Saved in:
Bibliographic Details
Main Authors: Zhang, Kexun, Zhou, Shang, Wang, Danqing, Wang, William Yang, Li, Lei
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2410.22480
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914997158805504
author Zhang, Kexun
Zhou, Shang
Wang, Danqing
Wang, William Yang
Li, Lei
author_facet Zhang, Kexun
Zhou, Shang
Wang, Danqing
Wang, William Yang
Li, Lei
contents Sampling is a basic operation in many inference-time algorithms of large language models (LLMs). To scale up inference efficiently with a limited compute, it is crucial to find an optimal allocation for sample compute budgets: Which sampling configurations (model, temperature, language, etc.) do we use? How many samples do we generate in each configuration? We formulate these choices as a learning problem and propose OSCA, an algorithm that Optimizes Sample Compute Allocation by finding an optimal mix of different inference configurations. Our experiments show that with our learned mixed allocation, we can achieve accuracy better than the best single configuration with 128x less compute on code generation and 25x less compute on 4 reasoning tasks. OSCA is also shown to be effective in agentic workflows beyond single-turn tasks, achieving a better accuracy on SWE-Bench with 3x less compute than the default configuration. Our code and generations are released at https://github.com/LeiLiLab/OSCA.
format Preprint
id arxiv_https___arxiv_org_abs_2410_22480
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Scaling LLM Inference with Optimized Sample Compute Allocation
Zhang, Kexun
Zhou, Shang
Wang, Danqing
Wang, William Yang
Li, Lei
Computation and Language
Artificial Intelligence
Sampling is a basic operation in many inference-time algorithms of large language models (LLMs). To scale up inference efficiently with a limited compute, it is crucial to find an optimal allocation for sample compute budgets: Which sampling configurations (model, temperature, language, etc.) do we use? How many samples do we generate in each configuration? We formulate these choices as a learning problem and propose OSCA, an algorithm that Optimizes Sample Compute Allocation by finding an optimal mix of different inference configurations. Our experiments show that with our learned mixed allocation, we can achieve accuracy better than the best single configuration with 128x less compute on code generation and 25x less compute on 4 reasoning tasks. OSCA is also shown to be effective in agentic workflows beyond single-turn tasks, achieving a better accuracy on SWE-Bench with 3x less compute than the default configuration. Our code and generations are released at https://github.com/LeiLiLab/OSCA.
title Scaling LLM Inference with Optimized Sample Compute Allocation
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2410.22480