TF32 Quantization

[NightWinkle]

Hi !

From what I've seen, there is no TF32 option in the quantization options of a module. How is TF32 math enabled ? (or disabled)

Also, I thought ConvStrategy was an available setting but I realised it's not exposed. Since the default is autotune, would that mean that by default convs on cuda platform are TF32 implicit wmma ?

Thanks !

[laggui]

TF32 support is automatically detected based on the device properties, and so is MMA/WMMA. Autotune will select the best algo.

This is not exposed at a higher level (burn) as it is backend-specific, instead we rely on cubek/cubecl for that.