HarMoEny: Efficient multi-GPU inference of Mixture of Experts models by scheduling experts across accelerators
Bibliographic record
Abstract
Mixture-of-Experts (MoE) models enhance computational efficiency during inference by selectively activating specialized experts for each request.This approach facilitates efficient model scaling on multi-GPU systems leveraging expert parallelism, without compromising performance.However, uneven load distribution across experts often creates imbalances across GPUs, increasing latency due to idle waiting times.To address this, we propose HarMoEny, a novel solution that employs two simple techniques: (i) dynamic token redistribution to underutilized GPUs and (ii) asynchronous prefetching of experts from system to GPU memory.These methods achieve near-perfect load balancing across GPUs and amortize the cost of executing experts not resident in GPU memory, thereby reducing waiting times and latency.We implement HarMoEny and evaluate its performance against four state-of-the-art baselines using real-world and artificially imbalanced datasets.Under heavy load imbalance, HarMoEny improves throughput by 66.7%-68.6%and reduces time to first token by 31.8%-37.2%,compared to the next-best baseline.Furthermore, our ablation study demonstrates that HarMoEny's scheduling policy reduces the GPU idle time by up to 99% relative to the baseline policies.i This endeavour is not endurable without the help of an amazing group of people.These people have written this thesis, figuratively, just as much as I have.I would like to firstly thank my supervisor Prof. Oana Balmau who has been my buoy for the past 3½ years of my life.This time has been filled with growth mostly thanks to her attention, work-ethic, and flexibility.Thanks for the constant push for me to achieve and for going through this thesis enumerable times to improve it to its current state.The lessons learned here will serve for decades to come.I would like to then thank Prof. Anne-
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 | 0.002 | 0.005 |
| Meta-epidemiology (narrow) | 0.001 | 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 | 0.002 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
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".