[Help wanted] Support TensorRT

[csukuangfj]

TODO

• Support GPU via TensorRT

See https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html

[yuekaizhang]

I would like take on this.

• Support the Onnxruntime CUDA provider.

[manickavela29]

Hi @csukuangfj , @yuekaizhang

Observed that currently only CUDA EP support is there and TensorRT EP support is not there for onnxruntime.
is there ay active developments going on for TensorRT GPU backend?

[csukuangfj]

is there ay active developments going on for TensorRT GPU backend?

We don't have a plan to support it in the near future. Would you like to contribute?

[manickavela29]

I tried adding triggering onnxruntime's tensorrt ep for zipfromer but the model performance was very bad,
debugging further with standalone onnxruntime in python for Encoder models, will update if I see some good results.

[manickavela29]

Hi @csukuangfj,
TensorRT has several parameters, and these will be only valid if TensorRT provider is chosen,
so I need your suggestion on either of below 2.

1. Putting TRT configs as part of the model-config.cc file model-config.cc
2. Creating a new config for TRT and exposing the required parameters from it.

Thank you

[csukuangfj]

Could you create a new config for tensorrt and add this config as a member field of OnlineModelConfig and OfflineModelConfig?

You can set the default values of this config as the one used in

sherpa-onnx/sherpa-onnx/csrc/session.cc

Lines 137 to 150 in b714817

	std::vector<const char*> option_values = {
	"0",
	"2147483648",
	"10",
	"5",
	"0",
	"0",
	"0",
	"1",
	"1",
	"1",
	".",
	"1",
	".", // can be same as the engine cache folder

[manickavela29]

yes, I will send the PR for configs separately in some time.

[manickavela29]

Current perf Cuda vs Trt

csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 1.930044 ms
csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.034984 ms
csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.034912 ms
csrc/online-websocket-server-impl.cc:Run:256 Warm up completed : 3 times.
csrc/online-websocket-server.cc:main:79 Started!
csrc/online-websocket-server.cc:main:80 Listening on: 6007
csrc/online-websocket-server.cc:main:81 Number of work threads: 8

csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.535651 ms
csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.187492 ms
csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.187698 ms

Apart from this, with Trt there is a huge session creation time.
which is expected, only way to handle is to cache the engine images.

[yuekaizhang]

Current perf Cuda vs Trt

csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 1.930044 ms csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.034984 ms csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.034912 ms csrc/online-websocket-server-impl.cc:Run:256 Warm up completed : 3 times. csrc/online-websocket-server.cc:main:79 Started! csrc/online-websocket-server.cc:main:80 Listening on: 6007 csrc/online-websocket-server.cc:main:81 Number of work threads: 8

csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.535651 ms csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.187492 ms csrc/online-zipformer2-transducer-model.cc:RunEncoder:445 Encoder Duration : 0.187698 ms

Apart from this, with Trt there is a huge session creation time. which is expected, only way to handle is to cache the engine images.

May I know the results for CPU provider if you have？ Also, could you explain why there are three lines for each block? e.g. 0.535651 ms 0.187492 ms 0.187698 ms. @manickavela29

[manickavela29]

I can try to get for CPU numbers, but i don't have any high performance CPU,

(in between someone can add support for dnnl ep 🙂)

But here the focus itself is towards GPU with Cuda Vs Trt, is CPU benchmarking relevant?

Code blocks are just performance log which I added for zipformer. Those are not part of the patch