跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.25971 · v1

Kameleo: In-Situ Activation-Precision Selection for Bit-Serial LLM Accelerators

首次提交 2026-11-19 19:55
版本更新 2026-11-19 19:55
雷达观测 2026-11-20 16:53
arXiv ↗PDF ↗

摘要

Weight-only quantization keeps the activations of large language models (LLMs) in a higher-precision format, pushing the on-chip inference cost to the activation arithmetic. Bit-serial, variable-length accelerators offer a promising path to reduce this cost by lowering activation precision while preserving model quality. They process only the mantissa bits that each layer needs. The length has to be chosen per layer and operator, and today it is chosen by offline calibration: a search driven by the perplexity of the full model on a representative dataset. The search is repeated for every model and workload, and its result degrades silently on other data. This paper presents Kameleo, an accelerator that selects the activation precision at runtime, on the accelerator, from the data it is processing. Kameleo replaces the global perplexity by a local proxy that compares the output of each GEMM at reduced and at full precision. It exploits the fact that a bit-serial datapath produces this comparison almost for free, because the reduced-precision result is an intermediate state of the full-precision computation. Two passes over the first tokens of a workload set the precision of every GEMM. On four LLMs and seven datasets, Kameleo stays within a 1% perplexity budget at a 2.11-2.34x reduction of prefill cycles on the matrix multiplication unit (MXU), close to the reduction of configurations calibrated offline on the evaluation data itself. Calibrated configurations transferred to other data, in contrast, often violate the budget. The selection hardware adds only 3.7% to the accelerator area. Kameleo thereby makes variable-length mantissas feasible in real deployments, where offline calibration for every model and workload is not.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-19 19:55 · 最新观测版本

Kameleo: In-Situ Activation-Precision Selection for Bit-Serial LLM Accelerators

Weight-only quantization keeps the activations of large language models (LLMs) in a higher-precision format, pushing the on-chip inference cost to the activation arithmetic. Bit-serial, variable-length accelerators offer a promising path to reduce this cost by lowering activation precision while preserving model quality. They process only the mantissa bits that each layer needs. The length has to be chosen per layer and operator, and today it is chosen by offline calibration: a search driven by the perplexity of the full model on a representative dataset. The search is repeated for every model and workload, and its result degrades silently on other data. This paper presents Kameleo, an accelerator that selects the activation precision at runtime, on the accelerator, from the data it is processing. Kameleo replaces the global perplexity by a local proxy that compares the output of each GEMM at reduced and at full precision. It exploits the fact that a bit-serial datapath produces this comparison almost for free, because the reduced-precision result is an intermediate state of the full-precision computation. Two passes over the first tokens of a workload set the precision of every GEMM. On four LLMs and seven datasets, Kameleo stays within a 1% perplexity budget at a 2.11-2.34x reduction of prefill cycles on the matrix multiplication unit (MXU), close to the reduction of configurations calibrated offline on the evaluation data itself. Calibrated configurations transferred to other data, in contrast, often violate the budget. The selection hardware adds only 3.7% to the accelerator area. Kameleo thereby makes variable-length mantissas feasible in real deployments, where offline calibration for every model and workload is not.

查看此版本解读 →查看此版本原文 ↗