Created by: NHZlX

If we set CUDA_ARCH_NAME to All and run the bert using paddle-trt fp16. There will be a bug of

FatalError: use EmbEltwiseLayernormPluginDynamic FP16, but GPU doesn't have FP16. at (/workspace/paddle/fluid/inference/tensorrt/convert/emb_eltwise_layernorm.cc:96)