RuntimeError: CUDA error: an illegal memory access was encountered Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Open zxk72 opened this issue 1 year ago • 4 comments

RuntimeError: CUDA error: an illegal memory access was encountered Compile with TORCH_USE_CUDA_DSA to enable device-side assertions. How do I solve this problem while running?

Oct 14 '24 12:10 zxk72

Have you solved this problem?

Oct 16 '24 09:10 1-2-3-4-0

Hello, @1-2-3-4-0 @zxk72 someone also report this problem to me, and i'm looking for solutions. Can YOU please also look at this issue?https://github.com/pytorch/pytorch/issues/21819

Oct 16 '24 09:10 Adamdad

It might be because you set the number of groups to 32? I encountered this issue when I tried setting the number of groups to 32. So a new question arises: why can't the number of groups be set to 32, given that the condition in the .h file states that less than 32 is valid

Dec 05 '24 12:12 bravoYJ

@bravoYJ Hello, my intuition is that the group number should not be set too large, as it may negate the parameter-saving benefits. I have not tested scenarios where the group number is excessively large.

Dec 05 '24 16:12 Adamdad