How to enable GPU T4 x 2 in API?

[Vander-Bilt]

kaggle-api is great.
But how to enable GPU T4 x 2 in API?

[stevemessick]

@Vander-Bilt Thanks!

Sorry to say, that isn't implemented yet. A work-around is to use kaggle.com to enable TPU in the notebook. Then go back to your CLI to get the updated kernel with kaggle kernels pull .... Then the setting will be preserved.

@jplotts FYI

[Vander-Bilt]

Sorry, let me clarify my question. Currently I use api.kernels_push... in my Python code to start my notebook with config file kernel-metadata.json as below:

{
  "id": "robaaay/onetest",
  "title": "onetest,
  "code_file": "onetest.ipynb",
  "language": "python",
  "kernel_type": "notebook",
  "is_private": true,
  "enable_gpu": true,
  "enable_internet": true,
  "dataset_sources": [],
  "competition_sources": [],
  "kernel_sources": [],
  "model_sources": []
}

You can see, "enable_gpu": true, as result, the notebook used P100 GPU as default, but I want to use T4 GPU for my notebook execution.

[stevemessick]

Using the work-around I mentioned, I see that the GPU type is saved in the .ipynb file as "accelerator":"nvidiaTeslaT4".

Use kaggle kernels pull stevemessick/blank -m to get the files for my kernel, then try modifying your own kernel files to see if you can get it to use T4s.

It might be easier to just use the work-around on your own kernel to get started. Good luck!

[Vander-Bilt]

Thanks @stevemessick

I did more tests, looks like it issue is still there.

GPU type nvidiaTeslaT4 is there in my .ipynb file:
"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[],"dockerImageVersionId":31041,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}}

If GPU type and TPU type are set in kernel-metadata.json as below:

  "enable_gpu": false,
  "enable_tpu": true,

Then I can see the information on kaggle.com after push the kernel:

blank
Version #5 with TPU VM v3-8
Successful

It works for switching to TPU.

If GPU type and TPU type are set in kernel-metadata.json as below:

  "enable_gpu": true,
  "enable_tpu": false,

Then I can see the information on kaggle.com after push the kernel:

blank
Version #4 with GPU P100
Successful

Looks like it not working for switching to GPU T4, it just keeping default GPU P100, whatever you set in .ipynb file.

[stevemessick]

Thanks, you've helped discover a bug in the server code. We understand what's wrong, and we'll get it fixed in the future, but it may be awhile due to other arcs of work in progress.