So you should either roll much more than needed and sha256 the result or you should use a debiasing procedure like the one in the codex32 booklet:
https://www.secretcodex32.com/Even if the dice is very biased, throwing it 99 times, as xx879 says, will suffice. In
this post, I demonstrate that even if there is a 50% chance of getting a 6 (and 10% for each 1, 2, 3, 4, 5), it will still produce 216 bits of entropy. As a disclaimer, I had just applied
Shannon's equation, and I'm by no means an expert in that branch of math.
I saw Andrew Poelstra talking about this on a "Bitcoin Layer" podcast recently. I have casino grade dice, with the sharp edges. Are these really biased? I assumed the casinos would demand some controls during manufacturing to try un-bias the dice.
All physical dice are biased to an extent. I think casino dice are known for approaching perfection.