Post-Training Quantization of Autoregressive Weather Models

Ananyo Bhattacharya; Swastik Bhattacharya; Christiane Jablonowski

cs.LGOct 1, 2026

Post-Training Quantization of Autoregressive Weather Models

Authors: Ananyo Bhattacharya, Swastik Bhattacharya, Christiane Jablonowski

Organizations: University of Michigan, Ann Arbor, USA · University of Texas at Dallas, Dallas, USA

Abstract

Advancements in high-resolution numerical weather prediction (NWP) and data assimilation (DA) have shaped the developments in deep learning (DL) architectures emulating atmospheric dynamics. Emulators for weather forecasting exhibit forecast quality comparable to physics based models at forecast horizon scaling from few days to subseasonal time scales. The emulators are driven by hardware-accelerated matrix multiplication in autoregressive inferences, significantly reducing the computation time and resources required for NWP. Optimization of the matrix multiplication processes in GPU architectures provides opportunities to scale towards high-resolution domain, and offers implementation of out of the box solutions. Post-training quantization (PTQ) has been demonstrated across multiple DL architectures to accelerate and increase the number of computations in unit time while consuming less power, enabling applications on edge hardware. In this study, we investigate the effect of PTQ on pre-trained AI emulators for global-scale weather forecasting. We implement PTQ algorithms in Deep Learning Weather Prediction (DLWP) and FourCastNet (FCN) models as a proof of concept for geophysical fluid dynamics applications. We systematically investigate the effect of PTQ on emulator inferences over short-range forecast horizons. Evaluation of PTQ configurations using simulated quantization hints at qualitatively meaningful forecasts over short-time horizons. These results provide a first benchmark of PTQ for autoregressive weather emulators and a basis for quantization-based optimization of DL models for dynamical systems.

Figures & tables

Figure 1: Illustration of the PTQ implementation and model evaluation in DLWP and FCN models using NVIDIA Earth2Studio and Model Optimizer (ModelOpt). Earth2Studio calls pre-trained checkpoints for each model available on Hugging Face. The full precision models undergo weight and activation quantization based on inference based calibration. ECMWF ERA5 serve as the initial conditions to drive emulator inferences and evaluate their performance as a function of forecast hour. All PTQ experiments are executed using the configuration settings provided by ModelOpt.

Model Name	Architecture	Input Type	Output Variables
DLWP	U-Net (cubed sphere)	ECMWF ERA5	$z_{300}$ , $z_{500}$ , $z_{700}$ , $z_{1000}$ , $t_{850}$ , $t_{2m}$ , $tcwv$
FCN	AFNO (ViT backbone)	ECMWF ERA5	$z_{500}$ , $z_{1000}$ , $t_{850}$ , $t_{2m}$ , $u_{10m}$ , $v_{10m}$ , $tcwv$

Table 1: A summary of the deep learning atmospheric dynamics emulators, the back-end architecture, input data type and output variables considered for evaluation in this study.

Name	Weight	Activation	Notes
FP32	FP32	FP32	Full precision baseline
W8A8	INT8	INT8	Tensor weight scaling-based quantization
W8A32	INT8	FP32	INT8 weights with FP32 activations
W4A32	INT4	FP32	INT4 weights with FP32 activations
W2A32	INT2	FP32	INT2 weights with FP32 activations
INT8 SmoothQuant	INT8	INT8	Activation smoothing using SmoothQuant

Table 2: A list of the PTQ methods described with respective weight precision, activation precision and notes for the quantization algorithm applied to pre-trained checkpoints on full-precision based DLWP and FCN emulators

Figure 2: RMSE forecast error as a function of lead time for DLWP inference runs relative to ERA5. The growth of error in PTQ configurations are compared against the full precision (black dotted) baseline in the above panels: z500 ( $\mathrm{m^{2}\,s^{-2}}$ ), z300 ( $\mathrm{m^{2}\,s^{-2}}$ ), z700 ( $\mathrm{m^{2}\,s^{-2}}$ ), z1000 ( $\mathrm{m^{2}\,s^{-2}}$ ), t850 (K), t2m (K) and tcwv ( $\mathrm{kg\,m^{-2}}$ ). RMSE errors also highlight the variations within a lead time for selected initial condition representing seasonal variability in forecast error.

Figure 3: ACC forecast skill score as a function of forecast lead time in DLWP inference runs relative to ERA5 as truth and WeatherBench2 climatology. Change in ACC skill highlights the quality of PTQ experiments relative to the FP32 baseline.

Figure 4: RMSE forecast error as a function of lead time for FCN inference runs relative to ERA5. The growth of error in PTQ configurations are compared against the full precision (black dotted) baseline in the above panels: z500 ( $\mathrm{m^{2}\,s^{-2}}$ ), z1000 ( $\mathrm{m^{2}\,s^{-2}}$ ), t850 (K), t2m (K), tcwv ( $\mathrm{kg\,m^{-2}}$ ), u10m ( $\mathrm{m\,s^{-1}}$ ) and v10m ( $\mathrm{m\,s^{-1}}$ ). RMSE errors also highlight the variations within a lead time for selected initial condition representing seasonal variability in forecast error.

Figure 5: ACC forecast skill score as a function of forecast lead time in FCN inference runs relative to ERA5 as truth and WeatherBench2 climatology. Change in ACC skill highlights the quality of PTQ experiments relative to the FP32 baseline.

Figure 6: Two-dimensional spatial maps for 2 m Temperature at the end of 60-hour inference runs for the DLWP model calibrated to different PTQ configurations. ERA5 is the reference truth, and FP 32 is the reference for precision for DLWP capability.

Figure 7: Two-dimensional spatial maps for 2 m Temperature at the end of 60-hour inference runs for the FCN model calibrated to different PTQ configurations. ERA5 is the reference truth, and FP32 is the reference for precision for FCN capability.

Figure 8: Mean power consumption (W) of model inferences on V100 and A100 GPU nodes for different PTQ configurations implemented in (a) DLWP and (b) FCN models through the NVIDIA Earth2Studio package

Figure 9: Total inference time per forecast (s) on V100 and A100 GPU nodes for different PTQ configurations implemented in (a) DLWP and (b) FCN models through NVIDIA Earth2Studio package

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10: RMSE error relative to FP32 (black dotted) inference as a function of forecast lead time for DLWP model configured to PTQ experiments. Growth in forecast error relative to FP32 highlights the degradation of forecast skill for different cases of weight and activation quantization.

Figure 11: RMSE error relative to FP32 (black dotted) inference as a function of forecast lead time for FCN model configured to PTQ experiments. Growth in forecast error relative to FP32 highlights the degradation of forecast skill for different cases of weight and activation quantization.

Figure 12: Two-dimensional spatial maps for total column water vapor at the end of 60-hour inference runs for the DLWP model calibrated to different PTQ configurations. ERA5 is the reference truth, and FP 32 is the reference for precision for DLWP capability.

Figure 13: Two-dimensional spatial maps for total column water vapor at the end of 60-hour inference runs for the FCN model calibrated to different PTQ configurations. ERA5 is the reference truth, and FP32 is the reference for precision for FCN capability.

Post-Training Quantization

Post-Training Quantization compresses trained neural networks into low-bit weights, activations, or caches to reduce memory use and inference cost.

See all 369 papers

Jul 30, 2026

Weather Emulators at the Frontier of Heat Extremes Predictability

Atmospheric predictability declines rapidly beyond the next ten days, such that forecasts at longer lead times primarily convey large-scale trends rather than specific states. Yet in a warming world, improving early warnings of extreme heat is an increasingly critical challenge. Here we evaluate six state-of-the-art deep learning weather emulators - Pangu-Weather, FuXi, ArchesWeather, AIFS, GraphCast and Aurora - alongside leading dynamical systems and statistical baselines in forecasting global near-surface temperature and extreme heat at lead times of 10-15 days. We find that several emulators rival or even surpass physics-based forecasts in deterministic temperature skill, but do so at the cost of reduced spectral fidelity, in a process widely known as blurring. While all models show some degree of predictive skill for extreme heat, most emulators under-represent peak intensities, and IFS recall is greater than that of any of the emulators. These results highlight both the emerging potential of AI to enhance extended range temperature prediction, and the remaining challenges in delivering reliable, actionable early warnings in a changing climate.

Cas Decancq, Thomas Mortier, Jessica Keune +1

1Hydro-Climate Extremes Lab (H-CEL), Department of Environment, Ghent University, Ghent, Belgium. · 2European Centre for Medium-Range Weather Forecasts (ECMWF), Reading, United Kingdom.

Weather Forecasting Extreme Weather Forecasting

Ask PDF

Oct 23, 2026

MPQ: A Message-Passing View of Post-Training Quantization

Post-training quantization (PTQ) is a key technique for efficiently deploying large language models, as it compresses weights into low-bit representations without retraining. Existing PTQ solvers, including GPTQ, LDLQ, and block coordinate descent, commit each weight block to a single codeword before its neighbors are considered, leaving near-optimal alternatives unexplored. In this work, we cast row-wise PTQ as posterior inference over discrete codebooks, in which existing solvers correspond to a zero-temperature, hard-decision limit retaining only the top codeword per block. We propose Message-Passing Quantization (MPQ), which preserves soft posterior information across blocks during inference and commits to hard codewords only at the output stage. MPQ first uses vector approximate message passing to compute approximate blockwise posteriors, then refines a hard-decision initialization by exact block coordinate descent. Across 2-bit lattice and 3- and 4-bit scalar quantization on dense language models from Llama, Qwen, and Mistral families, MPQ improves perplexity and zero-shot accuracy over strong hard-decision baselines, and composes favorably with preprocessing techniques such as activation-aware scaling and incoherence processing. Our implementation is available at https://github.com/poseidonchan/MPQ.

Yanshuo Chen, Chenxi Liu, Reza Shirkavand +1

Department of Computer Science, University of Maryland, College Park, MD 20742

Quantization LLM Quantization Post-Training Quantization Low-Bit Quantization

Ask PDF

Mar 20, 2026

DAQ: Delta-Aware Quantization for Post-Training LLM Weight Compression

We introduce Delta-Aware Quantization (DAQ), a data-free post-training quantization framework that preserves the knowledge acquired during post-training. Standard quantization objectives minimize reconstruction error but are agnostic to the base model, allowing quantization noise to disproportionately corrupt the small-magnitude parameter deltas (

ΔW

) that encode post-training behavior -- an effect we analyze through the lens of quantization as implicit regularization. DAQ replaces reconstruction-based objectives with two delta-aware metrics -- Sign Preservation Rate and Cosine Similarity -- that directly optimize for directional fidelity of

ΔW

, requiring only the base and post-trained weight matrices. In a pilot FP8 study, DAQ recovers style-specific capabilities lost under standard quantization while maintaining general performance.

Xiaoming Yu, Shize Tang, Guanghua Yu +4

Yuanbao & Hunyuan AI Infra Team

Quantization LLM Quantization Neural Network Quantization Post-Training Quantization Weight-Only Quantization

Ask PDF

Abstract

Post-Training Quantization

Explore similar work

Weather Emulators at the Frontier of Heat Extremes Predictability

MPQ: A Message-Passing View of Post-Training Quantization

DAQ: Delta-Aware Quantization for Post-Training LLM Weight Compression