Register ENABLE_FP4 as a cmake option

System Info

x86_64
cuda 12.4

Who can help?

@tongyuantongyu @byshiue

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

install cuda 12.4
cmake ..
make

Expected behavior

compilation success

actual behavior

[  2%] Building CXX object tensorrt_llm/common/CMakeFiles/common_src.dir/opUtils.cpp.o
In file included from cpp/tensorrt_llm/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h:40,
                 from cpp/tensorrt_llm/kernels/cutlass_kernels/fpA_intB_gemm/bf16_int4_gemm_fg_scalebias.cu:17:
 cpp/tensorrt_llm/kernels/cutlass_kernels/cutlass_type_conversion.h:31:10: 
 fatal error: cuda_fp4.h: No such file or directory
   31 | #include <cuda_fp4.h>
      |          ^~~~~~~~~~~~
compilation terminated.

It s just because cutlass_type_conversion.h does not listen to the ENABLE_FP4 define.

additional notes

Would you mind if I register ENABLE_FP4 as a cmake option (default ON) ?

@tongyuantongyu

Start with cpp/tensorrt_llm/kernels/cutlass_kernels/cutlass_type_conversion.h and inspect how ENABLE_FP4 is handled during the cmake .. configuration. Locate the CMake entry point that defines build options, register ENABLE_FP4 there, then run cmake .. and make to verify compilation no longer unconditionally requires cuda_fp4.h.