Manual warmup per model instance / specify warmup config dynamically using c api

[asaff1]

Is your feature request related to a problem? Please describe.
I'm having a problem with warming up the model.
Currently, the config.pbtxt provides a warmup section, but I'd like to warmup the model for many different batch sizes, which will make the configpb.txt very large. I use the C API, so I'd like to do it either manually in code - by sending InferRequests after model load, or somehow dynamically providing the model config to LoadModel function (currently, not possible).

Doing it manually in code be like:

        auto request = tds::InferRequest::Create(tds::InferOptions(model_name_));
        for (auto& inp : model_inputs_) {
            request->AddInput(inp, begin, end, tds::DataType::FP32, input_shape, tds::MemoryType::CPU, 0);
        }

        auto result = server_->Infer(*request);

The problem here, I cannot specify the device (the system has multiple GPUs) or more precisely, the instance of the model to send the InferRequest to. Is that possible using the c++ api? If yes, it will solve my issue.

The other way, could be using the ModelWarmup config, it will do it per model instance which is good. But I cannot specify it in any other way than the config.pbtxt. I'd like to specify it on LoadModel, or in other dynamic way with c code - is that possible with c++ api, without editing the config.pbtxt file?

Describe the solution you'd like
A way to warmup the model, for all instances (that runs on different devices). Either by the warmup feature, or manually specifying the model instance of the InferRequest.