Eelco Hoogendoorn

Noticed I needed another jit after vmap-ing a jitted function; that makes the vmap-ed and non-vectorized timings pretty close, so eliminating that mystery. Now you can see some impact of...

> Note that JIT compilation will happen once for each unique input shape. Ah! No I did not know that, I was assuming it was cached on type only like...

Yeah I was on OSX py38, jax 0.2.16, jaxlib 0.1.67; I went with the conda install which doesnt have the newer jaxlib yet. Upgrading to 0.1.68 does indeed give a...

Somewhat related to the above; do you consider it feasible/possible/desirable to be able to dispatch on the basis of (a computable subset of) the value of arguments? Im trying to...

Yeah... im not seeing any other libraries that dispatch in any way on object value; just purely on type (this example of dispatching based on array dimension is the first...

Hey @wesselb ; thats looking pretty neat. The ability to specify some kind of computable matching function is quite critical for my actual application though. Though I suppose that would...

I was wondering if this is considered bad style, to dispatch based on the object value; but according to wikipedia: https://en.wikipedia.org/wiki/Multiple_dispatch > Multiple dispatch or multimethods is a feature of...

> Would something like how the example works suffice, or would you need something more flexible? Yeah; id need something more flexible. > types are hashable Yeah; in my example...

I dont fully appreciate the internal mechanisms (i suppose it requires some extra complexity to shoehorn it into the existing mechanisms of plum); but yeah the external API indeed looks...

I suppose it would suffice; but considering the number of different kinds of matching functions that I need, and the fact that they dont really have a ton of reuse;...