Announcement on Compatibility between MinerU 2.5 and the New vLLM Version

[myhloli]

Recently, we officially released the MinerU 2.5 model with significantly improved performance, and migrated the inference acceleration framework from sglang to vLLM. This upgrade leverages vLLM’s richer ecosystem to enhance compatibility with mainstream platforms, enabling more users to easily access state-of-the-art document parsing capabilities.

In the latest vLLM 0.10.2 release, its V1 engine has finally added support for Turing architecture and earlier GPUs—an important step forward in expanding hardware compatibility. However, this version upgrades the required PyTorch version to 2.8.0. We previously observed compatibility issues between PyTorch 2.8.0 and our pipeline backend (see GitHub Discussion #3337), which led us to cap the PyTorch version at <2.8.0.

To enable as many users as possible to experience the powerful performance of MinerU 2.5, we have now removed the upper limit on the PyTorch version in our latest release. We have also invested significant effort into adapting the pipeline backend to PyTorch 2.8.0, aiming to minimize performance degradation while maintaining functional stability. Nevertheless, some compatibility issues remain unavoidable, such as:

• A slight decrease in OCR processing speed;
• Rare cases of incorrect page reading order.

To help you achieve the best possible experience, we provide the following recommendations based on your deployment method:

🔧 Users Installing via uv / pip

If you are primarily using the pipeline backend and encounter the above issues, we recommend downgrading PyTorch to version 2.7.1 to avoid compatibility problems:

uv pip install torch==2.7.1 torchvision

This version offers optimal compatibility with the current pipeline backend and restores previous performance levels.

🐳 Docker Users

The default base image in the current Dockerfile is vllm/vllm-openai:v0.10.1.1, which includes PyTorch 2.7.1. Therefore, it is not affected by the PyTorch 2.8.0 compatibility issues, and we recommend most users continue using this version.

Only switch to the vLLM 0.10.2 image if both of the following conditions apply:

1. You are using a Turing or older GPU architecture;
2. You require vLLM to accelerate VLM inference.

In such cases, update your base image to:

vllm/vllm-openai:v0.10.2

---

We are committed to achieving the best possible balance among performance, compatibility, and stability. Thank you for your understanding and support. We welcome your feedback via GitHub to help us further improve the project.

— The MinerU Team

[sdfu9cjuhdf]

佬，我想知道我在quay.io/ascend/vllm-ascend:v0.10.2rc1镜像容器里，下载mineru[core]，然后怎么才能用到vlm加速呢？之前我是在华为的mindie镜像中用fastapi启动的。用的pipeline模式。mineru-api --host 0.0.0.0 --port 8009 现在的话，我理解应该backend要传vlm-pipeline。
提问重点：我如何能使用到vllm加速呢？

[eziohzy]

vllm加速不应该pipeline模型， 应该用backend=vllm吧

[MaiZiPiaoPiao]

当你在镜像容器里安装了mineru[core]之后，直接启动mineru-api --host 0.0.0.0 --port 8009，就可以使用了。通过镜像建立的容器，默认就安好了vLLM