Fix cuda compile error with bf16

Differential Revision: D73562284

Pull Request resolved: #2122

Author: metascroy

torchao/csrc/cuda/fp6_llm/utils_parallel_dequant.cuh

@@ -25,9 +25,7 @@
 
 #include <cuda.h>
 #include <cuda_fp16.h>
-#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
 #include <cuda_bf16.h>
-#endif
 #include <cuda_runtime.h>
 
 /*