coffezhou opened a new issue, #18600: URL: https://github.com/apache/tvm/issues/18600
### Expected behavior TVM should run the model correctly. ### Actual behavior For the following simple model, <img width="890" height="580" alt="Image" src="https://github.com/user-attachments/assets/cb3b5dd8-c2dd-43f8-b31e-b9b11e60622e"; /> it can be executed by onnxruntime and onnx's ReferenceEvaluator, the results are as follows: ``` [[[[-14.942039 25.242477 12.100965 ... 57.370975 43.9531 2.7949858 ] [ 10.37292 -10.740159 -38.21392 ... -25.30576 20.154476 -19.093735 ] [ 23.041002 41.820953 -34.108677 ... 13.982165 7.2081375 11.132856 ] ... [ -5.590187 -30.208424 9.572759 ... 32.32894 10.533618 -10.12186 ] [-11.91229 29.118868 -20.662777 ... -16.910463 -3.0499296 -8.469831 ] [ -1.340055 -17.950693 1.7485574 ... 44.479458 11.766148 -29.151772 ]] [[ 3.5306716 -17.110134 -12.246135 ... 3.6663613 10.713657 18.570526 ] [ 14.867607 -7.739101 -32.685593 ... -8.97093 -14.347578 32.1913 ] [ 7.462762 -14.961873 -18.268976 ... -16.221258 9.145517 22.871971 ] ... [ 58.016407 -28.310326 27.333632 ... 33.982685 -0.22384354 44.509457 ] [-27.212149 -17.831676 7.140829 ... 10.431597 33.523167 -36.84701 ] [ 4.2589498 -18.459444 9.673733 ... -21.87571 11.560403 -11.319146 ]] [[ 1.6221294 -24.357258 37.832687 ... -15.675017 17.112104 -19.541176 ] [ -1.9535408 -43.395298 -7.589774 ... 32.80785 39.018574 -9.826303 ] [ 23.802937 -22.675396 2.9990005 ... -29.012583 55.184933 16.124914 ] ... [ 13.447808 1.2474155 68.31904 ... -33.191654 7.361887 -14.480252 ] [ 8.724002 -44.378754 -3.259966 ... -43.776226 47.07514 4.846405 ] [-28.980995 0.75377727 -20.248524 ... -13.6773405 34.299976 -5.997253 ]] ... [[-33.108776 -16.460379 7.451367 ... -3.0716655 8.792357 9.199628 ] [-10.372536 -22.135944 -21.878239 ... 55.7764 -7.191108 -14.206541 ] [-53.47607 10.408365 -20.441448 ... 23.109488 1.2017591 6.3908386 ] ... [ -0.9444683 -29.720179 -36.938797 ... -32.23233 -35.5812 42.359478 ] [-28.41974 -0.1681884 -38.875988 ... 14.968706 -23.108547 -32.480797 ] [ 2.9518757 1.1772063 -21.35058 ... -24.29589 1.791322 -2.575415 ]] [[ -1.1625404 -6.98105 -10.189097 ... -26.975689 -14.790653 -8.432305 ] [ 17.315321 34.95646 -33.681236 ... -8.27439 9.273292 -20.079172 ] [ -9.413344 25.616909 -8.439555 ... -17.31526 18.547462 -18.068047 ] ... [ 0.11532002 -46.406166 -29.583313 ... -17.339268 17.583694 38.199493 ] [-49.912685 10.065314 -67.41777 ... -17.945366 -17.426779 24.320415 ] [ -9.248135 4.745124 -5.5740147 ... 33.417686 -3.3946886 4.773201 ]] [[-11.940619 1.2533135 -15.480752 ... 17.695894 23.379114 -6.3436594 ] [ -2.8251894 -1.4625757 12.186805 ... -25.161877 6.0388117 -8.681617 ] [ 37.58026 7.1358566 -53.616978 ... 49.06318 -9.981028 15.518544 ] ... [-19.931892 23.587257 -68.27191 ... -25.605322 -31.568523 0.6990657 ] [ 56.762306 -4.785701 3.260337 ... 2.4171884 -50.66727 -23.977222 ] [ 13.220358 -13.136884 18.563856 ... -9.229172 -32.588108 27.032162 ]]]] ``` However, the onnx frontend of TVM cannot import it: ``` File "/home/ubuntu/Documents/test1.py", line 49, in test tvm_model = from_onnx(onnx_model, opset=14, keep_params_in_input=True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/Documents/DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4260, in from_onnx return g.from_onnx(graph, opset) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/Documents/DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3890, in from_onnx self._construct_nodes(graph) File "/home/ubuntu/Documents/DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4071, in _construct_nodes raise err File "/home/ubuntu/Documents/DLCompilers/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 4068, in _construct_nodes op = self.bb.normalize(op) ^^^^^^^^^^^^^^^^^^^^^ File "/home/ubuntu/Documents/DLCompilers/tvm/python/tvm/relax/block_builder.py", line 672, in normalize return _ffi_api.BlockBuilderNormalize(self, expr) # type: ignore ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "python/tvm_ffi/cython/function.pxi", line 904, in tvm_ffi.core.Function.__call__ File "<unknown>", line 0, in tvm::relax::Normalizer::Normalize(tvm::RelaxExpr const&) File "<unknown>", line 0, in tvm::relax::ExprFunctor<tvm::RelaxExpr (tvm::RelaxExpr const&)>::VisitExpr(tvm::RelaxExpr const&) File "<unknown>", line 0, in tvm::relax::ExprFunctor<tvm::RelaxExpr (tvm::RelaxExpr const&)>::InitVTable()::{lambda(tvm::ffi::ObjectRef const&, tvm::relax::ExprFunctor<tvm::RelaxExpr (tvm::RelaxExpr const&)>*)#9}::_FUN(tvm::ffi::ObjectRef const&, tvm::relax::ExprFunctor<tvm::RelaxExpr (tvm::RelaxExpr const&)>*) File "<unknown>", line 0, in tvm::relax::Normalizer::VisitExpr_(tvm::relax::CallNode const*) File "<unknown>", line 0, in tvm::relax::Normalizer::InferStructInfo(tvm::relax::Call const&) File "<unknown>", line 0, in tvm::relax::InferStructInfoBroadcastArith(tvm::relax::Call const&, tvm::relax::BlockBuilder const&) File "<unknown>", line 0, in tvm::relax::StructInfo tvm::relax::InferStructInfoBroadcast<tvm::runtime::DataType (*)(tvm::relax::Call const&, tvm::relax::BlockBuilder const&, tvm::relax::StructInfo const&, tvm::relax::StructInfo const&)>(tvm::relax::Call const&, tvm::relax::BlockBuilder const&, tvm::runtime::DataType (*)(tvm::relax::Call const&, tvm::relax::BlockBuilder const&, tvm::relax::StructInfo const&, tvm::relax::StructInfo const&)) File "<unknown>", line 0, in tvm::relax::InferBinaryBroadcastShape(tvm::relax::Call const&, tvm::relax::BlockBuilder const&, tvm::ffi::Array<tvm::PrimExpr, void> const&, tvm::ffi::Array<tvm::PrimExpr, void> const&) File "<unknown>", line 0, in tvm::relax::BlockBuilderImpl::ReportFatal(tvm::Diagnostic const&) File "<unknown>", line 0, in tvm::runtime::detail::LogFatal::Entry::Finalize() tvm.error.InternalError: In Op(relax.add), the first input shape at dim 1 is T.int64(16) and the second input shape at dim 1 is T.int64(32), which are not broadcastable. [11:03:35] /home/ubuntu/Documents/DLCompilers/tvm/src/relax/ir/block_builder.cc:64: Warning: BlockBuilder destroyed with remaining blocks! ``` ### Environment OS: Ubuntu 20.04 TVM: 0.23.dev0 (https://github.com/apache/tvm/commit/f4e28d3153323ad97a7e74740c9fb22300fd6cd0) onnxruntime: 1.23.2 ### Steps to reproduce This bug can be reproduced by the following code with the model in the attachment. ```python from typing import Dict, List, Literal, Optional import sys import os import numpy as np import onnx from onnx.reference import ReferenceEvaluator import onnxruntime from onnx import ModelProto, TensorProto, helper import tvm import tvm.testing from tvm import relax from tvm.relax.frontend.onnx import from_onnx import argparse import pickle def test() -> None: onnx_model = onnx.load("11.onnx") # Configure model format. onnx_model.ir_version = 8 onnx_model.opset_import[0].version = 14 with open("inputs.pkl", 'rb') as fp: inputs = pickle.load(fp) # Run the model through onnx to get the expected result. try: ort_session = onnxruntime.InferenceSession( onnx_model.SerializeToString(), providers=["CPUExecutionProvider"] ) ort_output = ort_session.run([], inputs) except Exception as e: print(e) print("This model cannot be executed by onnxruntime!") sys.exit(1) print("onnxruntime:", ort_output[0]) # ReferenceEvaluator sess = ReferenceEvaluator("11.onnx") re_output = sess.run(None, inputs) print("ReferenceEvaluator:", re_output[0]) tvm.testing.assert_allclose(re_output[0], ort_output[0], rtol=0.1, atol=0.1) # TVM tvm_model = from_onnx(onnx_model, opset=14, keep_params_in_input=True) if __name__ == "__main__": test() ``` [testcase.zip](https://github.com/user-attachments/files/24303049/testcase.zip) ### Triage Please refer to the list of label tags [here](https://github.com/apache/tvm/wiki/Issue-Triage-Labels) to find the relevant tags and add them below in a bullet format (example below). * needs-triage -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
