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

CUDA: fix allignment on register spill for FA (#18815)

Browse Source
This commit is contained in:
Johannes Gäßler
2026-01-15 15:14:50 +01:00
committed by GitHub
parent 8cc0ba957b
commit 5c662d21a3
3 changed files with 25 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-vec.cuh
+2 -2
View File
@@ -132,7 +132,7 @@ static __global__ void flash_attn_ext_vec(
#ifdef V_DOT2_F32_F16_AVAILABLE
half2 Q_reg[ncols][(D/2)/nthreads_KQ]; // Will be initialized completely.
#else
float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
__align__(16) float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
#endif // V_DOT2_F32_F16_AVAILABLE
int Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
float2 Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
@@ -200,7 +200,7 @@ static __global__ void flash_attn_ext_vec(
for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
__align__(16) float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
if (ncols == 1 || ic0 + j < int(ne01.z)) {
ggml_cuda_memcpy_1<cpy_nb>(tmp, &Q_j[i]);
ggml_cuda_memcpy_1<cpy_nb>(tmp + cpy_ne/2, &Q_j[i + cpy_ne/2]);