Hi, I think I have met a bug about cublas. Here I use Apex fused_dense module: import torch from apex.fused_dense import FusedDense t = torch.rand((13824, 400), device="cuda", dtype=torch.float, requires_grad=True) dense = FusedDense(400, 1).to("cuda") out = dense(t - 0.5) loss = torch.sum(out) l…

Cublas Bug

mnicely June 20, 2022, 2:56pm 8

I suggest filing a bug and be sure to add your reproducer code.
How to report a bug