{-\gamma_N} (Q_b + d)^{\gamma_c}\)$ where \((Q_b + d)^{\gamma_c}\) acts as a quasi-polynomial term characterizing dynamic range truncation error. Because the terms are separable, fixing a target loss degradation budget \(\delta_{opt} = l\) yields a strictly linear log-scale trade-off: \(\log(N) = \frac{A}{\gamma_N} Q_r + \frac{\log(C) - \log(l)}{\gamma_N}\), providing direct mathematical proof that larger models accommodate substantially higher low-precision ratios under identical quality thresholds.

3. Sample-Efficient Calibration & Scale Extrapolation: Zero-Shot Transfer to Large Checkpoints Rather than chasing a universal predictor with family-agnostic constants, the law functions as a highly sample-efficient calibration tool. For any given model family (LLaMA, Qwen-1.5, Qwen-3) and quantization algorithm (MXINT, HQQ, GPTQ, Minifloat, MXFP), the functional form is invariant, requiring only a few small checkpoints (\(N \le 8\text{B}\), or custom pre-trained models between 60M and 1.1B) to fit the five constants (\(C, A, \gamma_N, d, \gamma_c\)). Once fitted, the law extrapolates zero-shot to 13B, 14B, and 32B checkpoints with high fidelity (\(R^2 \ge 0.96\)). Analytically, this yields two decisive deployment insights: first, for models exceeding 70B, \(Q_r\) can safely surpass 0.90 while keeping perplexity degradation below 0.5; second, for models beyond 50B, the absolute gap in loss degradation between fine (\(Q_b=32\)) and coarse (\(Q_b=128\)) blocks narrows to under 0.5, demonstrating that scale diminishes the marginal value of fine block granularity relative to hardware scale-factor memory overhead.

A Worked Example

Consider deploying a Qwen-3 32B model with matmul-wise MXINT-4 post-training quantization. From small-checkpoint calibration (\(N \le 8\text{B}\)), the weak scaling parameters are identified as \(C = 0.0409\), \(A = 6.0825\), and \(\gamma_N = 0.9994\). If the engineering requirement enforces an absolute perplexity degradation ceiling of \(l = 0.20\), we evaluate the linear trade-off relation: $\(\log(32) \approx 3.466 = \frac{6.0825}{0.9994} Q_r + \frac{\log(0.0409) - \log(0.20)}{0.9994}\)$ Solving for \(Q_r\) yields \(Q_r \approx 0.83\). The deployment team can immediately conclude that casting 83% of linear projections to INT4 and retaining 17% in BF16 is mathematically guaranteed to meet the 0.20 perplexity bound. This prediction eliminates the need for expensive combinatorial trials on the 32B model, executing in zero GPU time.

Loss & Training

The study operates strictly in the post-training quantization (PTQ) regime without post-quantization fine-tuning, isolating the direct degradation caused by discrete arithmetic rounding. To provide a rigorous baseline, the authors pre-trained the CLM model family across five parameter scales (60M, 200M, 400M, 600M, 1.1B) using TorchTitan on FineWeb-Edu, adhering strictly to Chinchilla-optimal token budgets (training tokens = 22 \(\times\) non-embedding parameters) with AdamW and linear warm-up. Evaluation is performed on 1,000 sequences of length 2,048 subsampled from SlimPajama, reporting cross-entropy perplexity degradation.

Key Experimental Results

Main Results

The main empirical evaluation spans 17 model scales from 60M to 32B parameters across more than 115,500 evaluated checkpoints, covering CLM, LLaMA, Qwen-1.5, and Qwen-3. The table below reports the goodness-of-fit for zero-shot scale extrapolation to 13B–32B models using coefficients fitted exclusively on checkpoints with \(N \le 8\text{B}\) (reproduced from Table 4 of the paper):

Granularity Target Model Model Size (\(N\)) \(\delta_{opt}\) Fit (\(R^2\)) \(\delta_{opt}\) Error (MSE) \(\mathbb{E}[\delta]\) Fit (\(R^2\)) \(\mathbb{E}[\delta]\) Error (MSE)
Layer-wise Qwen-1.5-14B 14B 0.9837 0.1662 0.9770 0.3785
Layer-wise Qwen-3-14B 14B 0.9786 0.2726 0.9783 0.4021
Layer-wise LLaMA-13B 13B 0.9646 0.7787 0.9944 0.1683
Layer-wise Qwen-3-32B 32B 0.9778 0.2596 0.9771 0.3908
Matmul-wise Qwen-3-14B 14B 0.9940 0.2138 0.9970 3.1558
Matmul-wise Qwen-1.5-32B 32B 0.9631 0.5597 0.9850 1.1031
Matmul-wise Qwen-3-32B 32B 0.9935 0.2050 0.9969 2.7820

Ablation Study

The ablation analysis examines varying numerical formats (MXINT-2, MXINT-4 weight-only, MXINT-4 weight-activation, HQQ-4, GPTQ, Minifloat, MXFP) and explores the parameter shifts between weak and strong scaling laws (compiled from Table 1 and Appendix F):

Setup & Numerical Format Target Stat Scaling Law Form Bit-width & Mechanism Size Exponent \(\gamma_N\) Ratio Exponent \(A\) Goodness-of-Fit (\(R^2\)) Key Characteristics
Qwen-3 MXINT-4 (W-only) \(\delta_{opt}\) Weak Law 4-bit weights, BF16 act. 1.1251 5.9150 0.9552 Smooth, stable parameter trajectory
Qwen-3 MXINT-4 (W-only) \(\delta_\mu\) Weak Law 4-bit weights, BF16 act. 1.9460 4.0875 0.9977 Exceptionally tight fit for mean degradation
Qwen-3 MXINT-4 (W-only) \(\delta_{opt}\) Strong Law 4-bit weights (\(Q_b \in [16, 256]\)) 0.8181 8.1566 0.7025 Granularity isolation re-allocates exponent weight
Qwen-3 MXINT-4 (W-only) \(\delta_\mu\) Strong Law 4-bit weights (\(Q_b \in [16, 256]\)) 1.5909 4.1025 0.8336 Granularity parameters \(\gamma_c=1.6736, d=-3.7594\)
CLM MXINT-4 (W-only) \(\delta_{opt}\) Weak Law 4-bit weights 0.6818 4.4940 0.9790 Canonical pre-trained model behavior
CLM HQQ 4-bit \(\delta_{opt}\) Weak Law 4-bit calibration-free 0.6471 2.1631 0.9668 Low \(A\) coefficient reflects high resilience
CLM MXINT-2 (W+A) \(\delta_{opt}\) Weak Law 2-bit weights & activations 0.0992 9.1365 0.9707 Extreme compression; \(\gamma_N \to 0\) limits scaling

Key Findings

  • Parameter scaling outpaces quantization penalty at scale: in the weak law, the shrinkage rate of \(N^{-\gamma_N}\) with increasing model size consistently overcomes the exponential expansion of \(e^{A Q_r}\). In models exceeding 70B parameters, the low-precision fraction \(Q_r\) can safely exceed 0.90 while maintaining degradation \(\delta < 0.5\).
  • Diminishing returns of fine block granularity: while finer block sizes \(Q_b\) mitigate dynamic range clipping, the absolute difference in loss degradation between \(Q_b = 32\) and \(Q_b = 128\) drops below 0.5 for models larger than 50B parameters. Larger parameter counts provide intrinsic noise averaging, making aggressive block scaling overhead unnecessary.
  • Boundary conditions under extreme compression and distillation: under severe 2-bit weight-activation compression (MXINT-2), loss degradation surges dramatically, leading to parameter instability (e.g., negative \(\gamma_N = -4.62\) in Qwen-3). Furthermore, distilled models occasionally exhibit negative degradation (slight perplexity gains post-quantization due to regularization effects), which violates the non-negativity premise and signals the boundary of the positive-scaling regime.

Highlights & Insights

  • Analytical formulation of mixed-precision deployment: cleanly condenses complex combinatorial bit assignments into a separable multiplicative equation over \(Q_r\), \(Q_b\), and \(N\), unifying post-training mixed-precision behavior with classic precision scaling laws.
  • Sample-efficient cross-scale extrapolation: dispels the notion that large-scale mixed-precision profiles require brute-force search on massive clusters; calibrating constants on \(\le 8\text{B}\) checkpoints accurately predicts 13B–32B behavior (\(R^2 \ge 0.96\)), saving thousands of GPU hours.
  • Actionable blueprint for inference silicon: provides theoretical justification for modern AI hardware trends (such as Blackwell MXFP formats), proving that hardware architectures should prioritize wider execution bandwidth for low-precision tensor operations rather than over-engineering fine-grained block scaling metadata pipelines.

Limitations & Future Work

  • Restricted to post-training quantization: the formulation models un-finetuned PTQ degradation and does not cover quantization-aware training (QAT) or post-quantization parameter recovery techniques like LoRA.
  • Instability at extreme low-bit limits: for sub-3-bit regimes (e.g., MXINT-2 weight-and-activation), severe non-linear degradation and outlier explosion violate the log-linear error accumulation assumption.
  • Sole reliance on language modeling perplexity: while perplexity tracks base foundation model health accurately, downstream capabilities involving complex multi-step reasoning, mathematical proof, or long-context retrieval may experience different degradation dynamics not fully captured by pre-training perplexity.
  • vs Kumar et al. (2024) (Precision Scaling Laws): Kumar et al. modeled uniform-precision post-training degradation across bit-widths but lacked any formulation for mixed-precision allocations or block granularity. The proposed law directly recovers Kumar's precision scaling law as a special case when \(Q_r = 0\), while expanding it to model heterogeneous bit ratios and hardware block sizes.
  • vs Frantar et al. (2025) (Compression Scaling Laws): Frantar et al. examined the joint scaling behavior of unstructured sparsity and quantization within a QAT training framework. In contrast, this paper tackles zero-retraining post-training quantization (PTQ) and resolves the combinatorial assignment space using random partition statistics.
  • vs Heuristic Mixed-Precision Quantizers (AWQ, SpQR, LLM-MQ): Prior mixed-precision quantizers rely on first-order or second-order gradient heuristics to identify salient layers individually. This work proves that per-layer rankings transfer poorly across scales (Spearman correlation \(\approx 0.5\)), whereas the macro-level quantization ratio \(Q_r\) adheres to a rigorous scaling law that provides analytical budget constraints for search algorithms.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Formulates the first unified, continuous scaling law for post-training mixed-precision quantization and block granularity.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Validated over 17 model scales, 115,500+ checkpoints, 6 arithmetic formats, and verified via 13B–32B zero-shot extrapolation.
  • Writing Quality: ⭐⭐⭐⭐⭐ Highly rigorous and structured exposition, moving from combinatorial formulation to sampling statistics, scaling laws, and engineering guidelines.
  • Value: ⭐⭐⭐⭐⭐ Delivers critical theoretical foundation and practical recipes for LLM inference serving, heterogeneous scheduling, and next-generation AI hardware design.