Vadim Kantorov

It appears that it is implemented in TRT as convolutions, which for n~several hundreds means large-receptive-size convs which are often inefficient and not optimized for?

window_size which often matches the fft order?

For small overlaps, it probably does not matter match And a more important thing is that for many frequencies, fast dft can be faster than convs - otherwise, dft/fft should...

One more example. Here in octave `eigs` completes in 30 seconds, and I couldn't wait for completion with scipy. Am I doing anything wrong? [L_octave.mat.zip](https://github.com/scipy/scipy/files/12609553/L_octave.mat.zip) ```matlab % bug.m % run...

Here are the lowest-magnitude eigvals of this laplacian matrix. As expected for a laplacian matrix, the lowest-magnitude eigval is ~0.0. But then the eigenvalues are also very small and close...

Here are octave codes for `eigs`: (Links removed)

Okay, I waited for completion. SciPy+ARPACK took ~250 seconds (while using eigsh), and octave+ARPACK took 11 seconds (using eigs). Quite a serious difference, given that both packages supposedly use ARPACK...

We observed that using pickle instead eliminates the boost. Will try to post a repro here.

So several build problems: - can't build any branch on `ubuntu-24.04` - can't build any branch on `ubuntu-22.04` when removing custom boost install (`sudo add-apt-repository -y ppa:mhier/libboost-latest && sudo apt-get...

haven't tried again - feel free to create new GitHub repo, add there `.github/workflows/buildtritoninferenceserver25.yml` with some workflow code inspired by OP https://github.com/triton-inference-server/server/issues/7997#issue-2839814892 (maybe make sure to try a more recent...