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

NCC_EVRF039

Error message: QuantizeMX custom call block_size must be 32.

The QuantizeMX custom call implements OCP MXFP-8 microscaling quantization (https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf). The OCP MXFP specification requires a block size of 32 elements per scaling factor. The compiler raises this error when a different block_size value is provided in the backend_config.

To fix this error, use block_size=32.

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