Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

clean up unused variables warnings (#23975)

Browse Source
This commit is contained in:
Anav Prasad
2026-06-02 10:38:37 +08:00
committed by GitHub
parent 210a6570ce
commit 1fd5f48037
6 changed files with 22 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-mma-f16.cuh
+3 -3
View File
@@ -568,7 +568,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
constexpr bool Q_in_reg = ggml_cuda_fattn_mma_get_Q_in_reg (DKQ, DV, ncols);
constexpr int nstages = ggml_cuda_fattn_mma_get_nstages (DKQ, DV, ncols1, ncols2);
constexpr int stride_tile_Q = DKQ/2 + 4;
constexpr int stride_tile_K = nbatch_K2 + 4;
constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
@@ -604,9 +603,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
#pragma unroll
for (int k0_start = (DKQ/2-1) - (DKQ/2-1) % nbatch_K2; k0_start >= 0; k0_start -= nbatch_K2) {
const int k0_stop = k0_start + nbatch_K2 < DKQ/2 ? k0_start + nbatch_K2 : DKQ/2;
const int k0_diff = k0_stop - k0_start;
if constexpr (nstages <= 1) {
const int k0_diff = k0_stop - k0_start;
constexpr bool use_cp_async = nstages == 1;
flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check>
(K_h2 + int64_t(k_VKQ_0)*stride_K + k0_start, tile_K, k0_diff, stride_K, k_VKQ_sup);
@@ -640,6 +639,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
}
}
} else {
constexpr int stride_tile_Q = DKQ/2 + 4;
#pragma unroll
for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -954,9 +954,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
static_assert(DV % (2*nbatch_V2) == 0, "bad loop size");
const int i0_stop = i0_start + 2*nbatch_V2;
const int i0_diff = i0_stop - i0_start;
if constexpr (nstages <= 1) {
const int i0_diff = i0_stop - i0_start;
if (!V_is_K_view || i0_stop > 2*nbatch_K2) {
constexpr bool use_cp_async = nstages == 1;
flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check>