Taming the 8-Billion Parameter Model on a low-cost GPU card for Retrosynthesis
5 affiliations
- 1Faculty of Chemistry
- 2University of Strasbourg
- 3Molecular Targets and Therapeutics Center
- 4Helmholtz Zentrum München
- 5BIGCHEM GmbH
Abstract
We explored the use of post-training quantization (PTQ) to enhance the efficiency of reasoning language models for chemical retrosynthesis in drug discovery. We specifically evaluated the RetroDFM-R model, demonstrating that reducing numerical precision to 4-bit significantly lowered GPU memory requirements while lowering Top-1 accuracy by only about 2 percentage points (3–4% relative) for NormalFloat-4 (nf4) and double-quantized nf4 (nf4dq). The accuracy for 8-bit quantization (int8_ao) was equivalent to the original model but reduced energy consumption by about 17% when only one prediction was required. By testing various quantization methods across different hardware configurations, the study revealed that these techniques allow complex chemical predictions to run on budget-friendly, consumer-grade GPUs. In the augmented setting the 4-bit models returned more distinct candidates, which raised Top-10 accuracy on the full test set by up to 0.8 percentage points; on reactions for which every model filled ten candidate positions, the original model remained the most accurate. In summary this study presents the first evaluation of PTQ applied to RetroDFM-R, a reasoning language model, for chemical retrosynthesis. Ultimately, the research suggests that quantization is a vital tool for making advanced chemoinformatics tools accessible to academic researchers who can run it on cards with smaller GPU memory and can reduce the energy per molecule by about 17 - 30% in the default setting, although it increases it by 30–35% in the augmented setting.
Related
- RetroAgent: Harnessing LLMs to Search Over Structured Memory for Agentic Retrosynthesis Planning
- Chemist-aligned retrosynthesis by ensembling diverse inductive bias models
- RetroMol: Knowledge Distillation of Energy-Based Retrosynthesis Models
- URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment
- Taming nitrogen stereogenicity: enantioselectivity control at a reluctant center.
BibTeX
@misc{hussein2026taming,
title = {{Taming the 8-Billion Parameter Model on a low-cost GPU card for Retrosynthesis}},
author = {Abdulwahab Hussein and Igor V. Tetko},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010933/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010933/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Taming the 8-Billion Parameter Model on a low-cost GPU card for Retrosynthesis
AU - Abdulwahab Hussein
AU - Igor V. Tetko
PY - 2026
DA - 2026/10/23
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010933/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010933/v1
AB - We explored the use of post-training quantization (PTQ) to enhance the efficiency of reasoning language models for chemical retrosynthesis in drug discovery. We specifically evaluated the RetroDFM-R model, demonstrating that reducing numerical precision to 4-bit significantly lowered GPU memory requirements while lowering Top-1 accuracy by only about 2 percentage points (3–4% relative) for NormalFloat-4 (nf4) and double-quantized nf4 (nf4dq). The accuracy for 8-bit quantization (int8_ao) was equivalent to the original model but reduced energy consumption by about 17% when only one prediction was required. By testing various quantization methods across different hardware configurations, the study revealed that these techniques allow complex chemical predictions to run on budget-friendly, consumer-grade GPUs. In the augmented setting the 4-bit models returned more distinct candidates, which raised Top-10 accuracy on the full test set by up to 0.8 percentage points; on reactions for which every model filled ten candidate positions, the original model remained the most accurate. In summary this study presents the first evaluation of PTQ applied to RetroDFM-R, a reasoning language model, for chemical retrosynthesis. Ultimately, the research suggests that quantization is a vital tool for making advanced chemoinformatics tools accessible to academic researchers who can run it on cards with smaller GPU memory and can reduce the energy per molecule by about 17 - 30% in the default setting, although it increases it by 30–35% in the augmented setting.
ER -