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][Easy] Annotate unused vars in the csrc files (#14798)
Signed-off-by: Lu Fang <lufang@fb.com>
Lu Fang committed
Mar 14, 2025 at 21:40 UTC
8c0d15d5c5658b74a70694124af2ac250fdc4e23
4 files changed
+14
-13
csrc/prepare_inputs/advance_step.cu
+1
-1
@@ -274,7 +274,7 @@ void advance_step_flashinfer(
274
cudaDeviceGetAttribute(&blocks, cudaDevAttrMultiProcessorCount, dev);
275
cudaDeviceGetAttribute(&threads, cudaDevAttrMaxThreadsPerBlock, dev);
276
277
- int block_tables_stride = block_tables.stride(0);
277
+ [[maybe_unused]] int block_tables_stride = block_tables.stride(0);
278
TORCH_CHECK((blocks * threads > num_queries),
279
"multi-step: not enough threads to map to num_queries = ",
280
num_queries, " block_tables.stride(0) = ", block_tables.stride(0),
csrc/quantization/fp8/amd/quant_utils.cuh
+1
-1
@@ -446,7 +446,7 @@ scaled_vec_conversion<uint16_t, uint8_t>(const uint8_t& a, float scale) {
446
template <>
447
__inline__ __device__ uint32_t
448
scaled_vec_conversion<uint32_t, uint16_t>(const uint16_t& a, float scale) {
449
- __half2_raw h2r =
449
+ [[maybe_unused]] __half2_raw h2r =
450
__hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret);
451
union {
452
__half2_raw h2r;
csrc/quantization/gptq/q_gemm.cu
+8
-8
@@ -206,8 +206,8 @@ __global__ void gemm_half_q_half_gptq_4bit_kernel(
206
int offset_m = blockIdx.y * m_count;
207
int offset_k = blockIdx.z * BLOCK_KN_SIZE;
208
209
- int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
210
- int end_m = min(offset_m + m_count, size_m);
209
+ [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
210
+ [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
211
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
212
213
int n = offset_n + t * 4;
@@ -344,8 +344,8 @@ __global__ void gemm_half_q_half_gptq_2bit_kernel(
344
int offset_m = blockIdx.y * m_count;
345
int offset_k = blockIdx.z * BLOCK_KN_SIZE;
346
347
- int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
348
- int end_m = min(offset_m + m_count, size_m);
347
+ [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
348
+ [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
349
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
350
351
int n = offset_n + t * 4;
@@ -465,8 +465,8 @@ __global__ void gemm_half_q_half_gptq_3bit_kernel(
465
int offset_m = blockIdx.y * m_count;
466
int offset_k = blockIdx.z * BLOCK_KN_SIZE;
467
468
- int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
469
- int end_m = min(offset_m + m_count, size_m);
468
+ [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
469
+ [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
470
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
471
472
int n = offset_n + t * 4;
@@ -593,8 +593,8 @@ __global__ void gemm_half_q_half_gptq_8bit_kernel(
593
int offset_m = blockIdx.y * m_count;
594
int offset_k = blockIdx.z * BLOCK_KN_SIZE;
595
596
- int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
597
- int end_m = min(offset_m + m_count, size_m);
596
+ [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
597
+ [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
598
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
599
600
int n = offset_n + t * 4;
csrc/rocm/attention.cu
+4
-3
@@ -308,8 +308,8 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
308
309
constexpr int GQA_RATIO4 = DIVIDE_ROUND_UP(GQA_RATIO, 4);
310
311
- __shared__ float shared_qk_max[NWARPS][16 + 1];
312
- __shared__ float shared_exp_sum[NWARPS][16 + 1];
311
+ [[maybe_unused]] __shared__ float shared_qk_max[NWARPS][16 + 1];
312
+ [[maybe_unused]] __shared__ float shared_exp_sum[NWARPS][16 + 1];
313
// shared_logits is used for multiple purposes
314
__shared__ _B16x4 shared_logits[NWARPS][4][16][4];
315
@@ -426,7 +426,8 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
426
const cache_t* k_ptr2 = k_ptr + kblock_number * kv_block_stride;
427
const int klocal_token_idx =
428
TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id;
429
- const int kglobal_token_idx = partition_start_token_idx + klocal_token_idx;
429
+ [[maybe_unused]] const int kglobal_token_idx =
430
+ partition_start_token_idx + klocal_token_idx;
431
const int kphysical_block_offset = klocal_token_idx % BLOCK_SIZE;
432
const cache_t* k_ptr3 = k_ptr2 + kphysical_block_offset * KX;
433