一句话：Upix 是首个把分治-合并（DCC）推断实现为框架的概率编程系统，可将通用模型拆分并编译到加速器上高效推断。
通用概率编程语言能描述具有随机支撑结构的模型，但这类模型的后验推断极为困难，也很难在现代硬件上加速。本文提出 Upix——首个将分治-合并（divide-conquer-combine, DCC）推断算法实现为框架的概率编程系统。Upix 自动把通用 PPL 表达的模型拆分为多个具有静态支撑结构的子模型，再用 JAX 编译到加速器硬件上执行；同时借鉴可编程推断文献中的既有概念，允许对推断算法进行大量定制。为评估该系统，作者在 Upix 中实现了两个已有 DCC 算法并实例化了三个新算法。结果显示：在相同时间预算内可实现多达 1070 倍的计算量，从而获得比现有方法更好的近似质量；在最多 64 个 CPU 核心与 8 块 GPU 的机器上，Upix 能将推断扩展到对 CPU 与先前方法而言慢到不实用的负载。
Universal probabilistic programming languages (PPLs) enable the specification of models with stochastic support structure. Posterior inference is notoriously hard for this class of models and remains difficult to accelerate on modern hardware. In response to these challenges, we introduce Upix - the first probabilistic programming system that realises the divide-conquer-combine (DCC) inference algorithm as a framework. In Upix, a model expressed in a universal PPL is automatically split into multiple sub-models with static support structure, which are then compiled with JAX for execution on accelerator hardware. The system allows extensive customisation of inference algorithms by incorporating established concepts from programmable inference literature. To evaluate our system, we implemented two existing DCC algorithms in Upix and instantiated three novel algorithms. We show that our implementation can result in better approximation quality compared to existing approaches by achieving up to 1070 times more computation within the same time budget. On machines with up to 64 CPU cores and 8 GPU devices, we demonstrate that Upix enables the scaling of inference algorithms to workloads that are impractically slow for CPUs and prior methods.
@article{paperbot12154,
title = {Probabilistic Programming with Programmable Divide-Conquer-Combine Inference on Modern Hardware},
author = {Markus Böck and Jürgen Cito},
journal = {Proceedings of the ACM on Programming Languages},
volume = {10},
number = {OOPSLA2},
year = {2026},
doi = {10.1145/3839533}
}