Spawning isolated processes for each test #67
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
meta-cla
bot
added
the
CLA Signed
msaroufim
reviewed
Aug 13, 2025
|
Add eval_multiprocessing.py to run each test in a separate subprocess for better CUDA error isolation. Log from test_adverse_cases.pyLog from test_eval_multiprocessing.py |
jiannanWang
changed the title
msaroufim
reviewed
Aug 13, 2025
|
@jiannanWang Also it'd be useful to add some timings for running op_info suite against aten here. You mentioned there was a large slowdown, so it's worth seeing how bad it is. Feel free to just use time {command} |
PaliC
reviewed
Aug 13, 2025
PaliC
reviewed
Aug 13, 2025
msaroufim
reviewed
Aug 14, 2025
msaroufim
reviewed
Aug 14, 2025
msaroufim
reviewed
Aug 14, 2025
msaroufim
reviewed
Aug 14, 2025
|
Nvidia-smi results showing this is multiprocessing. I'm not sure why device 0 got so many process. Runtime: 1 GPU: Main (no multiprocessing) It turns out that multiprocessing does not provide a speedup in our case. This is likely because our tests run very quickly, and most of the overhead comes from initialization, which multiprocessing cannot accelerate. For now, we might need to avoid using multiprocessing as the default solution. Instead, we can offer it as an alternative option for users who encounter frequent CUDA errors in their experiments. |
|
@jiannanWang these results make sense I think, but what command are you using for testing / what is the backend/suite. Unclear if we should be concerned with using 8 gpus. Also can you a similar time trial on main for record keeping. |
Added in the previous comment. I'm using |
msaroufim
reviewed
Aug 19, 2025
| @@ -37,4 +37,4 @@ jobs: | |||
| run: uv run python -m BackendBench.scripts.main --suite facto --backend aten --ops "add.Tensor" | |||
|
|
|||
| - name: Run pytest tests | |||
| run: uv run pytest test/ | |||
| run: uv run pytest test/ --deselect test/test_adverse_cases.py | |||
There was a problem hiding this comment.
let's not do this here cause then everyone would need to be aware of this test locally, you can add a skip test directly in the test file
msaroufim
reviewed
Aug 19, 2025
| import pytest | ||
| import torch | ||
|
|
||
| import BackendBench.multiprocessing_eval as multiprocessing_eval |
There was a problem hiding this comment.
can we merge this file into the the test_adverse_cases file?
msaroufim
approved these changes
Aug 19, 2025
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.
Fix #45 .
This PR introduces a
--num-workersflag to enable multiprocessing evaluation. The main purpose is to spawn isolated processes to prevent CUDA errors from waterfalling during benchmarking. Note that multiprocessing does not improve speed because tests run quickly and most overhead comes from initialization, which multiprocessing cannot accelerate. Therefore, multiprocessing is disabled by default and should only be enabled if CUDA errors block the experiment.