[Bug] LegalizeOps fails on valid onnx model: AssertionError: emit_te now only supports Tensor that has ShapeExpr shape

[coffezhou]

Expected behavior

TVM should compile the model correctly.

Actual behavior

When compiling the model with the default relax optimization pipeline, TVM crashes as follows:

Traceback (most recent call last):
  File "/home/carla/Documents/test/test.py", line 51, in <module>
    main()
  File "/home/carla/Documents/test/test.py", line 45, in main
    ex = relax.build(tvm_model, target="llvm", relax_pipeline=relax_pipeline)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/carla/Documents/tvm/python/tvm/relax/vm_build.py", line 253, in build
    mod = relax_pipeline(mod)
          ^^^^^^^^^^^^^^^^^^^
  File "/home/carla/Documents/tvm/python/tvm/ir/transform.py", line 238, in __call__
    return _ffi_transform_api.RunPass(self, mod)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "tvm/ffi/cython/./function.pxi", line 212, in tvm.ffi.core.Function.__call__
  File "tvm/ffi/cython/./function.pxi", line 265, in tvm.ffi.core.tvm_ffi_callback
  File "/home/carla/Documents/tvm/python/tvm/relax/backend/cpu_generic/pipeline.py", line 73, in _pipeline
    mod = seq(mod)
  File "/home/carla/Documents/tvm/python/tvm/ir/transform.py", line 238, in __call__
    return _ffi_transform_api.RunPass(self, mod)
  File "tvm/ffi/cython/./function.pxi", line 212, in tvm.ffi.core.Function.__call__
  File "tvm/ffi/cython/./function.pxi", line 265, in tvm.ffi.core.tvm_ffi_callback
  File "/home/carla/Documents/tvm/python/tvm/relax/transform/legalize_ops/nn.py", line 509, in _nn_gelu
    return bb.call_te(te_gelu, call.args[0], primfunc_name_hint="gelu")
  File "/home/carla/Documents/tvm/python/tvm/relax/block_builder.py", line 356, in call_te
    tir_func, call_args, output_sinfo, tir_vars = gen_call_tir_inputs(func, *args, **kwargs)
  File "/home/carla/Documents/tvm/python/tvm/relax/utils.py", line 351, in gen_call_tir_inputs
    te_args = _convert_te_arg(args)
  File "/home/carla/Documents/tvm/python/tvm/relax/utils.py", line 289, in _convert_te_arg
    new_arg = _convert_te_arg_helper(te_args)
  File "/home/carla/Documents/tvm/python/tvm/relax/utils.py", line 273, in <genexpr>
    return tuple(_convert_te_arg_helper(x) for x in arg)
  File "/home/carla/Documents/tvm/python/tvm/relax/utils.py", line 273, in <genexpr>
    return tuple(_convert_te_arg_helper(x) for x in arg)
  File "/home/carla/Documents/tvm/python/tvm/relax/utils.py", line 224, in _convert_te_arg_helper
    assert isinstance(
AssertionError: emit_te now only supports Tensor that has ShapeExpr shape

Environment

OS: Ubuntu 20.04 TVM: 0.21.dev0(bcb68b130) CUDA: 11.8

Steps to reproduce

This bug can be reproduced by the following code with the model in the attachment. As shown in the code, the model can be executed by onnxruntime. However, tvm failed to compile this model with the default pipeline.

import sys

import numpy as np
import onnx
import onnxruntime

import tvm
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx

import pickle

            
def main():
    onnx_model = onnx.load("a674.onnx")
    
    with open("inputs.pkl", "rb") as fp:
        inputs = pickle.load(fp)
    
    try:
        ort_session = onnxruntime.InferenceSession(
            onnx_model.SerializeToString(), providers=["CPUExecutionProvider"]
        )
        ort_output = ort_session.run([], inputs)
    except Exception as e:
        print(e)
        sys.exit(1)
    print(ort_output)    
    # Convert the onnx model into relax through the onnx importer.
    tvm_model = from_onnx(onnx_model, keep_params_in_input=True)
    # Convert operators for inference mode.
    tvm_model = relax.transform.DecomposeOpsForInference()(tvm_model)
    # Legalize any relax ops into tensorir.
    tvm_model = relax.transform.LegalizeOps()(tvm_model)

    # Separate model from parameters.
    tvm_model, params = relax.frontend.detach_params(tvm_model)
        
    # Compile the relax graph into a VM then run.
    #----------------------cpu-----------------------
    with tvm.transform.PassContext(opt_level=0):
        target = tvm.target.Target("llvm", host="llvm")
        relax_pipeline = relax.pipeline.get_default_pipeline(target)
        
        ex = relax.build(tvm_model, target="llvm", relax_pipeline=relax_pipeline)
       
if __name__ == "__main__":
    
    main()
       

testcase.zip

Triage

• needs-triage

[vacu9708]

I think I've fixed the bug. I'll commit soon.