Right now triton_model looks at the config.pbtxt in Triton to relay the expected itemsize to triton_block. Since there's no support for complex inputs and outputs and PyTorch always assumed interleaved floats or doubles for complex, there's no reason we cannot support complex blocks in GRC and Python.
I think the best way to do this is to add a parameter to the triton_block to explicitly state that the input and/or output is complex, effectively ignoring the information in the config.pbtxt only for the purposes of configuring appropriate block io_signatures.
Right now
triton_modellooks at the config.pbtxt in Triton to relay the expecteditemsizetotriton_block. Since there's no support for complex inputs and outputs and PyTorch always assumed interleaved floats or doubles for complex, there's no reason we cannot support complex blocks in GRC and Python.I think the best way to do this is to add a parameter to the triton_block to explicitly state that the input and/or output is complex, effectively ignoring the information in the config.pbtxt only for the purposes of configuring appropriate block io_signatures.