How to set platform_settings while expand new backend?

[SunCrazy]

Hi,
Could you please provide the complete platform settings that can be modified?

I want to expand a new backend to deploy model, and need know that what kind of valid arguments that i can set.

Thanks

[gushiqiao]

Firstly, you need to write the quantitative configuration of the new platform in the "dipoorlet/platform_settings.py" , using the "trt" platform as an example:

And write the "{new_platform_name: new_platform settings}" of in the dictionary "platform_setting_table".

Secondly, you need to create a file on generating new platform deployment configurations under the "dipoorlet/deploy/", with the file name "deploy_new_platform.py". At the same time, write a new function to output deployment configurations in this file and register the new platform through a registrar. Taking the "trt" platform as an example, the content in "dipoorlet/deploy/deploy_trt.py" is:

Finally, you need to import the new function into the “dipoorlet/deploy/deploy_base.py”:

At this point, a new platform is registered, and you can output the deployment configuration of the platform by using the "-D" and the name of the new platform.

[SunCrazy]

Sorry for not clarifying the requirements.
In fact, I want to know that all possible config parameters in trt_platform_settings dict. In other words, what kinds of parameters I can set while add a new backend.

[gushiqiao]

For example, the new platform can support quantization operator types ("quant_nodes"); hyperparameter (bit count, symmetric/asymmetric) of weighted quantization and activated quantization supported by new platform ("qi_params", "qw_params"), whether the platform supports quantization per channel ("per_channel"), etc. You need to determine these parameters based on the platform you want to add, and write the parameters of the new platform according to the parameter format of the existing platform (such as "trt_platform_settings").