Hipify self_multihead_attn

Enable HIP floa to hald conversion

apex/contrib/csrc/multihead_attn/encdec_multihead_attn_cuda.cu

 #include <vector>
 #include <iostream>
-
+//below lines enable hip float to half conversion which are disabled by default in hip_fp16.h
+#undef __HIP_NO_HALF_OPERATORS__
+#undef __HIP_NO_HALF_CONVERSIONS__
+//#endif
 #include <ATen/ATen.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <cuda.h>

apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_additive_mask_cuda.cu

 #include <vector>
 #include <math.h>
 #include <iostream>
-
+//below lines enable hip float to half conversion which are disabled by default in hip_fp16.h
+#undef __HIP_NO_HALF_OPERATORS__
+#undef __HIP_NO_HALF_CONVERSIONS__
+//#endif
 #include <cuda.h>
 #include <cuda_runtime.h>
 #include <cuda_fp16.h>

apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_cuda.cu

 #include <vector>
 #include <iostream>
-
+//below lines enable hip float to half conversion which are disabled by default in hip_fp16.h
+#undef __HIP_NO_HALF_OPERATORS__
+#undef __HIP_NO_HALF_CONVERSIONS__
+//#endif
 #include <ATen/ATen.h>
 #include <cuda.h>
 #include <cuda_runtime.h>

apex/contrib/csrc/multihead_attn/self_multihead_attn_cpp.cpp

@@ -3,7 +3,7 @@
 
 namespace multihead_attn {
 namespace self {
-namespace cublas_gemmex {
+namespace rocblas_gemm_ex {
 
 std::vector<torch::Tensor> fwd_cuda(
                                bool                 use_time_mask,  
@@ -121,12 +121,12 @@ std::vector<torch::Tensor> bwd(
                                 );
 }
 
-} // end namespace cublas_gemmex
+} // end namespace rocblas_gemm_ex
 } // end namespace self
 } // end namespace multihead_attn
 
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def("forward", &multihead_attn::self::cublas_gemmex::fwd, "Self Multihead Attention Forward.");
-  m.def("backward", &multihead_attn::self::cublas_gemmex::bwd, "Self Multihead Attention Backward.");
+  m.def("forward", &multihead_attn::self::rocblas_gemm_ex::fwd, "Self Multihead Attention Forward.");
+  m.def("backward", &multihead_attn::self::rocblas_gemm_ex::bwd, "Self Multihead Attention Backward.");
 }