Free running? Triggered? What repeat rate?
What logic did you try, 74LVC? Handles 5V. Maybe isn't quite under 1ns, or you need too many gates in parallel and then the skew kills it?
Potato Semi comes to mind, but it looks like they don't have a 5V model, only 3.3V. (AFAIK it's basically "unlocked" CMOS, rated to run at stupid frequencies.)
I suppose you could still make do, using a couple of those but merging their outputs with a pulse transformer (aka 0° power divider).
As for avalanche, yes, that'll do, but you'll need very low loss coax for the delay line, to get nearly as sharp a falling edge. Possibly some filtering or clamping could be done to deal with that. You've got plenty of signal to play with, you can afford to burn a few dozen volts here or there with a rather lossy filter.
A lossy and nonlinear filter might be, for example: coupling the pulse through a capacitor, the end of which is charged slightly negative (say -10V). 100V pulse comes along, raising the end to +90V. Use a series diode to direct the pulse into the output. Pulse ends, dropping to 10V, 5, 2, etc. over some nanoseconds ("drool" due to line loss); well, the cap end drops to 0V, -5, -8, etc. at the same time -- the "drool" has been chopped off by the combination of diode and reverse bias. Add some large value resistors to apply that bias, and you're done.
50 ohm coax is about 200m/us so you need at least 4 meters for the delay line. There's no reasonable way to vary the line length (electrical trombone lines do exist, but, eh..).
ZTX415 are rated for avalanche, and are good when higher power is needed. Regular types like MMBT3904 and MMBT2369 will also do, if less power is needed (which is certainly the case here). You may even find it worthwhile shopping for lower voltage transistors; the savings being in less supply voltage required. Maybe you can get away with a voltage multiplier, rather than an inductive supply.
All transistor types I've tested, will avalanche pulse at some operating condition (Vces, Ices and Rbe). The main difference is how wide that range is (particularly Rbe). I don't think I'd be confident enough to design a production device around a 3904 or 2369 abusing this property, but the ZTX415 and relatives are tested for this and can be designed in reliably. If you can afford an adjustment step, or accept low production yield, then others will do.
Point being, I remember once testing a, something RF PNP, a salvaged Japanese (2SAxxxx) type, which I think was rated around 30V 100mA, and avalanched around 50-60V. I think it even performed with B-E shorted (i.e., Rbe = 0), which is convenient. So, if lower supply voltage is an option, consider selecting a few RF parts to test, and see how it goes.
Still other options -- consider a wideband RF amplifier. This might be a PHEMT or whatever, in the usual bias circuit; or an MMIC; or something more complicated, say using GaN FETs, maybe coupling transformers, etc. Generate a pulse with say ~100ps edges using something simple and fast, PECL for instance, and amplify it up to 5V. The amplifier will be class A (more or less), so expect relatively high power dissipation. (Transistors and MMICs capable of a few watts are readily available, this is no problem.)
The downside is, most of those perform best with a tuned load; they'll do 10GHz happily, but they won't do 20ps edges. Some are better at pulses than others, and I forget which are best (PHEMTs, GaN FETs, SiGe HBTs..). The relevant parameter is output capacitance, or its equivalent -- most of these parts have none of the conventional ratings and you have to interpret the s-parameters instead. In that case, look for the phase and magnitude slope of s22, try to fit an equivalent capacitance to it.
You can of course get ready-made wideband amps, if you don't mind spending a few bucks. Likely the best option for a one-off.
Tim