fix cublas gemm call for bf16 input

cuda/utils/cublas_wrapper.h

@@ -122,12 +122,13 @@ inline cublasStatus_t cublasXgemm(cublasHandle_t handle,
     // TODO: Support bf16 for HIP
     assert(false);
 #else
+    const float alpha_fp32(*alpha), beta_fp32(*beta);
     return cublasSgemmEx(handle, transa, transb, m, n, k,
-            (const float*)alpha,
-            (const void*)A, CUDA_R_16F, lda,
-            (const void*)B, CUDA_R_16F, ldb,
-            (const float*)beta,
-            (void*)C, CUDA_R_16F, ldc);
+            (const float*)&alpha_fp32,
+            (const void*)A, CUDA_R_16BF, lda,
+            (const void*)B, CUDA_R_16BF, ldb,
+            (const float*)&beta_fp32,
+            (void*)C, CUDA_R_16BF, ldc);
 #endif
 }
 #endif  // CUBLAS_WRAPPER_H