Swapnil Patel
Swapnil Patel
this could be helpful for chat_input as well. Is there a known way to focus on chat_input today ?
I went through the latest code and seems like the updated code already takes care of running on cpu if cuda is unavailable. I believe the trainer should work out...
A solution I use is to perform some sort of search algorithm (eg. binary/golden-section) with range of -R values till there’s no more increase in achieved-RPS to find max
exported artifacts can be verified on run here : https://github.com/numba/numba/actions/runs/13822376198
I've created a gist documenting the steps to build numba for win-64 using GitHub Actions. You can find it here: [Steps to use GHA for win-64 numba builds](https://gist.github.com/swap357/d0b36dee3b0ab15687e0fcac4a9128b7)
The errors are because default channel is not implicit anymore. This PR should fix it - https://github.com/numba/llvmlite/pull/1241
added here - https://github.com/numba/numba/pull/10107
@lericson Thanks for the report. I was able to reproduce this issue.
This seems to be because `np_all` (or`np.any`) overload doesn't have scalar handling yet. The `np.nditer()` approach seems to be intended for arrays only, not scalars. https://github.com/numba/numba/blob/5e0ceb077fb219a21ac8a50e5234ca643aaf96f9/numba/np/arraymath.py#L801-L810
Triaged this issue with maintainers, here is the summary - - When OpenBLAS 0.3.30 runs kernels with Apple M4 SME instructions, it breaks NumPy. https://github.com/numpy/numpy/issues/29393 Minimal Working Reproducer can be...