{
  "id": 569982,
  "title": "bfloat16 for alpha3 clone proteinx",
  "url": "/competitions/stanford-rna-3d-folding/discussion/569982",
  "author_name": "",
  "post_date": "2025-03-25T09:51:27.287616200Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>alpha3 clone proteinx can run as bfloat16.<br>\nBut both kaggle GPU T4 and P100 does not support bf16 (they only supprt fp16)</p>\n<p>running in fp16 leads to nan.<br>\nanyone has any solution?</p>\n<p>e.g.</p>\n<ul>\n<li>you know of non nvidia cuda support for bfloat16 for old gpu</li>\n<li>way to finetune bf16 to fp16? or add some dynamic quantisation, etc</li>\n</ul>",
  "messages": [
    {
      "id": "3159118",
      "postDate": "03/25/2025 09:51:27",
      "content": "<p>alpha3 clone proteinx can run as bfloat16.<br>\nBut both kaggle GPU T4 and P100 does not support bf16 (they only supprt fp16)</p>\n<p>running in fp16 leads to nan.<br>\nanyone has any solution?</p>\n<p>e.g.</p>\n<ul>\n<li>you know of non nvidia cuda support for bfloat16 for old gpu</li>\n<li>way to finetune bf16 to fp16? or add some dynamic quantisation, etc</li>\n</ul>",
      "rawMarkdown": "alpha3 clone proteinx can run as bfloat16.\nBut both kaggle GPU T4 and P100 does not support bf16 (they only supprt fp16)\n\nrunning in fp16 leads to nan.\nanyone has any solution?\n\ne.g.\n- you know of non nvidia cuda support for bfloat16 for old gpu\n- way to finetune bf16 to fp16? or add some dynamic quantisation, etc",
      "votes": null
    },
    {
      "id": "3159678",
      "postDate": "03/25/2025 20:10:12",
      "content": "<p>Looks like it supports fp32 for inference?</p>\n<blockquote>\n  <p>dtype: data type used in inference. Valid options include \"bf16\" and \"fp32\".</p>\n</blockquote>",
      "rawMarkdown": "Looks like it supports fp32 for inference?\n>dtype: data type used in inference. Valid options include \"bf16\" and \"fp32\".",
      "votes": null
    },
    {
      "id": "3159941",
      "postDate": "03/26/2025 05:48:57",
      "content": "<p>we trace the code and NaN is the results of some matrix multilication. after some fix we can run proeinX on fp16 (instead bfp16)</p>",
      "rawMarkdown": "we trace the code and NaN is the results of some matrix multilication. after some fix we can run proeinX on fp16 (instead bfp16)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3159678,
      "author_name": "zacchaeus",
      "author_url": "",
      "post_date": "03/25/2025 20:10:12",
      "content": "<p>Looks like it supports fp32 for inference?</p>\n<blockquote>\n  <p>dtype: data type used in inference. Valid options include \"bf16\" and \"fp32\".</p>\n</blockquote>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3159941,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "03/26/2025 05:48:57",
      "content": "<p>we trace the code and NaN is the results of some matrix multilication. after some fix we can run proeinX on fp16 (instead bfp16)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3159118": "alpha3 clone proteinx can run as bfloat16.\nBut both kaggle GPU T4 and P100 does not support bf16 (they only supprt fp16)\n\nrunning in fp16 leads to nan.\nanyone has any solution?\n\ne.g.\n- you know of non nvidia cuda support for bfloat16 for old gpu\n- way to finetune bf16 to fp16? or add some dynamic quantisation, etc",
    "3159678": "Looks like it supports fp32 for inference?\n>dtype: data type used in inference. Valid options include \"bf16\" and \"fp32\".",
    "3159941": "we trace the code and NaN is the results of some matrix multilication. after some fix we can run proeinX on fp16 (instead bfp16)"
  },
  "source": "meta"
}