[Bug] Possible error in strided_slice_checker

[AreopagX]

Hello, I'm trying to use TensorRT for tuning my model. For that matter, I use partition_for_tensorrt before building the model. However, this function fails when processing strided_slice ops. This might be related to strided_slice ops, where fewer axes than tensor dimensions are sliced, e.g., %435 = strided_slice(%376, begin=[0i64], end=[1i64], strides=[1i64], axes=[0i64]) with input tensors of 4 dims. I suspect that these ops will fail since the for loop

tvm/python/tvm/relay/op/contrib/tensorrt.py

Lines 712 to 713 in 7e3dc45

	for i in range(0, len(args[0].checked_type.shape)):
	begin = int(attrs.begin[i])

iterates over the number of input dimensions while the strided_slice op does not necessarily cover all dimensions, e.g., in the above call only the 0th axis.

Expected behavior

I expected the partition_for_tensorrt function to return without error.

Actual behavior

IndexError in strided_slice_checker:

tvm/python/tvm/relay/op/contrib/tensorrt.py

Line 713 in 7e3dc45

begin = int(attrs.begin[i])

, error thrown at:

tvm/python/tvm/runtime/container.py

Line 57 in 7e3dc45

raise IndexError("Index out of range. size: {}, got index {}".format(length, idx))

Environment

OS: Ubuntu 20.04, Kernel 5.15
TVM: built from master branch, commit 0e046da

Triage

• needs-triage
• byoc:ethosu

[masahi]

You're right, the axes argument in strided_slices was added later. So when the TRT backend was developed, it was assumed that the size of begin / end is always equal to the rank of the input.

Can you send a PR to fix this?