PockEngine: Sparse and Efficient Fine-tuning in a Pocket
Name
3613424.3614307.pdf
Size
7.03 MB
Format
Adobe PDF
Checksum (MD5)
aefb063b69db39e7267802a89dde23da
Author(s)
Zhu, Ligeng
Hu, Lanxiang
Lin, Ji
Chen, Wei-Ming
Wang, Wei-Chen
Gan, Chuang
Han, Song
Date Issued
October 28, 2023
Publisher
ACM|56th Annual IEEE/ACM International Symposium on Microarchitecture
Citation
Zhu, Ligeng, Hu, Lanxiang, Lin, Ji, Chen, Wei-Ming, Wang, Wei-Chen et al. 2023. "PockEngine: Sparse and Efficient Fine-tuning in a Pocket."
Version
Final published version
Abstract
On-device learning and efficient fine-tuning enable continuous and privacy-preserving customization (e.g., locally fine-tuning large language models on personalized data). However, existing training frameworks are designed for cloud servers with powerful accelerators (e.g., GPUs, TPUs) and lack the optimizations for learning on the edge, which faces challenges of resource limitations and edge hardware diversity. We introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices. PockEngine supports sparse backpropagation: it prunes the backward graph and sparsely updates the model with measured memory saving and latency reduction while maintaining the model quality. Secondly, PockEngine is compilation first: the entire training graph (including forward, backward and optimization steps) is derived at compile-time, which reduces the runtime overhead and brings opportunities for graph transformations. PockEngine also integrates a rich set of training graph optimizations, thus can further accelerate the training cost, including operator reordering and backend switching. PockEngine supports diverse applications, frontends and hardware backends: it flexibly compiles and tunes models defined in PyTorch/TensorFlow/Jax and deploys binaries to mobile CPU/GPU/DSPs. We evaluated PockEngine on both vision models and large language models. PockEngine achieves up to 15 × speedup over off-the-shelf TensorFlow (Raspberry Pi), 5.6 × memory saving back-propagation (Jetson AGX Orin). Remarkably, PockEngine enables fine-tuning LLaMav2-7B on NVIDIA Jetson AGX Orin at 550 tokens/s, 7.9 × faster than the PyTorch.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
MIT-IBM Watson AI Lab
Terms of Use
Creative Commons Attribution
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1145/3613424.3614307
