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

NCC_EVRF051#

Error message: Data type F8E4M3FN is not supported on TRN1/TRN2.

The F8E4M3FN (8-bit floating point with 4-bit exponent and 3-bit mantissa) data type is natively supported on Trainium3 (Trn3) and later hardware generations. The compiler raises this error for an F8E4M3FN operand targeting Trn1 or Trn2 unless the experimental F8E4M3 conversion flag is enabled.

For QuantizeMX, the compiler also raises this error when backend_config selects "float8_e4m3fn" and the conversion flag is enabled, because MXFP operations do not support conversion to F8E4M3.

To fix this error for QuantizeMX, target Trn3 or later and do not enable --experimental-unsafe-fp8e4m3fn-as-fp8e4m3. For other operations targeting Trn1 or Trn2, the flag can be used to convert F8E4M3FN operands to F8E4M3.

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