[Feature] Handling onnxrt execution provider config for various models

[manickavela29]

Hi @csukuangfj,

With #992, config for backends are handled as arguments and it is done.

But there is an additional issue with arguments and models, as the suggested configs are not specific to models
In the case of zipformer,

• configs required for encoder model, need not be required for decoder and joiner
• Tensorrt has max_workspace_size which will limit the gpu memory usage for the session,
• Encoder model can take mode space, which is not necessary for decoder. but with our current arguments both are allocated same memory

Solution :
for zipformer case,
Adding encoder_config, decoder_config and joiner_config

sherpa-onnx/sherpa-onnx/csrc/online-transducer-model-config.h

Lines 14 to 17 in 3e4307e

	std::string encoder;
	std::string decoder;
	std::string joiner;

and a new overloaded function at providers and session
All the config will be hardcoded for specific model and not as argument while starting sherpa,

But if you have any better suggestions let me know

[csukuangfj]

Adding encoder_config, decoder_config and joiner_config

That would introduce too many command line arguments

---

How about creating separate sess_opts_ for the decoder and the joiner and hard-coding the config values if tensorrt is used?

[manickavela29]

Yes, I meant on similar lines,

By adding separate configs for encoder,decoder and joiner,
I actually meant they will be hardcoded and not exposed as an argument.

And as you suggested, they will have separate sess_opts_ which will be built with their hardcoded custom config

[manickavela29]

Actually, given the mode size of decoder and joiner,
we can as well just running with CUDA EP itself,
since encoder is the only heavy lifter here