Open yondonfu opened 6 months ago
@yondonfu This might be a bug related to torch.jit
. It tries to dump the source code location of the exception. However, unfortunately, the source code of the operation is missing since we have modified it dynamically.
I noticed that using a higher batch size with SVD that exceeds the available GPU VRAM triggers a Segmentation fault after the first warm up call that triggers compilation. If I re-run the same test without using stable-fast compilation, I get a CUDA OOM error instead. I was expecting to see the CUDA OOM error when there is insufficient GPU VRAM available while using stable-fast.
Here is the output when using stable-fast to compile the model:
Here is the output when not using stable-fast to compile the model:
Here is the test script:
The test was run on a RTX 4090.