The kernel appears to have died at the `CatBoost` fitting stage if incorrect `gpu_id` passed

[andrewsonin]

If I have only two available GPUs, I can still pass gpu_id='2' to the TabularAutoML without getting an error.

nvidia-smi output:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 465.19.01    Driver Version: 465.19.01    CUDA Version: 11.3     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA Tesla V1...  Off  | 00000000:3B:00.0 Off |                    0 |
| N/A   44C    P0    36W / 250W |   1581MiB / 32510MiB |     13%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  NVIDIA Tesla V1...  Off  | 00000000:86:00.0 Off |                    0 |
| N/A   47C    P0    47W / 250W |   1249MiB / 32510MiB |    100%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

Code:

automl = TabularAutoML(
    task=Task('multiclass', ...),
    ...,
    general_params={'use_algos': [['linear_l2', 'cb', 'cb_tuned']]},
    cb_params={'default_params': {'thread_count': 48}},
    gpu_ids='2',
    verbose=3
)
oof_predictions = automl.fit_predict(
    train,
    roles=...
)

The kernel appears to have died at the fit_predict stage when it tries to train the CatBoost model.

Changing gpu_ids='2' to gpu_ids='1' solves the problem.

Please add the corresponding check to the __init__ of the TabularAutoML.

[CrustaceanJ]

Hi, @andrewsonin. Thank you for your interest in our library.

We are planning to add a separate setup for NLP, CV and reporting (may be others) blocks. However, some will not have PyTorch from which one could easily snatch GPU IDs (as we do now, if you don't pass gpu_ids). This seems to be solved through parsing aforenamed nvidia-smi. I think that along with the pull request for a separate installation, we will add the exception you proposed.

[github-actions]

Stale issue message