Analysis of Quantized and Efficiently Adapted Protein Language Models
Authors: Ilan Yaniv Zeisler, Sebastian Clancy, Pouriya Bayat, Saaim Raad, Ivan Kraskov, Matthew Xie, Vivian White, Spencer Perkins, +3 more
Organizations: Leslie Dan Faculty of Pharmacy, University of Toronto, Toronto, ON, Canada. · Department of Electrical and Computer Engineering, University of Toronto, Toronto, Ontario, Canada. · Department of Computer Science, University of Toronto, Toronto, Ontario, Canada. · Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, Ontario, Canada. · Department of Computing and Software, McMaster University, Hamilton, ON, Canada.
Background: Protein language models (PLMs) are increasingly used for sequence generation and property prediction, but their size makes fine-tuning and deployment expensive. The effects of quantization and parameter efficient fine-tuning on performance, representations and generation remain insufficiently characterized. Results: We evaluated 4-bit quantization and low-rank adapter fine-tuning (QLoRA) across ESM-2, ESMC, ProtBERT, ProtT5, Ankh, Ankh3 and Profluent-E1. Across protein prediction tasks, many model-task pairs retained more than 90% of full fine-tuning performance. Peak GPU memory savings approached 90% for the largest models, although performance and efficiency varied by model, dataset and training configuration. QLoRA often preserved early-layer representations while inducing task-specific adaptations in middle and late layers, resembling full fine-tuning with smaller representational changes. Training speed and power effects were more varied. For unconditional generation with ProLLaMA, ProtGPT2, ProGen2, ProteinGLM and ESM3, 4-bit quantization largely preserved predicted structural and sequence-level properties, but token-level analysis revealed model-dependent shifts in autoregressive output distributions. Conclusion: QLoRA and 4-bit quantization reduce PLM computational requirements, particularly GPU memory usage. Our results support QLoRA as a first-pass strategy for memory limited adaptation, reserving full fine-tuning for challenging tasks, unstable architectures or low validation recovery. For generative PLMs, sequence-level and structural metrics should be complemented with distributional analysis, since downstream predictions alone may miss quantization-induced shifts. These approaches can broaden access to large-scale protein modelling while requiring model- and task-specific validation.
Figures & tables
Figure 1: Efficient use of pretrained protein language models for both property prediction and protein generation. a We evaluated PLMs using two different tasks. For fine-tuning (adaptation) to a specific dataset (left), the pretrained PLM backbone is quantized to 4-bit precision and augmented with trainable LoRA adapters, producing a QLoRA model. For protein generation (right), the PLM is quantized without added LoRA adapters. b To better understand the effects of quantization, we looked at several metrics for success, including performance (accuracy), training memory, power consumption, latency, and variability. Model performance was maintained, while memory and power improvements were observed. c The suggested practitioner workflow derived from benchmarking results. QLoRA is used as a first-pass strategy for memory-constrained supervised adaptation, with fallback to full fine-tuning for difficult tasks, unstable architectures or low-performance recovery. For quantized generation, aggregate sequence and structural metrics should be supplemented with distributional diagnostics to evaluate downstream use cases, falling back to the full precision model in some cases.
Figure 2: QLoRA fine-tuning performance ratios across models and tasks. Each value shows the relative performance (full model score / QLoRA model score) in large text, with standard deviation across three replicates ( n=3 seeds) shown below.
Figure 3: Relative QLoRA/full fine-tuning performance of ESM-2 models for different tasks by model size. Ratios are geometric means with 95% confidence interval error bars across three replicates ( n=3 seeds).
Figure 4: Layer-wise representational similarity between QLoRA and full fine-tuning, with higher mean CKA indicating higher similarity to the full model. a Mean CKA similarity averaged across datasets and three paired seeds for each model as a function of normalized layer depth. b Dataset-specific CKA profiles for ESM-2 8M and 3B, averaged across three paired seeds.
Figure 5: kNN task probe data averaged across six datasets. a Layer-wise task information by kNN probe scores across model depth for full fine-tuned models, averaged across three seeds. b Difference in probe scores between QLoRA and full fine-tuned models, averaged across three seed pairs. Red indicates QLoRA has lower task probe score, blue indicates full fine-tuning has lower task probe score.
Figure 6: Predicted properties of unconditionally generated sequences from full and quantized models. Profiles compare structural quality, structural similarity, secondary structure content and disorder for each model. Full lines indicate original models, dotted indicate quantized models.
Figure 7: Teacher-forced KL divergence between full and quantized autoregressive models. a Mean token-level divergence as a function of normalized sequence position and b full-model reference entropy. c Joint count density of tokens over per-token KL divergence and full-model reference entropy, showing where quantization-induced discrepancies are concentrated relative to the reference model’s confidence levels. Per-token KL is pooled over three generation seeds and three length bins per model. Lines: bin-wise mean ± 95% bootstrap CI (1,000 resamples); bins with < 10 tokens omitted. Hexbins: all tokens, shared log-count scale, axes clipped at the 99.5th percentile.
Figure 8: Peak training memory allocation across models and batch sizes. Values show percent differences for QLoRA relative to full-fine-tuning, with negative values indicating reduced memory usage for QLoRA. Standard deviation across three replicates is listed in smaller text.
Figure 9: Training throughput and power draw across models and batch sizes. a Percent difference in average training steps per second and b percent difference in average GPU power draw for QLoRA relative to full fine-tuning. All values include standard deviation across three replicates.
Figure 10: Peak memory, inference latency and average power consumption for unconditional generation.
Use case
Recommended starting point
Validation check
Signals for caution
Supervised protein property prediction
Start with QLoRA (especially when GPU memory is limiting)
Compare QLoRA against full fine-tuning on representative validation subset (ratio and uncertainty)
QLoRA recovery below practical threshold, typically < 90-95%*, large uncertainty introduced or latency degrades
Low signal or high variance tasks
Use QLoRA as initial screen, not as a default
Run multiple seeds and compare absolute performance and variance, not just QLoRA/full ratio
Baseline full fine-tuning performance is weak/highly variable
Hyperparameter sensitive models
Perform a small learning rate and batch size sweep before comparison
Check convergence stability across seeds
Results are unstable across random seeds, learning rates or batch sizes
Unconditional generation
Use 4-bit quantization for exploratory memory efficient generation
Compare sequence-level and structural metrics between full and quantized models
Avoid relying only on QLoRA/full performance ratio
Appendix
Table S1: Practical validation and reporting checklist for quantized PLM workflows.
Figure S1: Full fine-tuning and QLoRA performance across downstream tasks. Bars show absolute task performance, and points show QLoRA/full fine-tuning ratios with 95% confidence intervals.
Figure S2: Layer-wise representational similarity to the pretrained base model with a full fine-tuned models and b QLoRA fine-tuned models.
Figure S3: Layer-wise task information in pretrained and ESM-2 fine-tuned representations. a Mean kNN probe scores across normalized depth for pretrained base models and QLoRA fine-tuned models. b Dataset-specific kNN probe scores across normalized depth for ESM-2 8M and 3B after full fine-tuning and QLoRA fine-tuning.
Figure S4: Predicted properties of unconditionally generated sequences from full and quantized models stratified by sequence length bins. Bars compare full and quantized models across Foldseek, Boltz-2, secondary structure and disorder metrics.
Figure S5: Additional teacher-forced KL divergence analyses. a Cumulative distribution of per-token KL divergence for each autoregressive model. Dashed lines indicate 95% and 99% of accumulated tokens. b Mean per-token divergence by model and sequence-length bin. c Mean per-token divergence across normalized sequence position, stratified by model and sequence-length bin.
Fluorescence
Stability
Secondary Structure
Full
QLoRA
Full
QLoRA
Full
QLoRA
Model
BS
LR
BS
LR
BS
LR
BS
LR
BS
LR
BS
LR
ESM-2 8M
32
5E-05
32
5E-05
32
5E-05
32
5E-05
32
5E-05
32
5E-05
ESM-2 35M
32
5E-05
32
5E-05
32
5E-05
32
5E-05
32
5E-05
32
5E-05
ESM-2 150M
32
5E-05
32
5E-05
32
5E-05
32
5E-05
32
5E-06
32
5E-05
ESM-2 650M
16
5E-05
16
5E-05
16
5E-05
16
5E-05
16
5E-06
16
5E-05
Appendix
Table S2: Batch size and learning rate for full and QLoRA models fine-tuned on fluorescence, stability and secondary structure datasets. LoRA parameters were fixed with r=8 , α=32 , dropout=0.05 . Runs were completed on NVIDIA GeForce RTX 4090s, RTX 5090s and a DGX Spark (GB10 Grace Blackwell) depending on model size and availability.
Post-Training Quantization (PTQ) and Low-Rank Adaptation (LoRA) constitute the standard pipeline for efficient Large Language Model (LLM) deployment. However, applying them sequentially poses a problem: PTQ often leaves behind random noise that is spread out (across the model's weights) in a way LoRA can't easily fix, meaning that LoRA ends up wasting its limited capacity trying to fix uncorrectable noise instead of improving task performance. In this paper, we propose \textbf{ProjQ}, a novel framework for constraining quantization noise to the low-rank manifold via orthogonal subspace projection. We derive an efficient alternating algorithm that shapes the quantization noise into a low-rank structure, effectively offloading dominant error components to the subsequent adapter while minimizing the residual error in the orthogonal "uncorrectable" subspace. Our theoretical analysis demonstrates that ProjQ preserves strictly greater model plasticity for downstream tasks compared to standard PTQ. Extensive experiments on LLaMA-2, Qwen2.5 and Qwen3 confirm that ProjQ consistently outperforms existing methods in both quantization error compensation and downstream task fine-tuning, achieving up to 2× lower evaluation loss for compensation and matching the performance of standard 4-bit baselines on language modeling tasks with only 3 bits. The code is available on https://github.com/yy9301/ProjQ .
Wenya Yu, Chao Zhang, Li Wang +2
School of Mathematics and Statistics, Central South University, Changsha, China · Department of Computer Science, Central South University, Changsha, China · CRAN, Université de Lorraine, CNRS, Nancy, France +1
Looped language models provide a parameter-efficient way to scale iterative test-time computation by repeatedly executing a shared recurrent core. Post-training quantization (PTQ) can reduce the memory footprint and inference cost of looped language models, but errors introduced by a quantized shared core affect subsequent cores. Among PTQ methods, channel scaling and orthogonal rotations preserve the floating-point computation while producing representations with different quantization quality. We find that quantization configuration candidate rankings can change with recurrent depth, motivating configuration selection at the target deployment depth. However, evaluating every candidate over the full calibration set at this depth is costly. We therefore propose Loopy, a PTQ framework that formulates shared-core quantization through a recurrent-depth-aware objective, selecting shared low-bit representations by their final prediction loss at the target deployment depth. Channel scaling and orthogonal rotations parameterize the candidate representations. To approximately solve this selection problem efficiently, Loopy progressively allocates calibration windows to promising candidates while preserving complete target-depth execution, using only forward evaluations. Across eight settings, Loopy achieves the state-of-the-art results among different baselines. On Ouro-1.4B under W4A4, Loopy reduces LAMBADA perplexity by 36.5% relative to SpinQuant. Our code is available at https://github.com/Shameless0817/Loopy-review.git.
Zeyu LI, Yipu ZHANG, Jintao Chen +2
The Hong Kong University of Science and Technology · Duke Kunshan University
As increasingly large pre-trained models are released, deploying them on edge devices for privacy-preserving applications requires effective compression. Recent works combine quantization with the fine-tuning of high-precision LoRA adapters, which can substantially reduce model size while mitigating the accuracy loss from quantization. However, edge devices have inherently heterogeneous capabilities, while performing configuration-wise fine-tuning for every quantization setting is computationally prohibitive. In this paper, we propose CoA-LoRA, a method that dynamically adjusts the LoRA adapter to arbitrary quantization configurations (i.e., the per-layer bit-width choices of a pre-trained model) without requiring repeated fine-tuning. This is accomplished via a configuration-aware model that maps each configuration to its low-rank adjustments. The effectiveness of this model critically depends on the training configuration set, a collection of configurations chosen to cover different total bit-width budgets. However, constructing a high-quality configuration set is non-trivial. We therefore design a Pareto-based configuration search that iteratively optimizes the training configuration set, yielding more precise low-rank adjustments. Our experiments demonstrate that, unlike the state-of-the-art methods that require fine-tuning a separate LoRA adapter for each configuration, CoA-LoRA incurs no additional time cost while achieving comparable or even superior performance to those methods.
Rongguang Ye, Ming Tang, Edith C. H. Ngai
Southern University of Science and Technology · The University of Hong Kong