{
  "id": 209456,
  "title": "results are different between inference by cpu and gpu",
  "url": "/competitions/riiid-test-answer-prediction/discussion/209456",
  "author_name": "",
  "post_date": "2021-01-07T16:18:57.342957600Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I was trying to inference the public sakt model using cpu and gpu separately. Then I checked the output submission file, and I saw the different results. So, is it the test_dfs in iter are different every time or the inferences between cpu and gpu are different? </p>",
  "messages": [
    {
      "id": "1142847",
      "postDate": "01/07/2021 16:18:57",
      "content": "<p>I was trying to inference the public sakt model using cpu and gpu separately. Then I checked the output submission file, and I saw the different results. So, is it the test_dfs in iter are different every time or the inferences between cpu and gpu are different? </p>",
      "rawMarkdown": "I was trying to inference the public sakt model using cpu and gpu separately. Then I checked the output submission file, and I saw the different results. So, is it the test_dfs in iter are different every time or the inferences between cpu and gpu are different?",
      "votes": null
    },
    {
      "id": "1142885",
      "postDate": "01/07/2021 16:36:59",
      "content": "<p>I think it is just the difference between doing fp32 and fp64, and difference accumulates as the model grows bigger.</p>",
      "rawMarkdown": "I think it is just the difference between doing fp32 and fp64, and difference accumulates as the model grows bigger.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1142885,
      "author_name": "scaomath",
      "author_url": "",
      "post_date": "01/07/2021 16:36:59",
      "content": "<p>I think it is just the difference between doing fp32 and fp64, and difference accumulates as the model grows bigger.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1142847": "I was trying to inference the public sakt model using cpu and gpu separately. Then I checked the output submission file, and I saw the different results. So, is it the test_dfs in iter are different every time or the inferences between cpu and gpu are different?",
    "1142885": "I think it is just the difference between doing fp32 and fp64, and difference accumulates as the model grows bigger."
  },
  "source": "meta"
}