---
title: "ParVL：多模态大模型的并行扩展与可分配计算框架"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.04010v1"
permalink: "https://feed.astralor.com/items/itm_003d0e5b0c7d1613ffde5341"
published: "2026-08-04T17:59:58.000Z"
score: 72
channel: "tech"
tags: ["ParVL","多模态大模型","并行计算扩展","视觉Transformer","计算分配"]
---

# ParVL：多模态大模型的并行扩展与可分配计算框架

## AI 点评

> 这篇论文的价值在于它把扩展的视角从'加参数'转向了'重排计算'，对做多模态模型落地的人尤其值得关注。如果你关心部署成本或任务定制，ParVL 这种可分配机制可能会启发更高效的设计。

## 速览

现有大规模多模态模型的扩展策略通常只扩大模型参数或串行推理计算，导致高昂的显存或延迟开销，并且难以调整视觉Transformer与语言模型之间固定的计算分配。ParVL框架通过复用现有ViT和LLM骨干参数，在多个视觉与语言分支上并行扩展计算，并支持在固定骨干参数预算下灵活分配视觉与语言模态的计算资源。实验表明，ParVL在约130亿token上全参数微调后，整体性能优于同配置单分支基线，且最优视—语分配因任务而异。

## 关键点

- 提出ParVL框架，通过共享骨干复用实现并行计算扩展，而非增加参数或串行深度
- 系统研究固定参数预算下视觉与语言模态之间的计算分配权衡
- 在约130亿token上全参数微调，整体性能优于单分支基线，且最优分配因任务变化
- 开源代码已发布

## 全文

Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transformer and the Large Language Model components, limiting task-specific optimization. To address this, we introduce the Parallel Vision-Language (ParVL) scaling framework for MLLMs, which scales parallel computation by reusing the existing ViT and LLM backbone parameters across multiple vision and language branches. This framework raises a central question: given a fixed backbone parameter budget, how should additional shared-backbone computation be allocated between the vision and language modalities? We instantiate each parallel computational stream with branch-specific prefix parameters over a shared backbone, and train the entire model end-to-end via full-parameter supervised fine-tuning on roughly 13B tokens. We systematically study the computation-allocation trade-off between the ViT encoder and LLM decoder. ParVL improves overall multimodal performance over same-recipe single-branch baselines, and the best evaluated vision--language allocation varies across tasks. Code is available at https://github.com/YangYangGirl/ParVL.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_003d0e5b0c7d1613ffde5341