a Problem with 4xNomos8kSCHAT-L.onnx & safetensor

[petercham]

The 4xNomos8kSCHAT-L onnx model will show full nan results using CUDA inference, but using the CPU normal is just very slow. The exact same environment and inference code, other models and fp16 versions do not.

I wonder if it could be the wrong version of the download, and if it could be put on the hf so that it can be verified again.

[petercham]

Tests found that the latest safetensor version is half the size, but the inference speed is not significantly improved, with both 1152×864 image enlargements taking more than 17 seconds compared to the 300mb pth version. (Test platform: AMD7950 64GB NVIDIA 4090)