coreai_opt.quantization.spec.fp4_forward

coreai_opt.quantization.spec.fp4_forward(tensor)

Round to the nearest FP4 E2M1 value, returned in fp32.

The E2M1 grid is not uniform – 0, 0.5, 1, 1.5, 2, 3, 4, 6 – so a cast cannot do this. Ties go to the even encoding index, and a magnitude above the grid saturates to 6.0. A NaN is passed through unchanged instead.

Parameters:

tensor (torch.Tensor) – Values already divided by their scale, in fp32.

Returns:

The rounded values, in fp32, with any NaN preserved.

Return type:

torch.Tensor