The usual behavior for the "mean" operation in numerical frameworks is a reduction of a tensor to a single value. However, in GGML this operation instead calculates the mean per row. This is I think unexpected behavior and the naming is also inconsistent with GGML_SUM (per tensor) and GGML_SUM_ROWS (per row).
The usual behavior for the "mean" operation in numerical frameworks is a reduction of a tensor to a single value. However, in GGML this operation instead calculates the mean per row. This is I think unexpected behavior and the naming is also inconsistent with
GGML_SUM(per tensor) andGGML_SUM_ROWS(per row).