possible bug in block fp implementation

[a-maci]

goldeneye/src/num_sys_class.py

Line 351 in b9f53f1

mant_adj = mant / power_exp_diff

After shifting the mantissa bits in the above line, the values might become < 1. So the exponents after reconstruction are no longer the same (one can check this after extracting the exponents after line 363 of the reconstructed tensor).

You probably need mant_adj[mant_adj < 1.0] = 1.0 after the above line.