Jeremy Kun

What if we add a pass option for validation, that does something like count of operations that failed noise validation, and only err if that count exceeds a value given...

In my view, the binaries should be shipped with the PyPI package at release time, rather than fetched on the fly. This way the behavior of the frontend and the...

As I'm working on pip-installability, installing the nightlies as part of the frontend would just be better subsumed into using the pip-installed binaries. So I'll remove a TODO in the...

> libraries determine a unique path by computing the Non-Adjacent Form of the desired rotation offset. More specifically, it seems like the balanced NAF is the best approach. TIL! Great...

> Clearly, there are multiple ways to get to each number, but (afaik) libraries determine a unique path by computing the Non-Adjacent Form of the desired rotation offset. More specifically,...

I wrote a very naive (and slow) ILP that does this: https://github.com/j2kun/sdr/blob/main/sdr_ilp.py Basically, it has an integer variable for each "base" rotation in `[-N+1, N-1]` (2N variables), one for each...

Demonstrating slowness, the box_blur 64x64 test case gives a trivial solution and takes 850 seconds ``` Finished building model, starting solver Finished solving, extracting solution. Solution(objective=5.0, rotations=[63, 65, 127, 3968,...

The vos-vos-erkin algorithm (https://github.com/google/heir/pull/1286) I'm working on also requires all layout conversions to be broken down into power-of-two rotations. In the scope of this ticket, we should also revisit that...

Looks like that upstream PR will not get merged, and we'll need to roll our own parser/printer for dense attributes...

> Do we have an bazel option to default the openfhe target with opt flag It's not possible to have per-target differences in compilation mode with bazel, the compilation mode...