Request for Official ONNX Export + TensorRT Conversion Scripts for SAM3

Hello Team, Thank you for releasing SAM3 — it’s an excellent model, and we are exploring deployment on GPU servers and edge devices.

At the moment, there is no official documentation or code for exporting SAM3 to ONNX or converting it to TensorRT. Since ONNX/TensorRT are essential for production and edge deployment, official support would be greatly helpful.

Request

Could the team please provide:

1. Official ONNX export script(s) for SAM3

   • Full model export, or
   • Any recommended export structure (e.g., encoder/decoder separation) if applicable
   • Guidance on whether splitting the model is required or beneficial

2. TensorRT build instructions or scripts, including:

   • ONNX → TRT conversion steps
   • Requirements for dynamic/static shapes
   • Recommended precision modes (FP16, FP32)

3. Inference examples:

   • ONNX runtime example
   • TensorRT inference example
   • Any hybrid approach if that is the recommended pathway

4. Documentation on:

   • Preprocessing expected by the exported model
   • The exact shapes and formats of model inputs/outputs
   • Best practices for validating exported ONNX files

Why this is important

For edge and production systems (Jetson, DeepStream, inference servers, etc.), ONNX/TensorRT support is essential. Community-driven scripts work partially, but official guidance will ensure reliability and consistent deployment.

Additional Question

If splitting SAM3 into separate components (e.g., encoder/decoder) is recommended for export or inference — could the team clarify this approach?