Hi,
I’ve been using TensorRT since its early releases (2, 3). Back then, if you wanted to build using FP16 mode, you had to do 2 things:
parser->parse(...,DataType::kHALF);
builder->setHalf2Mode(true);
Now, you set FP16 via:
config->setFlag(BuilderFlag::kFP16);
My question is:
Should we stop using kHALF in the parser, and use kFLOAT always, regardless of INT8, FP16 or FP32 precision?
What happens if we have it enabled?