Loop operator with no initial values for loop-carried dependencies aborts

[mneilly]

The attached DLRM model uses ONNX loop operators which have only the trip count and condition inputs but no initial values for the loop-carried dependencies. This results in onnx2trt aborting as follows when it hits an empty std::vector:

$ onnx2trt -o dlrm_s_pytorch.trt dlrm_s_pytorch.onnx 
----------------------------------------------------------------
Input filename:   dlrm_s_pytorch.onnx
ONNX IR version:  0.0.6
Opset version:    11
Producer name:    pytorch
Producer version: 1.8
Domain:           
Model version:    0
Doc string:       
----------------------------------------------------------------
Parsing model
[2020-10-15 05:44:40 WARNING] [TRT]/local/tensorRT/onnx-tensorrt/onnx2trt_utils.cpp:220: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[2020-10-15 05:44:40 WARNING] [TRT]/local/tensorRT/onnx-tensorrt/onnx2trt_utils.cpp:246: One or more weights outside the range of INT32 was clamped
terminate called after throwing an instance of 'std::out_of_range'
  what():  vector::_M_range_check: __n (which is 0) >= this->size() (which is 0)
Aborted (core dumped)

This occurs when there are no initial values and an empty stateVars vector is accessed at line 1730 below:

onnx-tensorrt/builtin_op_importers.cpp

Lines 1722 to 1730 in a3a4e38

	// Add initial state inputs using recurrent layers.
	std::vector<nvinfer1::IRecurrenceLayer*> stateVars{};
	for (size_t i = 2; i < inputs.size(); ++i)
	{
	stateVars.emplace_back(loop->addRecurrence(convertToTensor(inputs[i], ctx)));
	ctx->loopTensors()[body.input(i).name()] = node.input(i);
	ctx->registerTensor(TensorOrWeights{stateVars.back()->getOutput(0)}, body.input(i).name());
	}
	ctx->registerLayer(stateVars.at(0), node.name());

I can't upload the version with embedded tensors since it is too large so I'm attaching the version with external tensors. #542 needs to be resolved or worked around to hit this issue when using external tensors.

dlrm-external-tensors.zip

[kevinch-nv]

Thanks for the report. I'll make a MR fixing this along with #542

[kevinch-nv]

Merged the fix - can you confirm that the fix works for you?

[mneilly]

I'll check it out. :)

[kevinch-nv]

@mneilly Did the fix work for you? Can this be closed?

[kevinch-nv]

Closing this issue.