TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2024
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910627004416000 |
|---|---|
| author | Li, Zonghang Feng, Wenjiao Guizani, Mohsen Yu, Hongfang |
| author_facet | Li, Zonghang Feng, Wenjiao Guizani, Mohsen Yu, Hongfang |
| contents | Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2410_00531 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices Li, Zonghang Feng, Wenjiao Guizani, Mohsen Yu, Hongfang Distributed, Parallel, and Cluster Computing Artificial Intelligence 68T50 I.2.11 Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models. |
| title | TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices |
| topic | Distributed, Parallel, and Cluster Computing Artificial Intelligence 68T50 I.2.11 |
| url | https://arxiv.org/abs/2410.00531 |