Robotics
ONNX
Safetensors
Gr00tN1d7

onnx model cuda compatible

#2
by michalin-nvidia - opened
No description provided.
michalin-nvidia changed pull request title from cuda_compat_model to onnx model cuda compatible
michalin-nvidia changed pull request status to open

The export was generated with CUDA 12.8, PyTorch 2.7.1+cu128, and LEAPP 0.5.2.

  1. Any more docs on the process to generate it, especially LEAPP, which we could reference?
  2. Any compatibility facts to point out between the CUDA version used for expert and the CUDA runtime version?

exports/g1-static-apple-b1-480x640/onnx/leapp-0.5.2/graph.png filter=lfs diff=lfs merge=lfs -text

nitpick: lfs for all png files vs this specific path?

NVIDIA org

What's the use of exports/g1-static-apple-b1-480x640/onnx/leapp-0.5.2/graph.png file?

What's the use of exports/g1-static-apple-b1-480x640/onnx/leapp-0.5.2/graph.png file?

I believe this is part of the exported bundle from LEAPP. It is the results from the compute graph trace.

exports/g1-static-apple-b1-480x640/onnx/leapp-0.5.2/graph.png filter=lfs diff=lfs merge=lfs -text

nitpick: lfs for all png files vs this specific path?

I will add the rule for all png files from a leapp export.

The export was generated with CUDA 12.8, PyTorch 2.7.1+cu128, and LEAPP 0.5.2.

  1. Any more docs on the process to generate it, especially LEAPP, which we could reference?
  2. Any compatibility facts to point out between the CUDA version used for expert and the CUDA runtime version?

I will add a README to indicate how the model was exported. In terms of the documentation on CUDA version, I think this is just for debug trace purposes that is added by LEAPP, but I don't think there is a compatibility dependency. As long as the system that is used to trace the compute graph has a CUDA version that supports the ONNX operators exported it should be able to finish the export. If it didn't I believe the export should at least fail.

pulkitg-nv changed pull request status to merged
pulkitg-nv deleted the refs/pr/2 ref

Sign up or log in to comment