Thanks a lot, it actually worked ! I knew I was close, but not that close 
Now, it is running, yet it is telling me that it is going to take about 300h…any idea why it is so much slower ?
I actually had tried really simple distributions earlier and using as_op, and noticed a considerable slowdown but not even by that much…
I guess the original code works directly on the logp, is it what makes the whole difference?