fix lm eval test reproducbility issues #1260
Open
+32
−8
Conversation
This file contains 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
Signed-off-by: Brian Dellabetta <bdellabe@redhat.com>
|
👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review. Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed. |
dsikka
reviewed
Mar 17, 2025
|
@brian-dellabetta The down_proj is the most likely to fail hessian inversion, since it is the weight with the largest input size. This is likely a normal hessian inevitability issue, which can be fixed by shuffling the dataset/ using an image dataset |
Signed-off-by: Brian Dellabetta <bdellabe@redhat.com>
Signed-off-by: Brian Dellabetta <bdellabe@redhat.com>
|
Hitting several issues around reproducibility and slowness of lm-eval (takes about a minute for each of the 30 samples to run). Will continue this next week after resolving more urgent tasks |
dsikka
pushed a commit
that referenced
this pull request
Mar 21, 2025
SUMMARY: multi-modal lm-eval tests are failing due to a non-reproducibility issue that still needs to be resolved. In the meantime, moving those tests to a skipped folder until resolution. Resolution can be tracked in #1260 TEST PLAN: no new source code Signed-off-by: Brian Dellabetta <bdellabe@redhat.com>
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Outdated suggestions cannot be applied.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.
SUMMARY:
lm-eval multimodal tests were failing to reproduce across different versions of compressed tensors. After upgrading the models from 2B to 7B, the tests appear to be reproducing across compressed tensors 0.9.1, 0.9.2 and nightly. I ran extensively for the fp8 config across different versions of CT, and it always returned the same result.
I also removed the random seed from the configs. after running several of each of the 3 configs, i did not see any change in result. this may cause errors during ci/cd testing but I'd like to see if it does, i feel that is a better e2e test anyway.
The Qwen2.5 7B model was erroring out during GPTQ with default
dampening_frac=0.01, during Hessian calculation of themodel.layers[_].mlp.down_projmatrix which is aLinear(in_features=18944, out_features=3584, bias=False)layer. This happened regardless of shuffling, but works consistently withdampening_frac=0.1. The llava model runs fine withdampening_frac=0.01.Tests take a long time to run, even with just 30 eval samples it can take 30 minutes to run the model in
vl_w4a16_actorder_weight.yaml(~50 minutes total with compression), compared to a couple minutes it takes to run the dense models through lm-eval directly. This will add roughly 2 hours to the weekly testing runtime. Is that expected and acceptable?TEST PLAN:
no new src code, just fixing tests