How does FP32 to FP16 conversion work in FP16 mode?

If the BuilderFlag::kFP16 is on, is the input of fp32 to fp16 auto-casted for an fp16 plugin?

I have a pure fp16 plugin at the top of a network, so the following code is needed to inject the input

config->setFlag(BuilderFlag::kFP16);  // fp16 on
.....                              // intemediate code
// Question here: should in_features_ be fp32 or fp16?
GPU_CHECK(cudaMemcpyAsync(in_buffers_[0], in_features_, size_in, cudaMemcpyDeviceToDevice, stream_));
context_->enqueueV2(in_buffers_, stream_, nullptr);     // expect FP16 running

Do I need explicitly convert the in_features_ to fp16 before the data transfer?