A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 18:30 UTC
[MISC] More AMD unused var clean up (#14926)
Signed-off-by: Lu Fang <lufang@fb.com>
Lu Fang committed
Mar 17, 2025 at 01:40 UTC
cd0cd85102e4b5971dd44109776942df5cdca70f
1 file changed
+4
-4
csrc/rocm/attention.cu
+4
-4
@@ -127,7 +127,7 @@ __device__ __forceinline__ T from_float(const float& inp) {
127
128
template <typename T>
129
__device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
130
- union tmpcvt {
130
+ [[maybe_unused]] union tmpcvt {
131
uint16_t u;
132
_Float16 f;
133
__hip_bfloat16 b;
@@ -160,7 +160,7 @@ __device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
160
template <typename T>
161
__device__ __forceinline__ _B16x4 addx4(const _B16x4& inp1,
162
const _B16x4& inp2) {
163
- union tmpcvt {
163
+ [[maybe_unused]] union tmpcvt {
164
uint16_t u;
165
_Float16 f;
166
__hip_bfloat16 b;
@@ -1273,9 +1273,9 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
1273
const int seq_idx = blockIdx.y;
1274
const int context_len = context_lens[seq_idx];
1275
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
1276
- constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
1276
+ [[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
1277
const int warpid = threadIdx.x / WARP_SIZE;
1278
- const int laneid = threadIdx.x % WARP_SIZE;
1278
+ [[maybe_unused]] const int laneid = threadIdx.x % WARP_SIZE;
1279
1280
__shared__ float shared_global_exp_sum;
1281
// max num partitions supported is warp_size * NPAR_LOOPS