is independent of the actual bit precision limitation?

[maxwellgodv]

is independent of the actual bit precision limitation. May I know how to understand this sentence？What is actual bit precision limitation?

[andrei-stoian-zama]

is independent of the actual bit precision limitation. May I know how to understand this sentence？What is actual bit precision limitation?

Since Concrete Python 2.0 the simulation itself requires that the circuit respect the bit-width constraint. This constraint is: all inputs and intermediary values in the function to be compiled must be <=16bit if and only if lookup tables are used or univariate functions. If there are no lookup tables or univariate functions, you can exceed 16b with an upper bound determined by the operation that you perform.

If you want to measure maximum bitwidth that the circuit would attain you can use "tracing" like this:

import numpy as np
from concrete import fhe

configuration = fhe.Configuration(
    enable_unsafe_features=True,
    use_insecure_key_cache=True,
    insecure_key_cache_location=".keys",
    single_precision=False,
    parameter_selection_strategy=fhe.ParameterSelectionStrategy.MULTI,
)

@fhe.compiler({"x": "encrypted", "y": "encrypted"})
def f(x, y):
    return x + y

inputset = [
    (
        np.random.randint(0, 10, size=()),
        np.random.randint(0, 10, size=()),
    )
    for _ in range(100)
]
circuit = f.trace(inputset, configuration)

print(circuit.maximum_integer_bit_width())

[RomanBredehoft]

Just a quick note what @andrei-stoian-zama said above : the current Concrete-ML public version does not use Concrete Python 2.0 yet, but it will soon ! So currently, you might expect different constraints. For example, simulating circuits with 16 bits is possible, which is what is independent of the actual bit precision limitation was referring to. However, this won't be the case in the near future, so I would recommend you not to take this as granted 😉

[maxwellgodv]

Can I measure maximum bitwidth like this？ What are the current constraints？

[RomanBredehoft]

Yes this is indeed how to measure the maximum bit-width reached in the model ! You cannot execute a model in FHE if the model has a bit-width of 17 bits or more.