Thomas Viehmann

I'm not entirely sure if it is relevant for this discussion but: It appears that the current code (after #2952) causes SIGILL on NVIDIA Jetson devices because the HWCAP_CPUID seems...

@martin-frbg Thank you for the pointer. When I try to look at their 0.3.14 version, it seems to be missing that patch, too?! I'll file a bug with Debian.

Yes, thank you. It'd be a shame if the upcoming Debian release would be missing the fix. I filed it as [Debian Bug 986996](https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=986996) on their system.

I'm glad I didn't just waste your time on something that was completely and obviously fixed. Thank you again for your work on OpenBLAS and your help with this.

Right, thank you for reporting! In the meantime, please change test to val.

Note that we don't do outlier detection but drift detection. So depending on the drift detector you take, the question being answered is whether a "batch" (loosely, possibly multiple accumulated...

Hi Vinayak, thank your for your report and interest in TorchDrift. Typically, this a an issue of numerical overflow or sampling if you use bootstrap. The main branch has some...

The p-value gives you the rate of false positives, so if you feed samples from exactly the same distribution as the reference N times, you can expect p * N...

So there is a partial re-enactment in the PyTorch community voices podcast ( https://www.youtube.com/watch?v=rV5BhoKILoE ).

@HammamWahab, please take this as intuition rather than hard maths: - We are operating on samples in some ℝⁿ . This means that one would need some smoothing to map...