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.

[arya0000001]

@stevemessick
I'm encountering this same issue. In my case, T4 is more important than other GPUs in kaggle. I really need to use API for T4. Do you know when this bug will be fixed? I noticed that someone already reported it a long time ago, back in 2024...

[hp0404]

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.

In the notebook, I set GPU to T4 x2:

then I pulled this notebook and "pushed" it which started the run:

So the CLI call overwrote the settings.

The notebook has "accelerator" key set to T4:

[morizin]

anybody with workaround

[hp0404]

anybody with workaround

nope, I'm starting the task manually once a day from the web

[morizin]

The latest api includes an additional keyword in the json on the number compute resources. I guess its some matter of time they can get it done

[psinger]

This is extremely frustrating, model_sources also doesnt seem to work.
At least make 2xt4 the default please

[morizin]

This is extremely frustrating, model_sources also doesnt seem to work. At least make 2xt4 the default please

🙌

[stevemessick]

This was fixed in #907, #916, #918. We plan to publish a new point release with these changes soon. The new option is --accelerator, which can be abbreviated to --acc (or perhaps shorter).