Trellis 2 is incredibly difficult to run on non-CUDA hardware, as CUDA dependencies are baked very deeply on Trellis 2 dependencies and dependencies of those dependencies, things like CuMesh and many others.
Microsoft is pushing for the ONNX format and runtime for LLM for the local copilot.
The ONNX runtime has shortcomings in performance under windows, but it does have a great advantage in portability compared to Pytorch safetensor. It's very easy to run.
This being a first party Microsoft model I would expect Microsoft to release ONNX version of Trellis 2, and use dependencies that don't depend so hard on CUDA to run.
Is Microsoft planning to continue development on ONNX, and release first party model like Trellis 2 in ONNX format?
Trellis 2 is incredibly difficult to run on non-CUDA hardware, as CUDA dependencies are baked very deeply on Trellis 2 dependencies and dependencies of those dependencies, things like CuMesh and many others.
Microsoft is pushing for the ONNX format and runtime for LLM for the local copilot.
The ONNX runtime has shortcomings in performance under windows, but it does have a great advantage in portability compared to Pytorch safetensor. It's very easy to run.
This being a first party Microsoft model I would expect Microsoft to release ONNX version of Trellis 2, and use dependencies that don't depend so hard on CUDA to run.
Is Microsoft planning to continue development on ONNX, and release first party model like Trellis 2 in ONNX format?