Installing a specific PyTorch build (f/e CPU-only) with Poetry

Question

I've recently found poetry to manage dependencies. In one project, we use PyTorch. How do I add this to poetry?

We are working on machines that have no access to a CUDA GPU (for simple on the road inferencing/testing) and workstations where we do have access to CUDA GPUs. Is it possible to use poetry to ensure every dev is using the same PyTorch version?

There seems to be no obvious way to decide which PyTorch version to install. I thought about adding the different installation instructions as extra dependencies, but I failed to find an option to get the equivalent settings like:

pip3 install torch==1.3.1+cpu torchvision==0.4.2+cpu -f https://download.pytorch.org/whl/torch_stable.html

I would be fine with setting the total path to the different online wheels, like: https://download.pytorch.org/whl/torch_stable.html/cpu/torch-1.3.1%2Bcpu-cp36-cp36m-win_amd64.whl

But I would rather not but them in git directly... The closest option I've seen in poetry is either downloading them manually and then using file = X command.

Blok

I wonder how it'll work if with poetry if we're using the nightly version of pytorch. – Beagle

Answer 1

Currently, Poetry doesn't have a -f option (there's an open issue and an open PR), so you can't use the pip instructions. You can install the .whl files directly:

poetry add https://download.pytorch.org/whl/torch_stable.html/cpu/torch-1.3.1%2Bcpu-cp36-cp36m-win_amd64.whl

or add the dependency directly to your .toml file:

[tool.poetry.dependencies]
torch = { url = "https://download.pytorch.org/whl/torch_stable.html/cpu/torch-1.3.1%2Bcpu-cp36-cp36m-win_amd64.whl" }

Republican

This is not a satisfying answer as this does not give any solution to the problem being: depending on the platform, point to the correct wheel. – Tales

Note: This won't work if you have dependencies that rely on pytorch, since poetry fails to resolve the dependency versions (github.com/python-poetry/poetry/issues/4231#issue-931746833) – Prolocutor

Answer 2

Since poetry 1.2, you can do this:

poetry source add --priority explicit pytorch_cpu https://download.pytorch.org/whl/cpu
poetry add --source pytorch_cpu torch torchvision

and it will install from the specified index-url.

(and this also works with https://download.pytorch.org/whl/cu118 )

Intercollegiate

This is the most appropriate answer at the time of writing: it exactly incorporates PyTorch's current official index-url for CPU-only installation and does not require any tools other than poetry itself. Thanks! – Mcgannon

The option "-p" does not exist with Poetry 1.4 – Detonator

Answer 3

After spending a couple of hours on this issue, I found a "solution" by combining Poetry and pip just for PyTorch. You don't need to specify the wheel URLs directly and thus remain cross-platform.

I'm using Poe The Poet, a nice task runner for Poetry that allows to run any arbitrary command.

[tool.poetry.dev-dependencies]
poethepoet = "^0.10.0"

[tool.poe.tasks]
force-cuda11 = "python -m pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html"

You can run:

poetry install

and then:

poe force-cuda11  # relies on pip and use PyTorch wheels repo

Overcheck

Answer 4

An updated solution from this issue in the Poetry github:

poetry add torch --platform linux --python "^3.7"

Lagunas

Good link to the issue. Note that this other related issue github.com/python-poetry/poetry/issues/1616 is still open, so this is still an issue and no stable solution is currently available – Tales

This sort of worked for me but I bumped into other (I suspect related) dependency issues further down the track. The Poe The Poet solution suggested by @Overcheck solved my issue though :-) – Fancie

How do you specify the cuda version? I'm assuming that command will default to cpu. – Redeploy

Answer 5

In late 2021, utilizing markers and multiple constraints should work.

$ poetry --version
Poetry version 1.1.11

# pyproject.toml
[tool.poetry.dependencies]
python = "~3.9"
torch = [
  {url = "https://download.pytorch.org/whl/cpu/torch-1.10.0%2Bcpu-cp39-cp39-linux_x86_64.whl", markers = "sys_platform == 'linux'"},
  {url = "https://download.pytorch.org/whl/cpu/torch-1.10.0%2Bcpu-cp39-cp39-win_amd64.whl", markers = "sys_platform == 'win32'", }
]
numpy = "^1.21.4"

[build-system]
requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"

$ poetry install
The currently activated Python version 3.8.12 is not supported by the project (~3.9).
Trying to find and use a compatible version. 
Using python3.9 (3.9.9)
Creating virtualenv machine-learning in /home/redqueen/machine_learning/.venv
Updating dependencies
Resolving dependencies... (36.0s)

Writing lock file

Package operations: 3 installs, 0 updates, 0 removals

  • Installing typing-extensions (4.0.1)
  • Installing numpy (1.21.4)
  • Installing torch (1.10.0+cpu https://download.pytorch.org/whl/cpu/torch-1.10.0%2Bcpu-cp39-cp39-linux_x86_64.whl)

NOTE: Numpy has to be listed. Otherwise you'll get an import error.

Without numpy:

$ python
Python 3.9.9 (main, Nov 23 2021, 00:34:08) 
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
/home/redqueen/machine_learning/.venv/lib/python3.9/site-packages/torch/package/_directory_reader.py:17: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at  ../torch/csrc/utils/tensor_numpy.cpp:68.)
  _dtype_to_storage = {data_type(0).dtype: data_type for data_type in _storages}
>>> quit()

With numpy:

$ python
Python 3.9.9 (main, Nov 23 2021, 00:34:08) 
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
False
>>> quit()

Reference:

https://python-poetry.org/docs/dependency-specification/#python-restricted-dependencies

Disclaimer

I do not have a Windows (or Mac) to test this on.

Skeleton

When I use this approach, Poetry will still download torch from this url every time I run poetry lock. Is there any way to avoid this, to your knowledge? – Tunstall

@Tunstall I believe that's a bug with an open issue. See here or here – Skeleton

Thanks. I'm able to "get around" this by running poetry lock --up-update – Tunstall

Answer 6

There is a fork that I am maintaining called relaxed-poetry It is a very young fork but it supports what you want with the following configuration:

# pyproject.toml

[tool.poetry.dependencies]
python = "^3.8"
torch = { version = "=1.90+cu111", source = "pytorch" }

[[tool.poetry.source]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cu111/"
secondary = true

Check it if you like, it can be installed side by side with poetry.

Ean

This maybe a game changer for ML. Very difficult to use Poetry with pytorch currently. – Overcheck

Adding the poetry source lets me install the right version of pytorch but then I can't install anything else (e.g. black) because poetry searches only in download.pytorch.org – Cohosh

With poetry 1.5.0 it works with priority = "supplemental" instead of secondary python-poetry.org/docs/repositories/… – Broody

Answer 7

I found this worked for me (with GOU running CUDA 12.1 and Python 3.11 on Windows) in my toml file. You can find other versions of PyTorch (torch, torchaudio, torchvideo) to run at https://download.pytorch.org/whl/cu121 that are for CPU only, running on Windows, Linux and Mac

[tool.poetry.dependencies]
torch = {url = "https://download.pytorch.org/whl/cu121/torch-2.2.1%2Bcu121-cp311-cp311-win_amd64.whl"}
torchaudio = {url = "https://download.pytorch.org/whl/cu121/torchaudio-2.2.1%2Bcu121-cp311-cp311-win_amd64.whl"}
torchvision = {url = "https://download.pytorch.org/whl/cu121/torchvision-0.17.1%2Bcu121-cp311-cp311-win_amd64.whl"}

Profluent

how can I find relevant links for my python 3.10 and macos ? – Intersex