CPU version of torchvision requires non-CPU version of torch

[iamhatesz]

🐛 Describe the bug

Hi!

The +cpu wheels from here requires torch instead of torch+cpu. This works fine when installing with pip:

pip install torch==1.10.2+cpu torchvision==0.11.3+cpu torchaudio==0.10.2+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html

but fails with more advanced dependency managers like Poetry:

[tool.poetry.dependencies]
torch = {url = "https://download.pytorch.org/whl/cpu/torch-1.10.2%2Bcpu-cp39-cp39-linux_x86_64.whl"}
torchvision = {url = "https://download.pytorch.org/whl/cpu/torchvision-0.11.3%2Bcpu-cp39-cp39-linux_x86_64.whl"}

 Because torchvision (0.11.3+cpu) depends on torch (1.10.2)
   and my-project depends on torch (1.10.2+cpu), torchvision is forbidden.

And when you look into this wheel, you can indeed see:

Requires-Dist: torch (==1.10.2)

The proper value should be torch==1.10.2+cpu. I believe this matters, as torch installed from PyPI is ~900 MB, while torch+cpu installed from PyTorch index is only ~200 MB.

Versions

Collecting environment information...
PyTorch version: 1.10.2+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.3 LTS (x86_64)
GCC version: (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0
Clang version: Could not collect
CMake version: version 3.16.3
Libc version: glibc-2.31

Python version: 3.9.6 (default, Sep  8 2021, 10:04:08)  [GCC 9.3.0] (64-bit runtime)
Python platform: Linux-5.13.0-28-generic-x86_64-with-glibc2.31
Is CUDA available: False
CUDA runtime version: 11.0.221
GPU models and configuration: GPU 0: NVIDIA GeForce GTX 1080
Nvidia driver version: 470.103.01
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A

Versions of relevant libraries:
[pip3] numpy==1.22.2
[pip3] torch==1.10.2+cpu
[pip3] torchaudio==0.10.2+cpu
[pip3] torchvision==0.11.3+cpu
[conda] Could not collect

[Kavan72]

@iamhatesz Yep, you are right. I'm getting the same error, the only thing is how PyTorch team is managing release and packaging for vision.

[Kavan72]

@iamhatesz check this python-poetry/poetry#4231 (comment)

[iamhatesz]

Hey! Any update on this? Seems like a simple thing to fix, but produces a lot of issues with CPU-only wheels on CI pipelines.

[seemethere]

This appears to be what you're looking for: pytorch/pytorch#26340 (comment)

We introduced pep503 compliant indices to download.pytorch.org and they are now the recommended way to install pytorch. I believe that these can also be used with poetry as well.

[seemethere]

Going to close this since pep503 compliant indices should resolve

Feel free to reopen if you feel like your issue has not been resolved

[Kavan72]

hey @seemethere thanks for the update,

I guess the problem is still there while installing CPU wheels, check this

I've added the PyTorch site to the extra source, and getting the same error
here is my pyproject.toml

[tool.poetry]
name = "poetry-test"
version = "0.1.0"
description = ""
authors = ["Your Name <you@example.com>"]
readme = "README.md"
packages = [{include = "poetry_test"}]

[tool.poetry.dependencies]
python = "^3.8"

[[tool.poetry.source]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
indexed = true
secondary = true

[build-system]
requires = ["poetry-core"]
build-backend = "poetry.core.masonry.api"

[chuanma]

@Kavan72 I also have the same issue. Before they find a proper way to fix this, I just add torchvision because torch will be added as a dependency. Note in the screenshot below that the CPU version of torch was correctly found where I also tried to add torchaudio. BTW, it worked with Poetry 1.2.0b3, but NOT with Poetry 1.1.14. Also I only tested it under linux.

poetry add torchvision --source pytorch-cpu

[josearangos]

To try used

pip install --find-links https://download.pytorch.org/whl/torch_stable.html
torch==1.10.0+cpu

[olsonperrensen]

If you need it in a req.txt file you should put -f https://download.pytorch.org/whl/torch_stable.html in a seperate line:

-f https://download.pytorch.org/whl/torch_stable.html
torch==2.0.1+cpu

(credit)