Fill a range with random numbers and execution policy

We can fill a range [first, last) using

std::mt19937 g;
std::uniform_real_distribution<> u;
std::generate(first, last, [&]() { return u(g); });

Theoretically, it would be more performant to execute std::generate with the execution policy std::execution::par_unseq. However, we could write

std::generate(std::execution::par_unseq, first, last, [&]() { return u(g); });

but is this really safe? I think, parallal access to g might be problematic. If that's actually true, how can we fix this? I've seen strange looking code like

std::generate(std::execution::par_unseq, first, last, []()
{
    thread_local std::mt19937 g;
    thread_local std::uniform_real_distribution<> u;
    return u(g);
});

But is thread_local really sensible here? And how should g be seeded here if the generated samples are supposed to be independent?