Uploading Docker image (via code submission) to TIRA is slow

PAN
1

Hello! I am trying to make a code submission for the multi-author style analysis shared task. I have two questions.

  1. My Docker image runs correctly on my local machine, but it has been taking a long time to upload to the site. I suspect it might be because the image is ~20GB, but I don’t know if something else is going wrong. Is there an expected Docker image size, and how long should pushing it (after all local inference is completed) take?

  2. As a test run, I submitted a smaller Docker image (with fewer files) and that successfully uploaded to TIRA. I see it when I click on “Submit”. Is there something else I need to do after the submission uploads to the site? In addition, when my full pipeline uploads to TIRA, how do I indicate which submission I want to be used for the leaderboard?

Thank you so much!

2

Hi @denizbt,

Thanks for reaching out!
We do not have a strict upper limit for the size of Docker images, the only hard limit that we have is 15 GB per layer. (if it is big because of some hugging face models, we have a --mount-hf-model flag that allows to mount models from hugging face read-only, more details on this are described here.)

If the upload is too slow, it might be an option that you invite us to the github repository (my account is mam10eks), so that we build the image on our side and upload it (also, please note that we can extend the deadline on an individual basis to ensure all your submissions are valid and work).

After the software is uploaded, please run it (for the multi-author analysis task) on the spot-check dataset and on the test dataset. The organizers will later review the outputs and will contact you in case there is some problem.

Please describe the approaches that you want to have on the leaderboard in your notebook paper that you will submit to easychair (deadline may 30), the organizers will review this and use this to write their overview paper which will then inform the leaderboard. (So it is no problem if you have submissions that you do not describe in your notebook, e.g., because they are re-uploads of the baseline to ensure everything works.)

I hope this answers your questions?

Best regards,

Maik

3

Hi! Thank you so much, that answers most of my questions. I will try the --mount-hf-model flag.

Another question I have is that I was able to submit a smaller version of my code but when I click “Run”, it doesn’t work, I get this output “Job did not run. (Maybe it is still submitted to the cluster or failed to start. It might take up to 5 minutes to submit a Job to the cluster.)”

I have tried re-running it multiple times. Is there a way for me to check whether the issue is with my submission or is the cluster simply busy?

Thanks!
Deniz

4

Hi, can look into this later, it might be some special case, e.g., that the docker image is build for a wrong architecture.

1 Like