Menos: Split Fine-Tuning Large Language Models with Efficient GPU Memory Sharing
Bibliographic record
Abstract
Fine-tuning of pre-trained large language models has become increasingly popular, yet existing fine-tuning methods are typically centralized, requiring users to send local data to centralized servers, or model owners to open-source their models. However, data and models are valuable assets that few enterprises and users wish to share. In this paper, we deviate from conventional wisdom and advocate the use of split learning for fine-tuning models with private data, local to each of the clients. The most formidable challenge to split fine-tuning is the size of large language models: when multiple clients start their fine-tuning tasks, their use of GPU memory will overwhelm a GPU-equipped server, especially as the number of clients scales up. To address this challenge, we present Menos, the first memory-efficient split fine-tuning framework designed to optimize the server GPU footprint through spatial and temporal sharing. Specifically, Menos utilizes the adapter-based nature of modern fine-tuning techniques, and proposes to spatially share the base model parameters among multiple clients. It also schedules memory-intensive operations during the communication gaps of split learning, thereby temporally sharing limited GPU memory at runtime. Comprehensive real-world evaluations using state-of-the-art large language models demonstrate the effectiveness of Menos, reducing GPU memory consumption by up to 72%, yet incurring negligible overhead.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.006 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.003 | 0.003 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.005 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".