{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "\n\n# Debugging Torch-TensorRT Compilation\n\nTensorRT conversion can perform many graph transformations and backend specific\noptimizations that are sometimes hard to inspect. Torch-TensorRT provides a\nDebugger utility to help visualize FX graphs around lowering passes, monitor\nengine building, and capture profiling or TensorRT API traces.\n\nIn this example, we demonstrate how to:\n\n    1. Enable the Torch-TensorRT Debugger context\n    2. Capture and visualize FX graphs before and/or after specific lowering passes\n    3. Configure logging directory and verbosity\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "import os\nimport tempfile\n\nimport numpy as np\nimport torch\nimport torch_tensorrt as torch_trt\nimport torchvision.models as models\n\ntemp_dir = os.path.join(tempfile.gettempdir(), \"torch_tensorrt_debugger_example\")\n\nnp.random.seed(0)\ntorch.manual_seed(0)\ninputs = [torch.rand((1, 3, 224, 224)).to(\"cuda\")]\n\n\nmodel = models.resnet18(pretrained=False).to(\"cuda\").eval()\nexp_program = torch.export.export(model, tuple(inputs))\nenabled_precisions = {torch.float}\nworkspace_size = 20 << 30\nmin_block_size = 0\nuse_python_runtime = False\ntorch_executed_ops = {}\n\nwith torch_trt.dynamo.Debugger(\n    log_level=\"debug\",\n    logging_dir=temp_dir,\n    engine_builder_monitor=False,  # whether to monitor the engine building process\n    capture_fx_graph_after=[\n        \"complex_graph_detection\"\n    ],  # fx graph visualization after certain lowering pass\n    capture_fx_graph_before=[\n        \"remove_detach\"\n    ],  # fx graph visualization before certain lowering pass\n):\n\n    trt_gm = torch_trt.dynamo.compile(\n        exp_program,\n        tuple(inputs),\n        use_python_runtime=use_python_runtime,\n        enabled_precisions=enabled_precisions,\n        min_block_size=min_block_size,\n        torch_executed_ops=torch_executed_ops,\n        immutable_weights=False,\n        reuse_cached_engines=False,\n    )\n\n    trt_output = trt_gm(*inputs)\n\n\n\"\"\"\nThe logging directory will contain the following files:\n- /tmp/torch_tensorrt_debugger_example/\n    torch_tensorrt_logging.log\n    - /lowering_passes_visualization/\n        after_complex_graph_detection.svg\n        before_remove_detach.svg\n\"\"\""
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.11.14"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}