[Inference] save_optimized_model_pass support tensorrt

[yuanlehome]

PR types

New features

PR changes

Others

Description

This PR's work:

1. save_optimized_model_pass support tensorrt, cpu/gpu/xpu has been supported and verified in [Inference] Save optimized model by pass #53696, [Inference] save_optimized_model_pass support gpu #55551.
2. Uniformly package all non runtime parameters of TenosrRTEngine to ConstructionParams.
3. Support load trt engine in runtime(in tensorrt_engine op run).
4. Remove some unused function and variable in trt related code.
5. Refine some unittest related to trt.
6. Add unittest for function of save_optimized_model_pass(gpu/trt).

How to use:

//
// Step 1: first run
//
// other code
config.SetModel(/*model dir*/);
// other code
config.EnableUseGpu(...); // for gpu
config.EnableTensorRtEngine(..., use_static=true, ...). // for tensorrt
// other code
config.SwitchIrOptim(true);
config.SetOptimCacheDir("./optim_cache");
config.EnableSaveOptimModel(true);
// other code
auto predictor = CreatePredictor(config);

// Optimized model will be save to ./optim_cache/_optimized.*

//
// Step 2: second run
//
// other code
config.SetModel(/*optimized model dir*/);
// other code
config.EnableUseGpu(...); // for gpu
config.EnableTensorRtEngine(..., use_static=true, ...). // for tensorrt
// other code
config.SwitchIrOptim(false);
// other code
auto predictor = CreatePredictor(config);
// other code
predictor->Run();
// other code

Note:

1. Not support EnableTensorRtEngine(..., use_cuda_graph=true).

Others

Pcard-71500

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[XieYunshen]

LGTM for set_tests_properties(test_save_optimized_model_pass PROPERTIES TIMEOUT 300)

[Aurelius84]

LGTM for ENFORCE MSG

[zhangjun]

LGTM

[XiaoguangHu01]

LGTM