Does imatrix work independent of the dtype used for calibration?

[bartowski1182]

IE, if I convert to F32 and BF16, calculate the imatrix from F32 (since it supports GPU offloading), but then apply it to the BF16 weights during quantization, will that work as expected or will the change in dtype affect it?

@ikawrakow sorry for the ping but after searching couldn't find any way to reach out to you directly and I imagine you're the most well versed on this subject..

Similarly, if F16 for imatrix and BF16 for quantization, curious how this would work