ViRel-Budget: Reliability-Constrained Visual-Token Budgeting for Green Vision-Language Inference
Year: 2026
Type: software
Abstract: ViRel-Budget: Camera-Ready Artifact Release This release accompanies the camera-ready version of ViRel-Budget: Reliability-Constrained Visual-Token Budgeting for Green Vision-Language Inference, accepted for an oral presentation at the ACM Multimedia 2026 GreenMM Workshop. ViRel-Budget evaluates how aggressively vision-language models can reduce visual-token processing while preserving task correctness, dense-answer agreement, and intervention-defined response behavior. The release separates prospective reliability evaluation from repeated-workload energy measurement and reports claims within the evaluated models, pruning implementations, workloads, and hardware. Release contents This release provides: the installable ViRel-Budget framework and backend adapters; prespecified development and prospective evaluation configurations; three group-disjoint 210-query workload draws for robustness analysis; prospective controller decisions and strict-label tables; compact raw repeated-workload records and NVIDIA telemetry; carbon, Pareto-frontier, controller, integrity, and feature-analysis summaries; deterministic Random-pruning support for the official FastV implementation; pinned FastV, SCOPE, and LLaVA-PruMerge source revisions; dataset manifests, provenance records, seeds, and checksums; reproducibility, measurement, methodology, and limitation documentation; and machine-readable citation metadata. Evaluation scope The released evaluation materials cover: 1,200 queries used for controller development; 900 label-hidden queries used for prospective reliability evaluation; and three group-disjoint 210-query workload draws used for repeated-workload robustness and energy analysis. The repeated workload draws are evaluations within the same study and are not presented as independent replications. Main empirical finding Across the separate 630-query repeated-workload population, SCOPE reduced measured call-window GPU energy relative to matched dense inference at both evaluated model scales: LLaVA-1.5-7B: 11.87% lower GPU energy, 95% CI [7.19%, 16.30%]; LLaVA-1.5-13B: 10.62% lower GPU energy, 95% CI [7.48%, 13.46%]. FastV and the deterministic Random baseline did not demonstrate consistent measured-energy savings on the evaluated implementation and hardware. These findings distinguish visual-token reduction and dense avoidance from realized device-energy reduction. Camera-ready corrections This release incorporates the camera-ready integrity and reproducibility revisions from PR #1, including: corrected repository, dataset, citation, author, and package metadata; clearer separation of prospective reliability and energy evidence; replacement of overstrong deployment and environmental terminology; accurate repeated-workload terminology in place of independent-replication language; documentation of dataset provenance and source-group isolation; explicit treatment and reporting of zero-eligible-intervention cases; clarification of artifact exclusions and redistribution boundaries; CodeCarbon 3.2.9 consistency; pinned third-party source revisions; and a qualified open data and code statement. Validation The release passed: 29/29 unit tests; the repository release-verification audit; package installation and command-line smoke testing; pyproject.toml and CITATION.cff parsing; shell syntax checks for all setup scripts; and Git diff and release-hygiene checks. Reproducibility boundary The repository supports verification of the released manifests, summaries, strict labels, compact raw records, and feature analysis. Full model-backed execution additionally requires compatible NVIDIA hardware, separately obtained model weights, official third-party implementations, and source-dataset images. The reported environmental quantities concern measured operational GPU energy combined with disclosed grid-intensity and PUE scenarios. They do not include embodied carbon or complete lifecycle emissions. License and citation The ViRel-Budget code is released under the Apache License 2.0. External datasets, images, model weights, and pruning implementations retain their original licenses and terms.
Authors: Wan, Sean, Ou, Shilin, Zhang, Luyao
Language: English
Topic: Green IT and Sustainability
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203782373