albanD
albanD
cc @jbschlosser @suo this was done on purpose in https://github.com/pytorch/pytorch/pull/80451
FYI @malfet
@pytorchbot rebase
@pytorchbot merge -f "was unlinked from old internal diff, we can ignore internal signal"
Looking into the CPU-side issue: - Good news: I found what is causing the issue: the call into the BLAS library gemm is causing these. - Bad news: This repros...
Here is a version that does fail for me (have to run it a couple times sometimes) ```python for ii in range(100): fa = torch.randn(499, 526, device='cpu') b = torch.randn(1,...
> @albanD not sure how you've tested with MKL (as there isn't one for M1), but one possible venue would be to switch from Accelerate to OpenBLAS, which some say...
> I am curious if we tried without Accelerate on arm platform to see if it helps Yes it works fine using our native mm kernels on ARM. @malfet was...
> Am kind of surprised that y’all keep bringing up Accelerate in an MPS context. There are two independent issues here actually: - CPU code generates random NaNs (this issue...
By the way, we also have some ongoing efforts to allow doing this out of tree: https://github.com/bdhirsh/pytorch_open_registration_example Did you consider this approach? And if so, what were the limitations? We...