xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-16 14:27:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[MISC] Remove unused variableds in C++ (#19609)

Browse Source 
Signed-off-by: Lu Fang <lufang@fb.com>
This commit is contained in:
Lu Fang 2025-06-16 11:05:28 +08:00 committed by GitHub
parent a5e7242d5f
commit c6703d1e0d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
6 changed files with 2 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
csrc/attention/paged_attention_v1.cu
View File

@ -65,9 +65,6 @@ void paged_attention_v1_launcher(
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
@ -193,4 +190,4 @@ void paged_attention_v1(
#undef WARP_SIZE
#undef MAX
#undef MIN
#undef DIVIDE_ROUND_UP
#undef DIVIDE_ROUND_UP

5
csrc/attention/paged_attention_v2.cu
View File

@ -66,9 +66,6 @@ void paged_attention_v2_launcher(
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
@ -203,4 +200,4 @@ void paged_attention_v2(
#undef WARP_SIZE
#undef MAX
#undef MIN
#undef DIVIDE_ROUND_UP
#undef DIVIDE_ROUND_UP

1
csrc/prepare_inputs/advance_step.cu
View File

@ -274,7 +274,6 @@ void advance_step_flashinfer(
cudaDeviceGetAttribute(&blocks, cudaDevAttrMultiProcessorCount, dev);
cudaDeviceGetAttribute(&threads, cudaDevAttrMaxThreadsPerBlock, dev);
[[maybe_unused]] int block_tables_stride = block_tables.stride(0);
TORCH_CHECK((blocks * threads > num_queries),
"multi-step: not enough threads to map to num_queries = ",
num_queries, " block_tables.stride(0) = ", block_tables.stride(0),

2
csrc/quantization/fp8/amd/quant_utils.cuh
View File

@ -446,8 +446,6 @@ scaled_vec_conversion<uint16_t, uint8_t>(const uint8_t& a, float scale) {
template <>
__inline__ __device__ uint32_t
scaled_vec_conversion<uint32_t, uint16_t>(const uint16_t& a, float scale) {
[[maybe_unused]] __half2_raw h2r =
__hip_cvt_fp8x2_to_halfraw2(a, fp8_type::__default_interpret);
union {
__half2_raw h2r;
uint32_t ui32;

8
csrc/quantization/gptq/q_gemm.cu
View File

@ -206,8 +206,6 @@ __global__ void gemm_half_q_half_gptq_4bit_kernel(
auto offset_m = blockIdx.y * m_count;
auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
[[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
[[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
int n = offset_n + t * 4;
@ -344,8 +342,6 @@ __global__ void gemm_half_q_half_gptq_2bit_kernel(
auto offset_m = blockIdx.y * m_count;
auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
[[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
[[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
int n = offset_n + t * 4;
@ -465,8 +461,6 @@ __global__ void gemm_half_q_half_gptq_3bit_kernel(
auto offset_m = blockIdx.y * m_count;
auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
[[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
[[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
int n = offset_n + t * 4;
@ -593,8 +587,6 @@ __global__ void gemm_half_q_half_gptq_8bit_kernel(
auto offset_m = blockIdx.y * m_count;
auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
[[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
[[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
int n = offset_n + t * 4;

20
csrc/rocm/attention.cu
View File

@ -136,11 +136,6 @@ __device__ __forceinline__ T from_float(const float& inp) {
template <typename T>
__device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
[[maybe_unused]] union tmpcvt {
uint16_t u;
_Float16 f;
__hip_bfloat16 b;
} t16;
_B16x4 ret;
if constexpr (std::is_same<T, _Float16>::value) {
union h2cvt {
@ -169,11 +164,6 @@ __device__ __forceinline__ _B16x4 from_floatx4(const floatx4& inp) {
template <typename T>
__device__ __forceinline__ _B16x4 addx4(const _B16x4& inp1,
const _B16x4& inp2) {
[[maybe_unused]] union tmpcvt {
uint16_t u;
_Float16 f;
__hip_bfloat16 b;
} t1, t2, res;
_B16x4 ret;
if constexpr (std::is_same<T, _Float16>::value) {
union h2cvt {
@ -325,8 +315,6 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
constexpr int GQA_RATIO4 = DIVIDE_ROUND_UP(GQA_RATIO, 4);
[[maybe_unused]] __shared__ float shared_qk_max[NWARPS][16 + 1];
[[maybe_unused]] __shared__ float shared_exp_sum[NWARPS][16 + 1];
// shared_logits is used for multiple purposes
__shared__ _B16x4 shared_logits[NWARPS][4][16][4];
@ -444,8 +432,6 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
const cache_t* k_ptr2 = k_ptr + kblock_number * kv_block_stride;
const int klocal_token_idx =
TOKENS_PER_WARP * warpid + token_depth * 16 + lane16id;
[[maybe_unused]] const int kglobal_token_idx =
partition_start_token_idx + klocal_token_idx;
const int kphysical_block_offset = klocal_token_idx % BLOCK_SIZE;
const cache_t* k_ptr3 = k_ptr2 + kphysical_block_offset * KX;
@ -1309,9 +1295,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
const int context_len = context_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
[[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const auto warpid = threadIdx.x / WARP_SIZE;
[[maybe_unused]] const auto laneid = threadIdx.x % WARP_SIZE;
__shared__ float shared_global_exp_sum;
// max num partitions supported is warp_size * NPAR_LOOPS
@ -2080,9 +2064,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
const int context_len = context_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
[[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warpid = 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
@ -2816,9 +2798,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
const int context_len = context_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
[[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warpid = 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