Communication-Efficient MoE Fine-Tuning with Locality-Aware Expert Placement
Bibliographic record
Abstract
With the prevailing Mixture-of-Experts (MoE) architecture pushing the performance of Large Language Models (LLMs) to new limits, fine-tuning MoE models presents a significant challenge due to their tremendous number of parameters and sparsely activated network structures. While expert parallelism has been proposed to train large-scale MoE models by distributing expert layers among multiple devices, it fails to exploit the unique communication patterns in fine-tuning pre-trained MoE models. In this paper, we demonstrate that expert layers are not uniformly accessed, but exhibit a stable locality, with some experts being accessed more frequently than others throughout the fine-tuning process. Based on this insight, we introduce Vela, a novel fine-tuning system for MoE architectures that leverages expert locality to reduce communication overhead. Specifically, Vela implements a novel training and communication framework that separates expert layers from the MoE model, and employs a locality-aware expert placement mechanism to minimize the communication overhead, thereby significantly improving the fine-tuning efficiency. Our extensive array of evaluations demonstrates that Vela decreases the communication overhead by up to 25%, consequently accelerating the fine-tuning process by up to 28% compared to conventional methods.
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) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.003 | 0.002 |
| 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".