evaluate
evaluate copied to clipboard
huggingface
Import issue
When I was using "import evaluate" library in Kaggle notebook. The import automatically fulled 15,4 GB of GPU memory, and I could not further train the model due to lack of memory. This happened only when I was using this import, so there must be something wrong with it.
This might be a tensorflow issue, but should have been fixed in https://github.com/huggingface/evaluate/commit/9b6cea3ef7f84dd619d65a6fd5a7da07f2386fae afaik.
What version of evaluate are you using?
Hi @LecbychMichal could you check whether this issue was resolved with the latest version? If not could you provide the version you are using?
Hi, I'm facing the same issue. It fills 9970MiB / 11264MiB (examined by nvidia-smi) just after importing it. On two computers I have evaluate=0.2.2 and this issue arises on only one of them.
I have a suspicion that it has to do something with hdf5 though I'm not sure. How to trace this issue?
Indeed, looks like there was some mismatch with hdf5 (or some other libraries?). Once I loaded the version 1.10.1, I suddenly didn't get a full GPU. I don't have any closer information but it may still help those in the future wondering why their GPU is getting full just after import evaluate.
I'm facing same issue, simply importing evaluate will consume all my GPU memory. The version of evaluate is 0.2.2
Code:
nvidia-smi report:
Closing this for now, let me know if you are still facing the issue with 0.3.0