arXiv cs.AI· Xingyang Li, Tianlang Zhao, Song Han, Ligeng Zhu·· 10 小时前AI 评分34
QDA:用智能体实现闭环 LLM 量化设计
QDA: Closed-Loop LLM Quantization Design with Agents
AI 导读
QDA 是一个由智能体为每个模型和精度自动设计量化配方的系统,在四个推理模型上,W4A16、W4A8、W3A16 平均准确率分别比最强公开配方提升 0.9、1.4、2.2 个百分点,每次搜索用 64 GPU-hours。
正文 · 原文
Abstract:Post-training quantization is the standard way to serve large language models at low cost. However, the best quantization recipe differs across models and precisions. As a result, experts must redesign it for every new model, which is time-consuming and often sub-optimal. Fortunately, an LLM agent can take over this design by refining the recipe over rounds of building and evaluating candidates, so that more rounds lead to better recipes. However, evaluating one quantized reasoning model takes tens of GPU-hours, and combining methods from different codebases means porting each to the new architecture by hand. We present QDA, a system in which an agent designs the recipe for each model and precision. To make evaluation fast, QDA scores each candidate in minutes by its output KL divergence from the original model on reasoning traces that the agent never sees, and checks promising ones with long generations to catch failures such as repetition loops. To make building fast, QDA provides a library of public PTQ methods that are compatible with one another and with new model architectures. Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint. Every QDA checkpoint also runs on existing serving kernels with no inference overhead. We will release the library, the evaluator and the harness.
| Comments: | 19 pages, 7 figures, 10 tables |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.23388 [cs.LG] |
| (or arXiv:2611.23388v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2611.23388 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xingyang Li XY Li [view email]
[v1]
Tue, 17 Nov 2026 19:00:09 UTC (939 KB)
来源:arXiv cs.AI · arxiv.org