Comparison of Quantization Error for Each Linear Operator ?

[Lenan22]

I would like to compare the error before and after quantization for each linear operator, specifically by calculating the cosine similarity between the quantized computation results and the bfloat16/float16/float32 computation results. Do you have an existing solution for this? If so, could you share it with me?

[jerryzh168]

we don't have an example right now in torchao, but we can try to provide an example, I can take a look a bit later if no one takes this up