pytorch

8ca1674c - trt engine inspector demo (#66683)

trt engine inspector demo (#66683) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/66683 Starting from TensorRT 8.2, we have this nice engine inspector which gives you much details of trt layer. Test Plan: ``` buck run mode/opt -c python.package_style=inplace scripts/yinghai/test:trt_engine_inspector ``` And you will see something like ``` {"Layers": [{ "Name": "PWN(PWN(relu_1), add_1)", "LayerType": "PointWiseV2", "Inputs": [ { "Name": "x", "Dimensions": [10,2], "Format/Datatype": "Row major linear FP16 format" }], "Outputs": [ { "Name": "(Unnamed Layer* 1) [ElementWise]_output", "Dimensions": [10,2], "Format/Datatype": "Row major linear FP16 format" }], "ParameterType": "PointWise", "ParameterSubType": "PointWiseExpression", "NbInputArgs": 1, "InputArgs": ["arg0"], "NbOutputVars": 1, "OutputVars": ["var1"], "NbParams": 0, "Params": [], "NbLiterals": 4, "Literals": ["0.000000e+00f", "1.000000e+00f", "0.000000e+00f", "0.000000e+00f"], "NbOperations": 2, "Operations": ["const auto var0 = pwgen::iMax(arg0, literal0);", "const auto var1 = pwgen::iPlus(arg0, var0);"], "TacticValue": "0x0" },{ "Name": "matmul_1", "LayerType": "MatrixMultiply", "Inputs": [ { "Name": "(Unnamed Layer* 1) [ElementWise]_output", "Dimensions": [10,2], "Format/Datatype": "Row major linear FP16 format" }, { "Name": "y", "Dimensions": [10,2], "Format/Datatype": "Row major linear FP16 format" }], "Outputs": [ { "Name": "output0", "Dimensions": [10], "Format/Datatype": "Row major linear FP16 format" }], "ParameterType": "MatrixMultiply", "MatrixOpA": "VECTOR", "MatrixOpB": "VECTOR", "Alpha": 1, "Beta": 0, "TacticValue": "0x1" }], "Bindings": ["x" ,"y" ,"output0" ]} ``` Reviewed By: RoshanPAN, wushirong Differential Revision: D31681405 fbshipit-source-id: 31f912c37812ac17c6421073e0c35e512463ba6e