Unable to install PyTorch with CUDA

[lingkerio]

When I use pdm add torch to install PyTorch, it installs version 2.3.1+cpu. However, I want to install the GPU version, and I haven't found any command that helps me specify the GPU version. Is there a different command or option I should use to install PyTorch with GPU support?

[pawamoy]

I suggest you check out the official docs to see how to install pytorch with GPU support: https://pytorch.org/get-started/locally/.

[lingkerio]

So i must use pip or conda?

[frostming]

did you use the cuda source for installation?

Refer to https://pdm-project.org/en/latest/usage/config/#configure-the-package-indexes for how to

[lingkerio]

did you use the cuda source for installation?

Refer to https://pdm-project.org/en/latest/usage/config/#configure-the-package-indexes for how to

Sorry, I don't quite understand what you mean. Maybe you mean I should install pytorch using something like "pdm add torch@example.com/torch.wheel", but using this method will cause me to get stuck for a long time every time I install a new package, as if I'm re-downloading pytorch.

[frostming]

According to the website, you need to specify a pytorch index for installing CUDA packages

The [[tool.pdm.source]] is the PDM equivalent to pip's --index-url, please configure it as the docs says.

[erklem]

Did you try this? #1645 (comment)

[lingkerio]

Did you try this? #1645 (comment)

Thank you! I had already configured the following setup a while ago based on suggestions from pawamoy and frostming:

[[tool.pdm.source]]
name = "private"
url = "https://download.pytorch.org/whl/cu121"
include_packages = ["torch", "torchvision", "torchaudio"]
exclude_packages = ["*"]

I checked the link you provided and found that it aligns with my approach, which helped confirm that I did it correctly. I really appreciate your help!

[ulic-youthlic]

i tried this, and when i used pdm add torch to add my torch package, it seemed that pdm would find torch vx.x.x+cpu firstly, how can i use torch with cuda package as first?

[lingkerio]

i tried this, and when i used pdm add torch to add my torch package, it seemed that pdm would find torch vx.x.x+cpu firstly, how can i use torch with cuda package as first?

Here is the configuration I provided:

[tool.pdm.resolution]
respect-source-order = true
# Don't let PDM install all these runtime libraries -- they add GB's of bloat!
excludes = [
  "nvidia-cublas-cu12",
  "nvidia-cuda-cupti-cu12",
  "nvidia-cuda-nvrtc-cu12",
  "nvidia-cuda-runtime-cu12",
  "nvidia-cudnn-cu12",
  "nvidia-cufft-cu12",
  "nvidia-curand-cu12",
  "nvidia-cusolver-cu12",
  "nvidia-cusparse-cu12",
  "nvidia-nccl-cu12",
  "nvidia-nvtx-cu12",
  "triton",
]

[[tool.pdm.source]]
name = "private-pypi"
url = "https://download.pytorch.org/whl/cu121"
include_packages = ["torch", "torchvision", "torchaudio"]

However, this configuration doesn't seem to behave much differently from the one I mentioned earlier when downloading torch. I suggest you try running pdm add torch -v to provide more detailed information.