Closed JohannesGaessler closed 1 month ago
Fixes https://github.com/ggerganov/ggml/issues/991 .
The problem is that the batch size is stored in ne[3] for 2D but in ne[2] for 1D. The CUDA code on master always retrieves the batch size from ne[3] so the result is incorrect for 1D Im2COL for batch sizes > 1.
ne[3]
ne[2]
Im2COL
Fixes https://github.com/ggerganov/ggml/issues/991 .
The problem is that the batch size is stored in
ne[3]for 2D but inne[2]for 1D. The CUDA code on master always retrieves the batch size fromne[3]so the result is incorrect for 1DIm2COLfor batch sizes > 1.