comparing models without fooling yourself with PyTorch

[batchfield55]

That matches what I was seeing. I ended up in about the same place using TensorRT-LLM. The biggest change for me was simplifying the stack instead of adding one more optimization layer.