Release of CmdStan 2.40.0

Announcements
1

We are very happy to announce that CmdStan 2.40.0 is out!

This release cycle brings new 1-D integration functions, quality of life improvements, and many bug fixes.

For details, see the blog post: Release of CmdStan 2.40 – The Stan Blog

2

Thanks a lot! The new to_(row)vector_array and to_matrix functions are very welcome.

3

Those were added by a first time contributor, fatimmajumder on github!

4

variadic integrate 1d functions… finally… great.

5

@seantalts, any chance we can get these new functions into stanli?

6

Oh nice, will do! Thanks for the ping.

Edit: out in the latest stanli release now. Lmk if you have any issues.

7

Are you using them in the model block, that is, would you benefit from faster gradients? There is a way to make the gradient computation faster, but priority of making PR depends on whether anyone cares

8

Not at the moment, but faster gradients for this function would be very useful for sure. I have somewhat avoided its use for rare cases given that it was not variadic - but that will change. Which issue or PR is it?

9

For integrate_1d_gauss_kronrod there is a PR integrate_1d_gauss_kronrod performance fix by avehtari · Pull Request #3403 · stan-dev/math · GitHub (which did not make to 2.40), which fixes one perfomance issue has big impact only sometimes.

That PR mentions possible follow-up

integrate_1d_adjoint runs one quadrature per var scalar and does a full reverse sweep at each node while keeping a single adjoint. Batching them onto one shared partition is a further ~8× on this model and needs no additional Boost change, but it touches adjoint code shared with integrate_1d, so maybe another PR?

The same batching for integrate_1d_double_exponential would be also possible, but would require modifying Boost code, so I did not propose that yet.