Rounding_threshold_bits

FHE Libraries Zama Concrete ML
1

I’m trying to understand the role of rounding_threshold_bits in Concrete-ML.

My understanding is that neural networks use quantization-aware training (QAT) by default. Is rounding_threshold_bits applicable only to neural networks, or can it be used with other model types as well?

Does it apply only to QAT, or is it also relevant for post-training quantization (PTQ)?

Also, how does it differ from n_bits, and how are the two parameters related?

Thanks

2

rounding_threshold_bits is a flag that enables a performance/accuracy trade-off in NN models. When using rounding the correctness guarantee in TFHE is relaxed (by default it is of 2^-128 probability of off-by-one error for one PBS = one application of an activation function). NN are robust to such errors up to a certain degree. `rounding_threshold_bits`is most useful in post-training quantization.

Built in neural networks use QAT by default and do not apply rounding_threshold_bits