We can serialize the trt engine, which we call plan file. and we can not use plan file cross GPU model, but haw can i get the GPU info from a plan file to avoid cross-use plan on different GPU model?
I’m personally not aware of such a tool from TensorRT. In our own applications, we write down a collection of invariants “on the side” (in our own file) when serializing models. We write the Cuda Compute Capability and TensorRT version. Unfortunately, I’m not aware of a complete set of invariants which allow one to say for sure whether a plan file is valid. I’ve asked this question here, and not received a complete answer: https://devtalk.nvidia.com/default/topic/1046137/serialized-engine-validity/