myelin::ir::tensor_t*& myelin::ir::operand_t::tensor()

[quintetoy]

I have got a problem like this,I was running pointpillars model-- pillarscatter, and when i transform an onnx to tensorrt ,this occured.
python: /root/gpgpu/MachineLearning/myelin/src/compiler/./ir/operand.h:166: myelin::ir::tensor_t*& myelin::ir::operand_t::tensor(): Assertion `is_tensor()' failed.

who can tell me how to solve this problem?

my environment :
cuda:10.2
tensorrt:8.2

[quintetoy]

pointpillars scatter ops caused this problem,does anyone tried this?

[yasserkhalil93]

Facing the same error. I have used index_select and index_put in my pytorch code which successfully translated to onnx. However, gave me the following when creating tensorrt:

python: /root/gpgpu/MachineLearning/myelin/src/compiler/./ir/operand.h:166: myelin::ir::tensor_t*& myelin::ir::operand_t::tensor(): Assertion is_tensor()' failed.
Aborted (core dumped)`

Any help is appreciated.

my environment :
cuda:10.2
tensorrt:8.2
onnx:1.9.0
pytorch: 1.5.1

[ttyio]

Hello @quintetoy @yasserkhalil93 , sounds bug in TensorRT, could you share us the onnx the model to debug? thanks!

[Keysmis]

@ttyio @yasserkhalil93 hello~, any update now? is it succeed?

[ttyio]

Hello @Keysmis , we did not get the repro to debug the issue. could you try 8.2 release? send us repro if it still failed. thanks!

[KeyKy]

I also get this error! when using 8.2 tensorrt.

python: /root/gpgpu/MachineLearning/myelin/src/compiler/./ir/operand.h:166: myelin::ir::tensor_t*&myelin::ir::operand_t::tensor(): Assertion `is_tensor()' failed.

For my model, it works well in 8.0 tensorrt

[aeoleader]

I am having the same issue with 8.2.2.1 version.

[ttyio]

@KeyKy @aeoleader , could you provide us repro to debug the issue? thanks!

[aeoleader]

@ttyio
I have done a little bit of testing. It seems like a problem in the slice assignment.

Repro steps:

1. Convert model to onnx

def test_func(x):
    t = torch.zeros(x.shape)

    # is_tensor error
    t[0] = x[0]
    return t

class MyModule(torch.nn.Module):
    def __init__(self):
        super(MyModule, self).__init__()
    
    def forward(self, x):
        return test_func(x)

if __name__ == '__main__':
    model = MyModule()
    input_shape = (1, 16, 4, 224, 224)
    model.cpu().eval()
    model.to('cuda:0')
    input_tensor = torch.randn(input_shape).to('cuda:0')
    output_file = 'debug.onnx'
    torch.onnx.export(
    model,
    input_tensor,
    output_file,
    export_params=True,
    keep_initializers_as_inputs=True,
    verbose=False,
    opset_version=11)

2. Use the trtexec to compile with the following command:

trtexec --onnx=debug.onnx --saveEngine=debug.trt --best --buildOnly --workspace=4096 --verbose

[ttyio]

Thank you @aeoleader for the repro, I am able to repro the failure and create internal bug to track this.

[aeoleader]

@ttyio
Can you also try this as well? Constant padding seems broken as well.