logsumexp-plus ?

[srush]

Is it possible to add logsumexp-plus (or am I missing it) ? Would be really useful, but seems a bit more complex than the other to do.

[thakkarV]

@srush Hi! Thanks a lot! Surely is doable, and you may even be able to just use cuASR out of the box with it. But I can add it in with cuASR 1.2 release that will support Ampere. BTW, have you used cuASR? If so, may I ask what you are using it for?

[srush]

Hi @thakkarV,

I haven't used it yet, but I am planning on using it as the backend of https://github.com/harvardnlp/pytorch-struct . Which currently has memory inefficient semiring implementations. CuASR max/+ will let us efficiently run Viterbi algorithms on very large models (like HMMs). If it had log-space (forward and backward) we could also train with it.

However wrapping the library for use in pytorch is proving harder than I thought. Need to figure that out first.

[thakkarV]

Ah I have forever wanted to use cuASR to run Viterbi. The original reason to support pluggable operators was so that we could do something like that (instead of just shortest path problems, which was our original goal). Can you please email me at [email protected] ? I can share with you some more details on the timeline I have planned for features (Ampere uArch support and some changes I have planned to the operator struct API) and perhaps give you a hand with integration.