请提高 PyTorch 版本上限

[hongbo-miao]

Description of the bug | 错误描述

我之前发了个用 uv 安装的方法在 #1374 (comment)

但我发现里面默认会用 paddlepaddle，我现在希望安装 paddlepaddle-gpu：

pyproject.toml

[project]
name = "mineru"
version = "1.0.0"
requires-python = "~=3.12.0"
dependencies = [
  "paddlepaddle-gpu==3.0.0b2",
]

[project.optional-dependencies]
build = [
  "setuptools",
  "torch",
]
compile = [
  "detectron2",
  "magic-pdf[full]==1.1.0",
]

[tool.uv]
package = false
prerelease = "allow"
no-build-isolation-package = ["detectron2"]

[tool.uv.sources]
paddlepaddle-gpu = { index = "paddlepaddle-gpu" }
detectron2 = { git = "https://github.com/facebookresearch/detectron2.git", branch = "main" }

[[tool.uv.index]]
name = "paddlepaddle-gpu"
url = "https://www.paddlepaddle.org.cn/packages/stable/cu123"
explicit = true

在跑 uv sync --extra=build 时会报错

warning: Missing version constraint (e.g., a lower bound) for `setuptools`
warning: Missing version constraint (e.g., a lower bound) for `torch`
  × No solution found when resolving dependencies for split (python_full_version >= '3.12.0' and python_full_version < '3.13' and platform_machine != 'aarch64' and platform_system == 'Linux'):
  ╰─▶ Because torch>=2.2.2,<=2.3.1 depends on nvidia-nvtx-cu12{platform_machine == 'x86_64' and platform_system == 'Linux'}==12.1.105 and only the following versions of torch are available:
          torch<=2.2.2
          torch==2.3.0
          torch>=2.3.1
      we can conclude that torch>=2.2.2,<=2.3.1 depends on nvidia-nvtx-cu12{platform_machine == 'x86_64' and platform_system == 'Linux'}==12.1.105.
      And because magic-pdf[full]==1.1.0 depends on torch>=2.2.2,<=2.3.1, we can conclude that magic-pdf[full]==1.1.0 depends on nvidia-nvtx-cu12{platform_machine == 'x86_64' and platform_system == 'Linux'}==12.1.105.
      And because paddlepaddle-gpu==3.0.0b2 depends on nvidia-nvtx-cu12{platform_machine == 'x86_64' and platform_system == 'Linux'}==12.4.127 and mineru[compile] depends on magic-pdf[full]==1.1.0, we can conclude that mineru[compile] and paddlepaddle-gpu==3.0.0b2 are
      incompatible.
      And because mineru[compile] depends on paddlepaddle-gpu==3.0.0b2 and your project requires mineru[compile], we can conclude that your project's requirements are unsatisfiable.
error: Recipe `uv-install-dependencies` failed on line 8 with exit code 1

但由于 magic-pdf 依赖的 PyTorch 版本太低：

MinerU/setup.py

Lines 40 to 41 in 1e4d4b5

	"torch>=2.2.2,<=2.3.1", # torch2.4.0及之后版本未测试，先卡住版本上限
	"torchvision>=0.17.2,<=0.18.1", # torchvision 受torch版本约束

这个旧的 PyTorch 又依赖旧的 nvidia-nvtx-cu12。
而 paddlepaddle-gpu==3.0.0b2 需要更新的 nvidia-nvtx-cu12。

请提高 PyTorch 版本上限，谢谢！

How to reproduce the bug | 如何复现

如上

Operating system | 操作系统

Linux

Python version | Python 版本

3.12

Software version | 软件版本 (magic-pdf --version)

1.1.0

Device mode | 设备模式

cuda

[myhloli]

paddlegpu可以装3.0.0b1的cu118版本

[hongbo-miao]

@myhloli 多谢！改成

[[tool.uv.index]]
name = "paddlepaddle-gpu"
url = "https://www.paddlepaddle.org.cn/packages/stable/cu118"
explicit = true

确实可以安装成功了，不过还想希望将来有时间能提高 PyTorch 上限 ☺️

[hongbo-miao]

（最新用 uv 的安装方法发在了 https://github.com/opendatalab/MinerU/discussions/1374）