{
  "id": 583679,
  "title": "My Whitepaper: GRU-Based Runtime Prediction Pipeline using TPUGraphs Dataset",
  "url": "/competitions/predict-ai-model-runtime/discussion/583679",
  "author_name": "Suma Mallapragada",
  "post_date": "2025-06-08T14:01:34.231000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi all,</p>\n<p>I recently published a whitepaper titled <strong>\"Runtime Prediction of AI Model Operations Using a GRU-Based Neural Network\"</strong> based on my participation in this Kaggle competition.</p>\n<p>📄 Zenodo DOI: <a href=\"https://doi.org/10.5281/zenodo.15618294\" target=\"_blank\">https://doi.org/10.5281/zenodo.15618294</a>  </p>\n<p>This work builds directly on the TPUGraphs dataset provided in the contest, and proposes a GRU-based model instead of classification-based approaches. The model predicts runtime using:</p>\n<ul>\n<li>Opcode-weighted node feature scaling</li>\n<li>Edge dependency embeddings</li>\n<li>Configurable node flagging (for layout and tile)</li>\n<li>Configuration convolution</li>\n<li>Sequential modeling with GRUs</li>\n</ul>\n<p>🔍 Key difference: It performs <strong>regression</strong> to predict actual runtimes, rather than classify into runtime classes. This helps smooth out class boundary issues and handles close configurations better.</p>\n<p>I’d love to know if anyone has explored GRU/LSTM-based methods during the competition, or whether a sequential modeling idea could benefit future compiler prediction problems.</p>\n<p>Would appreciate feedback, benchmarking ideas, or collaboration suggestions.</p>",
  "messages": [
    {
      "id": 3219918,
      "postDate": "2025-06-08T14:01:34.233Z",
      "content": "<p>Hi all,</p>\n<p>I recently published a whitepaper titled <strong>\"Runtime Prediction of AI Model Operations Using a GRU-Based Neural Network\"</strong> based on my participation in this Kaggle competition.</p>\n<p>📄 Zenodo DOI: <a href=\"https://doi.org/10.5281/zenodo.15618294\" target=\"_blank\">https://doi.org/10.5281/zenodo.15618294</a>  </p>\n<p>This work builds directly on the TPUGraphs dataset provided in the contest, and proposes a GRU-based model instead of classification-based approaches. The model predicts runtime using:</p>\n<ul>\n<li>Opcode-weighted node feature scaling</li>\n<li>Edge dependency embeddings</li>\n<li>Configurable node flagging (for layout and tile)</li>\n<li>Configuration convolution</li>\n<li>Sequential modeling with GRUs</li>\n</ul>\n<p>🔍 Key difference: It performs <strong>regression</strong> to predict actual runtimes, rather than classify into runtime classes. This helps smooth out class boundary issues and handles close configurations better.</p>\n<p>I’d love to know if anyone has explored GRU/LSTM-based methods during the competition, or whether a sequential modeling idea could benefit future compiler prediction problems.</p>\n<p>Would appreciate feedback, benchmarking ideas, or collaboration suggestions.</p>",
      "rawMarkdown": "Hi all,\n\nI recently published a whitepaper titled **\"Runtime Prediction of AI Model Operations Using a GRU-Based Neural Network\"** based on my participation in this Kaggle competition.\n\n📄 Zenodo DOI: https://doi.org/10.5281/zenodo.15618294  \n\nThis work builds directly on the TPUGraphs dataset provided in the contest, and proposes a GRU-based model instead of classification-based approaches. The model predicts runtime using:\n\n- Opcode-weighted node feature scaling\n- Edge dependency embeddings\n- Configurable node flagging (for layout and tile)\n- Configuration convolution\n- Sequential modeling with GRUs\n\n🔍 Key difference: It performs **regression** to predict actual runtimes, rather than classify into runtime classes. This helps smooth out class boundary issues and handles close configurations better.\n\nI’d love to know if anyone has explored GRU/LSTM-based methods during the competition, or whether a sequential modeling idea could benefit future compiler prediction problems.\n\nWould appreciate feedback, benchmarking ideas, or collaboration suggestions.\n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3219918": "Hi all,\n\nI recently published a whitepaper titled **\"Runtime Prediction of AI Model Operations Using a GRU-Based Neural Network\"** based on my participation in this Kaggle competition.\n\n📄 Zenodo DOI: https://doi.org/10.5281/zenodo.15618294  \n\nThis work builds directly on the TPUGraphs dataset provided in the contest, and proposes a GRU-based model instead of classification-based approaches. The model predicts runtime using:\n\n- Opcode-weighted node feature scaling\n- Edge dependency embeddings\n- Configurable node flagging (for layout and tile)\n- Configuration convolution\n- Sequential modeling with GRUs\n\n🔍 Key difference: It performs **regression** to predict actual runtimes, rather than classify into runtime classes. This helps smooth out class boundary issues and handles close configurations better.\n\nI’d love to know if anyone has explored GRU/LSTM-based methods during the competition, or whether a sequential modeling idea could benefit future compiler prediction problems.\n\nWould appreciate feedback, benchmarking ideas, or collaboration suggestions.\n"
  }
}