Skip to content

DEV: pin CUDA variant for PyTorch #186

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lucascolley merged 2 commits into from
Mar 26, 2025
Merged

DEV: pin CUDA variant for PyTorch #186

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dd5772f
DEV: remove redundant CUDA pins for JAX
crusaderky Mar 26, 2025
a1e2885
Revert JAX and make torch explicit too
crusaderky Mar 26, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion pixi.lock
View file
Open in desktop

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

7 changes: 3 additions & 4 deletions pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,6 +132,7 @@ python = "~=3.13.0"
numpy = "=1.22.0"

# Backends that can run on CPU-only hosts
# Note: JAX and PyTorch will install CPU variants.
[tool.pixi.feature.backends.dependencies]
pytorch = "*"
dask = "*"
Expand All @@ -154,24 +155,22 @@ jax = "*"
# jax = "*" # unavailable

# Backends that require a GPU host and a CUDA driver
# Note: JAX and PyTorch automatically install CUDA variants
# thanks to the `system-requirements` below.
[tool.pixi.feature.cuda-backends]
system-requirements = { cuda = "12" }

[tool.pixi.feature.cuda-backends.target.linux-64.dependencies]
cupy = "*"
jaxlib = { version = "*", build = "cuda12*" }
Copy link
Member

@lucascolley lucascolley Mar 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@traversaro could I double check that this is correct?

Copy link

@traversaro traversaro Mar 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a bit tricky. It is correct that having system-requirements = { cuda = "12" } and jax and pytorch as dependencies will result in CUDA-enabled jaxlib and pytorch being installed, as in both cases the CUDA variant have higher priority (see https://github.com/conda-forge/pytorch-cpu-feedstock/blob/46274b8459ee640a9f90a90d75ac931b770673ed/recipe/meta.yaml#L5-L10 and https://github.com/conda-forge/jaxlib-feedstock/blob/8d785af5387be376037f0210fc6dc2f5da95613c/recipe/meta.yaml#L4-L6). However, have a constraint on the build string ensures that this will always be true in the future.

Let's make an example. If you remove the jaxlib = { version = "*", build = "cuda*" } constraint, and let's pretend that a new jaxlib 0.100.0 (this is a made up number) is released. Let's also pretend that jaxlib 0.100.0 was released without cuda-version==12 compatible builds, for example as cuda 12 was dropped and only cuda 13 is supported, or as there was a regression in cuda builds for 0.100.0, that were marked as broken while cpu builds were still available. In that case, if you do not have jaxlib = { version = "*", build = "cuda*" } , if you refresh the lockfile the solver will just silently install a cpu-only version of jaxlib, while an error will be printed if jaxlib = { version = "*", build = "cuda*" } is present. In a nutshell I find having jaxlib = { version = "*", build = "cuda*" } a bit more robust, even if someone could disagree as it is not super-clear if build string structure is part of the public interface of the conda packages. The long term clean solution for this problem is probably flags (see conda/ceps#111 and https://prefix.dev/blog/ceps_2025), but that still needs to be discussed at the CEP-level.

Copy link
Contributor Author

@crusaderky crusaderky Mar 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

for example as cuda 12 was dropped and only cuda 13 is supported [...] the solver will just silently install a cpu-only version of jaxlib

This is a strong argument against this PR. I'll make it explicit again.

Copy link
Contributor Author

@crusaderky crusaderky Mar 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@lucascolley ready for review again

Copy link
Member

@lucascolley lucascolley Mar 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

thanks for the fast help as always @traversaro !! much appreciated


[tool.pixi.feature.cuda-backends.target.osx-64.dependencies]
# cupy = "*" # unavailable
# jaxlib = { version = "*", build = "cuda12*" } # unavailable

[tool.pixi.feature.cuda-backends.target.osx-arm64.dependencies]
# cupy = "*" # unavailable
# jaxlib = { version = "*", build = "cuda12*" } # unavailable

[tool.pixi.feature.cuda-backends.target.win-64.dependencies]
cupy = "*"
# jaxlib = { version = "*", build = "cuda12*" } # unavailable

[tool.pixi.environments]
default = { features = ["py313"], solve-group = "py313" }
Expand Down
Loading