Missing warmup process as k2-sherpa

[jingzhaoou]

Missing warmup process as k2-sherpa

[csukuangfj]

Could you help test the online or offline websocket server in sherpa-onnx and see if the second request is much faster than the very first request?

---

By the way, we have similar methods as k2-fsa/sherpa in sherpa-onnx

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

Line 61 in aa71087

virtual std::vector<Ort::Value> GetEncoderInitStates() = 0;

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

Lines 44 to 45 in aa71087

	virtual std::vector<Ort::Value> StackStates(
	const std::vector<std::vector<Ort::Value>> &states) const = 0;

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

Lines 54 to 55 in aa71087

	virtual std::vector<std::vector<Ort::Value>> UnStackStates(
	const std::vector<Ort::Value> &states) const = 0;

It is not difficult to add warmup to sherpa-onnx if you find it is helpful to reduce the latency of the first request.

---

It may be tricky to save the model initial states during ONNX export I guess.

You don't need to save the initial states during model exporting. The initial state is constructed in the code in sherpa-onnx.
We only need some model parameters to construct the initial states.

[jingzhaoou]

As I get better understanding of the code base, I realized that sherpa-onnx rewrites the Python initialization code in C++. The warm-up process is actually already in place. I am closing this PR.

[uni-sagar-raikar]

@csukuangfj @jingzhaoou Could you point me to code where model initialization is rewritten in C++. Didn't quite understand the point from previous comment. We see that, Warmup is still not part of the online websocket server.

[csukuangfj]

There is no warmup in sherpa-onnx, I think.