Goran Flegar

I agree with @thoasm's last comment. Making an eigensolver a `LinOp` definitely does not make sense, simply because the result of the operation is definitely not an application of some...

@hartwiganzt with "ILU preconditioner" I was talking about the linear operator `U^-1 L^-1` (this is the actual operator that is applied at every iteration), so `|| U^-1 L^-1 ||`, which...

@tcojean #141 does something different - it computes the norm of each column of the matrix, so the result is a row-vector of norms. This issue is a feature request...

We also noticed this issue on our internal version of Triton proper, running a newer LLVM version. `test_core.py` crashes after trying to emit too many warnings about `tt.divisibility` being unhandeled...

The "fix" I used internally for now is https://github.com/openai/triton/commit/e3259831bfcbaef86143524bab1ce05bf5186387. Would this be something you're interested in upstreaming? This allows us to also speed up compile-times since we're not processing all...

#3147 puts diagnostics behind an environment variable

I don't think this should cause performance regressions. The only thing this does is uses all the warps we have available in a block, instead of potentially keeping some of...

> would ensure that the implementation of the Krylov solver in question is robust I wouldn't call _correct_ code "_robust_" :smile: > it might be difficult to give accurate names...

> I do not argue that what you suggest is not a good idea. In fact, it sure is very nice! The problem I see is that this will practically...

> I believe anoting tricks (such as the even and odd iterations in TFQMR) will happen due to point 1. I didn't get this one, could you explain? I guess...