hyenal

Did you try to change the `SMVERSIONFLAGS` in common.mk to sm_20 ? Actually support for sm_12 architecture was dropped in CUDA 7.0 so it might a reason why you get...

I would like to add support for this feature as well for the reason mentioned by @myaseen-oraclase. I work with generative models and this would make evolution tracking easier. This...

I could reproduce the error on an Apple M3 Pro chip and `torch==2.2.1`. The issue boils down to an error in torch at [this line](https://github.com/Lightning-AI/torchmetrics/blob/master/src/torchmetrics/functional/classification/precision_recall_curve.py#L69). It turns out that padding...

To make things faster I tried running on a GPU ```bash RUN_SLOW=1 pytest tests/models/audio_spectrogram_transformer/ tests/models/deit/ tests/models/videomae/ tests/models/vision_encoder_decoder/ tests/models/vision_text_dual_encoder/ tests/models/vit/ tests/models/vit_mae/ tests/models/vit_msn/ tests/models/yolos/ ``` So far I am getting a few...

@amyeroberts I rebased and ran the pipeline as indicated. The last one should have failed (I know yolo and the encoder/decoder are not ready yet) so I am not sure...

Thanks you @amyeroberts I will fix the tests then and request a new SLOW run when things are fixed :)

@amyeroberts when you have some time could you run the latest slow run I pushed ? I fixed most of the issues but there are 3 failures (`ViT2BertModelTest.test_real_model_save_load_from_pretrained` , `NougatModelIntegrationTest.test_forward_pass`,...

The MR is now ready, 3 slow tests are failing but I am unable to find the source of it (a precision error due to SDPA ?) if possible I...

@amyeroberts I am afraid that I cannot find a direct link between this PR and the current failures: - `tests/models/vision_encoder_decoder/test_modeling_vision_encoder_decoder.py::ViT2BertModelTest::test_real_model_save_load_from_pretrained`: these tests also fails for me on `main`. It seems...

> Last thing to do is add performance numbers for the models e.g. [like here for Mistral](https://huggingface.co/docs/transformers/v4.40.2/en/model_doc/mistral#expected-speedups). It's not necessary to run for all of the models (although this would...