Make -DENABLE_FP8 conditional on CMake option

[martin-ueding]

The definition -DENABLE_FP8 is added only conditional on the used CUDA version. There is an additional CMake option ENABLE_FP8 which is supposed to control whether FP8 is used. The option only controlled the message, but not the actual definition.

This change makes the compile option dependent on the CMake option like the message suggests it should be.

Signed-off-by: Martin Ueding [email protected]

[comaniac]

Upvote for this PR. I got the following build failure with CUDA 11.8 without this PR:

/tmp/FasterTransformer/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h(1743): error: identifier "printf" is undefined

/tmp/workspace/FasterTransformer/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h(1852): error: identifier "printf" is undefined

2 errors detected in the compilation of "/tmp/workspace/FasterTransformer/src/fastertransformer/kernels/unfused_attention_kernels.cu".

[maxhgerlach]

build failure with CUDA 11.8

Somehow the header isn't included in the .cu files to provide printf. A Work around to build with ENABLE_FP8: CUDAFLAGS="-include stdio.h" cmake ...