Roy Frostig
Roy Frostig
We're in agreement here by and large. This is something that we've thought about improving before, whether at the JAX or XLA level. I can't find an open issue for...
@minqi – the thinking hasn't changed much since this issue was last active. Although there's a fundamental puzzle regarding whether/how to do better, for now we're still producing `select` when...
> @inversecrime would sequential_vmap be similar to unrolling the batch using a for loop? It generates a `jax.lax.map`, which bottoms out in a (rolled) XLA loop.
We believe that we don't need `random_gamma_grad_p` as a primitive at all any more, as proposed in #27628.