{
  "id": 416967,
  "title": "TPU problem",
  "url": "/competitions/benetech-making-graphs-accessible/discussion/416967",
  "author_name": "",
  "post_date": "2023-06-13T15:52:35.314639900Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Has anyone tried running training with tpu? There is no import of packages (just installed), the same transformers for example. After assigning the device, there is no error, but nothing more than a NameError. Are there any known ways to solve the problem?</p>",
  "messages": [
    {
      "id": "2301079",
      "postDate": "06/13/2023 15:52:35",
      "content": "<p>Has anyone tried running training with tpu? There is no import of packages (just installed), the same transformers for example. After assigning the device, there is no error, but nothing more than a NameError. Are there any known ways to solve the problem?</p>",
      "rawMarkdown": "Has anyone tried running training with tpu? There is no import of packages (just installed), the same transformers for example. After assigning the device, there is no error, but nothing more than a NameError. Are there any known ways to solve the problem?",
      "votes": null
    },
    {
      "id": "2302183",
      "postDate": "06/14/2023 11:49:58",
      "content": "<p>I have tried but get different errors than you.  I could not get it to run with more than 1 process.  I do hope after this competition is over someone will publish code on how to train on a TPU, as this was my first attempt.</p>",
      "rawMarkdown": "I have tried but get different errors than you.  I could not get it to run with more than 1 process.  I do hope after this competition is over someone will publish code on how to train on a TPU, as this was my first attempt.",
      "votes": null
    },
    {
      "id": "2302417",
      "postDate": "06/14/2023 14:38:29",
      "content": "<p>If you mean 1 core, then as I understand it, it should only work with one core. This is VM TPU. The option with 8 cores, as I understand it, was removed quite recently. Last night the error disappeared, it is possible that the problem was due precisely to the fact that instead of two TPU options, only one was left and there were problems associated with this change.</p>",
      "rawMarkdown": "If you mean 1 core, then as I understand it, it should only work with one core. This is VM TPU. The option with 8 cores, as I understand it, was removed quite recently. Last night the error disappeared, it is possible that the problem was due precisely to the fact that instead of two TPU options, only one was left and there were problems associated with this change.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2302183,
      "author_name": "danielphalen",
      "author_url": "",
      "post_date": "06/14/2023 11:49:58",
      "content": "<p>I have tried but get different errors than you.  I could not get it to run with more than 1 process.  I do hope after this competition is over someone will publish code on how to train on a TPU, as this was my first attempt.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2302417,
          "author_name": "zaakciiru",
          "author_url": "",
          "post_date": "06/14/2023 14:38:29",
          "content": "<p>If you mean 1 core, then as I understand it, it should only work with one core. This is VM TPU. The option with 8 cores, as I understand it, was removed quite recently. Last night the error disappeared, it is possible that the problem was due precisely to the fact that instead of two TPU options, only one was left and there were problems associated with this change.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2301079": "Has anyone tried running training with tpu? There is no import of packages (just installed), the same transformers for example. After assigning the device, there is no error, but nothing more than a NameError. Are there any known ways to solve the problem?",
    "2302183": "I have tried but get different errors than you.  I could not get it to run with more than 1 process.  I do hope after this competition is over someone will publish code on how to train on a TPU, as this was my first attempt.",
    "2302417": "If you mean 1 core, then as I understand it, it should only work with one core. This is VM TPU. The option with 8 cores, as I understand it, was removed quite recently. Last night the error disappeared, it is possible that the problem was due precisely to the fact that instead of two TPU options, only one was left and there were problems associated with this change."
  },
  "source": "meta"
}