Use pytorch utils to detect ninja #7687
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Emrys-Merlin
requested review from
jomayeri,
loadams and
tjruwase
as code owners
Signed-off-by: Tim Adler <[email protected]>
Emrys-Merlin
force-pushed
the
refactor/ninja_discovery
branch
from
1a7f682 to
e826740
Compare
|
To add, not only this dependency creates complexity to these environments, it is a blocker for creating standalone conda packages depending on DS - for example for openfold-3, as described in the issue. Thanks a lot for your consideration. |
Collaborator
|
@Emrys-Merlin thanks for this contribution. What is the earliest pytorch version can detect ninja? |
|
I have looked a bit and at least it is in old pytorch versions (1.8.1 and likely earlier). Does deepspeed aims at supporting very old versions of pytorch? I think to be in the safe side we should then add this simple logic here, maybe completely replacing using pytorch's or as a fallback for old versions if import fails. What do you think? |
Collaborator
|
@sdvillal thanks for the clarification. No, we don't plan to support very old pytorch versions. This addresses my concerns. |
sfc-gh-truwase
approved these changes
Nov 12, 2025
rraminen
pushed a commit
to rraminen/DeepSpeed
that referenced
this pull request
Dec 1, 2025
This PR replaces the ninja detection based on `import ninja` with the corresponding pytorch utils to detect ninja. The motivation behind this change is twofold. 1. The detection based on importing the python interface works in the pip-world, but fails in the conda-world, as the corresponding ninja conda package does not include the python interface. The implication on detection is [known](conda-forge/ninja-feedstock#26) and the recommended solution is to use the `subprocess` module to detect ninja. This approach is followed by, e.g., [meson](https://github.com/mesonbuild/meson-python/blob/1c8092dc477cbc7e1e4d40913608d9daae75f793/mesonpy/__init__.py#L1077-L1088) and [pytorch](https://github.com/pytorch/pytorch/blob/d33d125c9413c5043aa5f74fad909a576288242d/torch/utils/cpp_extension.py#L2312-L2325). As the `subprocess`-based check works in both the pip- and the conda-world, I think, it would make sense to switch over. 2. As ninja is only invoked through pytorch, it should be sufficient to check if pytorch can find ninja. Hence, instead of reimplementing the `subprocess` check, I think it would be better to use the pytorch utils (which implement said test anyways). Without this refactor, every conda environment that depends on DeepSpeed will need to install ninja as a PyPI dependency, even though the conda version of ninja would be sufficient for the compilation. In my opinion, this adds unnecessary complexity to these environments. I tried to keep the changes minimal. As some additional context, @sdvillal and I stumbled over this issue while working on packaging aqlaboratory/openfold-3 for conda-forge. Signed-off-by: Tim Adler <[email protected]> Co-authored-by: Olatunji Ruwase <[email protected]> Signed-off-by: rraminen <[email protected]>
5 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR replaces the ninja detection based on
import ninjawith the corresponding pytorch utils to detect ninja.The motivation behind this change is twofold.
subprocessmodule to detect ninja. This approach is followed by, e.g., meson and pytorch. As thesubprocess-based check works in both the pip- and the conda-world, I think, it would make sense to switch over.subprocesscheck, I think it would be better to use the pytorch utils (which implement said test anyways).Without this refactor, every conda environment that depends on DeepSpeed will need to install ninja as a PyPI dependency, even though the conda version of ninja would be sufficient for the compilation. In my opinion, this adds unnecessary complexity to these environments.
I tried to keep the changes minimal.
As some additional context, @sdvillal and I stumbled over this issue while working on packaging aqlaboratory/openfold-3 for conda-forge.