Wouldn't it be a lot easier to just use a part that's inherently
flatter? How about a Sirenza SBB-4089, for example.

If you really want to compensate a part that has some rolloff, you
could try, along your 50 ohm transmission line, putting in series a
shunt combination of an R and a C. The R adds some attenuation, and
the C shorts it out at high frequencies. You need to use tiny parts to
avoid problems with parasitic C and L. 0603 or 0402 should work OK.
You need to take into account the driving amplifier's output impedance,
and the load's impedance, and their variation with frequency. If you
use multiple sections like that, you can tailor the response better
than you can do with a single section. You could add in some
inductance in series with resistance, shunt to ground, if you could
find some inductance that behaves nicely at 2.5GHz. Maybe a shorted
stub? That would let you keep a more constant load impedance versus
frequency for the driving amplifier.

And what use would a 6 bit wideband system be anyway? What are you
receiving with it that doesn't require more dynamic range than you'll
get with 6 bits? And if you're only sampling at 1GHz, why are you
worried about anything more than about 400MHz bandwidth?

Much more interesting to be looking at 14 or more bits...I keep telling
the ADC manufacturers that I'd like 18 bits, with noise and distortion
to match. Haven't been demanding a 5Gs/s rate yet though. Wouldn't
quite know what to do with the data stream.

Cheers,
Tom