I’m trying to tune a model that has a combination of convolutions and group convolutions. However, it seems that autotvm.task.extract_from_program is having an issue with the group convolutions as it returns no workloads. For example, one such extracted task is Task(func_name=topi_nn_group_conv2d_nchw, args=(('TENSOR', (1, 768, 11, 19), 'float32'), ('TENSOR', (24, 128, 1, 1), 'float32'), (1, 1), (0, 0), (1, 1), 6, 'float32'), kwargs={}, workload=None). Regular convolution tasks have proper workloads as expected. As far as I can tell, autotvm should be able to support group convolution extraction so I’m pretty confused where this issue is coming from. Anyone have any ideas?
For posterity, the issue turned out to be that tuning group convolutions is currently only supported for cuda. When attempting to tune on x86 or other platforms, tvm sees that group conv technically can be autotuned but because no search space is registered to the configuration it yields this error. We should consider adding a check to task extraction that provides a clearer failure.