MobiLLM: Enabling On-Device Fine-Tuning of Billion-Sized LLMs via Server-Assisted Side-Tuning
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
On-device fine-tuning of large language models (LLMs) has attracted a lot of attention because of its tailoring personalized models while retaining user data locally on the mobile device. However, it faces significant challenges due to prohibitive memory requirements and slow training speeds. In this paper, we propose MobiLLM, a novel scheme enabling memory-efficient LLM fine-tuning on a single mobile device via server-assisted side-tuning. Particularly, MobiLLM strategically offloads backpropagation computations to an edge server while allowing the resource-constrained mobile device to retain merely a pretrained backbone model with frozen parameters during finetuning. It constructs a backpropagation bypass via parallel adapters decoupled from the backbone. During forward propagation, the device employs low bitwidth quantization for transmitting intermediate activations to the server to reduce communication overhead. The advantage of MobiLLM lies in: 1) confining training data strictly to the mobile device, and 2) eliminating on-device backpropagation while overlapping local computations with server execution. Collectively, MobiLLM ensures the data never leaves the local mobile device while significantly reducing mobile memory and computational burdens. We implement MobiLLM on several popular mobile devices, including NVIDIA Jetson Xavier NX and CPU-only laptops. Extensive experimental results demonstrate that MobiLLM can enable a resource-constrained mobile device to fine-tune billion-sized LLMs, achieving up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$4\times$</tex-math></inline-formula> memory reduction and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$2.3\times$</tex-math></inline-formula> faster convergence as compared to state-of-the-art baselines.
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.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.002 | 0.010 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.002 |
| Open science | 0.002 | 0.000 |
| Research integrity | 0.000 | 0.002 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
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 it