huggingface/pytorch_block_sparse

Only blockshape=(32, 32) is supported

Routhleck opened this issue · 0 comments

When I set the blockshape to (64, 64) or (128, 128), the result is quite wrong