[Bug] TVM produces wrong result with the "default" optimization pipeline

[coffezhou]

Expected behavior

For the following onnx model, the output of "v6_0" should be 0.

[array(518, dtype=int64), 
array([ True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True]), 
array(0, dtype=int64), 
array([25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
       25, 25, 25, 25, 25, 25, 25], dtype=uint8)]

Actual behavior

However, when we compile the onnx model using relax.build with the default optimization pipeline, tvm produces -1 of v6_0.

 (<tvm.nd.NDArray shape=(), cpu(0)>
array(518), <tvm.nd.NDArray shape=(24,), cpu(0)>
array([ True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True,  True]), 
<tvm.nd.NDArray shape=(), cpu(0)>
array(-1), 
<tvm.nd.NDArray shape=(24,), cpu(0)>
array([25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
       25, 25, 25, 25, 25, 25, 25], dtype=uint8))

Environment

OS: Ubuntu 20.04 TVM: 0.20.dev0 (f6236ce41)

Steps to reproduce

This bug can be reproduced by the following code with the model in the attachment.

from typing import Dict, List, Literal, Optional

import numpy as np
import onnx
import onnxruntime
from onnx import ModelProto, TensorProto, helper, mapping

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

import pickle

def get_oracle(oracle_path):
    with open(oracle_path, 'rb') as f:
        oracle = pickle.load(f)
  
    return oracle['input'], oracle['output']

def check(
    model: ModelProto,
    inputs: Optional[Dict[str, np.ndarray]] = None,
) -> None:

    # Run the model through onnx to get the expected result.
    try:
        ort_session = onnxruntime.InferenceSession(
            model.SerializeToString(), providers=["CPUExecutionProvider"]
        )
        ort_output = ort_session.run([], inputs)
    except:
        print("This model cannot be executed by onnxruntime!")
        sys.exit(1)
        
    print("onnxrumtime: ", ort_output)    
    # Convert the onnx model into relax through the onnx importer.
    tvm_model = from_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)
    
    # Prepare inputs.
    input_list = [
        inputs[key.name_hint] for key in tvm_model["main"].params if key.name_hint in inputs
    ]
    if params:
        input_list += params["main"]
        
    # Compile the relax graph into a VM then run.
    #----------------------cpu-----------------------
    with tvm.transform.PassContext(opt_level=0):
        
        ex = relax.build(tvm_model, target="llvm", relax_pipeline="default")
        vm = relax.VirtualMachine(ex, tvm.cpu())
    
        # Run model and check outputs.
        vm.set_input("main", *input_list)
        vm.invoke_stateful("main")
        tvm_cpu_output = vm.get_outputs("main")
    #----------------------cpu-----------------------
    print("tvm cpu", tvm_cpu_output)

            
def main(model_path, oracle_path):
    onnx_model = onnx.load(model_path)
    inputs, outputs = get_oracle(oracle_path)
    
    check(onnx_model, inputs=inputs)
    
main("model.onnx", "oracle.pkl")

testcase.zip

Triage

Please refer to the list of label tags here to find the relevant tags and add them below in a bullet format (example below).

• needs-triage