fix: enable fp4 compilation on cu128 (#5286)

sgl-kernel/CMakeLists.txt

@@ -145,7 +145,7 @@ if (SGL_KERNEL_ENABLE_FP8)
     )
 endif()
 
-if (SGL_KERNEL_ENABLE_FP4)
+if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_FP4)
     list(APPEND SGL_KERNEL_CUDA_FLAGS
         "-DENABLE_NVFP4=1"
     )
@@ -253,10 +253,7 @@ if (BUILD_FA3)
         FLASHATTENTION_DISABLE_SM8x
         FLASHATTENTION_DISABLE_BACKWARD
         FLASHATTENTION_DISABLE_DROPOUT
-        # FLASHATTENTION_DISABLE_ALIBI
-        # FLASHATTENTION_DISABLE_SOFTCAP
         FLASHATTENTION_DISABLE_UNEVEN_K
-        # FLASHATTENTION_DISABLE_LOCAL
         FLASHATTENTION_VARLEN_ONLY
     )
 endif()