Better guards for DeepSpeed imports #3351
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
lewtun
commented
Apr 24, 2025
| @@ -167,6 +163,8 @@ def iter_params(module, recurse=False): | |||
|
|
|||
| def add_hooks(model: "DeepSpeedEngine") -> None: | |||
| """Adds the optimizer hooks from a DeepSpeed ZeRO-3 model.""" | |||
| import deepspeed | |||
There was a problem hiding this comment.
I realise doing this import every time we generate is sub-optimal, but DS generation is slow anyway and I don't think it's used much vs vllm
There was a problem hiding this comment.
Actually it's not sub-optimal. Python's import is effectively conditional—when you import, Python checks if the module is already loaded (via sys.modules) and reuses it if so. So it won't reload the module or cause significant overhead.
5 tasks
|
I'm not sure with the doc builder CI is failing with the installation and Have you seen this before @qgallouedec ? |
it seems like we're having it in all our PR since yesterday |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
The Command to repro: cc @kashif if you can take a look at the GKD issue when you're back |
|
The Command to repro: Not caused by this PR since the error is also on |
|
@qgallouedec I've tested all scripts related to the changes in this PR and apart from the ones that were already broken on |
qgallouedec
approved these changes
Apr 25, 2025
There was a problem hiding this comment.
Perfect. My only concern is that future me, who will forget about this, might not check the git blame and might move it back to top-level in the course of an innocent PR 😇. Can you add a little comment like this?
def my_func():
import deepspeed # local import (instead of top-level) to avoid DS init interfering with other backends (like vllm)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What does this PR do?
Whenever we run
import deepspeed, CUDA is initialised which is (a) slow and (b) interferes with thevllmserver when DP=8 and TP=1 (see vllm-project/vllm#17079):This PR fixes the issue on the TRL side by:
deepspeedonly when needed_prepare_deepspeed()in the trainers for the single-use function in the utils.Tested with
To fully solve thePR now merged.vllmserver issue will require an additional refactor on thetransformersside, but this PR is done to keep things self-contained.Update: transformers PR here huggingface/transformers#37755
TODO
Check the following trainers with Z3:
Before submitting
Pull Request section?
to it if that's the case.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.