Prompt-specific input_ids baked into ONNX graph forces re-export when changing from “dog” to “cat”

[yogurfrul]

In the current Sam3ONNXWrapper implementation, input_ids (derived from the text prompt such as “dog”) is exported as a constant node inside the ONNX file.
Because the tensor is frozen, the only way to switch the prompt to “cat” (or any other text) is to:

Re-build the entire ONNX model with the new prompt.
Re-build the TensorRT engine from the updated ONNX.

it‘s terrible！

[dataplayer12]

Absolutely. This will be the next feature I will add. It's a bit complicated because I still want to keep C++ implementation without implementing the tokenizer in cpp. Thanks for raising the issue! Please wait a few days

[yogurfrul]

tokenizer can be c++ implementation using tokenizers-cpp，then sam3_app could be an end-2-end delopy solution