[MISC] More AMD unused var clean up (#14926)

Signed-off-by: Lu Fang <[email protected]>

Author: houseroad

csrc/rocm/attention.cu

@@ -127,7 +127,7 @@ __device__ __forceinline__ T from_float(const float& inp) {
 
 template <typename T>
 __device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
-  union tmpcvt {
+  [[maybe_unused]] union tmpcvt {
     uint16_t u;
     _Float16 f;
     __hip_bfloat16 b;
@@ -160,7 +160,7 @@ __device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
 template <typename T>
 __device__ __forceinline__ _B16x4 addx4(const _B16x4& inp1,
                                         const _B16x4& inp2) {
-  union tmpcvt {
+  [[maybe_unused]] union tmpcvt {
     uint16_t u;
     _Float16 f;
     __hip_bfloat16 b;
@@ -1273,9 +1273,9 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
   const int seq_idx = blockIdx.y;
   const int context_len = context_lens[seq_idx];
   const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
-  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  [[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   const int warpid = threadIdx.x / WARP_SIZE;
-  const int laneid = threadIdx.x % WARP_SIZE;
+  [[maybe_unused]] const int laneid = threadIdx.x % WARP_SIZE;
 
   __shared__ float shared_global_exp_sum;
   // max num partitions supported is warp_size * NPAR_LOOPS