Pick up anything

[AryahP]

Hi!

I was trying to run the model associated with Pick up anything by following the instructions you have given on the GalaxeaR1Lite. However, after running the following commands:

cd ~/g0plus_ros2
./g0plus_hs_start_v1.sh

I get the following error, which indicates a hardware compatibility issue between my desktop and the .engine files, my desktop using v12.0 and the model using 8.6.

A potential solution I came across was to rebuild the .engine files using the model weights stored as .onnx files. However, I was unable to find these large files in the Pick Up Anything G0_VLA/G0Plus_PP_CKPT directory or elsewhere.

Would you be able to point me in the correct direction to solve this issue?

Thanks

[EastWindBlows]

Hi @OpenGalaxea team,

I have the exact same problem as described in this issue (TensorRT engine incompatibility).

Official prefill.fp16.engine / decode.fp16.engine only work on RTX 4090 (sm_89). On my RTX 3090 (sm_86) I get architecture mismatch error. I also cannot find any ONNX export script or the source code of gemma_rmsnorm.so.

Could you please offer some solutions? For example, directly provide pre-built TensorRT engines for RTX 3090 (sm_86) + matching gemma_rmsnorm.so, or provide the complete ONNX export script (that can handle custom Gemma RMSNorm), or share the plugin source code + build instructions so we can compile it ourselves. Any other recommended way to build the engine on 30-series GPUs would also be extremely helpful.

My environment:

• GPU: NVIDIA RTX 3090 24GB
• CUDA: 12.8
• TensorRT: 10.13.0
• Checkpoint: G0-VLA / G0Plus_PP_CKPT

Thank you! Looking forward to your help. 🙏