This document is relevant for: Inf2, Trn1, Trn2, Trn3

NCC_EVRF046

Error message: ScaledMatmul custom call LHS tensor must have rank >= 2.

The __op$block_scaled_dot custom call performs matrix multiplication on MXFP8-quantized tensors. Matrix operations require at least 2-dimensional tensors (matrices). The LHS (left-hand side) operand must have rank >= 2. The compiler raises this error when the LHS tensor is a scalar or 1-D vector.

To fix this error, reshape the LHS to have at least 2 dimensions.

This document is relevant for: Inf2, Trn1, Trn2, Trn3