---
id: 20260818-T0-04
title: "INT8量化推理：交换GPU内核竟导致模型输出剧变"
title_en: "Integer Alibi: Cross-Kernel Divergence in INT8 LLM Inference"
url: https://ai.daily.yangsir.net/daily/20260818-T0-04
issue_date: 2026-08-18
publish_date: 2026-08-17T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2608.13756
---

# INT8量化推理：交换GPU内核竟导致模型输出剧变

一项新的AI研究发现，在INT8量化的大语言模型推理中，即使保持相同的配置（如检查点、硬件、提示词），仅更换实现相同接口的GPU内核，就可能导致模型输出发生显著偏差。这项名为'Integer Alibi'的研究对当前将不同内核视为可互换的假设提出了挑战，为量化推理的可复现性敲响警钟。

## English Version

**Integer Alibi: Cross-Kernel Divergence in INT8 LLM Inference**

New research shows that swapping GPU kernels that implement the same INT8 GEMM interface can cause divergent outputs in quantized LLM inference, even when all other settings are identical. The 'Integer Alibi' study challenges the assumption that these kernels are interchangeable, highlighting reproducibility issues in quantization.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2608.13756)

**详情页**：https://ai.daily.yangsir.net/daily/20260818-T0-04

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*